tests : reduce FA test sizes (#28842)

- Clamp the -j parallelism to min(nproc, 2) so a single-core runner
  uses -j 1 and multi-core runners use at most -j 2, instead of
  unconditionally using $(nproc).
- Add a 3600s timeout to both test-backend-ops runs (the high-perf CPU
  path and the default path) so a hung test cannot stall CI indefinitely.
- Note a TODO to reduce the timeout to 1800s in the future.

Assisted-by: pi:llama.cpp/Qwen3.8-27B

.github/workflows/build-cpu.yml

@@ -221,7 +221,6 @@ jobs:
       #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
       #     $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
       #     cd build
-      #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
       #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900
 
       - name: ccache-clear

.github/workflows/build-openvino.yml

@@ -33,7 +33,7 @@ env:
   LLAMA_ARG_LOG_PREFIX: 1
   LLAMA_ARG_LOG_TIMESTAMPS: 1
   # TODO: fix failing tests on OpenVINO backend
-  CTEST_EXCLUDE: "test-llama-archs|^test-recurrent-state-|test-backend-ops|test-save-load-state"
+  CTEST_EXCLUDE: "test-llama-archs|^test-recurrent-state-|test-save-load-state"
 
 jobs:
   ubuntu-24-openvino:

.github/workflows/build-self-hosted.yml

@@ -160,10 +160,10 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          DAWN_VERSION="v20260317.182325"
+          DAWN_VERSION="v20260908.214631"
           DAWN_OWNER="google"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-ubuntu-latest-Release"
           echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
           curl -L -o artifact.tar.gz \
             "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
@@ -246,10 +246,10 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          DAWN_VERSION="v20260317.182325"
+          DAWN_VERSION="v20260908.214631"
           DAWN_OWNER="google"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-macos-latest-Release"
+          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-macos-latest-Release"
           echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
           curl -L -o artifact.tar.gz \
             "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
@@ -395,7 +395,11 @@ jobs:
       - name: Test
         id: ggml-ci
         run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          LLAMA_ARG_THREADS=$(nproc) \
+          GG_BUILD_HIGH_PERF=1 \
+          GG_BUILD_NO_BF16=1 \
+          GG_BUILD_EXTRA_TESTS_0=1 \
+          bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
   cpu-arm64-graviton4-kleidiai:
     runs-on: ah-ubuntu_22_04-c8g_8x
@@ -434,6 +438,8 @@ jobs:
       - name: Test
         id: ggml-ci
         run: |
+          LLAMA_ARG_THREADS=$(nproc) \
           GG_BUILD_KLEIDIAI=1 \
           GG_BUILD_EXTRA_TESTS_0=1 \
-          bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          GG_BUILD_HIGH_PERF=1 \
+          bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

.github/workflows/build-vulkan.yml

@@ -164,9 +164,7 @@ jobs:
           export GGML_VK_VISIBLE_DEVICES=0
           export GGML_VK_DISABLE_F16=1
           export GGML_VK_DISABLE_COOPMAT=1
-          # This is using llvmpipe and runs slower than other backends
-          # test-backend-ops is too slow on llvmpipe, skip it
-          ctest -L main -E test-backend-ops --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   windows:
     runs-on: windows-2025

.github/workflows/build-wasm.yml

@@ -68,7 +68,7 @@ jobs:
 
       - name: Fetch emdawnwebgpu
         run: |
-          DAWN_TAG="v20260317.182325"
+          DAWN_TAG="v20260908.214631"
           EMDAWN_PKG="emdawnwebgpu_pkg-${DAWN_TAG}.zip"
           echo "Downloading ${EMDAWN_PKG}"
           curl -L -o emdawn.zip \

.github/workflows/build-webgpu.yml

@@ -77,10 +77,10 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          DAWN_VERSION="v20260317.182325"
+          DAWN_VERSION="v20260908.214631"
           DAWN_OWNER="google"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-macos-latest-Release"
+          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-macos-latest-Release"
           echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
           curl -L -o artifact.tar.gz \
             "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
@@ -147,10 +147,10 @@ jobs:
         id: dawn-depends
         run: |
           sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v20260317.182325"
+          DAWN_VERSION="v20260908.214631"
           DAWN_OWNER="google"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-ubuntu-latest-Release"
           echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
           curl -L -o artifact.tar.gz \
             "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
@@ -190,6 +190,4 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          # This is using llvmpipe and runs slower than other backends
-          # test-backend-ops is too slow on llvmpipe, skip it
-          ctest -L main -E test-backend-ops --verbose --timeout 900
+          ctest -L main --verbose --timeout 900

.github/workflows/fusion.yml

@@ -0,0 +1,67 @@
+name: Fusion
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/fusion.yml',
+      'ggml/**',
+      'tests/fusion/**',
+      'tests/test-fusion.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/fusion.yml',
+      'ggml/**',
+      'tests/fusion/**',
+      'tests/test-fusion.cpp'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  # TODO: add jobs for other backends as they adopt the fusion debug API
+  metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_BLAS=OFF \
+            -DGGML_METAL=ON
+          time cmake --build build --config Release --target test-llama-archs -j $(sysctl -n hw.logicalcpu)
+          time cmake --build build --config Release --target test-fusion -j $(sysctl -n hw.logicalcpu)
+
+      - name: Generate models
+        id: generate_models
+        run: |
+          rm -rf build-ci-models && mkdir -p build-ci-models
+          ./build/bin/test-llama-archs -o build-ci-models
+
+      - name: Test fusion
+        id: test_fusion
+        run: |
+          ./build/bin/test-fusion --models build-ci-models --device MTL0 --check tests/fusion/MTL.csv

.github/workflows/server-sanitize.yml

@@ -75,7 +75,7 @@ jobs:
       - name: ccache-buckets-restore
         uses: ./.github/actions/ccache-buckets
         with:
-          key: server-sanitize
+          key: server-sanitize-${{ matrix.sanitizer }}
           folder: llama.cpp
           hf_bucket: ggml-org/cache
 
@@ -99,7 +99,7 @@ jobs:
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
         with:
-          key: server-sanitize
+          key: server-sanitize-${{ matrix.sanitizer }}
           folder: llama.cpp
           evict-old-files: 1d
           hf_bucket: ggml-org/cache
@@ -116,7 +116,6 @@ jobs:
         run: |
           source .venv/bin/activate
           cd tools/server/tests
-          export ${{ matrix.extra_args }}
           PYTEST_WORKERS=1 ./tests.sh
 
       - name: Slow tests
@@ -125,5 +124,4 @@ jobs:
         run: |
           source .venv/bin/activate
           cd tools/server/tests
-          export ${{ matrix.extra_args }}
           PYTEST_WORKERS=1 SLOW_TESTS=1 ./tests.sh

CMakeLists.txt

@@ -197,6 +197,16 @@ llama_option_depr(WARNING     LLAMA_CURL)
 include("cmake/license.cmake")
 license_add_file("llama.cpp" "LICENSE")
 
+#
+# compile options
+#
+
+# clang stores the modification time of the precompiled header sources inside the
+# header and rejects it when they differ, so the timestamp is left out of it
+add_compile_options(
+    "$<$<COMPILE_LANG_AND_ID:C,Clang,IntelLLVM>:SHELL:-Xclang -fno-pch-timestamp>"
+    "$<$<COMPILE_LANG_AND_ID:CXX,Clang,IntelLLVM>:SHELL:-Xclang -fno-pch-timestamp>")
+
 #
 # 3rd-party
 #

ci/run.sh

@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"
 
     # TODO: fix failing tests on OpenVINO backend
-    CTEST_EXTRA="-E test-llama-archs|^test-recurrent-state-|test-backend-ops|test-save-load-state"
+    CTEST_EXTRA="-E test-llama-archs|^test-recurrent-state-|test-save-load-state"
 fi
 
 ## helpers
@@ -250,7 +250,7 @@ function gg_run_ctest_debug {
     (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time cmake --build . --config Debug -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest -C Debug --output-on-failure -L main -E "test-opt|test-backend-ops|test-llama-archs" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest -C Debug --output-on-failure -L main -E "test-opt|test-llama-archs" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -334,6 +334,35 @@ function gg_sum_test_llama_archs_tensor_split {
     gg_printf '```\n'
 }
 
+# test_llama_archs_models
+
+function gg_run_test_llama_archs_models {
+    cd ${SRC}
+
+    set -e
+
+    # TODO: fix and re-enable `test-llama-archs` on OpenVINO
+    # TODO: the `test-llama-archs` currently does not build on Windows, so we check if the binary exists
+    if [ -z ${GG_BUILD_OPENVINO} ] && [ -f ./build-ci-release/bin/test-llama-archs ]; then
+        rm -rf build-ci-models && mkdir -p build-ci-models
+
+        # generate the dummy models used by the model-dependent tests
+        ./build-ci-release/bin/test-llama-archs -o build-ci-models 2>&1
+    fi
+
+    set +e
+}
+
+function gg_sum_test_llama_archs_models {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Generates the dummy models used by the model-dependent tests\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}.log)"
+    gg_printf '```\n'
+}
+
 # test_scripts
 
 function gg_run_test_scripts {
@@ -739,25 +768,48 @@ function gg_check_build_requirements {
     fi
 }
 
-function gg_run_test_backend_ops_cpu {
+function gg_run_test_backend_ops {
     cd ${SRC}
 
     cd build-ci-release
 
     set -e
 
-    (time ./bin/test-backend-ops -b CPU ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops-cpu.log
+    local n_jobs=$(nproc)
+    if [ "${n_jobs}" -gt 2 ]; then
+        n_jobs=2
+    fi
+    local args_extra="-j ${n_jobs}"
+
+    # TODO: fix multi-threaded for ROCm
+    #       https://github.com/ggml-org/llama.cpp/actions/runs/34576278519/job/103297889044?pr=28740#step:3:4865
+    if [ ! -z ${GG_BUILD_ROCM} ]; then
+        args_extra=""
+    fi
+
+    # TODO: MoltenVK bug?
+    #       https://github.com/ggml-org/llama.cpp/actions/runs/34611260059/job/103302413736?pr=28740#step:3:5897
+    if [ ! -z "${GG_BUILD_VULKAN}" ] && [ "$(uname -s)" = "Darwin" ]; then
+        args_extra=""
+    fi
+
+    # TODO: reduce the test-backend-ops timeout to 1800s
+    if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
+        (time timeout 3600 ./bin/test-backend-ops ${args_extra} -b CPU) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
+    else
+        (time timeout 3600 ./bin/test-backend-ops ${args_extra}       ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
+    fi
 
     set +e
 }
 
-function gg_sum_test_backend_ops_cpu {
+function gg_sum_test_backend_ops {
     gg_printf '### %s\n\n' "${ci}"
 
-    gg_printf 'Runs test-backend-ops for CPU backend\n'
+    gg_printf 'Runs test-backend-ops\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-test-backend-ops-cpu.log)"
+    gg_printf '%s\n' "$(cat $OUT/${ci}-test-backend-ops.log)"
     gg_printf '```\n'
     gg_printf '\n'
 }
@@ -790,11 +842,10 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
-test $ret -eq 0 && gg_run test_llama_archs_tensor_split
+test $ret -eq 0 && gg_run test_backend_ops
 
-if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
-    test $ret -eq 0 && gg_run test_backend_ops_cpu
-fi
+test $ret -eq 0 && gg_run test_llama_archs_models
+test $ret -eq 0 && gg_run test_llama_archs_tensor_split
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small

common/CMakeLists.txt

@@ -84,6 +84,8 @@ add_library(${TARGET}
     imatrix-loader.cpp
     imatrix-loader.h
     json-schema-to-grammar.cpp
+    json-schema.cpp
+    json-schema.h
     json.cpp
     json.h
     llguidance.cpp
@@ -134,6 +136,8 @@ set_target_properties(${TARGET} PROPERTIES
 target_include_directories(${TARGET} PUBLIC .)
 target_link_libraries     (${TARGET} PUBLIC vendor::nlohmann vendor::sheredom)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
+target_precompile_headers (${TARGET} PRIVATE common.h)
+target_precompile_headers (${TARGET} PRIVATE chat.h)
 
 if (LLAMA_SUBPROCESS)
     target_compile_definitions(${TARGET} PUBLIC LLAMA_SUBPROCESS)

common/arg.cpp

@@ -2277,14 +2277,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sampling());
     add_opt(common_arg(
         {"-j", "--json-schema"}, "SCHEMA",
-        "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        "JSON schema to constrain generations (https://json-schema.org/), e.g. `{\"type\": \"object\"}` for any JSON object",
         [](common_params & params, const std::string & value) {
             params.sampling.grammar = {COMMON_GRAMMAR_TYPE_OUTPUT_FORMAT, json_schema_to_grammar(json::parse(value))};
         }
     ).set_sampling());
     add_opt(common_arg(
         {"-jf", "--json-schema-file"}, "FILE",
-        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{\"type\": \"object\"}` for any JSON object",
         [](common_params & params, const std::string & value) {
             std::ifstream file(value);
             if (!file) {
@@ -3875,7 +3875,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--no-log-jsonl"},
         "Log as JSONL (one JSON object per line) to stdout, this also disables colored logging (default: disabled)",
         [](common_params &, bool value) {
-            common_log_set_jsonl(common_log_main(), value);
+            common_log_set_jsonl(value);
         }
     ).set_env("LLAMA_ARG_LOG_JSONL"));
     add_opt(common_arg(

common/chat-auto-parser-generator.cpp

@@ -5,6 +5,7 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
+#include "parsers/parsers.h"
 #include "peg-parser.h"
 
 #include <stdexcept>
@@ -12,16 +13,6 @@
 
 using json = common_json;
 
-// Helper to iterate over tools/functions
-static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
-    for (const auto & tool : tools) {
-        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
-            continue;
-        }
-        fn(tool);
-    }
-}
-
 namespace autoparser {
 
 parser_build_context::parser_build_context(common_chat_peg_builder & p, const generation_params & inputs) :
@@ -87,15 +78,6 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 
@@ -312,7 +294,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_json(parser_build_context
     foreach_function(inputs.tools, [&](const json & tool) {
         const auto & func   = tool.at("function");
         std::string  name   = func.at("name");
-        const auto & schema = func.contains("parameters") ? func.at("parameters") : json::object();
+        const auto   schema = common_chat_tool_parameters(func);
 
         // Build call_id parser based on position (if supported)
         bool have_call_id = false;
@@ -383,43 +365,31 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     common_peg_parser tool_choice = p.choice();
 
     foreach_function(inputs.tools, [&](const json & tool) {
-        const auto &          func       = tool.at("function");
-        std::string           name       = func.at("name");
-        auto                  params     = func.contains("parameters") ? func.at("parameters") : json::object();
-        const auto &          properties = params.contains("properties") ? params.at("properties") : json::object();
-
-        std::set<std::string> required;
-        if (params.contains("required")) {
-            required = params.at("required").get<std::set<std::string>>();
-        }
-
-        auto schema_info = common_schema_info();
-        schema_info.resolve_refs(params);
+        const auto & func = tool.at("function");
+        std::string  name = func.at("name");
 
         // Build parser for each argument, separating required and optional
         std::vector<common_peg_parser> required_parsers;
         std::vector<common_peg_parser> optional_parsers;
-        for (const auto & [param_name, param_schema] : properties.items()) {
-            bool is_required = required.find(param_name) != required.end();
-
+        foreach_parameter(func, [&](const common_chat_schema_property & param, const common_chat_schema_document_ptr & doc) {
             auto arg =
-                p.tool_arg(p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
+                p.tool_arg(p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param.name)) +
                                            arguments.name_suffix) +
                            arguments.value_prefix +
-                           (schema_info.resolves_to_string(param_schema) ?
+                           (param.schema->may_be_string() ?
                                 p.ac(p.tool_arg_string_value(until_suffix) +
                                     p.tool_arg_close(p.literal(arguments.value_suffix)), arguments.value_suffix) :
                                 (p.tool_arg_json_value(p.schema(
-                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
+                                    p.json(), "tool-" + name + "-arg-" + param.name + "-schema", doc, *param.schema)) +
                                     p.tool_arg_close(p.literal(arguments.value_suffix)))));
 
-            auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-            if (is_required) {
+            auto named_arg = p.rule("tool-" + name + "-arg-" + param.name, arg);
+            if (param.required) {
                 required_parsers.push_back(named_arg);
             } else {
                 optional_parsers.push_back(named_arg);
             }
-        }
+        });
 
         // Build required arg sequence in definition order
         common_peg_parser args_seq = p.eps();

common/chat-peg-parser.cpp

@@ -488,7 +488,7 @@ common_peg_parser common_chat_peg_builder::standard_constructed_tools(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+        ordered_json   params   = common_chat_tool_parameters(function);
 
         // Build argument parsers
         auto args = eps();
@@ -565,7 +565,7 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+        ordered_json   params   = common_chat_tool_parameters(function);
 
         auto args = eps();
         if (params.contains("properties") && !params["properties"].empty()) {
@@ -640,7 +640,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+        ordered_json   params   = common_chat_tool_parameters(function);
 
         // Build inner object fields
         std::vector<common_peg_parser> inner_fields;
@@ -726,7 +726,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+        ordered_json   params   = common_chat_tool_parameters(function);
 
         auto nested_name = literal("\"" + nested_name_field + "\"") + space() + literal(":") + space() +
                           atomic(literal("\"") + tool_name(literal(name)) + literal("\""));
@@ -795,7 +795,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+        ordered_json   params   = common_chat_tool_parameters(function);
 
         auto tool_name_ = name_key_parser + space() + literal(":") + space() +
                          atomic(literal("\"") + tool_name(literal(name)) + literal("\""));

common/chat.cpp

@@ -574,6 +574,16 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
     return result;
 }
 
+json common_chat_tool_parameters(const json & function) {
+    if (function.contains("parameters")) {
+        const auto & params = function.at("parameters");
+        if (!params.is_null() && !(params.is_object() && params.empty())) {
+            return params;
+        }
+    }
+    return json{{"type", "object"}, {"properties", json::object()}};
+}
+
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 

common/chat.h

@@ -360,6 +360,9 @@ common_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg
 
 common_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
+// The parameters schema of a function tool. A tool without parameters, or with an empty {}, takes zero arguments.
+common_json common_chat_tool_parameters(const common_json & function);
+
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
 

common/fit.cpp

@@ -1,5 +1,6 @@
 #include "fit.h"
 
+#include "json.h"
 #include "log.h"
 
 #include "../src/llama-ext.h"
@@ -915,6 +916,9 @@ void common_memory_breakdown_print(const struct llama_context * ctx) {
 
     std::vector<std::array<std::string, 9>> table_data;
     table_data.reserve(devices.size());
+
+    // same data as the table below, for --log-jsonl consumers
+    common_json rows = common_json::array();
     const std::string template_header = "%s: | %s | %s   %s    %s   %s   %s   %s    %s |\n";
     const std::string template_gpu    = "%s: | %s | %s = %s + (%s = %s + %s + %s) + %s |\n";
     const std::string template_other  = "%s: | %s | %s   %s    %s = %s + %s + %s    %s |\n";
@@ -989,6 +993,19 @@ void common_memory_breakdown_print(const struct llama_context * ctx) {
             std::to_string(mb.context / MiB),
             std::to_string(mb.compute / MiB),
             std::to_string(unaccounted / static_cast<int64_t>(MiB))});
+
+        rows.push_back({
+            {"kind",        "device"},
+            {"name",        name},
+            {"description", desc},
+            {"total",       total / MiB},
+            {"free",        free / MiB},
+            {"self",        self / MiB},
+            {"model",       mb.model / MiB},
+            {"context",     mb.context / MiB},
+            {"compute",     mb.compute / MiB},
+            {"unaccounted", unaccounted / static_cast<int64_t>(MiB)},
+        });
     }
 
     // print memory breakdown for host:
@@ -1004,6 +1021,15 @@ void common_memory_breakdown_print(const struct llama_context * ctx) {
             std::to_string(mb_host.context / MiB),
             std::to_string(mb_host.compute / MiB),
             ""}); // unaccounted
+
+        rows.push_back({
+            {"kind",    "host"},
+            {"name",    "Host"},
+            {"self",    self / MiB},
+            {"model",   mb_host.model / MiB},
+            {"context", mb_host.context / MiB},
+            {"compute", mb_host.compute / MiB},
+        });
     }
 
     // print memory breakdown for all remaining buffer types:
@@ -1025,6 +1051,16 @@ void common_memory_breakdown_print(const struct llama_context * ctx) {
             std::to_string(mb.context / MiB),
             std::to_string(mb.compute / MiB),
             ""}); // unaccounted
+
+        rows.push_back({
+            {"kind",    "buffer_type"},
+            {"name",    name},
+            {"self",    self / MiB},
+            {"model",   mb.model / MiB},
+            {"context", mb.context / MiB},
+            {"compute", mb.compute / MiB},
+        });
+
         seen_buffer_types.insert(buft);
     }
 
@@ -1042,6 +1078,11 @@ void common_memory_breakdown_print(const struct llama_context * ctx) {
             __func__, td[1].c_str(), td[2].c_str(), td[3].c_str(), td[4].c_str(), td[5].c_str(),
             td[6].c_str(), td[7].c_str(), td[8].c_str());
     }
+
+    LOG_JSON("fit_memory_breakdown", common_json({
+        {"unit", "MiB"},
+        {"rows", rows},
+    }));
 }
 
 void common_fit_print(

common/jinja/runtime.cpp

@@ -842,6 +842,12 @@ value member_expression::execute_impl(context & ctx) {
         } else {
             property = this->property->execute(ctx);
         }
+    } else if (is_stmt<integer_literal>(this->property)) {
+        // syntax: obj.index
+        property = mk_val<value_int>(cast_stmt<integer_literal>(this->property)->val);
+        if (property->as_int() < 0) {
+            throw std::runtime_error("Static member property cannot be negative");
+        }
     } else {
         // syntax: obj.prop
         if (!is_stmt<identifier>(this->property)) {

common/json-schema-to-grammar.cpp

@@ -1,5 +1,7 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
+#include "trie.h"
+#include "unicode.h"
 
 #include <algorithm>
 #include <limits>
@@ -336,18 +338,20 @@ static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
     return 2 + n_hex;
 }
 
-class common_schema_converter {
+class common_chat_schema_converter {
 private:
-    friend class common_schema_info;
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
-    std::function<json(const std::string &)> _fetch_json;
     bool _dotall;
     std::map<std::string, std::string> _rules;
-    std::unordered_map<std::string, json> _refs;
     std::unordered_set<std::string> _refs_being_resolved;
     std::vector<std::string> _errors;
     std::vector<std::string> _warnings;
 
+    template <typename T>
+    static const T & as(const common_chat_schema & node) {
+        return static_cast<const T &>(node);
+    }
+
     std::string _add_rule(const std::string & name, const std::string & rule) {
         std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-");
         if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
@@ -363,11 +367,11 @@ private:
         return key;
     }
 
-    std::string _generate_union_rule(const std::string & name, const std::vector<json> & alt_schemas) {
+    std::string _generate_union_rule(const std::string & name, const std::vector<common_chat_schema_ptr> & alt_schemas) {
         std::vector<std::string> rules;
         rules.reserve(alt_schemas.size());
         for (size_t i = 0; i < alt_schemas.size(); i++) {
-            rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
+            rules.push_back(visit(*alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
         }
         return string_join(rules, " | ");
     }
@@ -634,85 +638,68 @@ private:
             -> ["] ( [a] ([l] ([s] ([o] char+ | [^"o] char*) | [^"s] char*) | [n] ([d] char+ | [^"d] char*) | [^"ln] char*) | [^"a] char* )? ["]
     */
     std::string _not_strings(const std::vector<std::string> & strings) {
-
-        struct TrieNode {
-            std::map<char, TrieNode> children;
-            bool is_end_of_string;
-
-            TrieNode() : is_end_of_string(false) {}
-
-            void insert(const std::string & string) {
-                auto *node = this;
-                for (char c : string) {
-                    node = &node->children[c];
-                }
-                node->is_end_of_string = true;
-            }
-        };
-
-        TrieNode trie;
-        for (const auto & s : strings) {
-            trie.insert(s);
-        }
+        common_trie trie(strings);
 
         std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
         std::ostringstream out;
         out << "[\"] ( ";
-        std::function<void(const TrieNode &)> visit = [&](const TrieNode & node) {
-            std::ostringstream rejects;
+        std::function<void(size_t)> visit = [&](size_t idx) {
+            const auto & node = trie.nodes[idx];
+            std::string rejects;
             auto first = true;
-            for (const auto & kv : node.children) {
-                rejects << kv.first;
+            for (const auto & [cpt, child] : node.children) {
+                std::string c = common_unicode_cpt_to_utf8(cpt);
+                rejects += c;
                 if (first) {
                     first = false;
                 } else {
                     out << " | ";
                 }
-                out << "[" << kv.first << "]";
-                if (!kv.second.children.empty()) {
+                out << "[" << c << "]";
+                if (!trie.nodes[child].children.empty()) {
                     out << " (";
-                    visit(kv.second);
+                    visit(child);
                     out << ")";
-                } else if (kv.second.is_end_of_string) {
+                } else {
                     out << " " << char_rule << "+";
                 }
             }
             if (!node.children.empty()) {
-                if (!first) {
-                    out << " | ";
-                }
-                out << "[^\"" << rejects.str() << "] " << char_rule << "*";
+                out << " | [^\"" << rejects << "] " << char_rule << "*";
             }
         };
-        visit(trie);
+        visit(0);
 
         out << " )";
-        if (!trie.is_end_of_string) {
+        if (trie.nodes[0].pattern < 0) {
             out << "?";
         }
         out << " [\"]";
         return out.str();
     }
 
-    std::string _resolve_ref(const std::string & ref) {
-        auto it = ref.find('#');
-        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
+    std::string _resolve_ref(const common_chat_schema_ref & schema) {
+        auto it = schema.ref.find('#');
+        std::string ref_fragment = it != std::string::npos ? schema.ref.substr(it + 1) : schema.ref;
         static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
         std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
-        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
-            _refs_being_resolved.insert(ref);
-            json resolved = _refs[ref];
-            ref_name = visit(resolved, ref_name);
-            _refs_being_resolved.erase(ref);
+        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(schema.ref) == _refs_being_resolved.end()) {
+            if (!schema.target) {
+                _errors.push_back("Unresolved $ref " + schema.ref);
+                return "";
+            }
+            _refs_being_resolved.insert(schema.ref);
+            ref_name = visit(*schema.target, ref_name);
+            _refs_being_resolved.erase(schema.ref);
         }
         return ref_name;
     }
 
     std::string _build_object_rule(
-        const std::vector<std::pair<std::string, json>> & properties,
+        const std::vector<std::pair<std::string, const common_chat_schema *>> & properties,
         const std::unordered_set<std::string> & required,
         const std::string & name,
-        const json & additional_properties)
+        const common_chat_schema * additional_properties)
     {
         std::vector<std::string> required_props;
         std::vector<std::string> optional_props;
@@ -722,7 +709,7 @@ private:
             const auto &prop_name = kv.first;
             const auto &prop_schema = kv.second;
 
-            std::string prop_rule_name = visit(prop_schema, name + (name.empty() ? "" : "-") + prop_name);
+            std::string prop_rule_name = visit(*prop_schema, name + (name.empty() ? "" : "-") + prop_name);
             prop_kv_rule_names[prop_name] = _add_rule(
                 name + (name.empty() ? "" : "-") + prop_name + "-kv",
                 format_literal(json(prop_name).dump()) + " space \":\" space " + prop_rule_name
@@ -734,10 +721,10 @@ private:
             }
             prop_names.push_back(prop_name);
         }
-        if ((additional_properties.is_boolean() && additional_properties.get<bool>()) || additional_properties.is_object()) {
+        if (additional_properties) {
             std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
             std::string value_rule =
-                additional_properties.is_object() ? visit(additional_properties, sub_name + "-value")
+                additional_properties->kind() != common_chat_schema::KIND_ANY ? visit(*additional_properties, sub_name + "-value")
                 : _add_primitive("value", PRIMITIVE_RULES.at("value"));
 
             auto key_rule =
@@ -825,267 +812,163 @@ private:
     }
 
 public:
-    common_schema_converter(
-        const std::function<json(const std::string &)> & fetch_json,
-        bool dotall)
-          : _fetch_json(fetch_json), _dotall(dotall)
-    {
+    explicit common_chat_schema_converter(bool dotall) : _dotall(dotall) {
         _rules["space"] = SPACE_RULE;
     }
 
-    void resolve_refs(json & schema, const std::string & url) {
-        /*
-        * Resolves all $ref fields in the given schema, fetching any remote schemas,
-        * replacing each $ref with absolute reference URL and populates _refs with the
-        * respective referenced (sub)schema dictionaries.
-        */
-        std::function<void(json &)> visit_refs = [&](json & n) {
-            if (n.is_array()) {
-                for (auto & x : n) {
-                    visit_refs(x);
-                }
-            } else if (n.is_object()) {
-                if (n.contains("$ref")) {
-                    std::string ref = n["$ref"];
-                    if (_refs.find(ref) == _refs.end()) {
-                        json target;
-                        if (ref.find("https://") == 0) {
-                            std::string base_url = ref.substr(0, ref.find('#'));
-                            auto it = _refs.find(base_url);
-                            if (it != _refs.end()) {
-                                target = it->second;
-                            } else {
-                                // Fetch the referenced schema and resolve its refs
-                                auto referenced = _fetch_json(ref);
-                                resolve_refs(referenced, base_url);
-                                _refs[base_url] = referenced;
-                            }
-                            if (ref.find('#') == std::string::npos || ref.substr(ref.find('#') + 1).empty()) {
-                                return;
-                            }
-                        } else if (ref.find("#/") == 0) {
-                            target = schema;
-                            n["$ref"] = url + ref;
-                            ref = url + ref;
-                        } else {
-                            _errors.push_back("Unsupported ref: " + ref);
-                            return;
-                        }
-                        std::string pointer = ref.substr(ref.find('#') + 1);
-                        std::vector<std::string> tokens = string_split(pointer, "/");
-                        for (size_t i = 1; i < tokens.size(); ++i) {
-                            const std::string& sel = tokens[i];
-                            if (target.is_object() && target.contains(sel)) {
-                                target = target[sel];
-                            } else if (target.is_array()) {
-                                size_t sel_index;
-                                try {
-                                    sel_index = std::stoull(sel);
-                                } catch (const std::invalid_argument & e) {
-                                    sel_index = target.size();
-                                }
-                                if (sel_index >= target.size()) {
-                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
-                                    return;
-                                }
-                                target = target[sel_index];
-                            } else {
-                                _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
-                                return;
-                            }
-                        }
-                        _refs[ref] = target;
-                    }
-                } else {
-                    for (const auto & kv : n.items()) {
-                        visit_refs(kv.value());
-                    }
-                }
-            }
-        };
-
-        visit_refs(schema);
+    std::string add_schema(const std::string & name, const common_chat_schema & schema) {
+        return visit(schema, name);
     }
 
     static std::string _generate_constant_rule(const json & value) {
         return format_literal(value.dump());
     }
 
-    std::string visit(const json & schema, const std::string & name) {
-        json schema_type = schema.contains("type") ? schema["type"] : json();
-        std::string schema_format = schema.contains("format") ? schema["format"].get<std::string>() : "";
-        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
+    std::string _visit_primitive(const std::string & rule_name, const std::string & type) {
+        return _add_primitive(rule_name == "root" ? "root" : type, PRIMITIVE_RULES.at(type));
+    }
 
-        if (schema.contains("$ref")) {
-            return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
-        }
-        if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            const json & alts = schema.contains("oneOf") ? schema.at("oneOf") : schema.at("anyOf");
-            std::vector<json> alt_schemas;
-            for (const auto & alt : alts) {
-                alt_schemas.push_back(alt);
-            }
-            return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
-        }
-        if (schema_type.is_array()) {
-            std::vector<json> schema_types;
-            for (const auto & t : schema_type) {
-                json schema_copy(schema);
-                schema_copy["type"] = t;
-                schema_types.push_back(schema_copy);
-            }
-            return _add_rule(rule_name, _generate_union_rule(name, schema_types));
-        }
-        if (schema.contains("const")) {
-            return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
-        }
-        if (schema.contains("enum")) {
-            std::vector<std::string> enum_values;
-            for (const auto & v : schema["enum"]) {
-                enum_values.push_back(_generate_constant_rule(v));
-            }
-            return _add_rule(rule_name, "(" + string_join(enum_values, " | ") + ")");
-        }
-        if ((schema_type.is_null() || schema_type == "object")
-                && (schema.contains("properties") ||
-                    (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
-            std::unordered_set<std::string> required;
-            if (schema.contains("required") && schema["required"].is_array()) {
-                for (const auto & item : schema["required"]) {
-                    if (item.is_string()) {
-                        required.insert(item.get<std::string>());
+    std::string _visit_all_of(const common_chat_schema_all_of & schema, const std::string & name, const std::string & rule_name) {
+        std::unordered_set<std::string> required;
+        std::vector<std::pair<std::string, const common_chat_schema *>> properties;
+        std::map<std::string, size_t> enum_values;
+        std::function<void(const common_chat_schema &, bool)> add_component = [&](const common_chat_schema & comp, bool is_required) {
+            if (comp.kind() == common_chat_schema::KIND_REF) {
+                if (const auto * target = as<common_chat_schema_ref>(comp).target) {
+                    add_component(*target, is_required);
+                }
+            } else if (comp.kind() == common_chat_schema::KIND_OBJECT) {
+                for (const auto & prop : as<common_chat_schema_object>(comp).properties) {
+                    properties.emplace_back(prop.name, prop.schema.get());
+                    if (is_required) {
+                        required.insert(prop.name);
                     }
                 }
-            }
-            std::vector<std::pair<std::string, json>> properties;
-            if (schema.contains("properties")) {
-                for (const auto & prop : schema["properties"].items()) {
-                    properties.emplace_back(prop.key(), prop.value());
+            } else if (comp.kind() == common_chat_schema::KIND_ENUM) {
+                for (const auto & v : as<common_chat_schema_enum>(comp).values) {
+                    enum_values[_generate_constant_rule(v)] += 1;
                 }
             }
-            return _add_rule(rule_name,
-                _build_object_rule(
-                    properties, required, name,
-                    schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
+        };
+        for (const auto & child : schema.children) {
+            if (child->kind() == common_chat_schema::KIND_ANY_OF) {
+                for (const auto & alt : as<common_chat_schema_any_of>(*child).children) {
+                    add_component(*alt, false);
+                }
+            } else {
+                add_component(*child, true);
+            }
         }
-        if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
-            std::unordered_set<std::string> required;
-            std::vector<std::pair<std::string, json>> properties;
-            std::map<std::string, size_t> enum_values;
-            const std::string& hybrid_name = name;
-            std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
-                if (comp_schema.contains("$ref")) {
-                    add_component(_refs[comp_schema["$ref"]], is_required);
-                } else if (comp_schema.contains("properties")) {
-                    for (const auto & prop : comp_schema["properties"].items()) {
-                        properties.emplace_back(prop.key(), prop.value());
-                        if (is_required) {
-                            required.insert(prop.key());
-                        }
-                    }
-                } else if (comp_schema.contains("enum")) {
-                    for (const auto & v : comp_schema["enum"]) {
-                        const auto rule = _generate_constant_rule(v);
-                        if (enum_values.find(rule) == enum_values.end()) {
-                            enum_values[rule] = 0;
-                        }
-                        enum_values[rule] += 1;
-                    }
-                } else {
-                  // todo warning
-                }
-            };
-            for (const auto & t : schema["allOf"]) {
-                if (t.contains("anyOf")) {
-                    for (const auto & tt : t["anyOf"]) {
-                        add_component(tt, false);
-                    }
-                } else {
-                    add_component(t, true);
+        if (!enum_values.empty()) {
+            std::vector<std::string> enum_intersection;
+            for (const auto & p : enum_values) {
+                if (p.second == schema.children.size()) {
+                    enum_intersection.push_back(p.first);
                 }
             }
-            if (!enum_values.empty()) {
-                std::vector<std::string> enum_intersection;
-                for (const auto & p : enum_values) {
-                    if (p.second == schema["allOf"].size()) {
-                        enum_intersection.push_back(p.first);
-                    }
-                }
-                if (!enum_intersection.empty()) {
-                    return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ")");
-                }
+            if (!enum_intersection.empty()) {
+                return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ")");
             }
-            return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
         }
-        if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
-            json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
-            if (items.is_array()) {
+        return _add_rule(rule_name, _build_object_rule(properties, required, name, nullptr));
+    }
+
+    std::string visit(const common_chat_schema & schema, const std::string & name) {
+        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
+        std::string sub_name  = name + (name.empty() ? "" : "-");
+
+        switch (schema.kind()) {
+            case common_chat_schema::KIND_REF:
+                return _add_rule(rule_name, _resolve_ref(as<common_chat_schema_ref>(schema)));
+            case common_chat_schema::KIND_ANY_OF:
+                return _add_rule(rule_name, _generate_union_rule(name, as<common_chat_schema_any_of>(schema).children));
+            case common_chat_schema::KIND_ALL_OF:
+                return _visit_all_of(as<common_chat_schema_all_of>(schema), name, rule_name);
+            case common_chat_schema::KIND_CONST:
+                return _add_rule(rule_name, _generate_constant_rule(as<common_chat_schema_const>(schema).value));
+            case common_chat_schema::KIND_ENUM: {
+                std::vector<std::string> enum_values;
+                for (const auto & v : as<common_chat_schema_enum>(schema).values) {
+                    enum_values.push_back(_generate_constant_rule(v));
+                }
+                return _add_rule(rule_name, "(" + string_join(enum_values, " | ") + ")");
+            }
+            case common_chat_schema::KIND_OBJECT: {
+                const auto & obj = as<common_chat_schema_object>(schema);
+                if (obj.properties.empty() && obj.additional_properties && obj.additional_properties->kind() == common_chat_schema::KIND_ANY) {
+                    return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
+                }
+                std::vector<std::pair<std::string, const common_chat_schema *>> properties;
+                std::unordered_set<std::string> required;
+                for (const auto & prop : obj.properties) {
+                    properties.emplace_back(prop.name, prop.schema.get());
+                    if (prop.required) {
+                        required.insert(prop.name);
+                    }
+                }
+                return _add_rule(rule_name, _build_object_rule(properties, required, name, obj.additional_properties.get()));
+            }
+            case common_chat_schema::KIND_TUPLE: {
+                const auto & items = as<common_chat_schema_tuple>(schema).items;
                 std::string rule = "\"[\" space ";
                 for (size_t i = 0; i < items.size(); i++) {
                     if (i > 0) {
                         rule += " \",\" space ";
                     }
-                    rule += visit(items[i], name + (name.empty() ? "" : "-") + "tuple-" + std::to_string(i));
+                    rule += visit(*items[i], sub_name + "tuple-" + std::to_string(i));
                 }
                 rule += " space \"]\"";
                 return _add_rule(rule_name, rule);
             }
-            std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
-            int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
-            json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
-            int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : std::numeric_limits<int>::max();
-
-            return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, min_items, max_items, "\",\" space") + " space \"]\"");
-        }
-        if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
-            return _visit_pattern(schema["pattern"], rule_name);
-        }
-        if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
-            return _add_primitive(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
-        }
-        if ((schema_type.is_null() || schema_type == "string") && STRING_FORMAT_RULES.find(schema_format + "-string") != STRING_FORMAT_RULES.end()) {
-            auto prim_name = schema_format + "-string";
-            return _add_rule(rule_name, _add_primitive(prim_name, STRING_FORMAT_RULES.at(prim_name)));
-        }
-        if (schema_type == "string" && (schema.contains("minLength") || schema.contains("maxLength"))) {
-            std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
-            int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
-            int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
-            return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\"");
-        }
-        if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
-            int64_t min_value = std::numeric_limits<int64_t>::min();
-            int64_t max_value = std::numeric_limits<int64_t>::max();
-            if (schema.contains("minimum")) {
-                min_value = schema["minimum"].get<int64_t>();
-            } else if (schema.contains("exclusiveMinimum")) {
-                min_value = schema["exclusiveMinimum"].get<int64_t>() + 1;
+            case common_chat_schema::KIND_ARRAY: {
+                const auto & arr = as<common_chat_schema_array>(schema);
+                if (arr.items->kind() == common_chat_schema::KIND_ANY && arr.min_items == 0 && arr.max_items < 0) {
+                    return _visit_primitive(rule_name, "array");
+                }
+                std::string item_rule_name = visit(*arr.items, sub_name + "item");
+                int max_items = arr.max_items < 0 ? std::numeric_limits<int>::max() : arr.max_items;
+                return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, arr.min_items, max_items, "\",\" space") + " space \"]\"");
             }
-            if (schema.contains("maximum")) {
-                max_value = schema["maximum"].get<int64_t>();
-            } else if (schema.contains("exclusiveMaximum")) {
-                max_value = schema["exclusiveMaximum"].get<int64_t>() - 1;
+            case common_chat_schema::KIND_STRING: {
+                const auto & str = as<common_chat_schema_string>(schema);
+                if (!str.pattern.empty()) {
+                    return _visit_pattern(str.pattern, rule_name);
+                }
+                if (str.format == common_chat_schema::FORMAT_UUID) {
+                    return _visit_primitive(rule_name, "uuid");
+                }
+                if (str.format != common_chat_schema::FORMAT_NONE) {
+                    std::string prim_name = std::string(str.format == common_chat_schema::FORMAT_DATE ? "date" : str.format == common_chat_schema::FORMAT_TIME ? "time" : "date-time") + "-string";
+                    return _add_rule(rule_name, _add_primitive(prim_name, STRING_FORMAT_RULES.at(prim_name)));
+                }
+                if (str.min_length > 0 || str.max_length >= 0) {
+                    std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
+                    int max_len = str.max_length < 0 ? std::numeric_limits<int>::max() : str.max_length;
+                    return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, str.min_length, max_len) + " \"\\\"\"");
+                }
+                return _visit_primitive(rule_name, "string");
             }
-            std::stringstream out;
-            out << "(";
-            build_min_max_int(min_value, max_value, out);
-            out << ")";
-            return _add_rule(rule_name, out.str());
+            case common_chat_schema::KIND_INTEGER: {
+                const auto & i = as<common_chat_schema_integer>(schema);
+                if (i.minimum == std::numeric_limits<int64_t>::min() && i.maximum == std::numeric_limits<int64_t>::max()) {
+                    return _visit_primitive(rule_name, "integer");
+                }
+                std::stringstream out;
+                out << "(";
+                build_min_max_int(i.minimum, i.maximum, out);
+                out << ")";
+                return _add_rule(rule_name, out.str());
+            }
+            case common_chat_schema::KIND_NUMBER:
+                return _visit_primitive(rule_name, "number");
+            case common_chat_schema::KIND_BOOLEAN:
+                return _visit_primitive(rule_name, "boolean");
+            case common_chat_schema::KIND_NULL:
+                return _visit_primitive(rule_name, "null");
+            case common_chat_schema::KIND_ANY:
+                return _add_rule(rule_name, _add_primitive("value", PRIMITIVE_RULES.at("value")));
         }
-        if (schema.empty() || schema_type == "object") {
-            return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
-        }
-        if (schema_type.is_null() && schema.is_object()) {
-            // No type constraint and no recognized structural keywords (e.g. {"description": "..."}).
-            // Per JSON Schema semantics this is equivalent to {} and accepts any value.
-            return _add_rule(rule_name, _add_primitive("value", PRIMITIVE_RULES.at("value")));
-        }
-        if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
-            _errors.push_back("Unrecognized schema: " + schema.dump());
-            return "";
-        }
-        // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-        return _add_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+        return "";
     }
 
     void check_errors() {
@@ -1106,134 +989,6 @@ public:
     }
 };
 
-// common_schema_info implementation (pimpl)
-
-common_schema_info::common_schema_info()
-    : impl_(std::make_unique<common_schema_converter>(
-        [](const std::string &) { return json(); },
-        false)) {}
-
-common_schema_info::~common_schema_info() = default;
-
-common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
-common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
-
-void common_schema_info::resolve_refs(common_json & schema) {
-    impl_->resolve_refs(schema, "");
-}
-
-// Determines if a JSON schema can resolve to a string type through any path.
-// Some models emit raw string values rather than JSON-encoded strings for string parameters.
-// If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
-// true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const common_json & schema) {
-    std::unordered_set<std::string> visited_refs;
-
-    std::function<bool(const json &)> check = [&](const json & s) -> bool {
-        if (!s.is_object()) {
-            return false;
-        }
-
-        // Handle $ref
-        if (s.contains("$ref")) {
-            const std::string & ref = s["$ref"];
-            if (visited_refs.find(ref) != visited_refs.end()) {
-                // Circular reference, assume not a string to be safe
-                return false;
-            }
-            visited_refs.insert(ref);
-            auto it = impl_->_refs.find(ref);
-            if (it != impl_->_refs.end()) {
-                return check(it->second);
-            }
-            return false;
-        }
-
-        // Check type field
-        if (s.contains("type")) {
-            const json & schema_type = s["type"];
-            if (schema_type.is_string()) {
-                if (schema_type == "string") {
-                    return true;
-                }
-            } else if (schema_type.is_array()) {
-                // Type can be an array like ["string", "null"]
-                for (const auto & t : schema_type) {
-                    if (t == "string") {
-                        return true;
-                    }
-                }
-            }
-        }
-
-        // Check oneOf/anyOf - if any alternative can be a string
-        if (s.contains("oneOf")) {
-            for (const auto & alt : s["oneOf"]) {
-                if (check(alt)) {
-                    return true;
-                }
-            }
-        }
-        if (s.contains("anyOf")) {
-            for (const auto & alt : s["anyOf"]) {
-                if (check(alt)) {
-                    return true;
-                }
-            }
-        }
-
-        // Check allOf - all components must be compatible with string type
-        if (s.contains("allOf")) {
-            bool all_string = true;
-            for (const auto & component : s["allOf"]) {
-                if (!check(component)) {
-                    all_string = false;
-                    break;
-                }
-            }
-            if (all_string) {
-                return true;
-            }
-        }
-
-        // Check const - if the constant value is a string
-        if (s.contains("const")) {
-            if (s["const"].is_string()) {
-                return true;
-            }
-        }
-
-        // Check enum - if any enum value is a string
-        if (s.contains("enum")) {
-            for (const auto & val : s["enum"]) {
-                if (val.is_string()) {
-                    return true;
-                }
-            }
-        }
-
-        // String-specific keywords imply string type
-        if (s.contains("pattern") || s.contains("minLength") || s.contains("maxLength")) {
-            return true;
-        }
-
-        // Check format - many formats imply string
-        if (s.contains("format")) {
-            const std::string & fmt = s["format"];
-            if (fmt == "date" || fmt == "time" || fmt == "date-time" ||
-                fmt == "uri" || fmt == "email" || fmt == "hostname" ||
-                fmt == "ipv4" || fmt == "ipv6" || fmt == "uuid" ||
-                fmt.find("uuid") == 0) {
-                return true;
-            }
-        }
-
-        return false;
-    };
-
-    return check(schema);
-}
-
 std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
@@ -1242,25 +997,29 @@ std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf)
 #else
     (void)force_gbnf;
 #endif // LLAMA_USE_LLGUIDANCE
-    return build_grammar([&](const common_grammar_builder & callbacks) {
-        auto copy = schema;
-        callbacks.resolve_refs(copy);
-        callbacks.add_schema("", copy);
-    });
+    try {
+        return json_schema_to_grammar(common_chat_schema_from_json(schema));
+    } catch (const std::runtime_error & e) {
+        throw std::invalid_argument(std::string("JSON schema conversion failed:\n") + e.what());
+    }
+}
+
+std::string json_schema_to_grammar(const common_chat_schema_document & schema) {
+    common_chat_schema_converter converter(false);
+    converter.visit(*schema.root, "");
+    converter.check_errors();
+    return converter.format_grammar();
 }
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
-    common_schema_converter converter([&](const std::string &) { return json(); }, options.dotall);
+    common_chat_schema_converter converter(options.dotall);
     common_grammar_builder builder {
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);
         },
-        /* .add_schema = */ [&](const std::string & name, const common_json & schema) {
-            return converter.visit(schema, name == "root" ? "" : name);
+        /* .add_schema = */ [&](const std::string & name, const common_chat_schema & schema) {
+            return converter.add_schema(name == "root" ? "" : name, schema);
         },
-        /* .resolve_refs = */ [&](common_json & schema) {
-            converter.resolve_refs(schema, "");
-        }
     };
     cb(builder);
     converter.check_errors();

common/json-schema-to-grammar.h

@@ -1,37 +1,17 @@
 #pragma once
 
+#include "json-schema.h"
 #include "json.h"
 
 #include <functional>
-#include <memory>
 #include <string>
 
-std::string json_schema_to_grammar(const common_json & schema,
-                                   bool force_gbnf = false);
-
-class common_schema_converter;
-
-// Probes a JSON schema to extract information about its structure and type constraints.
-class common_schema_info {
-    std::unique_ptr<common_schema_converter> impl_;
-
-  public:
-    common_schema_info();
-    ~common_schema_info();
-
-    common_schema_info(const common_schema_info &) = delete;
-    common_schema_info & operator=(const common_schema_info &) = delete;
-    common_schema_info(common_schema_info &&) noexcept;
-    common_schema_info & operator=(common_schema_info &&) noexcept;
-
-    void resolve_refs(common_json & schema);
-    bool resolves_to_string(const common_json & schema);
-};
+std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf = false);
+std::string json_schema_to_grammar(const common_chat_schema_document & schema);
 
 struct common_grammar_builder {
-    std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const common_json &)> add_schema;
-    std::function<void(common_json &)> resolve_refs;
+    std::function<std::string(const std::string &, const std::string &)>    add_rule;
+    std::function<std::string(const std::string &, const common_chat_schema &)> add_schema;
 };
 
 struct common_grammar_options {

common/json-schema.cpp

@@ -0,0 +1,514 @@
+#include "json-schema.h"
+#include "common.h"
+
+#include <cmath>
+#include <map>
+#include <stdexcept>
+#include <string>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+class common_chat_schema_builder {
+    const common_json &      root_;
+    common_chat_schema_document & doc_;
+
+    // the targets built here, moved into doc_ once the whole schema is built
+    std::map<std::string, common_chat_schema_ptr> refs_;
+
+    // ref nodes get their target once every $ref is built, a cycle would otherwise need it too early
+    std::vector<common_chat_schema_ref *> pending_;
+
+    [[noreturn]] static void fail(const std::string & path, const std::string & msg) {
+        throw std::runtime_error("JSON schema error at " + path + ": " + msg);
+    }
+
+    static int get_count(const common_json & schema, const std::string & key, const std::string & path, int def) {
+        if (!schema.contains(key)) {
+            return def;
+        }
+        const common_json & value = schema.at(key);
+        if (!value.is_number_integer() || value.get<int>() < 0) {
+            fail(path, key + " must be a non-negative integer");
+        }
+        return value.get<int>();
+    }
+
+    // a fractional bound is rounded inwards, towards the integers it still admits
+    static int64_t get_bound(const common_json & schema, const std::string & key, const std::string & path, bool round_up) {
+        const common_json & value = schema.at(key);
+        if (value.is_number_integer()) {
+            return value.get<int64_t>();
+        }
+        if (!value.is_number()) {
+            fail(path, key + " must be a number");
+        }
+        double d = value.get<double>();
+        return (int64_t) (round_up ? std::ceil(d) : std::floor(d));
+    }
+
+    static common_chat_schema::string_format get_format(const common_json & schema, const std::string & path) {
+        if (!schema.contains("format")) {
+            return common_chat_schema::FORMAT_NONE;
+        }
+        const common_json & value = schema.at("format");
+        if (!value.is_string()) {
+            fail(path, "format must be a string");
+        }
+        std::string format = value.get<std::string>();
+        if (format == "date") {
+            return common_chat_schema::FORMAT_DATE;
+        }
+        if (format == "time") {
+            return common_chat_schema::FORMAT_TIME;
+        }
+        if (format == "date-time") {
+            return common_chat_schema::FORMAT_DATE_TIME;
+        }
+        if (format == "uuid" || (format.size() == 5 && format.compare(0, 4, "uuid") == 0 && format[4] >= '1' && format[4] <= '5')) {
+            return common_chat_schema::FORMAT_UUID;
+        }
+        return common_chat_schema::FORMAT_NONE;
+    }
+
+    const common_json & resolve_ref(const std::string & ref, const std::string & path) {
+        const common_json * target = &root_;
+        auto tokens = string_split(ref.substr(1), "/");
+        for (size_t i = 1; i < tokens.size(); i++) {
+            const std::string & sel = tokens[i];
+            if (target->is_object() && target->contains(sel)) {
+                target = &target->at(sel);
+            } else if (target->is_array()) {
+                size_t idx;
+                try {
+                    idx = std::stoull(sel);
+                } catch (const std::logic_error &) {
+                    idx = target->size();
+                }
+                if (idx >= target->size()) {
+                    fail(path, "cannot resolve $ref " + ref + ", " + sel + " is out of range");
+                }
+                target = &target->at(idx);
+            } else {
+                fail(path, "cannot resolve $ref " + ref + ", " + sel + " not found");
+            }
+        }
+        return *target;
+    }
+
+    common_chat_schema_ptr build_ref(const common_json & value, const std::string & path) {
+        if (!value.is_string()) {
+            fail(path, "$ref must be a string");
+        }
+        std::string ref = value.get<std::string>();
+        if (ref.compare(0, 2, "#/") != 0) {
+            fail(path, "unsupported $ref " + ref + ", only references into the same document are supported");
+        }
+        if (refs_.find(ref) == refs_.end()) {
+            // reserve the key first, so that a cycle back to this $ref stops here
+            refs_[ref] = nullptr;
+            refs_[ref] = build_node(resolve_ref(ref, path), ref);
+        }
+        auto node = std::make_unique<common_chat_schema_ref>(ref);
+        pending_.push_back(node.get());
+        return node;
+    }
+
+    template <typename T>
+    common_chat_schema_ptr build_alternatives(const common_json & alts, const std::string & path) {
+        if (!alts.is_array()) {
+            fail(path, "must be an array of schemas");
+        }
+        if (alts.empty()) {
+            fail(path, "must not be empty");
+        }
+        auto node = std::make_unique<T>();
+        size_t i = 0;
+        for (const auto & alt : alts) {
+            node->children.push_back(build_node(alt, path + "/" + std::to_string(i++)));
+        }
+        return node;
+    }
+
+    common_chat_schema_ptr build_object(const common_json & schema, const std::string & path) {
+        auto node = std::make_unique<common_chat_schema_object>();
+
+        std::unordered_set<std::string> required;
+        if (schema.contains("required") && schema.at("required").is_array()) {
+            for (const auto & name : schema.at("required")) {
+                if (name.is_string()) {
+                    required.insert(name.get<std::string>());
+                }
+            }
+        }
+
+        if (schema.contains("properties")) {
+            const common_json & properties = schema.at("properties");
+            if (!properties.is_object()) {
+                fail(path, "properties must be an object");
+            }
+            for (const auto & [name, prop] : properties.items()) {
+                node->properties.push_back({name, build_node(prop, path + "/properties/" + name), required.count(name) > 0});
+            }
+        }
+
+        if (schema.contains("additionalProperties")) {
+            const common_json & additional = schema.at("additionalProperties");
+            if (additional.is_boolean()) {
+                if (additional.get<bool>()) {
+                    node->additional_properties = std::make_unique<common_chat_schema_any>();
+                }
+            } else if (additional.is_object()) {
+                node->additional_properties = build_node(additional, path + "/additionalProperties");
+            } else {
+                fail(path, "additionalProperties must be a boolean or a schema");
+            }
+        } else if (!schema.contains("properties")) {
+            // {"type": "object"} on its own accepts any object
+            node->additional_properties = std::make_unique<common_chat_schema_any>();
+        }
+
+        return node;
+    }
+
+    common_chat_schema_ptr build_array(const common_json & schema, const std::string & path) {
+        auto node = std::make_unique<common_chat_schema_array>();
+        if (schema.contains("items") || schema.contains("prefixItems")) {
+            // "items" wins when both are present; as in the converter, a schema instead of an array is the item schema
+            const std::string key = schema.contains("items") ? "items" : "prefixItems";
+            const common_json & items = schema.at(key);
+            if (items.is_array()) {
+                auto tuple = std::make_unique<common_chat_schema_tuple>();
+                size_t i = 0;
+                for (const auto & item : items) {
+                    tuple->items.push_back(build_node(item, path + "/" + key + "/" + std::to_string(i++)));
+                }
+                return tuple;
+            }
+            node->items = build_node(items, path + "/" + key);
+        } else {
+            node->items = std::make_unique<common_chat_schema_any>();
+        }
+        node->min_items = get_count(schema, "minItems", path, 0);
+        node->max_items = get_count(schema, "maxItems", path, -1);
+        return node;
+    }
+
+    common_chat_schema_ptr build_string(const common_json & schema, const std::string & path) {
+        auto node = std::make_unique<common_chat_schema_string>();
+        if (schema.contains("pattern")) {
+            const common_json & pattern = schema.at("pattern");
+            if (!pattern.is_string()) {
+                fail(path, "pattern must be a string");
+            }
+            node->pattern = pattern.get<std::string>();
+        }
+        node->format     = get_format(schema, path);
+        node->min_length = get_count(schema, "minLength", path, 0);
+        node->max_length = get_count(schema, "maxLength", path, -1);
+        return node;
+    }
+
+    common_chat_schema_ptr build_integer(const common_json & schema, const std::string & path) {
+        auto node = std::make_unique<common_chat_schema_integer>();
+        if (schema.contains("minimum")) {
+            node->minimum = get_bound(schema, "minimum", path, /* round_up */ true);
+        } else if (schema.contains("exclusiveMinimum")) {
+            node->minimum = get_bound(schema, "exclusiveMinimum", path, /* round_up */ false) + 1;
+        }
+        if (schema.contains("maximum")) {
+            node->maximum = get_bound(schema, "maximum", path, /* round_up */ false);
+        } else if (schema.contains("exclusiveMaximum")) {
+            node->maximum = get_bound(schema, "exclusiveMaximum", path, /* round_up */ true) - 1;
+        }
+        return node;
+    }
+
+    common_chat_schema_ptr build_node(const common_json & schema, const std::string & path) {
+        if (!schema.is_object()) {
+            fail(path, "schema must be an object");
+        }
+        if (schema.contains("$ref")) {
+            return build_ref(schema.at("$ref"), path);
+        }
+        if (schema.contains("oneOf") || schema.contains("anyOf")) {
+            const std::string key = schema.contains("oneOf") ? "oneOf" : "anyOf";
+            return build_alternatives<common_chat_schema_any_of>(schema.at(key), path + "/" + key);
+        }
+
+        common_json type;
+        if (schema.contains("type")) {
+            type = schema.at("type");
+        }
+        if (type.is_array()) {
+            // {"type": ["a", "b"], ...} is {"anyOf": [{"type": "a", ...}, {"type": "b", ...}]}
+            if (type.empty()) {
+                fail(path, "type must not be empty");
+            }
+            auto node = std::make_unique<common_chat_schema_any_of>();
+            size_t i = 0;
+            for (const auto & t : type) {
+                common_json alt = schema;
+                alt["type"] = t;
+                node->children.push_back(build_node(alt, path + "/type/" + std::to_string(i++)));
+            }
+            return node;
+        }
+        if (schema.contains("const")) {
+            return std::make_unique<common_chat_schema_const>(schema.at("const"));
+        }
+        if (schema.contains("enum")) {
+            const common_json & values = schema.at("enum");
+            if (!values.is_array() || values.empty()) {
+                fail(path, "enum must be a non-empty array");
+            }
+            auto node = std::make_unique<common_chat_schema_enum>();
+            for (const auto & value : values) {
+                node->values.push_back(value);
+            }
+            return node;
+        }
+        if (!type.is_null() && !type.is_string()) {
+            fail(path, "type must be a string or an array of strings");
+        }
+
+        const std::string type_name = type.is_string() ? type.get<std::string>() : "";
+        const bool has_properties = schema.contains("properties") ||
+            (schema.contains("additionalProperties") && schema.at("additionalProperties") != true);
+
+        if (type_name.empty()) {
+            // without a type the structural keywords decide, in the same order as the converter
+            if (has_properties) {
+                return build_object(schema, path);
+            }
+            if (schema.contains("allOf")) {
+                return build_alternatives<common_chat_schema_all_of>(schema.at("allOf"), path + "/allOf");
+            }
+            if (schema.contains("items") || schema.contains("prefixItems")) {
+                return build_array(schema, path);
+            }
+            if (schema.contains("pattern") || schema.contains("minLength") || schema.contains("maxLength") || get_format(schema, path) != common_chat_schema::FORMAT_NONE) {
+                return build_string(schema, path);
+            }
+            return std::make_unique<common_chat_schema_any>();
+        }
+        if (type_name == "object") {
+            if (!has_properties && schema.contains("allOf")) {
+                return build_alternatives<common_chat_schema_all_of>(schema.at("allOf"), path + "/allOf");
+            }
+            return build_object(schema, path);
+        }
+        if (type_name == "string") {
+            if (schema.contains("allOf")) {
+                return build_alternatives<common_chat_schema_all_of>(schema.at("allOf"), path + "/allOf");
+            }
+            return build_string(schema, path);
+        }
+        if (type_name == "array") {
+            return build_array(schema, path);
+        }
+        if (type_name == "integer") {
+            return build_integer(schema, path);
+        }
+        if (type_name == "number") {
+            return std::make_unique<common_chat_schema_number>();
+        }
+        if (type_name == "boolean") {
+            return std::make_unique<common_chat_schema_boolean>();
+        }
+        if (type_name == "null") {
+            return std::make_unique<common_chat_schema_null>();
+        }
+        fail(path, "unrecognized type " + type_name);
+    }
+
+  public:
+    common_chat_schema_builder(const common_json & root, common_chat_schema_document & doc) : root_(root), doc_(doc) {}
+
+    common_chat_schema_ptr build() {
+        auto node = build_node(root_, "#");
+        for (auto & entry : refs_) {
+            doc_.refs[entry.first] = std::move(entry.second);
+        }
+        for (auto * ref : pending_) {
+            ref->target = doc_.refs.at(ref->ref).get();
+        }
+        return node;
+    }
+};
+
+common_chat_schema_document common_chat_schema_from_json(const common_json & schema) {
+    common_chat_schema_document doc;
+    doc.root = common_chat_schema_builder(schema, doc).build();
+    return doc;
+}
+
+static common_chat_schema::value_type json_type(const common_json & value) {
+    if (value.is_null()) {
+        return common_chat_schema::TYPE_NULL;
+    }
+    if (value.is_boolean()) {
+        return common_chat_schema::TYPE_BOOLEAN;
+    }
+    if (value.is_number_integer()) {
+        return common_chat_schema::TYPE_INTEGER;
+    }
+    if (value.is_number()) {
+        return common_chat_schema::TYPE_NUMBER;
+    }
+    if (value.is_string()) {
+        return common_chat_schema::TYPE_STRING;
+    }
+    if (value.is_array()) {
+        return common_chat_schema::TYPE_ARRAY;
+    }
+    return common_chat_schema::TYPE_OBJECT;
+}
+
+static common_chat_schema::type_set value_types_impl(const common_chat_schema & s, std::unordered_set<const common_chat_schema *> & visited) {
+    switch (s.kind()) {
+        case common_chat_schema::KIND_ANY:
+            return common_chat_schema::type_set::all();
+        case common_chat_schema::KIND_NULL:
+            return { common_chat_schema::TYPE_NULL };
+        case common_chat_schema::KIND_BOOLEAN:
+            return { common_chat_schema::TYPE_BOOLEAN };
+        case common_chat_schema::KIND_NUMBER:
+            return { common_chat_schema::TYPE_NUMBER, common_chat_schema::TYPE_INTEGER };
+        case common_chat_schema::KIND_INTEGER:
+            return { common_chat_schema::TYPE_INTEGER };
+        case common_chat_schema::KIND_STRING:
+            return { common_chat_schema::TYPE_STRING };
+        case common_chat_schema::KIND_ARRAY:
+        case common_chat_schema::KIND_TUPLE:
+            return { common_chat_schema::TYPE_ARRAY };
+        case common_chat_schema::KIND_OBJECT:
+            return { common_chat_schema::TYPE_OBJECT };
+        case common_chat_schema::KIND_CONST:
+            return { json_type(static_cast<const common_chat_schema_const &>(s).value) };
+        case common_chat_schema::KIND_ENUM: {
+            common_chat_schema::type_set types;
+            for (const auto & value : static_cast<const common_chat_schema_enum &>(s).values) {
+                types.add(json_type(value));
+            }
+            return types;
+        }
+        case common_chat_schema::KIND_REF: {
+            const auto * target = static_cast<const common_chat_schema_ref &>(s).target;
+            if (!target || !visited.insert(target).second) {
+                // a cycle contributes no type, to be safe
+                return {};
+            }
+            auto types = value_types_impl(*target, visited);
+            visited.erase(target);
+            return types;
+        }
+        case common_chat_schema::KIND_ANY_OF: {
+            common_chat_schema::type_set types;
+            for (const auto & child : static_cast<const common_chat_schema_any_of &>(s).children) {
+                types |= value_types_impl(*child, visited);
+            }
+            return types;
+        }
+        case common_chat_schema::KIND_ALL_OF: {
+            auto types = common_chat_schema::type_set::all();
+            for (const auto & child : static_cast<const common_chat_schema_all_of &>(s).children) {
+                types &= value_types_impl(*child, visited);
+            }
+            return types;
+        }
+    }
+    return {};
+}
+
+common_chat_schema::type_set common_chat_schema::value_types() const {
+    std::unordered_set<const common_chat_schema *> visited;
+    return value_types_impl(*this, visited);
+}
+
+static bool may_be_string_impl(const common_chat_schema & s, std::unordered_set<const common_chat_schema *> & visited) {
+    switch (s.kind()) {
+        case common_chat_schema::KIND_STRING:
+            return true;
+        case common_chat_schema::KIND_CONST:
+            return static_cast<const common_chat_schema_const &>(s).value.is_string();
+        case common_chat_schema::KIND_ENUM:
+            for (const auto & v : static_cast<const common_chat_schema_enum &>(s).values) {
+                if (v.is_string()) {
+                    return true;
+                }
+            }
+            return false;
+        case common_chat_schema::KIND_REF: {
+            // a cycle is taken as not a string, to be safe
+            const auto * target = static_cast<const common_chat_schema_ref &>(s).target;
+            if (!target || !visited.insert(target).second) {
+                return false;
+            }
+            bool result = may_be_string_impl(*target, visited);
+            visited.erase(target);
+            return result;
+        }
+        case common_chat_schema::KIND_ANY_OF:
+            for (const auto & child : static_cast<const common_chat_schema_any_of &>(s).children) {
+                if (may_be_string_impl(*child, visited)) {
+                    return true;
+                }
+            }
+            return false;
+        case common_chat_schema::KIND_ALL_OF: {
+            // every child must allow a string, an any child constrains nothing
+            bool any_string = false;
+            for (const auto & child : static_cast<const common_chat_schema_all_of &>(s).children) {
+                if (child->kind() == common_chat_schema::KIND_ANY) {
+                    continue;
+                }
+                if (!may_be_string_impl(*child, visited)) {
+                    return false;
+                }
+                any_string = true;
+            }
+            return any_string;
+        }
+        default:
+            return false;
+    }
+}
+
+bool common_chat_schema::may_be_string() const {
+    std::unordered_set<const common_chat_schema *> visited;
+    return may_be_string_impl(*this, visited);
+}
+
+const char * common_chat_schema::kind_name(node_kind kind) {
+    switch (kind) {
+        case KIND_ANY:     return "any";
+        case KIND_REF:     return "ref";
+        case KIND_ANY_OF:  return "anyOf";
+        case KIND_ALL_OF:  return "allOf";
+        case KIND_CONST:   return "const";
+        case KIND_ENUM:    return "enum";
+        case KIND_NULL:    return "null";
+        case KIND_BOOLEAN: return "boolean";
+        case KIND_NUMBER:  return "number";
+        case KIND_INTEGER: return "integer";
+        case KIND_STRING:  return "string";
+        case KIND_ARRAY:   return "array";
+        case KIND_TUPLE:   return "tuple";
+        case KIND_OBJECT:  return "object";
+    }
+    return "?";
+}
+
+const char * common_chat_schema::type_name(value_type type) {
+    switch (type) {
+        case TYPE_NULL:    return "null";
+        case TYPE_BOOLEAN: return "boolean";
+        case TYPE_NUMBER:  return "number";
+        case TYPE_INTEGER: return "integer";
+        case TYPE_STRING:  return "string";
+        case TYPE_ARRAY:   return "array";
+        case TYPE_OBJECT:  return "object";
+    }
+    return "?";
+}

common/json-schema.h

@@ -0,0 +1,198 @@
+#pragma once
+
+#include "json.h"
+
+#include <cstdint>
+#include <initializer_list>
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+// JSON schema, covering the subset that json_schema_to_grammar() can convert.
+
+struct common_chat_schema {
+    enum node_kind {
+        KIND_ANY,
+        KIND_REF,
+        KIND_ANY_OF,
+        KIND_ALL_OF,
+        KIND_CONST,
+        KIND_ENUM,
+        KIND_NULL,
+        KIND_BOOLEAN,
+        KIND_NUMBER,
+        KIND_INTEGER,
+        KIND_STRING,
+        KIND_ARRAY,
+        KIND_TUPLE,
+        KIND_OBJECT,
+    };
+
+    enum value_type {
+        TYPE_NULL,
+        TYPE_BOOLEAN,
+        TYPE_NUMBER,
+        TYPE_INTEGER,
+        TYPE_STRING,
+        TYPE_ARRAY,
+        TYPE_OBJECT,
+    };
+
+    enum string_format {
+        FORMAT_NONE,
+        FORMAT_UUID,  // uuid, uuid1 .. uuid5
+        FORMAT_DATE,
+        FORMAT_TIME,
+        FORMAT_DATE_TIME,
+    };
+
+    class type_set {
+        uint32_t mask_ = 0;
+
+      public:
+        type_set() = default;
+        type_set(std::initializer_list<value_type> types) {
+            for (auto type : types) {
+                add(type);
+            }
+        }
+
+        static type_set all() {
+            return { TYPE_NULL, TYPE_BOOLEAN, TYPE_NUMBER, TYPE_INTEGER, TYPE_STRING, TYPE_ARRAY, TYPE_OBJECT };
+        }
+
+        void add(value_type type) { mask_ |= 1u << type; }
+
+        bool has(value_type type) const { return (mask_ & (1u << type)) != 0; }
+        bool is_only(value_type type) const { return mask_ == (1u << type); }
+        bool empty() const { return mask_ == 0; }
+
+        type_set & operator|=(const type_set & other) { mask_ |= other.mask_; return *this; }
+        type_set & operator&=(const type_set & other) { mask_ &= other.mask_; return *this; }
+
+        bool operator==(const type_set & other) const { return mask_ == other.mask_; }
+        bool operator!=(const type_set & other) const { return mask_ != other.mask_; }
+    };
+
+    virtual ~common_chat_schema() = default;
+    virtual node_kind kind() const = 0;
+
+    type_set value_types() const;
+
+    // Whether a value matching the schema may be a string, through any branch of it.
+    bool may_be_string() const;
+
+    static const char * kind_name(node_kind kind);
+    static const char * type_name(value_type type);
+};
+
+using common_chat_schema_ptr = std::unique_ptr<common_chat_schema>;
+
+struct common_chat_schema_any : common_chat_schema {
+    node_kind kind() const override { return KIND_ANY; }
+};
+
+// {"$ref": "#/..."}, only references into the same document are supported
+struct common_chat_schema_ref : common_chat_schema {
+    std::string                ref;
+    const common_chat_schema * target = nullptr;  // owned by common_chat_schema_document::refs
+
+    explicit common_chat_schema_ref(std::string ref) : ref(std::move(ref)) {}
+
+    node_kind kind() const override { return KIND_REF; }
+};
+
+// oneOf / anyOf, or a "type" array expanded to one alternative per type
+struct common_chat_schema_any_of : common_chat_schema {
+    std::vector<common_chat_schema_ptr> children;
+
+    node_kind kind() const override { return KIND_ANY_OF; }
+};
+
+struct common_chat_schema_all_of : common_chat_schema {
+    std::vector<common_chat_schema_ptr> children;
+
+    node_kind kind() const override { return KIND_ALL_OF; }
+};
+
+struct common_chat_schema_const : common_chat_schema {
+    common_json value;
+
+    explicit common_chat_schema_const(common_json value) : value(std::move(value)) {}
+
+    node_kind kind() const override { return KIND_CONST; }
+};
+
+struct common_chat_schema_enum : common_chat_schema {
+    std::vector<common_json> values;
+
+    node_kind kind() const override { return KIND_ENUM; }
+};
+
+struct common_chat_schema_null : common_chat_schema {
+    node_kind kind() const override { return KIND_NULL; }
+};
+
+struct common_chat_schema_boolean : common_chat_schema {
+    node_kind kind() const override { return KIND_BOOLEAN; }
+};
+
+struct common_chat_schema_number : common_chat_schema {
+    node_kind kind() const override { return KIND_NUMBER; }
+};
+
+// bounds are inclusive, exclusiveMinimum / exclusiveMaximum are folded in
+struct common_chat_schema_integer : common_chat_schema {
+    int64_t minimum = INT64_MIN;  // INT64_MIN for unbounded
+    int64_t maximum = INT64_MAX;  // INT64_MAX for unbounded
+
+    node_kind kind() const override { return KIND_INTEGER; }
+};
+
+struct common_chat_schema_string : common_chat_schema {
+    std::string   pattern;  // empty when absent
+    string_format format     = FORMAT_NONE;
+    int           min_length = 0;
+    int           max_length = -1;  // -1 for unbounded
+
+    node_kind kind() const override { return KIND_STRING; }
+};
+
+struct common_chat_schema_array : common_chat_schema {
+    common_chat_schema_ptr items;  // a common_chat_schema_any when "items" is absent
+    int                    min_items = 0;
+    int                    max_items = -1;  // -1 for unbounded
+
+    node_kind kind() const override { return KIND_ARRAY; }
+};
+
+struct common_chat_schema_tuple : common_chat_schema {
+    std::vector<common_chat_schema_ptr> items;
+
+    node_kind kind() const override { return KIND_TUPLE; }
+};
+
+struct common_chat_schema_property {
+    std::string            name;
+    common_chat_schema_ptr schema;
+    bool                   required = false;
+};
+
+struct common_chat_schema_object : common_chat_schema {
+    std::vector<common_chat_schema_property> properties;             // in schema order
+    common_chat_schema_ptr                   additional_properties;  // null when not allowed
+
+    node_kind kind() const override { return KIND_OBJECT; }
+};
+
+struct common_chat_schema_document {
+    common_chat_schema_ptr                        root;
+    std::map<std::string, common_chat_schema_ptr> refs;
+};
+
+// A document shared by the PEG parsers built from its nodes, which it keeps alive
+using common_chat_schema_document_ptr = std::shared_ptr<const common_chat_schema_document>;
+
+// Throws std::runtime_error when the schema falls outside the supported subset.
+common_chat_schema_document common_chat_schema_from_json(const common_json & schema);

common/log.cpp

@@ -37,6 +37,16 @@ void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
+static bool common_log_jsonl = false;
+
+bool common_log_get_jsonl(void) {
+    return common_log_jsonl;
+}
+
+void common_log_set_jsonl(bool jsonl) {
+    common_log_jsonl = jsonl;
+}
+
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -87,6 +97,7 @@ struct common_log_entry {
     bool is_end       { false }; // signals the worker thread to stop
     bool prefix       { false };
     bool jsonl        { false };
+    bool is_json      { false }; // msg already holds a serialized JSON object
 
     common_log_entry(size_t size = 256) : msg(size) { }
 
@@ -107,6 +118,12 @@ struct common_log_entry {
         }
 
         if (jsonl) {
+            if (is_json) {
+                fprintf(fcur, "%s\n", msg.data());
+                fflush(fcur);
+                return;
+            }
+
             common_json obj = {
                 {"type",  "log"},
                 {"time",  timestamp},
@@ -156,7 +173,6 @@ struct common_log {
         file       = nullptr;
         prefix     = false;
         timestamps = false;
-        jsonl      = false;
         running    = false;
         t_start    = t_us();
 
@@ -184,7 +200,6 @@ private:
 
     bool prefix;
     bool timestamps;
-    bool jsonl;
     bool running;
 
     int64_t t_start;
@@ -273,7 +288,8 @@ public:
         entry.is_end    = false;
         entry.level     = level;
         entry.prefix    = prefix;
-        entry.jsonl     = jsonl;
+        entry.jsonl     = common_log_jsonl;
+        entry.is_json   = false;
         entry.timestamp = 0;
         if (timestamps) {
             entry.timestamp = t_us() - t_start;
@@ -283,6 +299,42 @@ public:
         cv_new.notify_one();
     }
 
+    void add_json(const char * type, const common_json & obj) {
+        const common_json full = {
+            {"type", type},
+            {"data", obj},
+        };
+
+        const std::string text = full.dump_safe();
+
+        std::unique_lock<std::mutex> lock(mtx);
+
+        // block if the queue is full
+        cv_full.wait(lock, [this]() { return !running || !is_full(); });
+
+        if (!running) {
+            // discard messages while the worker thread is paused
+            return;
+        }
+
+        auto & entry = queue[tail];
+
+        if (entry.msg.size() < text.size() + 1) {
+            entry.msg.resize(text.size() + 1);
+        }
+        memcpy(entry.msg.data(), text.c_str(), text.size() + 1);
+
+        entry.is_end    = false;
+        entry.level     = GGML_LOG_LEVEL_NONE;
+        entry.prefix    = false;
+        entry.jsonl     = true;
+        entry.is_json   = true;
+        entry.timestamp = 0;
+
+        tail = (tail + 1) % queue.size();
+        cv_new.notify_one();
+    }
+
     void resume() {
         std::lock_guard<std::mutex> lock(mtx);
 
@@ -388,12 +440,6 @@ public:
 
         this->timestamps = timestamps;
     }
-
-    void set_jsonl(bool jsonl) {
-        std::lock_guard<std::mutex> lock(mtx);
-
-        this->jsonl = jsonl;
-    }
 };
 
 //
@@ -440,6 +486,14 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
     va_end(args);
 }
 
+void common_log_add_json(struct common_log * log, const char * type, const common_json & obj) {
+    if (!common_log_jsonl) {
+        return;
+    }
+
+    log->add_json(type, obj);
+}
+
 void common_log_set_file(struct common_log * log, const char * file) {
     log->set_file(file);
 }
@@ -467,10 +521,6 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
-void common_log_set_jsonl(struct common_log * log, bool jsonl) {
-    log->set_jsonl(jsonl);
-}
-
 void common_log_flush(struct common_log * log) {
     log->pause();
     log->resume();

common/log.h

@@ -43,6 +43,10 @@ int  common_log_get_verbosity_thold(void);
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
+bool common_log_get_jsonl(void);
+
+void common_log_set_jsonl(bool jsonl); // not thread-safe
+
 int common_log_get_verbosity(enum ggml_log_level level);
 
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
@@ -91,7 +95,6 @@ void common_log_set_file      (struct common_log * log, const char * file); // n
 void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
 void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
 void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
-void common_log_set_jsonl     (struct common_log * log, bool jsonl);        // print each log as a JSON object on one line, not thread-safe
 void common_log_flush         (struct common_log * log);                    // flush all pending log messages
 
 // helper macros for logging
@@ -127,3 +130,17 @@ void common_log_flush         (struct common_log * log);                    // f
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)
 #define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
 #define LOG_CNTV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  verbosity, __VA_ARGS__)
+
+class common_json; // defined in common/json.h
+
+// helper allows different types of json output
+// no-op if --log-jsonl is not set
+void common_log_add_json(struct common_log * log, const char * type, const common_json & data);
+
+// will only print if --log-jsonl is set
+#define LOG_JSON(type, data) \
+    do { \
+        if (common_log_get_jsonl()) { \
+            common_log_add_json(common_log_main(), type, data); \
+        } \
+    } while (0)

common/parsers/cohere2moe.cpp

@@ -129,15 +129,6 @@ common_chat_params common_chat_params_init_cohere2moe(const common_chat_template
     if (include_grammar) {
         data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/deepseek.cpp

@@ -149,39 +149,28 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
                 std::string  name     = function.at("name");
-                auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
-                const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
-
-                std::set<std::string> required;
-                if (params.contains("required")) {
-                    required = params.at("required").get<std::set<std::string>>();
-                }
-
-                auto schema_info = common_schema_info();
-                schema_info.resolve_refs(params);
 
                 std::vector<common_peg_parser> required_parsers;
                 std::vector<common_peg_parser> optional_parsers;
-                for (const auto & [param_name, param_schema] : props.items()) {
-                    bool is_required = required.find(param_name) != required.end();
-                    bool is_string   = schema_info.resolves_to_string(param_schema);
+                foreach_parameter(function, [&](const common_chat_schema_property & param, const common_chat_schema_document_ptr & doc) {
+                    bool is_string = param.schema->may_be_string();
 
                     auto arg = p.tool_arg(
-                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param_name)) +
+                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param.name)) +
                                         p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
                         (is_string ?
                              p.tool_arg_string_value(p.until(PARAM_END)) :
-                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param_name + "-schema",
-                                                            param_schema, false))) +
+                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param.name + "-schema",
+                                                            doc, *param.schema))) +
                         p.tool_arg_close(p.literal(PARAM_END)));
 
-                    auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-                    if (is_required) {
+                    auto named_arg = p.rule("tool-" + name + "-arg-" + param.name, arg);
+                    if (param.required) {
                         required_parsers.push_back(named_arg);
                     } else {
                         optional_parsers.push_back(named_arg);
                     }
-                }
+                });
 
                 common_peg_parser args_seq = p.eps();
                 for (size_t i = 0; i < required_parsers.size(); i++) {
@@ -266,15 +255,6 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
     if (include_grammar) {
         data.grammar_lazy = has_tools && !require_tools;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/functionary-v3-2.cpp

@@ -45,7 +45,7 @@ common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_te
         foreach_function(inputs.tools, [&](const json & tool) {
             const auto & function = tool.at("function");
             std::string  name     = function.at("name");
-            const auto & schema   = function.at("parameters");
+            const auto   schema   = common_chat_tool_parameters(function);
 
             // Tool format: >>>function_name\n{json_args}
             auto tool_parser = p.tool(
@@ -82,11 +82,6 @@ common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_te
         data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
 
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/gemma4.cpp

@@ -291,15 +291,6 @@ common_chat_params common_chat_params_init_gemma4(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/gigachat-v3.cpp

@@ -33,7 +33,7 @@ common_chat_params common_chat_params_init_gigachat_v3(
             for (const auto & tool : inputs.tools) {
                 const auto & function = tool.at("function");
                 std::string name = function.at("name");
-                const auto & schema = function.at("parameters");
+                const auto  schema = common_chat_tool_parameters(function);
 
                 auto tool_name = p.json_member("name", "\"" + p.tool_name(p.literal(name)) + "\"");
                 auto tool_args = p.json_member("arguments", p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
@@ -65,11 +65,6 @@ common_chat_params common_chat_params_init_gigachat_v3(
         data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
 
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto schema = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/gpt-oss.cpp

@@ -109,7 +109,7 @@ common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
                 std::string  name     = function.at("name");
-                const auto & params   = function.at("parameters");
+                const auto   params   = common_chat_tool_parameters(function);
 
                 auto func_name  = p.literal(" to=functions.") + p.tool_name(p.literal(name));
                 auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
@@ -143,15 +143,6 @@ common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/kimi-k2.cpp

@@ -82,7 +82,7 @@ common_chat_params common_chat_params_init_kimi_k2(const common_chat_template &
         foreach_function(inputs.tools, [&](const json & tool) {
             const auto & function = tool.at("function");
             std::string  name     = function.at("name");
-            const auto & schema   = function.at("parameters");
+            const auto   schema   = common_chat_tool_parameters(function);
 
             // Match: functions.<name>:<digits>
             // Capture the full call id (functions.<name>:<digits>) using tool_id tag
@@ -116,11 +116,6 @@ common_chat_params common_chat_params_init_kimi_k2(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/kimi-k3.cpp

@@ -98,7 +98,7 @@ common_chat_params common_chat_params_init_kimi_k3(const common_chat_template &
         foreach_function(inputs.tools, [&](const json & tool) {
             const auto & function = tool.at("function");
             std::string  name     = function.at("name");
-            const json   schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+            const json   schema   = common_chat_tool_parameters(function);
 
             // arguments come one tag per key, with the JSON type in a type="..."
             // attribute. the type is taken from the tool schema instead, as it tells
@@ -155,13 +155,6 @@ common_chat_params common_chat_params_init_kimi_k3(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                if (function.contains("parameters")) {
-                    auto schema = function.at("parameters");
-                    builder.resolve_refs(schema);
-                }
-            });
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/lfm2.cpp

@@ -98,15 +98,6 @@ common_chat_params common_chat_params_init_lfm2(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/minicpm5.cpp

@@ -71,32 +71,27 @@ common_chat_params common_chat_params_init_minicpm5(const common_chat_template &
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto &      function = tool.at("function");
                 const std::string name     = function.at("name");
-                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                std::vector<common_peg_parser> arg_rules;
+                foreach_parameter(function, [&](const common_chat_schema_property & prop, const common_chat_schema_document_ptr & doc) {
+                    auto value_parser = p.eps();
+                    if (prop.schema->may_be_string()) {
+                        value_parser = string_value;
+                    } else {
+                        value_parser = p.tool_arg_json_value(
+                                p.schema(p.json(), "tool-" + name + "-arg-" + prop.name + "-schema", doc, *prop.schema)
+                            ) + p.tool_arg_close(p.literal("</param>"));
+                    }
+
+                    arg_rules.push_back(p.tool_arg(
+                        p.tool_arg_open(p.literal("<param name=\"") + p.tool_arg_name(p.literal(prop.name)) + p.literal("\">")) +
+                        value_parser
+                    ));
+                });
 
                 auto args = p.eps();
-                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
-                    auto schema_info = common_schema_info();
-                    schema_info.resolve_refs(params);
-
-                    auto arg_choice = p.choice();
-                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
-                        auto value_parser = p.eps();
-                        if (schema_info.resolves_to_string(prop_schema)) {
-                            value_parser = string_value;
-                        } else {
-                            value_parser = p.tool_arg_json_value(
-                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false)
-                                ) + p.tool_arg_close(p.literal("</param>"));
-                        }
-
-                        auto arg_rule = p.tool_arg(
-                            p.tool_arg_open(p.literal("<param name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
-                            value_parser
-                        );
-
-                        arg_choice |= arg_rule;
-                    }
-                    args = p.zero_or_more(arg_choice + p.space());
+                if (!arg_rules.empty()) {
+                    args = p.zero_or_more(p.choice(arg_rules) + p.space());
                 }
 
                 auto tool_parser = p.tool(
@@ -123,15 +118,6 @@ common_chat_params common_chat_params_init_minicpm5(const common_chat_template &
     if (include_grammar) {
         data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/minimax-m3.cpp

@@ -84,29 +84,18 @@ common_chat_params common_chat_params_init_minimax_m3(const common_chat_template
             return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
 
-        auto alternatives_of = [](const json & schema) -> std::optional<json> {
-            for (const auto * keyword : { "oneOf", "anyOf" }) {
-                if (schema.contains(keyword) && schema.at(keyword).is_array() && !schema.at(keyword).empty()) {
-                    return schema.at(keyword);
-                }
-            }
-            return std::nullopt;
-        };
-
         auto tool_choice = p.choice();
         foreach_function(inputs.tools, [&](const json & tool) {
             const auto & function = tool.at("function");
             std::string  name     = function.at("name");
-            auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-            auto schema_info = common_schema_info();
-            schema_info.resolve_refs(params);
+            auto         params   = common_chat_tool_parameters(function);
+            auto         doc      = std::make_shared<const common_chat_schema_document>(common_chat_schema_from_json(params));
 
             // The template expands argument values recursively in XML (see the to_xml() macro)
-            std::function<common_peg_parser(const json &, const std::string &, const std::string &)> value_of;
-            std::function<common_peg_parser(const json &, const std::string &)>                      members_of;
+            std::function<common_peg_parser(const common_chat_schema &, const std::string &, const std::string &)> value_of;
+            std::function<common_peg_parser(const common_chat_schema_object &, const std::string &)>              members_of;
 
-            auto element_of = [&](const std::string & tag, const json & schema, const std::string & rule_name) {
+            auto element_of = [&](const std::string & tag, const common_chat_schema & schema, const std::string & rule_name) {
                 const std::string close = NS + "</" + tag + ">";
                 return p.rule(rule_name,
                     p.tool_arg(
@@ -117,69 +106,57 @@ common_chat_params common_chat_params_init_minimax_m3(const common_chat_template
                         value_of(schema, rule_name, close)));
             };
 
-            value_of = [&](const json & schema,
+            value_of = [&](const common_chat_schema & schema,
                            const std::string & rule_name,
                            const std::string & close) -> common_peg_parser {
                 auto close_tag = p.tool_arg_close(p.literal(close));
 
                 // A string accepts anything, so a union with a string alternative is a string
-                if (schema_info.resolves_to_string(schema)) {
+                if (schema.may_be_string()) {
                     return p.ac(p.tool_arg_string_value(p.until(close)) + close_tag, close);
                 }
 
-                if (auto alternatives = alternatives_of(schema)) {
+                if (schema.kind() == common_chat_schema::KIND_ANY_OF) {
                     std::vector<common_peg_parser> choices;
 
                     size_t index = 0;
-                    for (const auto & alternative : *alternatives) {
+                    for (const auto & alternative : static_cast<const common_chat_schema_any_of &>(schema).children) {
                         const std::string alt_name = rule_name + "-" + std::to_string(index++);
 
                         // There is a risk that this breaks streaming deltas, but that's a risk we
                         // assume to provide tool arg streaming.
-                        choices.push_back(value_of(alternative, alt_name, close));
+                        choices.push_back(value_of(*alternative, alt_name, close));
                     }
 
                     return p.choice(choices);
                 }
 
-                const std::string type = schema.contains("type") && schema.at("type").is_string()
-                                             ? schema.at("type").get<std::string>()
-                                             : "";
-
-                if (type == "object" && schema.contains("properties")) {
-                    return p.tag(mm3::TOOL_ARG_OBJECT, members_of(schema, rule_name)) + p.space() + close_tag;
+                if (schema.kind() == common_chat_schema::KIND_OBJECT) {
+                    const auto & object = static_cast<const common_chat_schema_object &>(schema);
+                    if (!object.properties.empty()) {
+                        return p.tag(mm3::TOOL_ARG_OBJECT, members_of(object, rule_name)) + p.space() + close_tag;
+                    }
                 }
 
-                if (type == "array" && schema.contains("items")) {
+                if (schema.kind() == common_chat_schema::KIND_ARRAY) {
                     const std::string item_close = NS + "</item>";
                     auto item = p.rule(rule_name + "-item",
                         p.tag(mm3::TOOL_ARG_ITEM,
                               p.literal(NS + "<item>") +
-                                  value_of(schema.at("items"), rule_name + "-item", item_close)));
+                                  value_of(*static_cast<const common_chat_schema_array &>(schema).items, rule_name + "-item", item_close)));
                     return p.tag(mm3::TOOL_ARG_ARRAY, p.repeat(p.space() + item, 0, -1)) + p.space() + close_tag;
                 }
 
-                return p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", schema, false)) + close_tag;
+                return p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", doc, schema)) + close_tag;
             };
 
             // Required properties in schema order, then any number of optional ones in any order.
-            members_of = [&](const json & schema, const std::string & rule_prefix) -> common_peg_parser {
-                const auto & props = schema.at("properties");
-
-                std::set<std::string> required;
-                if (schema.contains("required")) {
-                    required = schema.at("required").get<std::set<std::string>>();
-                }
-
+            members_of = [&](const common_chat_schema_object & object, const std::string & rule_prefix) -> common_peg_parser {
                 std::vector<common_peg_parser> required_elements;
                 std::vector<common_peg_parser> optional_elements;
-                for (const auto & [key, key_schema] : props.items()) {
-                    auto element = element_of(key, key_schema, rule_prefix + "-" + key);
-                    if (required.find(key) != required.end()) {
-                        required_elements.push_back(element);
-                    } else {
-                        optional_elements.push_back(element);
-                    }
+                for (const auto & prop : object.properties) {
+                    auto element = element_of(prop.name, *prop.schema, rule_prefix + "-" + prop.name);
+                    (prop.required ? required_elements : optional_elements).push_back(element);
                 }
 
                 common_peg_parser members = p.eps();
@@ -201,8 +178,10 @@ common_chat_params common_chat_params_init_minimax_m3(const common_chat_template
                 return members;
             };
 
-            common_peg_parser invoke_body =
-                params.contains("properties") ? members_of(params, "tool-" + name + "-arg") : p.eps();
+            common_peg_parser invoke_body = p.eps();
+            if (doc->root->kind() == common_chat_schema::KIND_OBJECT) {
+                invoke_body = members_of(static_cast<const common_chat_schema_object &>(*doc->root), "tool-" + name + "-arg");
+            }
 
             auto func_parser = p.tool(
                 p.tool_open(p.literal(NS + "<invoke name=\"") +
@@ -238,15 +217,6 @@ common_chat_params common_chat_params_init_minimax_m3(const common_chat_template
     if (include_grammar) {
         data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/ministral3.cpp

@@ -89,7 +89,7 @@ common_chat_params common_chat_params_init_ministral_3(const common_chat_templat
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
                 std::string  name     = function.at("name");
-                const auto & schema   = function.at("parameters");
+                const auto   schema   = common_chat_tool_parameters(function);
 
                 tool_choice |=
                     p.rule("tool-" + name, p.tool_open(p.tool_name(p.literal(name)) + "[ARGS]") +
@@ -114,15 +114,6 @@ common_chat_params common_chat_params_init_ministral_3(const common_chat_templat
         data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
 
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/parsers/muse-glimmer.cpp

@@ -74,31 +74,26 @@ common_chat_params common_chat_params_init_muse_glimmer(const common_chat_templa
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto &      function = tool.at("function");
                 const std::string name     = function.at("name");
-                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                std::vector<common_peg_parser> arg_rules;
+                foreach_parameter(function, [&](const common_chat_schema_property & prop, const common_chat_schema_document_ptr & doc) {
+                    auto value_parser = p.eps();
+                    if (prop.schema->may_be_string()) {
+                        value_parser = string_value;
+                    } else {
+                        value_parser = p.tool_arg_json_value(
+                                p.schema(p.json(), "tool-" + name + "-arg-" + prop.name + "-schema", doc, *prop.schema))
+                            + p.tool_arg_close(p.literal("</atem:parameter>"));
+                    }
+
+                    arg_rules.push_back(p.tool_arg(
+                        p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop.name)) + p.literal("\">")) +
+                        value_parser));
+                });
 
                 auto args = p.eps();
-                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
-                    auto schema_info = common_schema_info();
-                    schema_info.resolve_refs(params);
-
-                    auto arg_choice = p.choice();
-                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
-                        auto value_parser = p.eps();
-                        if (schema_info.resolves_to_string(prop_schema)) {
-                            value_parser = string_value;
-                        } else {
-                            value_parser = p.tool_arg_json_value(
-                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false))
-                                + p.tool_arg_close(p.literal("</atem:parameter>"));
-                        }
-
-                        auto arg_rule = p.tool_arg(
-                            p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
-                            value_parser);
-
-                        arg_choice |= arg_rule;
-                    }
-                    args = p.zero_or_more(arg_choice + p.space());
+                if (!arg_rules.empty()) {
+                    args = p.zero_or_more(p.choice(arg_rules) + p.space());
                 }
 
                 auto tool_parser = p.tool(
@@ -131,11 +126,6 @@ common_chat_params common_chat_params_init_muse_glimmer(const common_chat_templa
     if (include_grammar) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
             parser.build_grammar(builder, data.grammar_lazy);
         });
         data.grammar_triggers = {

common/parsers/parsers.cpp

@@ -2,8 +2,6 @@
 
 #include "log.h"
 
-#include <set>
-
 void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
     for (const auto & tool : tools) {
         if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
@@ -14,21 +12,14 @@ void foreach_function(const json & tools, const std::function<void(const json &)
     }
 }
 
-void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn) {
-    if (!function.contains("parameters") || !function.at("parameters").is_object()) {
+void foreach_parameter(const json & function, const std::function<void(const common_chat_schema_property &, const common_chat_schema_document_ptr &)> & fn) {
+    auto         params = common_chat_tool_parameters(function);
+    auto         doc    = std::make_shared<const common_chat_schema_document>(common_chat_schema_from_json(params));
+    const auto * object = dynamic_cast<const common_chat_schema_object *>(doc->root.get());
+    if (!object) {
         return;
     }
-    const auto & params = function.at("parameters");
-    if (!params.contains("properties") || !params.at("properties").is_object()) {
-        return;
-    }
-    const auto & props = params.at("properties");
-    std::set<std::string> required;
-    if (params.contains("required") && params.at("required").is_array()) {
-        required = params.at("required").get<std::set<std::string>>();
-    }
-    for (const auto & [name, prop] : props.items()) {
-        bool is_required = (required.find(name) != required.end());
-        fn(name, prop, is_required);
+    for (const auto & prop : object->properties) {
+        fn(prop, doc);
     }
 }

common/parsers/parsers.h

@@ -20,8 +20,8 @@ using json = common_json;
 // iterate over the function tools of an OpenAI-style tools array
 void foreach_function(const json & tools, const std::function<void(const json &)> & fn);
 
-// iterate over the parameters of a function tool, flagging the ones listed as required
-void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn);
+// iterate over the parameters of a function tool, with the document that owns them
+void foreach_parameter(const json & function, const std::function<void(const common_chat_schema_property &, const common_chat_schema_document_ptr &)> & fn);
 
 // render a template; the override arguments let a parser feed in messages, tools or context it has rewritten
 std::string common_chat_template_direct_apply_impl(

common/parsers/qwen3-coder.cpp

@@ -93,28 +93,49 @@ common_chat_params common_chat_params_init_qwen3_coder(const common_chat_templat
 
             auto tool_choice = p.choice();
             foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function   = tool.at("function");
-                std::string  name       = function.at("name");
-                auto         parameters = function.contains("parameters") ? function.at("parameters") : json::object();
-
-                auto schema_info = common_schema_info();
-                schema_info.resolve_refs(parameters);
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
 
                 std::vector<common_peg_parser> required_args;
                 std::vector<common_peg_parser> optional_args;
 
-                foreach_parameter(function, [&](const std::string & param_name, const json & param_schema, bool is_required) {
-                    auto rule_name = "tool-" + name + "-arg-" + param_name;
+                foreach_parameter(function, [&](const common_chat_schema_property & param, const common_chat_schema_document_ptr & doc) {
+                    auto rule_name = "tool-" + name + "-arg-" + param.name;
 
-                    auto arg_open = p.tool_arg_open("<parameter=" + p.tool_arg_name(p.literal(param_name)) + ">\n");
+                    auto arg_open = p.tool_arg_open("<parameter=" + p.tool_arg_name(p.literal(param.name)) + ">\n");
 
-                    auto arg_value = schema_info.resolves_to_string(param_schema) ?
-                        arg_string :
-                        p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", param_schema)) + arg_close;
+                    auto types = param.schema->value_types();
+
+                    auto arg_value = p.eps();
+                    if (!types.has(common_chat_schema::TYPE_STRING)) {
+                        arg_value = p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", doc, *param.schema)) + arg_close;
+                    } else if (types.is_only(common_chat_schema::TYPE_STRING)) {
+                        arg_value = arg_string;
+                    } else {
+                        // The string alternative accepts any text, so the grammar only keeps the raw string
+                        // rule. The parser still tries the JSON alternatives first to type the value.
+                        auto json_value = p.choice();
+                        if (types.has(common_chat_schema::TYPE_OBJECT)) {
+                            json_value |= p.json_object();
+                        }
+                        if (types.has(common_chat_schema::TYPE_ARRAY)) {
+                            json_value |= p.json_array();
+                        }
+                        if (types.has(common_chat_schema::TYPE_NUMBER) || types.has(common_chat_schema::TYPE_INTEGER)) {
+                            json_value |= p.json_number();
+                        }
+                        if (types.has(common_chat_schema::TYPE_BOOLEAN)) {
+                            json_value |= p.json_bool();
+                        }
+                        if (types.has(common_chat_schema::TYPE_NULL)) {
+                            json_value |= p.json_null();
+                        }
+                        arg_value = p.gbnf(p.atomic(p.tool_arg_json_value(json_value) + arg_close) | arg_string, "xml-arg-string");
+                    }
 
                     auto arg_rule = p.rule(rule_name, p.tool_arg(arg_open + arg_value));
 
-                    (is_required ? required_args : optional_args).push_back(arg_rule);
+                    (param.required ? required_args : optional_args).push_back(arg_rule);
                 });
 
                 // Accept required arguments in any order, as Qwen does not always adhere to the
@@ -158,15 +179,6 @@ common_chat_params common_chat_params_init_qwen3_coder(const common_chat_templat
         data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
 
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 

common/peg-parser.cpp

@@ -953,7 +953,7 @@ std::string common_peg_arena::dump_impl(common_peg_parser_id
         } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
             return "Until(" + string_join(p.delimiters, " | ") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
-            return "Schema(" + dump_impl(p.child, visited) + ", " + (p.schema ? p.schema->dump() : "null") + ")";
+            return "Schema(" + dump_impl(p.child, visited) + ", " + (p.node ? common_chat_schema::kind_name(p.node->kind()) : "null") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
             return "Rule(" + p.name + ", " + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_ref_parser>) {
@@ -1119,8 +1119,13 @@ common_peg_parser common_peg_parser_builder::chars(const std::string & classes,
     return wrap(arena_.add_parser(common_peg_chars_parser{classes, ranges, negated, min, max}));
 }
 
+common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, common_chat_schema_document_ptr doc, const common_chat_schema & node, bool raw) {
+    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::move(doc), &node, raw}));
+}
+
 common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw) {
-    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<common_json>(schema), raw}));
+    auto doc = std::make_shared<const common_chat_schema_document>(common_chat_schema_from_json(schema));
+    return this->schema(p, name, doc, *doc->root, raw);
 }
 
 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
@@ -1573,30 +1578,9 @@ static std::set<std::string> collect_reachable_rules(
 
 // GBNF generation implementation
 void common_peg_arena::build_grammar(const common_grammar_builder & builder, bool lazy) const {
+    // A raw string value is parsed by the child rather than constrained by the schema
     auto schema_delegates = [](const common_peg_schema_parser & s) -> bool {
-        if (!s.schema) {
-            return true;
-        }
-        if (s.raw && s.schema->contains("type")) {
-            const auto & type_val = s.schema->at("type");
-            if (type_val.is_string() && type_val == "string") {
-                return true;
-            }
-            // Handle nullable types like ["string", "null"] - delegate when the
-            // non-null type is string, since the tagged format uses raw text
-            if (type_val.is_array()) {
-                for (const auto & t : type_val) {
-                    if (t.is_string() && t.get<std::string>() != "null") {
-                        return t.get<std::string>() == "string";
-                    }
-                }
-            }
-        }
-        // Delegate for enum schemas in raw mode - enum values are literal strings
-        if (s.raw && !s.schema->contains("type") && s.schema->contains("enum")) {
-            return true;
-        }
-        return false;
+        return !s.node || (s.raw && s.node->may_be_string());
     };
 
     // Unwrap the parser so we can properly check if it's a sequence or choice
@@ -1731,7 +1715,7 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                 if (schema_delegates(p)) {
                     return to_gbnf(p.child);
                 }
-                return builder.add_schema(p.name, *p.schema);
+                return builder.add_schema(p.name, *p.node);
             } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
                 return p.name;
             } else if constexpr (std::is_same_v<T, common_peg_ref_parser>) {
@@ -1859,7 +1843,6 @@ static common_json serialize_parser_variant(const common_peg_parser_variant & va
                 {"type", "schema"},
                 {"child", p.child},
                 {"name", p.name},
-                {"schema", p.schema ? *p.schema : json(nullptr)},
                 {"raw", p.raw}
             };
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1999,15 +1982,12 @@ static common_peg_parser_variant deserialize_parser_variant(const common_json &
         return common_peg_until_parser{j["delimiters"].get<std::vector<std::string>>()};
     }
     if (type == "schema") {
-        if (!j.contains("child") || !j.contains("name") || !j.contains("schema") || !j.contains("raw")) {
+        if (!j.contains("child") || !j.contains("name") || !j.contains("raw")) {
             throw std::runtime_error("schema parser missing required fields");
         }
         common_peg_schema_parser parser;
         parser.child = j["child"].get<common_peg_parser_id>();
         parser.name = j["name"];
-        if (!j["schema"].is_null()) {
-            parser.schema = std::make_shared<common_json>(j["schema"]);
-        }
         parser.raw = j["raw"].get<bool>();
         return parser;
     }

common/peg-parser.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include "json-schema.h"
 #include "json.h"
 
 #include <memory>
@@ -245,7 +246,8 @@ struct common_peg_until_parser {
 struct common_peg_schema_parser {
     common_peg_parser_id child;
     std::string name;
-    std::shared_ptr<common_json> schema;
+    common_chat_schema_document_ptr doc;  // owns node
+    const common_chat_schema * node = nullptr;
 
     // Indicates if the GBNF should accept a raw string that matches the schema.
     bool raw;
@@ -488,8 +490,10 @@ class common_peg_parser_builder {
     // A marker, i.e. text delimited by a pair of <> or []
     common_peg_parser marker();
 
-    // Wraps a parser with JSON schema metadata for grammar generation.
-    // Used internally to convert JSON schemas to GBNF grammar rules.
+    // Wraps a parser with the schema its GBNF is generated from, a node of the document that owns it
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, common_chat_schema_document_ptr doc, const common_chat_schema & node, bool raw = false);
+
+    // Parses the JSON schema into a document of its own
     common_peg_parser schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw = false);
 
     // Creates a named rule, stores it in the grammar, and returns a ref.

common/speculative.cpp

@@ -296,7 +296,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
             drafting[seq_id] = true;
             common_sampler_reset(smpls[seq_id].get());
 
-            common_batch_add(batch, dp.id_last, dp.n_past, { seq_id }, true);
+            common_batch_add(batch, dp.id_last, dp.pos0, { seq_id }, true);
         }
 
         int ret = llama_decode(ctx_dft, batch);
@@ -355,7 +355,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
                     continue;
                 }
 
-                common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                common_batch_add(batch, id, dp.pos0 + i + 1, { seq_id }, true);
             }
 
             if (batch.n_tokens == 0) {
@@ -1197,7 +1197,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             common_sampler_reset(smpls[seq_id].get());
 
-            const int32_t n = (int32_t) dp.n_past;
+            const int32_t n = (int32_t) dp.pos0;
 
             const int32_t n_draft = params.n_max;
 
@@ -1493,7 +1493,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
         const int32_t n_tokens = batch_in.n_tokens;
 
-        // remember the frist and last batch index for each sequence
+        // remember the first and last batch index for each sequence
         std::fill(i_batch_beg.begin(), i_batch_beg.end(), -1);
         std::fill(i_batch_end.begin(), i_batch_end.end(), -1);
 
@@ -1621,7 +1621,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
             drafting[seq_id] = true;
             common_sampler_reset(smpls[seq_id].get());
 
-            common_batch_add(batch, dp.id_last, dp.n_past, { seq_id }, true);
+            common_batch_add(batch, dp.id_last, dp.pos0, { seq_id }, true);
             std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, pending_h[seq_id].data(), row_bytes);
 
             i_last[seq_id] = batch.n_tokens - 1;
@@ -1635,16 +1635,16 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
         while (n_drafting > 0) {
             // each step decodes under a different head, i.e. a different decoder layer, and
-            // KV is per layer. process() filled this layer's KV only for positions < n_past
+            // KV is per layer. process() filled this layer's KV only for positions < pos0
             // (prompt + accepted prefix) — nothing in the draft region yet. so reset the
-            // draft region (the seq_rm lower bound is n_past, leaving the prompt KV intact)
+            // draft region (the seq_rm lower bound is pos0, leaving the prompt KV intact)
             // and select head i so it rebuilds its own layer's KV there; decoding just the
             // latest token would leave its attention reading cells only another head wrote.
             if (chain_heads) {
                 auto * mem_dft = llama_get_memory(ctx_dft);
                 for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
                     if (drafting[seq_id]) {
-                        llama_memory_seq_rm(mem_dft, seq_id, dparams[seq_id].n_past, -1);
+                        llama_memory_seq_rm(mem_dft, seq_id, dparams[seq_id].pos0, -1);
                     }
                 }
                 llama_set_nextn_layer_offset(ctx_dft, i);
@@ -1710,17 +1710,17 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                     const int n_rows = (int) result.size() + 1; // id_last + tokens drafted so far
                     for (int t = 0; t < n_rows; ++t) {
                         const llama_token tok = (t == 0) ? dp.id_last : result[t - 1];
-                        common_batch_add(batch, tok, dp.n_past + t, { seq_id }, t == n_rows - 1);
+                        common_batch_add(batch, tok, dp.pos0 + t, { seq_id }, t == n_rows - 1);
                         std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd,
                                     chain_h[seq_id].data() + (size_t) t * n_embd, row_bytes);
                     }
                 } else if (is_mem_shared) {
                     // note: with shared memory (e.g. Gemma4 assistants) we use the same position for all draft tokens
                     // ref: https://github.com/huggingface/transformers/blob/effde20942e3f82a1b97449f60b3a48c5ff96145/docs/source/en/model_doc/gemma4_assistant.md?plain=1#L36-L37
-                    common_batch_add(batch, id, dp.n_past, { seq_id }, true);
+                    common_batch_add(batch, id, dp.pos0, { seq_id }, true);
                     std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                 } else {
-                    common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                    common_batch_add(batch, id, dp.pos0 + i + 1, { seq_id }, true);
                     std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                 }
 

common/speculative.h

@@ -61,7 +61,7 @@ struct common_speculative_draft_params {
     // can be used to constraint the max draft based on the remaining context size
     int32_t n_max = -1;
 
-    llama_pos   n_past;
+    llama_pos   pos0;
     llama_token id_last;
 
     // TODO: remove in the future by keeping track of the prompt from the _begin() call and the consecutive accept calls

docs/backend/snapdragon/README.md

@@ -188,7 +188,7 @@ llama_memory_breakdown_print: |   - Host               |                  439 =
 Op test for MUL_MAT:
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --hex-hostbuf 0 --devices HTP0:0 -- test-backend-ops -b HTP0:0 -o MUL_MAT
+~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --devices HTP0:0 -- test-backend-ops -b HTP0:0 -o MUL_MAT
 ...
 Backend 2/3: HTP0:0
 Device description: Hexagon
@@ -213,14 +213,109 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
 | llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 |  tg64 |  51.54 ± 1.13 |
 ```
 
+## Multi-Device Execution Modes
+
+The Hexagon backend supports multiple execution and partitioning modes to accommodate different model sizes, memory
+constraints, and single- or multi-NPU hardware topologies:
+
+### 1. Single-Device Mode with Dynamic Buffer Mapping
+
+Runs the model on a single NPU session (e.g. `HTP0` or `HTP0:0`).
+
+A single NPU session provides ~3.5GB of available virtual address space. For models larger than 3.5GB, the backend
+automatically maps and unmaps weight buffers during graph execution. This allows large models to run on a single NPU
+without manual configuration:
+
+```bash
+./scripts/snapdragon/run.py --target adb --devices HTP0:0 -- \
+    llama-cli -m models/Llama-3.2-3B-Instruct-Q4_0.gguf -ngl 99 -p "Hello"
+```
+
+### 2. Layer-Split Mode across Virtual Sessions (`HTP0,HTP1,...` or `HTP0:0,HTP0:1,...`)
+
+Partitions model layers at load time across multiple virtual sessions hosted on a single physical NPU.
+
+Each virtual session acts as an independent backend device from llama.cpp's perspective (similar to multiple GPUs).
+Because layers are permanently distributed across sessions, each session's allocated weights remain within its private 3.5GB
+address space window, eliminating runtime buffer re-mapping overhead.
+
+Here is an example of running the GPT-OSS-20B model on a Snapdragon device using 4 virtual sessions on a single NPU:
+
+```bash
+./scripts/snapdragon/run.py --target adb \
+    --devices HTP0:0,HTP0:1,HTP0:2,HTP0:3 -- \
+    llama-cli --load-mode none -m /data/local/tmp/gguf/gpt-oss-20b-Q4_0.gguf -t 4 \
+    --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 -no-cnv -f surfing.txt
+```
+
+Log output snippet:
+
+```
+...
+llama_model_loader: - type  f32:  289 tensors
+llama_model_loader: - type q4_0:   96 tensors
+llama_model_loader: - type q8_0:    2 tensors
+llama_model_loader: - type mxfp4:  72 tensors
+...
+load_tensors: offloaded 25/25 layers to GPU
+load_tensors:          CPU model buffer size =  1182.09 MiB
+load_tensors:       HTP0:1 model buffer size =  2512.58 MiB
+load_tensors:       HTP0:3 model buffer size =  2093.83 MiB
+load_tensors:       HTP0:0 model buffer size =  2931.34 MiB
+load_tensors:       HTP0:2 model buffer size =  2512.58 MiB
+...
+llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
+llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.39 ms per token, 18.39 tokens per second)
+llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0:0 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP0:1 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP0:2 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP0:3 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
+```
+
+### 3. Tensor-Split Mode across Physical Devices (`HTP0:0,HTP1:0,...`)
+
+Distributes model tensors across distinct physical NPU hardware cores using llama.cpp's tensor parallelism
+(`--split-mode tensor`).
+
+Tensors are partitioned across physical NPUs for parallel execution (proportions are distributed equally by default without
+needing an explicit `--tensor-split` option):
+
+```bash
+./scripts/snapdragon/run.py --target adb \
+    --devices HTP0:0,HTP1:0 -- \
+    llama-cli -m models/Llama-3.2-3B-Instruct-Q4_0.gguf --split-mode tensor -ngl 99 -p "Hello"
+```
+
+### 4. Row-Split Multi-Device Mode via Device Grouping (`HTP0[0-1]`)
+
+Groups multiple physical NPU cores into a single logical device using bracket notation (`HTP0[0-1]` or `HTP0[0,1]`).
+
+Unlike host-level tensor-splitting, row-splitting is executed entirely inside the Hexagon backend:
+
+```bash
+./scripts/snapdragon/run.py --target adb \
+    --devices 'HTP0[0-1]' -- \
+    llama-cli -m models/Llama-3.2-3B-Instruct-Q4_0.gguf -ngl 99 -p "Hello"
+```
+
+You can also combine row-splitting with layer-splitting across multiple grouped devices (e.g. `--devices 'HTP0[0-1],HTP1[2-3]'`
+on 4 physical NPUs, or `--devices 'HTP0[0-1:0],HTP1[0-1:1]'` on 2 physical NPUs using virtual sessions 0 and 1).
+
 ## Environment variables
 
 - `GGML_HEXAGON_DEVICES` (default: not set, defaults to HTP0 session)
-  Controls which NPU devices and sessions to allocate. Can be configured as:
-  - A single integer `N`: Allocates `N` sessions named `HTP0`, `HTP1`, ..., `HTP<N-1>` (behaves identically to `GGML_HEXAGON_NDEV=N`).
-  - A comma-separated list of device names in `HTP<physical_idx>:<virtual_idx>` format (or legacy `HTP<idx>` format). For example, `HTP0:0,HTP0:1` creates two virtual
-    sessions on the first physical NPU (useful for memory limits). `HTP0:0,HTP1:0` allocates one session on each of the two physical NPUs
-    on a dual-NPU device.
+  Controls which NPU devices and sessions to allocate. Configurable via `--devices` in `run.py`:
+  - `N` (single integer): Allocates `N` virtual sessions named `HTP0`, `HTP1`, ..., `HTP<N-1>` on physical NPU 0.
+  - `HTP<phys>:<virt>,...`: Comma-separated list of individual devices specifying physical and virtual index:
+    - `HTP0:0,HTP0:1`: Two virtual sessions on physical NPU 0 (layer-split on single NPU).
+    - `HTP0:0,HTP1:0`: One session on physical NPU 0 and one on physical NPU 1 (tensor-split across physical cores).
+  - `HTP<name>[<phys_spec>]`: Device grouping syntax for row-split multi-device execution:
+    - `HTP0[0-1]`: A single logical device `HTP0` that groups physical cores 0 and 1.
+    - `HTP0[0-1],HTP1[2-3]`: Two layer-split devices across 4 physical NPUs (cores 0-1 and 2-3).
+    - `HTP0[0-1:0],HTP1[0-1:1]`: Two layer-split devices across 2 physical NPUs using virtual sessions 0 and 1.
 
 - `GGML_HEXAGON_NDEV` (deprecated)
   Replaced by `GGML_HEXAGON_DEVICES`. Controls the number of virtual sessions to allocate on physical NPU `0`.
@@ -229,9 +324,8 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
 - `GGML_HEXAGON_NHVX=0`
   Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
 
-- `GGML_HEXAGON_HOSTBUF=1`
-  Controls whether the Hexagon backend allocates host buffers. By default, all buffers except for REPACK are host buffers.
-  This option is required for testing Ops that require REPACK buffers (MUL_MAT and MUL_MAT_ID).
+- `GGML_HEXAGON_HOSTBUF=1` (default: 0, disabled)
+  Enables allocating host buffers for debugging. By default, host buffers are disabled.
 
 - `GGML_HEXAGON_VERBOSE=1`
   Enables verbose logging of Ops from the backend. Example output:
@@ -246,23 +340,26 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
   ```
 
 - `GGML_HEXAGON_PROFILE=1`
-  Enables Op profiling:
+  Enables Op profiling (configurable via `--hex-profile` in `run.py`):
 
-  - `1` Basic profile with per-op `usecs` and `cycles` counters
-  - `2` Extended profile with per-op `usecs`, `cycles` and default PMU counter data
-  - `0x1,...,0x8` Extended profile with per-op `usecs`, `cycles` and custom PMU counter data
+  - `1`: Basic profile with per-op `usecs` and `cycles` counters
+  - `2`: Extended profile with per-op `usecs`, `cycles` and default PMU counter data
+  - `0x1,...,0x8`: Extended profile with per-op `usecs`, `cycles` and custom PMU counter data
 
-  The logging output can be either saved into a file for post-processing or it can be piped directly into the post-processing tool
-  to generate the report.
-  Examples:
+  The logging output can be saved to a file or piped directly into the post-processing script:
 
-      `GGML_HEXAGON_PROFILE=1 ./scripts/snapdragon/run.py --target adb -- llama-cli ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`
+  ```bash
+  ./scripts/snapdragon/run.py --target adb --hex-profile 1 -- llama-cli ... |& \
+      ./scripts/snapdragon/ggml-hexagon-profile.py -
+  ```
 
 - `GGML_HEXAGON_OPFILTER=regex`
-  Allows filtering (disabling) Ops that match the regex pattern:
+  Filters (disables) Ops matching the regex pattern (configurable via `--hex-opfilter` in `run.py`):
 
-  Examples:
-
-      `GGML_HEXAGON_OPFILTER="FLASH_ATTN_EXT" ./scripts/snapdragon/run.py --target adb -- llama-cli ...` - Disable Flash Attention on Hexagon (falls back to CPU or GPU)
-      `GGML_HEXAGON_OPFILTER="ADD\|SUB" ./scripts/snapdragon/run.py --target adb -- llama-cli ...` - Disable ADD and SUB on Hexagon (fall back to CPU or GPU)
+  ```bash
+  # Disable Flash Attention on Hexagon (falls back to CPU or GPU)
+  ./scripts/snapdragon/run.py --target adb --hex-opfilter "FLASH_ATTN_EXT" -- llama-cli ...
 
+  # Disable ADD and SUB on Hexagon (fall back to CPU or GPU)
+  ./scripts/snapdragon/run.py --target adb --hex-opfilter "ADD|SUB" -- llama-cli ...
+  ```

docs/backend/snapdragon/developer.md

@@ -2,16 +2,16 @@
 
 ## Backend libraries
 
-The Hexagon backend consist of two parts:
+The Hexagon backend consists of two parts:
 
   - `libggml-hexagon`
-    This is the regular CPU-side GGML backend library, either shared or statically linked
+    This is the regular CPU-side GGML backend library, either shared or statically linked.
 
   - `libggml-htp-vNN`
     This is the NPU-side (HTP stands for Hexagon Tensor Processor) shared library that contains the Op dispatcher and kernels.
     The correct library is selected automatically at runtime based on the HW version.
 
-Here is an example of the build artifacts
+Here is an example of the build artifacts:
 
 ```
 ~/src/llama.cpp$ ls -l pkg-adb/llama.cpp/lib/libggml*
@@ -26,75 +26,307 @@ pkg-adb/llama.cpp/lib/libggml-htp-v81.so
 
 ## Memory buffers
 
-Hexagon NPU backend takes advantage of the Snapdragon's unified memory model where all buffers are fully accessible by the CPU and GPU.
-The NPU does have a dedicated tightly-coupled memory called VTCM but that memory is used only for intermediate data (e.g. dynamically
-quantized tensors) or temporary data (chunks of the weight tensors fetched via DMA).
-
-Please note that currently the Hexagon backend does not implement SET/GET_ROWS Ops because there is no advantage in offloading those
-to the NPU at this point.
-
-The backend does allocates non-host buffers for the tensors with datatypes that require repacking: Q4_0, Q8_0, MXFP4.
-From the MMU perspective these buffers are still regular buffers (normal access by the CPU) they are marked as non-host simply to force
-the repacking.
+The Hexagon NPU backend takes advantage of Snapdragon unified memory where all DDR buffers are accessible by CPU, GPU, and NPU.
+The NPU has dedicated tightly-coupled memory called VTCM (Vector Tightly-Coupled Memory). VTCM is used for intermediate data (such as
+dynamically quantized activations) and streaming buffers (chunks of weight and activation tensors fetched via DMA).
 
 ## Large model handling
 
-Hexagon NPU sessions (aka Process Domains (PD) in the Hexagon SDK) are limited to a maximum memory mapping window of around 3.5GB.
+Hexagon NPU sessions have a 32-bit virtual address space window of around 3.5GB.
 In llama.cpp/GGML, each Hexagon session is mapped to a single GGML backend device (e.g., `HTP0:0`, `HTP0:1`, etc. when using
 `GGML_HEXAGON_DEVICES`, or `HTP0`, `HTP1` in legacy mode).
 
-To support running models larger than 3.5GB on a single device, the Hexagon backend dynamically maps and unmaps execution buffers
-during the graph execution cycle to stay within the Process Domain window. This enables large models to run successfully on a single
-NPU device.
+To support running models larger than 3.5GB on a single device, the Hexagon backend dynamically maps and unmaps buffers:
+- Buffers are allocated in shared DDR (RPCMEM) via file descriptors (`fastrpc_mmap` using `FASTRPC_MAP_FD_DELAYED`).
+- Pinned buffers (such as KV cache and active compute buffers) remain mapped throughout execution.
+- Inactive weight buffers are dynamically mapped into the NPU session via `HAP_mmap()` during batch buffer preparation
+  (`prep_op_bufs()` in `htp/main.c`) and unmapped via `htp_iface_munmap()` when no longer needed by the active batch.
+- This dynamic sliding window allows a single NPU session to execute models that exceed the 3.5GB window.
 
-Alternatively, users can choose to use standard llama.cpp/GGML layer-splitting mode to partition and split the model across
-multiple Hexagon devices or virtual sessions (which behave like multiple GPUs from the offload and splitting perspective).
+Alternatively, users can partition and split the model across multiple virtual sessions or physical NPUs using layer-splitting,
+tensor-splitting, or row-splitting modes. For user-facing execution modes and examples, see the
+[Snapdragon user guide](README.md#multi-device-execution-modes).
 
-Here is an example of running GPT-OSS-20B model on a Snapdragon device using 4 virtual sessions on a single NPU (physical index 0).
+## Op and Kernel Development Guidelines
+
+Writing high-performance operators for Hexagon requires following specific guidelines.
+
+### DDR -> DMA -> VTCM Execution Pipeline
+
+- Strongly prefer the `DDR -> DMA -> VTCM -> compute (HVX/HMX) -> VTCM -> DMA -> DDR` data flow.
+- Direct HVX reads/writes from/to DDR are less efficient and should only be used as a fallback.
+- The DMA queue is a strict FIFO where operations must be pushed and popped in strict order.
+- Follow the pipelined multi-buffering sequence properly (typically 2x to 16x buffering) so every push has a corresponding pop:
+
+  1. In the prologue, push initial DDR -> VTCM transfers to prime the pipeline.
+  2. In the loop body, wait for buffer N via DMA pop, launch HVX/HMX compute on buffer N, push VTCM -> DDR writeback of result N,
+     and push DDR -> VTCM prefetch of buffer N+2.
+  3. In the epilogue, pop all remaining in-flight transfers to drain the pipeline.
+
+- Because every push must be matched by a pop, `dma_queue_flush()` is not required when the pipeline sequence is followed
+  properly. Flushing is only used in rare exceptions where a batch of operations is pushed without individual pops.
+- Use the DMA queue interface from [`dma-queue.h`](../../../ggml/src/ggml-hexagon/htp/dma-queue.h)
+  (`dma_queue_push_ddr_to_vtcm`, `dma_queue_pop`, `dma_queue_push_vtcm_to_ddr`).
+  See [`cumsum-ops.c`](../../../ggml/src/ggml-hexagon/htp/cumsum-ops.c) and
+  [`act-ops.c`](../../../ggml/src/ggml-hexagon/htp/act-ops.c) for reference implementations.
+
+### Avoid Scalar Reads and Writes to VTCM
+
+- Access VTCM data using DMA transfers or HVX/HMX vector instructions rather than scalar reads and writes.
+
+### Avoid Scalar Division in Inner Loops
+
+- Hexagon cores do not have hardware division instructions.
+- For recurring divisions across iterations or threads, use `fastdiv` from
+  [`hex-fastdiv.h`](../../../ggml/src/ggml-hexagon/htp/hex-fastdiv.h) with precomputed divisors (such as
+  `octx->ctx->mdev.count_div` or `octx->n_threads_div`).
+- Do not call `init_fastdiv_values()` for single-use divisions; use standard compiler division (`/`) instead.
+
+### Host-Side Precomputation via `kernel_params`
+
+- Precompute tensor shapes, strides, scale conversions, tiling layouts, and validation checks on the host CPU during graph
+  preparation in [`ggml-hexagon.cpp`](../../../ggml/src/ggml-hexagon/ggml-hexagon.cpp).
+- Pack precomputed parameters into the operator's fixed `kernel_params` structure in `htp_op_node` (such as
+  `htp_mm_kernel_params`, `htp_unary_kernel_params`, `htp_fa_kernel_params`, `htp_get_rows_kernel_params`).
+- The NPU executes directly using `octx->kernel_params` without redundant runtime metadata extraction or validation.
+- **Strict Host-Kernel Alignment**:
+  - Verify that parameters calculated by the host CPU are strictly honored by the NPU kernel.
+  - Ensure the kernel does not ignore host-computed fields (for example, falling back to `octx->n_threads` instead of
+    using `kparams->n_threads`, or ignoring precomputed `tasks_per_thread` and chunk counts).
+  - Both human developers and coding agents must audit both sides of the interface: ensure fields populated in `kernel_params`
+    in [`ggml-hexagon.cpp`](../../../ggml/src/ggml-hexagon/ggml-hexagon.cpp) are actively and consistently utilized by the
+    corresponding operator entry point and worker threads in `htp/*-ops.c`.
+
+### Tracing Instrumentation
+
+- All kernels must include trace events for performance profiling and timeline visualization in Perfetto
+  ([`hex-profile.h`](../../../ggml/src/ggml-hexagon/htp/hex-profile.h)).
+- Surround compute sections with `htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) info)` and
+  `htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) info)`.
+- Use specific event types for major phases:
+  - `HTP_TRACE_EVT_HVX_COMP`: Vector compute execution.
+  - `HTP_TRACE_EVT_DMA`: DMA transfer wait or poll cycles.
+  - `HTP_TRACE_EVT_FENCE`: Multi-device fence barrier synchronization.
+  - `HTP_TRACE_EVT_L2FLUSH`: L2 cache cleaning operations.
+- Pass meaningful progress metrics (such as row index, chunk index, or token index) in the 16-bit `info` parameter.
+
+### Work Queue and Threading
+
+- Distribute parallel work across NPU worker threads using the thread pool work queue:
+
+  ```c
+  work_queue_run(ctx->work_queue, worker_func, &op_ctx, n_threads);
+  ```
+
+- Keep worker functions independent and re-entrant. Worker threads should only operate on their designated chunk of rows or elements.
+
+### Avoid Redundant Defensive NULL Checks
+
+- Do not add defensive NULL checks or assertions for internal framework pointers or required graph operands and outputs.
+  Internal pointers include `ctx`, `octx`, local context structs like `*ctx`, `kparams`, and worker callback `data`.
+- These pointers are architectural invariants during kernel execution and host-side graph preparation.
+  Graph compute receives allocated nodes with valid required `node->src[N]` and `node->data` pointers.
+- Do not turn an invariant violation into an unsupported operation or missed fusion.
+  Checks such as `if (!octx || !octx->ctx)` clutter the code, obscure intent, and hide upstream errors.
+- **Distinction**: `octx->src[N]` pointers *can* be NULL by design and must be checked when optional.
+  Examples include attention masks, optional bias or weights in fused kernels, and frequency factors.
+
+### Multiline Macro Formatting
+
+- Keep trailing backslashes in multiline `#define` macros cleanly aligned to a consistent column.
+- Avoid trailing whitespace after macro backslashes.
+- Use [`scripts/snapdragon/ggml-hexagon-align-macros.py`](../../../scripts/snapdragon/ggml-hexagon-align-macros.py) to inspect, diff,
+  or automatically align macro definitions across Hexagon kernel sources:
+
+  ```bash
+  # Check for misaligned macros
+  python3 scripts/snapdragon/ggml-hexagon-align-macros.py ggml/src/ggml-hexagon/htp/
+
+  # Fix misaligned macros in-place
+  python3 scripts/snapdragon/ggml-hexagon-align-macros.py --fix ggml/src/ggml-hexagon/htp/
+  ```
+
+## Multi-Device Partitioning (mdev)
+
+Multi-device (mdev) mode enables row-level tensor parallel execution across multiple physical NPU cores or virtual NPU
+sessions.
+
+### 128-Byte Cache Line Alignment
+
+- Shared tensor buffers reside in DDR (RPCMEM) with a 128-byte cache line granularity
+  (`HEX_L2_LINE_SIZE` = 128 bytes, `HTP_TENSOR_MDEV_LINE_SIZE`).
+- **Rule**: Multi-device work partitions must align destination write regions to 128-byte cache line boundaries so distinct
+  devices never share or overwrite the same cache line.
+
+### Partitioning Helpers in `htp-tensor.h`
+
+Common partitioning logic is factored into reusable inline helpers in
+[`htp-tensor.h`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h):
+
+1. [`htp_tensor_mdev_rows_per_chunk`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h#L67):
+   Determines the minimum number of rows per chunk so that the chunk byte size is a multiple of 128 bytes:
+
+   ```
+   rows_per_chunk = 128 / hex_gcd_u32(row_size, 128)
+   ```
+
+   If row stride `nb[1]` is already a multiple of 128 bytes, `rows_per_chunk = 1`.
+   Returns `false` if the tensor cannot be safely row-partitioned (such as unaligned base pointer, permuted layout,
+   or non-128-byte aligned outer strides).
+
+2. [`htp_tensor_mdev_partition`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h#L94):
+   Calculates the per-device work range `struct htp_tensor_mdev_range { uint32_t start; uint32_t count; }` given
+   `total_units`, `units_per_chunk`, `mdev_idx`, `mdev_count`, and the precomputed `mdev_count_div`.
+   Handles chunk distribution across devices, assigns remainder units to the last device, and automatically triggers
+   single-device fallback when partitioning is unsafe.
+
+### Row-Partitioned Operators
+
+For row-wise operators
+(such as activations in [`act-ops.c`](../../../ggml/src/ggml-hexagon/htp/act-ops.c),
+binary ops in [`binary-ops.c`](../../../ggml/src/ggml-hexagon/htp/binary-ops.c),
+unary ops in [`unary-ops.c`](../../../ggml/src/ggml-hexagon/htp/unary-ops.c), and
+sameshape copies in [`cpy-ops.c`](../../../ggml/src/ggml-hexagon/htp/cpy-ops.c)):
+
+```c
+const uint32_t total_rows   = ne01 * ne02 * ne03;
+const size_t   dst_row_size = dst->ne[0] * elem_size;
+
+uint32_t row_start = 0;
+uint32_t nrows     = total_rows;
+
+if (octx->ctx->mdev.count > 1) {
+    uint32_t rows_per_chunk = 0;
+    htp_tensor_mdev_rows_per_chunk(dst, elem_size, (uint32_t) dst_row_size, &rows_per_chunk);
+    const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(
+        total_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+    row_start = range.start;
+    nrows     = range.count;
+}
+
+if (nrows == 0) {
+    return HTP_STATUS_OK;
+}
+```
+
+### Element-Partitioned Operators
+
+For flat element-wise operations (such as reshape copies in
+[`cpy-ops.c`](../../../ggml/src/ggml-hexagon/htp/cpy-ops.c)):
+- Partition total linear elements N = ne0 * ne1 * ne2 * ne3 in 128-byte cache line chunks (`elems_per_line = (elem_size == 4) ? 32 : 64`).
+- Requires strict 1D contiguity:
+  [`htp_tensor_is_contiguous(dst, elem_size)`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h#L28)
+  and 128-byte aligned destination pointer
+  [`htp_tensor_mdev_data_aligned(dst)`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h#L47).
+- If contiguous and aligned, pass `elems_per_line` to
+  [`htp_tensor_mdev_partition`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h#L94);
+  otherwise pass 0 to trigger Device 0 fallback.
+
+### Single-Device Fallback (Device 0)
+
+- Fallback to Device 0 (`mdev.idx == 0`) when partitioning would cause cache line tearing or when work cannot be evenly distributed.
+- Triggers:
+  1. Destination tensor cannot be safely partitioned (`rows_per_chunk == 0` or non-contiguous/unaligned buffer).
+  2. Total aligned chunks < `mdev_count`.
+- Device 0 processes the entire tensor `[0, total_units)`.
+- Devices 1 ... N-1 receive `count = 0` and return `HTP_STATUS_OK` immediately.
+
+### Flatten Outer Dimensions Globally
+
+- **Never partition solely on `ne01` (dimension 1).**
+- Partitioning only on `ne01` repeats the device boundary across every 2D slice (`ne02`, `ne03`). If each 2D slice is small,
+  false sharing occurs repeatedly throughout the tensor.
+- Always flatten outer dimensions globally: `total_rows = ne01 * ne02 * ne03` and partition once across the combined row space.
+
+### Stateless Starting Coordinates
+
+- Do not use incremental state variables across slices that assume the thread or device starts at index 0.
+- Precompute starting multidimensional coordinates at `r = row_start` (or `e = elem_start`) once using `fastdiv`.
+- In inner loops, step base pointers directly (`ptr += stride`) or reset/wrap coordinates explicitly (`if (++i01 == ne01) { ... }`).
+
+### Clean Range Encapsulation
+
+- Initialize single-device default ranges at declaration:
+
+  ```c
+  uint32_t row_start = 0;
+  uint32_t nrows     = total_rows;
+  ```
+
+- Encapsulate all multi-device logic inside `if (octx->ctx->mdev.count > 1)`. If the block is omitted or compiled out,
+  the operator runs standard single-device execution untouched.
+- Do not propagate `mdev_` prefixes to worker functions or context structs. Worker threads are device-agnostic and
+  should only receive standard range parameters (`ctx.row_start`, `ctx.nrows`).
+- In worker threads, calculate row intervals using standard arithmetic:
+
+  ```c
+  const uint32_t ir0 = ctx->row_start + dr * ith;
+  const uint32_t ir1 = MIN(ir0 + dr, ctx->row_start + ctx->nrows);
+  ```
+
+  In single-device mode (`row_start == 0`), this naturally simplifies to `dr * ith` and `MIN(ir0 + dr, ctx->nrows)` with zero overhead.
+
+## Multi-Device Synchronization
+
+Multi-device execution synchronizes worker sessions across devices using explicit barriers and tensor cache flushing.
+
+### Synchronization Fence Protocol
+
+Multi-device execution synchronizes worker sessions through atomic fence slots and barriers defined in
+[`htp-fence.h`](../../../ggml/src/ggml-hexagon/htp/htp-fence.h):
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --devices HTP0:0,HTP0:1,HTP0:2,HTP0:3 -- llama-cli --load-mode none -m /data/local/tmp/gguf/gpt-oss-20b-Q4_0.gguf -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 -no-cnv -f surfing.txt
-...
-llama_model_loader: - type  f32:  289 tensors
-llama_model_loader: - type q4_0:   96 tensors
-llama_model_loader: - type q8_0:    2 tensors
-llama_model_loader: - type mxfp4:  72 tensors
-...
-load_tensors: offloaded 25/25 layers to GPU
-load_tensors:          CPU model buffer size =  1182.09 MiB
-load_tensors:       HTP0:1 model buffer size =  2512.58 MiB
-load_tensors:       HTP0:3 model buffer size =  2093.83 MiB
-load_tensors:       HTP0:0 model buffer size =  2931.34 MiB
-load_tensors:       HTP0:2 model buffer size =  2512.58 MiB
-...
-llama_context: n_ctx_per_seq (8192) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
-llama_context:        CPU  output buffer size =     0.77 MiB
-llama_kv_cache_iswa: creating non-SWA KV cache, size = 8192 cells
-llama_kv_cache:     HTP0:1 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:3 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:0 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:2 KV buffer size =    25.50 MiB
-llama_kv_cache: size =  102.00 MiB (  8192 cells,  12 layers,  1/1 seqs), K (q8_0):   51.00 MiB, V (q8_0):   51.00 MiB
-llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
-llama_kv_cache:     HTP0:1 KV buffer size =     0.80 MiB
-llama_kv_cache:     HTP0:3 KV buffer size =     0.53 MiB
-llama_kv_cache:     HTP0:0 KV buffer size =     1.06 MiB
-llama_kv_cache:     HTP0:2 KV buffer size =     0.80 MiB
-llama_kv_cache: size =    3.19 MiB (   256 cells,  12 layers,  1/1 seqs), K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
-llama_context:     HTP0:0 compute buffer size =    16.06 MiB
-llama_context:     HTP0:1 compute buffer size =    16.06 MiB
-llama_context:     HTP0:2 compute buffer size =    16.06 MiB
-llama_context:     HTP0:3 compute buffer size =    16.06 MiB
-llama_context:        CPU compute buffer size =    98.19 MiB
-...
-llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
-llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.39 ms per token, 18.39 tokens per second)
-llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
-llama_perf_context_print:    graphs reused =         30
-llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
-llama_memory_breakdown_print: |   - HTP0:0 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:1 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:2 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:3 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
+[NPU Session 0]                              [NPU Session 1]
+       |                                            |
+  (Input Prep)                                 (Input Prep)
+       |                                            |
+  Pre-Op Barrier ----------------------------- Pre-Op Barrier
+  (mdev_sync_fence)                            (mdev_sync_fence)
+       |                                            |
+  Kernel Execution                             Kernel Execution
+  (Output Slice 0)                             (Output Slice 1)
+       |                                            |
+  Tensor Cache Flush                           Tensor Cache Flush
+  (htp_tensor_flush_all)                       (htp_tensor_flush_all)
+       |                                            |
+  Post-Op/Batch Barrier ---------------------- Post-Op/Batch Barrier
+  (htp_mdev_group_barrier)                     (htp_mdev_group_barrier)
+       |                                            |
+  Return Response to Host                      Return Response to Host
 ```
+
+### Atomic Fence Slots and Cache Invalidation
+
+- Fence synchronization operates on dedicated RPCMEM shared memory mapped across all participating sessions (`ctx->mdev.fence_base`).
+- Each device owns a dedicated 128-byte cache-line aligned fence slot:
+
+  ```c
+  atomic_uint * my_fence = htp_mdev_fence_slot(fence_base, mdev_idx);
+  ```
+
+- **Writing to fence ([`htp_fence_write`](../../../ggml/src/ggml-hexagon/htp/htp-fence.h#L18))**:
+  Stores `seq` and `status`, issues a `syncht` thread synchronization barrier, and flushes/invalidates the line
+  using `Q6_dccleaninva_A(fence)`.
+- **Reading from peer fence ([`htp_fence_read`](../../../ggml/src/ggml-hexagon/htp/htp-fence.h#L26))**:
+  Executes `Q6_dccleaninva_A(fence)` and `syncht` before reading atomic values to ensure fresh data from DDR.
+
+### Deterministic Monotonic Sequence Numbers
+
+- Barrier fences use monotonically increasing sequence numbers:
+
+  ```c
+  const uint32_t seq = ++ctx->mdev.fence_seq;
+  ```
+
+- Comparing sequence numbers with signed arithmetic `(int32_t)(peer_seq - seq) >= 0` prevents race conditions or
+  misaligned barrier arrivals across iterations.
+- If any peer reports an error status (`peer_status > HTP_STATUS_OK`), the barrier propagates the error and unblocks immediately.
+
+### Tensor Cache Flush and Pipeline Completion
+
+- In the kernel, ensure all pushed DMA operations have been popped in strict FIFO order to drain the queue.
+- Use [`htp_tensor_flush_all()`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h) to flush specific dirty tensors back to DDR:
+  - [`htp_tensor_flush_all()`](../../../ggml/src/ggml-hexagon/htp/htp-tensor.h) flushes only modified tensor address ranges,
+    ensuring peer devices and the host CPU observe consistent data in DDR.
+- Never signal completion before all DMA transfers are drained and dirty tensor flushes have completed.
+

docs/build-profiling.md

@@ -0,0 +1,122 @@
+## Build profiling
+This page is a working document for analyzing the current build and try to
+identify ways to improve the build time.
+
+### Requirements
+The profiling script requires clang to be used as the compiler tool chain and
+also requires that ClangBuildAnalyzer is installed.
+
+Mac:
+```console
+brew install clang-build-analyzer
+```
+
+Linux:
+```console
+git clone https://github.com/aras-p/ClangBuildAnalyzer.git
+cd ClangBuildAnalyzer
+cmake -B build -DCMAKE_BUILD_TYPE=Release
+cmake --build build -j$(nproc)
+sudo cp build/ClangBuildAnalyzer /usr/local/bin/
+```
+
+Windows: install LLVM/clang and Ninja (e.g. via the
+[LLVM releases page](https://github.com/llvm/llvm-project/releases) and
+`winget install Ninja-build.Ninja`), then build ClangBuildAnalyzer the same
+way as on Linux:
+```console
+git clone https://github.com/aras-p/ClangBuildAnalyzer.git
+cd ClangBuildAnalyzer
+cmake -B build -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+cmake --build build --config Release
+```
+Then add `ClangBuildAnalyzer\build` to `PATH`.
+
+### Usage
+Mac/Linux:
+```console
+$ ./scripts/build-profile.sh
+```
+
+Windows:
+```console
+> .\scripts\build-profile.ps1
+```
+
+Both accept `--full`/`-Full` (include Server, Tools, and Tests) and a jobs
+override (`-jN` / `-Jobs N`).
+
+Note: on Windows, `cmake` defaults to the Visual Studio generator, which
+ignores `CMAKE_C_COMPILER`/`CMAKE_CXX_COMPILER` and silently falls back to
+MSVC. `build-profile.ps1` passes `-G Ninja` so clang is actually used, this
+is required on ARM64.
+
+### Linux (Ubuntu 24.04)
+
+Environment:
+- Clang:     18.1.3 (Ubuntu clang version 18.1.3 (1ubuntu1))
+- libstdc++: GCC 13.3.0 (Ubuntu 13.3.0-6ubuntu2~24.04.1)
+- Target:    x86_64-pc-linux-gnu
+
+```console
++------------------------+-----+------------+------------+------------+
+| Build                  | TUs | Frontend   | Backend    | Total      |
++------------------------+-----+------------+------------+------------+
+| Minimal, master        | 249 |   468.2 s  |   270.3 s  |   738.5 s  |
+| Minimal, with PCH      | 253 |   177.1 s  |   265.8 s  |   442.9 s  |
+| Full,    master        | 396 |   811.0 s  |   692.2 s  | 1,503.2 s  |
+| Full,    with PCH      | 405 |   380.0 s  |   664.7 s  | 1,044.7 s  |
+| Full,    with PCH + UB | 264 |   357.7 s  |   635.7 s  |   993.4 s  |
++------------------------+-----+------------+------------+------------+
+
+PCH  = precompiled header.
+Full = includes building Server, Tools, and Tests.
+UB   = unity build for models
+```
+Note that the number of translation units (TUs) increases when using precompiled
+headers — each PCH target adds one extra TU for the precompilation step itself.
+
+### Mac (Apple M3)
+
+Environment:
+- Clang:  Apple clang version 17.0.0 (clang-1700.3.19.1)
+- libc++: ships with Apple clang 17.0.0 (Xcode toolchain)
+- Target: arm64-apple-macosx15.6
+
+```console
++------------------------+-----+------------+------------+------------+
+| Build                  | TUs | Frontend   | Backend    | Total      |
++------------------------+-----+------------+------------+------------+
+| Minimal, master        | 256 |   154.5 s  |    94.8 s  |   249.3 s  |
+| Minimal, with PCH      | 261 |    65.9 s  |    90.0 s  |   155.9 s  |
+| Full,    master        | 407 |   265.7 s  |   209.7 s  |   475.4 s  |
+| Full,    with PCH      | 414 |   154.6 s  |   197.5 s  |   352.1 s  |
+| Full,    with PCH + UB | 274 |   143.0 s  |   192.2 s  |   335.2 s  |
++------------------------+-----+------------+------------+------------+
+
+PCH = precompiled header.
+Full = includes building Server, Tools, and Tests.
+UB   = unity build for models
+```
+
+### Windows (ARM64)
+
+Environment:
+- Clang:  clang version 22.1.8 (LLVM, `C:\Program Files\LLVM`)
+- STL:    MSVC STL (Visual Studio 2022 Build Tools 14.44.35207)
+- Target: aarch64-pc-windows-msvc
+
+```console
++------------------------+-----+------------+------------+------------+
+| Build                  | TUs | Frontend   | Backend    | Total      |
++------------------------+-----+------------+------------+------------+
+| Minimal, master        | 249 |   159.4 s  |    82.2 s  |   241.6 s  |
+| Full,    master        | 373 |   337.2 s  |   167.4 s  |   504.6 s  |
+| Minimal, with PCH + UB | 113 |    62.3 s  |    82.4 s  |   144.7 s  |
+| Full,    with PCH + UB | 240 |   233.0 s  |   185.1 s  |   418.1 s  |
++------------------------+-----+------------+------------+------------+
+
+PCH = precompiled header.
+Full = includes building Server, Tools, and Tests.
+UB   = unity build for models
+```

docs/build.md

@@ -806,7 +806,7 @@ To read documentation for how to build on Android, [click here](./android.md)
 
 ## WebGPU
 
-The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `18eb229`.
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `94c3c9c`.
 
 In the llama.cpp directory, build with CMake:
 

docs/development/parsing.md

@@ -28,7 +28,7 @@ auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
     for (const auto & tool : tools) {
         const auto & function = tool.at("function");
         std::string name = function.at("name");
-        const auto & schema = function.at("parameters");
+        const auto   schema = common_chat_tool_parameters(function);
 
         auto tool_name = p.json_member("name", "\"" + p.literal(name) + "\"");
         auto tool_args = p.json_member("arguments", p.schema(p.json(), "tool-" + name + "-schema", schema));
@@ -108,6 +108,7 @@ For a more complete example, see `test_example_native()` in
 - **`rule(name, p, trigger)`** - Creates a named rule and returns a reference
 - **`trigger_rule(name, p)`** - Creates a trigger rule (entry point for lazy grammar generation)
 - **`schema(p, name, schema, raw)`** - Wraps parser with JSON schema metadata for grammar generation
+- **`schema(p, name, doc, node, raw)`** - Same, for a node of a `common_chat_schema_document` built earlier, e.g. one tool parameter
 
 ### AST Control
 
@@ -121,9 +122,6 @@ some exceptions.
 
 ```cpp
 data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-    foreach_function(params.tools, [&](const json & fn) {
-        builder.resolve_refs(fn.at("parameters"));
-    });
     parser.build_grammar(builder, data.grammar_lazy);
 });
 ```
@@ -151,7 +149,8 @@ implementation to generate the grammar instead of the underlying parser.
 
 The `raw` option emits a grammar suitable for a raw string instead of a JSON
 string. In other words, it won't be wrapped in quotes or require escaping
-quotes. It should only be used when `type == "string"`.
+quotes. It only takes effect when the schema may be a string, as reported by
+`common_chat_schema::may_be_string()`, otherwise the JSON grammar is used.
 
 The downside is that it can potentially lead to ambiguous grammars. For
 example, if a user provides the pattern `^.*$`, the following grammar may be

examples/json_schema_to_grammar.py

@@ -1,842 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import argparse
-import itertools
-import json
-import re
-import sys
-from typing import Any, List, Optional, Set, Tuple, Union
-
-def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
-
-    if max_items == 0:
-        return ""
-
-    if min_items == 0 and max_items == 1:
-        return f'{item_rule}?'
-
-    if not separator_rule:
-        if min_items == 1 and max_items is None:
-            return f'{item_rule}+'
-        elif min_items == 0 and max_items is None:
-            return f'{item_rule}*'
-        else:
-            return f'{item_rule}{{{min_items},{max_items if max_items is not None else ""}}}'
-
-    result = item_rule + ' ' + _build_repetition(f'({separator_rule} {item_rule})', min_items - 1 if min_items > 0 else 0, max_items - 1 if max_items is not None else None)
-    return f'({result})?' if min_items == 0 else result
-
-def _generate_min_max_int(min_value: Optional[int], max_value: Optional[int], out: list, decimals_left: int = 16, top_level: bool = True):
-    def digit_range(from_char: str, to_char: str):
-        out.append("[")
-        if from_char == to_char:
-            out.append(from_char)
-        else:
-            out.append(from_char)
-            out.append("-")
-            out.append(to_char)
-        out.append("]")
-
-    def more_digits(min_digits: int, max_digits: int):
-        out.append("[0-9]")
-        if min_digits == max_digits and min_digits == 1:
-            return
-        out.append("{")
-        out.append(str(min_digits))
-        if max_digits != min_digits:
-            out.append(",")
-            if max_digits != sys.maxsize:
-                out.append(str(max_digits))
-        out.append("}")
-
-    def uniform_range(from_str: str, to_str: str):
-        i = 0
-        while i < len(from_str) and from_str[i] == to_str[i]:
-            i += 1
-        if i > 0:
-            out.append("\"")
-            out.append(from_str[:i])
-            out.append("\"")
-        if i < len(from_str):
-            if i > 0:
-                out.append(" ")
-            sub_len = len(from_str) - i - 1
-            if sub_len > 0:
-                from_sub = from_str[i+1:]
-                to_sub = to_str[i+1:]
-                sub_zeros = "0" * sub_len
-                sub_nines = "9" * sub_len
-
-                to_reached = False
-                out.append("(")
-                if from_sub == sub_zeros:
-                    digit_range(from_str[i], chr(ord(to_str[i]) - 1))
-                    out.append(" ")
-                    more_digits(sub_len, sub_len)
-                else:
-                    out.append("[")
-                    out.append(from_str[i])
-                    out.append("] ")
-                    out.append("(")
-                    uniform_range(from_sub, sub_nines)
-                    out.append(")")
-                    if ord(from_str[i]) < ord(to_str[i]) - 1:
-                        out.append(" | ")
-                        if to_sub == sub_nines:
-                            digit_range(chr(ord(from_str[i]) + 1), to_str[i])
-                            to_reached = True
-                        else:
-                            digit_range(chr(ord(from_str[i]) + 1), chr(ord(to_str[i]) - 1))
-                        out.append(" ")
-                        more_digits(sub_len, sub_len)
-                if not to_reached:
-                    out.append(" | ")
-                    digit_range(to_str[i], to_str[i])
-                    out.append(" ")
-                    uniform_range(sub_zeros, to_sub)
-                out.append(")")
-            else:
-                out.append("[")
-                out.append(from_str[i])
-                out.append("-")
-                out.append(to_str[i])
-                out.append("]")
-
-    if min_value is not None and max_value is not None:
-        if min_value < 0 and max_value < 0:
-            out.append("\"-\" (")
-            _generate_min_max_int(-max_value, -min_value, out, decimals_left, top_level=True)
-            out.append(")")
-            return
-
-        if min_value < 0:
-            out.append("\"-\" (")
-            _generate_min_max_int(0, -min_value, out, decimals_left, top_level=True)
-            out.append(") | ")
-            min_value = 0
-
-        min_s = str(min_value)
-        max_s = str(max_value)
-        min_digits = len(min_s)
-        max_digits = len(max_s)
-
-        for digits in range(min_digits, max_digits):
-            uniform_range(min_s, "9" * digits)
-            min_s = "1" + "0" * digits
-            out.append(" | ")
-        uniform_range(min_s, max_s)
-        return
-
-    less_decimals = max(decimals_left - 1, 1)
-
-    if min_value is not None:
-        if min_value < 0:
-            out.append("\"-\" (")
-            _generate_min_max_int(None, -min_value, out, decimals_left, top_level=False)
-            out.append(") | [0] | [1-9] ")
-            more_digits(0, decimals_left - 1)
-        elif min_value == 0:
-            if top_level:
-                out.append("[0] | [1-9] ")
-                more_digits(0, less_decimals)
-            else:
-                more_digits(1, decimals_left)
-        elif min_value <= 9:
-            c = str(min_value)
-            range_start = '1' if top_level else '0'
-            if c > range_start:
-                digit_range(range_start, chr(ord(c) - 1))
-                out.append(" ")
-                more_digits(1, less_decimals)
-                out.append(" | ")
-            digit_range(c, "9")
-            out.append(" ")
-            more_digits(0, less_decimals)
-        else:
-            min_s = str(min_value)
-            length = len(min_s)
-            c = min_s[0]
-
-            if c > "1":
-                digit_range("1" if top_level else "0", chr(ord(c) - 1))
-                out.append(" ")
-                more_digits(length, less_decimals)
-                out.append(" | ")
-            digit_range(c, c)
-            out.append(" (")
-            _generate_min_max_int(int(min_s[1:]), None, out, less_decimals, top_level=False)
-            out.append(")")
-            if c < "9":
-                out.append(" | ")
-                digit_range(chr(ord(c) + 1), "9")
-                out.append(" ")
-                more_digits(length - 1, less_decimals)
-        return
-
-    if max_value is not None:
-        if max_value >= 0:
-            if top_level:
-                out.append("\"-\" [1-9] ")
-                more_digits(0, less_decimals)
-                out.append(" | ")
-            _generate_min_max_int(0, max_value, out, decimals_left, top_level=True)
-        else:
-            out.append("\"-\" (")
-            _generate_min_max_int(-max_value, None, out, decimals_left, top_level=False)
-            out.append(")")
-        return
-
-    raise RuntimeError("At least one of min_value or max_value must be set")
-
-class BuiltinRule:
-    def __init__(self, content: str, deps: list | None = None):
-        self.content = content
-        self.deps = deps or []
-
-# Constraining spaces to prevent model "running away".
-SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}'
-
-PRIMITIVE_RULES = {
-    'boolean'      : BuiltinRule('("true" | "false")', []),
-    'decimal-part' : BuiltinRule('[0-9]{1,16}', []),
-    'integral-part': BuiltinRule('[0] | [1-9] [0-9]{0,15}', []),
-    'number'       : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)?', ['integral-part', 'decimal-part']),
-    'integer'      : BuiltinRule('("-"? integral-part)', ['integral-part']),
-    'value'        : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
-    'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? space "}"', ['string', 'value']),
-    'array'        : BuiltinRule('"[" space ( value ("," space value)* )? space "]"', ['value']),
-    'uuid'         : BuiltinRule(r'"\"" [0-9a-fA-F]{8} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{12} "\""', []),
-    'char'         : BuiltinRule(r'[^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})', []),
-    'string'       : BuiltinRule(r'"\"" char* "\""', ['char']),
-    'null'         : BuiltinRule('"null"', []),
-}
-
-# TODO: support "uri", "email" string formats
-STRING_FORMAT_RULES = {
-    'date'            : BuiltinRule('[0-9]{4} "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
-    'time'            : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9]{3} )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
-    'date-time'       : BuiltinRule('date "T" time', ['date', 'time']),
-    'date-string'     : BuiltinRule('"\\"" date "\\""', ['date']),
-    'time-string'     : BuiltinRule('"\\"" time "\\""', ['time']),
-    'date-time-string': BuiltinRule('"\\"" date-time "\\""', ['date-time']),
-}
-
-DOTALL = '[\\U00000000-\\U0010FFFF]'
-DOT = '[^\\x0A\\x0D]'
-
-RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
-
-INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
-GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\\]')
-GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]', '\\': '\\\\'}
-
-NON_LITERAL_SET = set('|.()[]{}*+?')
-ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('^$.[]()|{}*+?')
-
-
-class SchemaConverter:
-    def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
-        self._prop_order = prop_order
-        self._allow_fetch = allow_fetch
-        self._dotall = dotall
-        self._raw_pattern = raw_pattern
-        self._rules = {
-            'space': SPACE_RULE,
-        }
-        self._refs = {}
-        self._refs_being_resolved = set()
-
-    def _format_literal(self, literal):
-        escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
-            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)) or m.group(0), literal
-        )
-        return f'"{escaped}"'
-
-    def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str:
-        '''
-            not_literal('a') -> '[^a]'
-            not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?'
-        '''
-        assert len(literal) > 0, 'Empty literal not supported'
-        def recurse(i: int):
-            c = literal[i]
-            if maybe_escaped_underscores and c == '_':
-                yield f'[^{c}\\\\]'
-                yield ' | '
-                yield f'"\\\\"? "{c}"'
-            else:
-                yield f'[^{c}]'
-            if i < len(literal) - 1:
-                yield ' | '
-                yield self._format_literal(c)
-                yield ' ('
-                yield from recurse(i + 1)
-                yield ')?'
-
-        return ''.join(('(', *recurse(0), ')'))
-
-    def _not_strings(self, strings):
-        class TrieNode:
-            def __init__(self):
-                self.children = {}
-                self.is_end_of_string = False
-
-            def insert(self, string):
-                node = self
-                for c in string:
-                    node = node.children.setdefault(c, TrieNode())
-                node.is_end_of_string = True
-
-        trie = TrieNode()
-        for s in strings:
-            trie.insert(s)
-
-        char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
-        out = ['["] ( ']
-
-        def visit(node):
-            rejects = []
-            first = True
-            for c in sorted(node.children.keys()):
-                child = node.children[c]
-                rejects.append(c)
-                if first:
-                    first = False
-                else:
-                    out.append(' | ')
-                out.append(f'[{c}]')
-                if child.children:
-                    out.append(f' (')
-                    visit(child)
-                    out.append(')')
-                elif child.is_end_of_string:
-                    out.append(f' {char_rule}+')
-            if node.children:
-                if not first:
-                    out.append(' | ')
-                out.append(f'[^"{"".join(rejects)}] {char_rule}*')
-        visit(trie)
-
-        out.append(f' ){"" if trie.is_end_of_string else "?"} ["]')
-        return ''.join(out)
-
-    def _add_rule(self, name, rule):
-        esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
-        if esc_name not in self._rules or self._rules[esc_name] == rule:
-            key = esc_name
-        else:
-            i = 0
-            while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule:
-                i += 1
-            key = f'{esc_name}{i}'
-        self._rules[key] = rule
-        return key
-
-    def resolve_refs(self, schema: dict, url: str):
-        '''
-            Resolves all $ref fields in the given schema, fetching any remote schemas,
-            replacing $ref with absolute reference URL and populating self._refs with the
-            respective referenced (sub)schema dictionaries.
-        '''
-        def visit(n: dict):
-            if isinstance(n, list):
-                return [visit(x) for x in n]
-            elif isinstance(n, dict):
-                ref = n.get('$ref')
-                if ref is not None and ref not in self._refs:
-                    if ref.startswith('https://'):
-                        assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)'
-                        import requests
-
-                        frag_split = ref.split('#')
-                        base_url = frag_split[0]
-
-                        target = self._refs.get(base_url)
-                        if target is None:
-                            target = self.resolve_refs(requests.get(ref).json(), base_url)
-                            self._refs[base_url] = target
-
-                        if len(frag_split) == 1 or frag_split[-1] == '':
-                            return target
-                    elif ref.startswith('#/'):
-                        target = schema
-                        ref = f'{url}{ref}'
-                        n['$ref'] = ref
-                    else:
-                        raise ValueError(f'Unsupported ref {ref}')
-
-                    for sel in ref.split('#')[-1].split('/')[1:]:
-                        assert target is not None, f'Error resolving ref {ref}: {sel} not in {target}'
-                        if isinstance(target, list):
-                            try:
-                                sel_index = int(sel)
-                            except ValueError:
-                                raise ValueError(f'Error resolving ref {ref}: {sel} not in {target}')
-                            assert 0 <= sel_index < len(target), f'Error resolving ref {ref}: {sel} not in {target}'
-                            target = target[sel_index]
-                        else:
-                            assert sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
-                            target = target[sel]
-
-                    self._refs[ref] = target
-                else:
-                    for v in n.values():
-                        visit(v)
-
-            return n
-        return visit(schema)
-
-    def _generate_union_rule(self, name, alt_schemas):
-        return ' | '.join((
-            self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}')
-            for i, alt_schema in enumerate(alt_schemas)
-        ))
-
-    def _visit_pattern(self, pattern, name):
-        '''
-            Transforms a regular expression pattern into a GBNF rule.
-
-            Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
-            Output: https://github.com/ggml-org/llama.cpp/blob/master/grammars/README.md
-
-            Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
-
-            Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which
-            we define sub-rules to keep the output lean.
-        '''
-
-        assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"'
-        pattern = pattern[1:-1]
-        sub_rule_ids = {}
-
-        i = 0
-        length = len(pattern)
-
-        def to_rule(s: tuple[str, bool]) -> str:
-            (txt, is_literal) = s
-            return "\"" + txt + "\"" if is_literal else txt
-
-        def transform() -> tuple[str, bool]:
-            '''
-                Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
-            '''
-            nonlocal i
-            nonlocal pattern
-            nonlocal sub_rule_ids
-
-            start = i
-            # For each component of this sequence, store its string representation and whether it's a literal.
-            # We only need a flat structure here to apply repetition operators to the last item, and
-            # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
-            # (GBNF's syntax is luckily very close to regular expressions!)
-            seq: list[tuple[str, bool]] = []
-
-            def get_dot():
-                if self._dotall:
-                    rule = DOTALL
-                else:
-                    # Accept any character... except \n and \r line break chars (\x0A and \xOD)
-                    rule = DOT
-                return self._add_rule(f'dot', rule)
-
-            def join_seq():
-                nonlocal seq
-                ret = []
-                for is_literal, g in itertools.groupby(seq, lambda x: x[1]):
-                    if is_literal:
-                        ret.append((''.join(x[0] for x in g), True))
-                    else:
-                        ret.extend(g)
-                if len(ret) == 1:
-                    return ret[0]
-                return (' '.join(to_rule(x) for x in seq), False)
-
-            while i < length:
-                c = pattern[i]
-                if c == '.':
-                    seq.append((get_dot(), False))
-                    i += 1
-                elif c == '(':
-                    i += 1
-                    if i < length:
-                        assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/'
-                    seq.append((f'({to_rule(transform())})', False))
-                elif c == ')':
-                    i += 1
-                    assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}'
-                    return join_seq()
-                elif c == '[':
-                    square_brackets = c
-                    i += 1
-                    while i < length and pattern[i] != ']':
-                        if pattern[i] == '\\':
-                            square_brackets += pattern[i:i+2]
-                            i += 2
-                        else:
-                            square_brackets += pattern[i]
-                            i += 1
-                    assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}'
-                    square_brackets += ']'
-                    i += 1
-                    seq.append((square_brackets, False))
-                elif c == '|':
-                    seq.append(('|', False))
-                    i += 1
-                elif c in ('*', '+', '?'):
-                    seq[-1] = (to_rule(seq[-1]) + c, False)
-                    i += 1
-                elif c == '{':
-                    curly_brackets = c
-                    i += 1
-                    while i < length and pattern[i] != '}':
-                        curly_brackets += pattern[i]
-                        i += 1
-                    assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}'
-                    curly_brackets += '}'
-                    i += 1
-                    nums = [s.strip() for s in curly_brackets[1:-1].split(',')]
-                    min_times = 0
-                    max_times = None
-                    try:
-                        if len(nums) == 1:
-                            min_times = int(nums[0])
-                            max_times = min_times
-                        else:
-                            assert len(nums) == 2
-                            min_times = int(nums[0]) if nums[0] else 0
-                            max_times = int(nums[1]) if nums[1] else None
-                    except ValueError:
-                        raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/')
-
-                    (sub, sub_is_literal) = seq[-1]
-
-                    if not sub_is_literal:
-                        id = sub_rule_ids.get(sub)
-                        if id is None:
-                            id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
-                            sub_rule_ids[sub] = id
-                        sub = id
-
-                    seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times), False)
-                else:
-                    literal = ''
-                    while i < length:
-                        if pattern[i] == '\\' and i < length - 1:
-                            next = pattern[i + 1]
-                            if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS:
-                                i += 1
-                                literal += pattern[i]
-                                i += 1
-                            else:
-                                literal += pattern[i:i+2]
-                                i += 2
-                        elif pattern[i] == '"' and not self._raw_pattern:
-                            literal += '\\"'
-                            i += 1
-                        elif pattern[i] not in NON_LITERAL_SET and \
-                                (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET):
-                            literal += pattern[i]
-                            i += 1
-                        else:
-                            break
-                    if literal:
-                        seq.append((literal, True))
-
-            return join_seq()
-
-        return self._add_rule(
-            name,
-            to_rule(transform()) if self._raw_pattern \
-                else "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"")
-
-
-    def _resolve_ref(self, ref):
-        ref_fragment = ref.split('#')[-1]
-        ref_name = 'ref' + re.sub(r'[^a-zA-Z0-9-]+', '-', ref_fragment)
-        if ref_name not in self._rules and ref not in self._refs_being_resolved:
-            self._refs_being_resolved.add(ref)
-            resolved = self._refs[ref]
-            ref_name = self.visit(resolved, ref_name)
-            self._refs_being_resolved.remove(ref)
-        return ref_name
-
-    def _generate_constant_rule(self, value):
-        return self._format_literal(json.dumps(value))
-
-    def visit(self, schema, name):
-        schema_type = schema.get('type')
-        schema_format = schema.get('format')
-        rule_name = name + '-' if name in RESERVED_NAMES else name or 'root'
-
-        if (ref := schema.get('$ref')) is not None:
-            return self._add_rule(rule_name, self._resolve_ref(ref))
-
-        elif 'oneOf' in schema or 'anyOf' in schema:
-            return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
-
-        elif isinstance(schema_type, list):
-            return self._add_rule(rule_name, self._generate_union_rule(name, [{**schema, 'type': t} for t in schema_type]))
-
-        elif 'const' in schema:
-            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']))
-
-        elif 'enum' in schema:
-            rule = '(' + ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) + ')'
-            return self._add_rule(rule_name, rule)
-
-        elif schema_type in (None, 'object') and \
-             ('properties' in schema or \
-              ('additionalProperties' in schema and schema['additionalProperties'] is not True)):
-            required = set(schema.get('required', []))
-            properties = list(schema.get('properties', {}).items())
-            return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
-
-        elif schema_type in (None, 'object', 'string') and 'allOf' in schema:
-            required = set()
-            properties = []
-            enum_sets = []
-            hybrid_name = name
-            def add_component(comp_schema, is_required):
-                if (ref := comp_schema.get('$ref')) is not None:
-                    comp_schema = self._refs[ref]
-
-                if 'properties' in comp_schema:
-                    for prop_name, prop_schema in comp_schema['properties'].items():
-                        properties.append((prop_name, prop_schema))
-                        if is_required:
-                            required.add(prop_name)
-
-                if 'enum' in comp_schema:
-                    enum_sets.append(set(comp_schema['enum']))
-
-            for t in schema['allOf']:
-                if 'anyOf' in t:
-                    for tt in t['anyOf']:
-                        add_component(tt, is_required=False)
-                else:
-                    add_component(t, is_required=True)
-
-            if enum_sets:
-                enum_intersection = enum_sets[0]
-                for s in enum_sets[1:]:
-                    enum_intersection &= s
-
-                if enum_intersection:
-                    rule = '(' + ' | '.join((self._generate_constant_rule(v) for v in sorted(enum_intersection))) + ')'
-                    return self._add_rule(rule_name, rule)
-
-            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None))
-
-        elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
-            items = schema.get('items', schema.get('prefixItems'))
-            if isinstance(items, list):
-                return self._add_rule(
-                    rule_name,
-                    '"[" space ' +
-                    ' "," space '.join(
-                        self.visit(item, f'{name}{"-" if name else ""}tuple-{i}')
-                        for i, item in enumerate(items)) +
-                    ' space "]"')
-            else:
-                item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
-                min_items = schema.get("minItems", 0)
-                max_items = schema.get("maxItems")
-                return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' space "]"')
-
-        elif schema_type in (None, 'string') and 'pattern' in schema:
-            return self._visit_pattern(schema['pattern'], rule_name)
-
-        elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''):
-            return self._add_primitive(
-                'root' if rule_name == 'root' else schema_format,
-                PRIMITIVE_RULES['uuid']
-            )
-
-        elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES:
-            prim_name = f'{schema_format}-string'
-            return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]))
-
-        elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema):
-            char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
-            min_len = schema.get('minLength', 0)
-            max_len = schema.get('maxLength')
-
-            return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\""')
-
-        elif schema_type in (None, 'integer') and \
-                ('minimum' in schema or 'exclusiveMinimum' in schema or 'maximum' in schema or 'exclusiveMaximum' in schema):
-            min_value = None
-            max_value = None
-            if 'minimum' in schema:
-                min_value = schema['minimum']
-            elif 'exclusiveMinimum' in schema:
-                min_value = schema['exclusiveMinimum'] + 1
-            if 'maximum' in schema:
-                max_value = schema['maximum']
-            elif 'exclusiveMaximum' in schema:
-                max_value = schema['exclusiveMaximum'] - 1
-
-            out = ["("]
-            _generate_min_max_int(min_value, max_value, out)
-            out.append(")")
-            return self._add_rule(rule_name, ''.join(out))
-
-        elif (schema_type == 'object') or (len(schema) == 0):
-            return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
-
-        elif schema_type is None and isinstance(schema, dict):
-            # No type constraint and no recognized structural keywords (e.g. {"description": "..."}).
-            # Per JSON Schema semantics this is equivalent to {} and accepts any value.
-            return self._add_rule(rule_name, self._add_primitive('value', PRIMITIVE_RULES['value']))
-
-        else:
-            assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
-            # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-            return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type])
-
-    def _add_primitive(self, name: str, rule: BuiltinRule):
-        n = self._add_rule(name, rule.content)
-
-        for dep in rule.deps:
-            dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep)
-            assert dep_rule, f'Rule {dep} not known'
-            if dep not in self._rules:
-                self._add_primitive(dep, dep_rule)
-        return n
-
-    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Optional[Union[bool, Any]]):
-        prop_order = self._prop_order
-        # sort by position in prop_order (if specified) then by original order
-        sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
-
-        prop_kv_rule_names = {}
-        for prop_name, prop_schema in properties:
-            prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
-            prop_kv_rule_names[prop_name] = self._add_rule(
-                f'{name}{"-" if name else ""}{prop_name}-kv',
-                fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}'
-            )
-        required_props = [k for k in sorted_props if k in required]
-        optional_props = [k for k in sorted_props if k not in required]
-
-        if additional_properties is not None and additional_properties != False:
-            sub_name = f'{name}{"-" if name else ""}additional'
-            value_rule = self.visit(additional_properties, f'{sub_name}-value') if isinstance(additional_properties, dict) else \
-                self._add_primitive('value', PRIMITIVE_RULES['value'])
-            key_rule = self._add_primitive('string', PRIMITIVE_RULES['string']) if not sorted_props \
-                else self._add_rule(f'{sub_name}-k', self._not_strings(sorted_props))
-
-            prop_kv_rule_names["*"] = self._add_rule(
-                f'{sub_name}-kv',
-                f'{key_rule} ":" space {value_rule}'
-            )
-            optional_props.append("*")
-
-        if not required_props and not optional_props:
-            return '"{" space "}"'
-
-        rule = '"{" space '
-        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)
-
-        if optional_props:
-            rule += ' ('
-            if required_props:
-                rule += ' "," space ( '
-
-            def get_recursive_refs(ks, first_is_optional):
-                [k, *rest] = ks
-                kv_rule_name = prop_kv_rule_names[k]
-                comma_ref = f'( "," space {kv_rule_name} )'
-                if first_is_optional:
-                    res = comma_ref + ('*' if k == '*' else '?')
-                else:
-                    res = kv_rule_name + (' ' + comma_ref + "*" if k == '*' else '')
-                if len(rest) > 0:
-                    res += ' ' + self._add_rule(
-                        f'{name}{"-" if name else ""}{k}-rest',
-                        get_recursive_refs(rest, first_is_optional=True)
-                    )
-                return res
-
-            rule += ' | '.join(
-                get_recursive_refs(optional_props[i:], first_is_optional=False)
-                for i in range(len(optional_props))
-            )
-            if required_props:
-                rule += ' )'
-            rule += ' )?'
-
-        rule += ' space "}"'
-
-        return rule
-
-    def format_grammar(self):
-        return '\n'.join(
-            f'{name} ::= {rule}'
-            for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0])
-        )
-
-
-def main(args_in = None):
-    parser = argparse.ArgumentParser(
-        description='''
-            Generates a grammar (suitable for use in ./llama-cli) that produces JSON conforming to a
-            given JSON schema. Only a subset of JSON schema features are supported; more may be
-            added in the future.
-        ''',
-    )
-    parser.add_argument(
-        '--prop-order',
-        default=[],
-        type=lambda s: s.split(','),
-        help='''
-            comma-separated property names defining the order of precedence for object properties;
-            properties not specified here are given lower precedence than those that are, and
-            are kept in their original order from the schema. Required properties are always
-            given precedence over optional properties.
-        '''
-    )
-    parser.add_argument(
-        '--allow-fetch',
-        action='store_true',
-        default=False,
-        help='Whether to allow fetching referenced schemas over HTTPS')
-    parser.add_argument(
-        '--dotall',
-        action='store_true',
-        default=False,
-        help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns')
-    parser.add_argument(
-        '--raw-pattern',
-        action='store_true',
-        default=False,
-        help='Treats string patterns as raw patterns w/o quotes (or quote escapes)')
-
-    parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
-    args = parser.parse_args(args_in)
-
-    if args.schema.startswith('https://'):
-        url = args.schema
-        import requests
-        schema = requests.get(url).json()
-    elif args.schema == '-':
-        url = 'stdin'
-        schema = json.load(sys.stdin)
-    else:
-        url = f'file://{args.schema}'
-        with open(args.schema) as f:
-            schema = json.load(f)
-    converter = SchemaConverter(
-        prop_order={name: idx for idx, name in enumerate(args.prop_order)},
-        allow_fetch=args.allow_fetch,
-        dotall=args.dotall,
-        raw_pattern=args.raw_pattern)
-    schema = converter.resolve_refs(schema, url)
-    converter.visit(schema, '')
-    print(converter.format_grammar())
-
-
-if __name__ == '__main__':
-    main()

examples/regex_to_grammar.py

@@ -1,20 +0,0 @@
-import json, subprocess, sys, os
-
-assert len(sys.argv) >= 2
-[_, pattern, *rest] = sys.argv
-
-print(subprocess.check_output(
-    [
-        "python",
-        os.path.join(
-        os.path.dirname(os.path.realpath(__file__)),
-        "json_schema_to_grammar.py"),
-        *rest,
-        "-",
-        "--raw-pattern",
-    ],
-    text=True,
-    input=json.dumps({
-        "type": "string",
-        "pattern": pattern,
-    }, indent=2)))

examples/speculative-simple/speculative-simple.cpp

@@ -188,7 +188,7 @@ int main(int argc, char ** argv) {
             common_speculative_get_draft_params(spec, seq_id) = {
                 /* .drafting   = */ true,
                 /* .n_max      = */ n_draft_max,
-                /* .n_past     = */ n_past,
+                /* .pos0       = */ n_past,
                 /* .id_last    = */ id_last,
                 /* .prompt     = */ &prompt_tgt,
                 /* .result     = */ &draft, // output

examples/ts-type-to-grammar.sh

@@ -1,28 +0,0 @@
-#!/usr/bin/env bash
-#
-# ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
-# python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json
-#
-set -euo pipefail
-
-readonly type="$1"
-
-# Create a temporary directory
-TMPDIR=""
-trap 'rm -fR "$TMPDIR"' EXIT
-TMPDIR=$(mktemp -d)
-
-DTS_FILE="$TMPDIR/type.d.ts"
-SCHEMA_FILE="$TMPDIR/schema.json"
-
-echo "export type MyType = $type" > "$DTS_FILE"
-
-# This is a fork of typescript-json-schema, actively maintained as of March 2024:
-# https://github.com/vega/ts-json-schema-generator
-npx ts-json-schema-generator --unstable --no-top-ref --path "$DTS_FILE" --type MyType -e none > "$SCHEMA_FILE"
-
-# Alternative, not actively maintained as of March 2024:
-# https://github.com/YousefED/typescript-json-schema
-# npx typescript-json-schema --defaultProps --required "$DTS_FILE" MyType | tee "$SCHEMA_FILE" >&2
-
-./examples/json_schema_to_grammar.py "$SCHEMA_FILE"

ggml/src/ggml-backend.cpp

@@ -1705,6 +1705,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                     ggml_tensor * ids_tensor = node->src[2];
                     ggml_backend_t ids_backend = split_backend;
 
+                    if (ggml_nelements(ids_tensor) == 0) {
+                        continue;
+                    }
+
                     // if the ids tensor is also an input of the split, it may not have been copied yet to the split backend
                     // in that case, we use the original ids tensor
                     for (int i = input_id + 1; i < split->n_inputs; i++) {

ggml/src/ggml-cpu/CMakeLists.txt

@@ -675,6 +675,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})
     target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
 
+    if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND NOT GGML_SYSTEM_ARCH STREQUAL "x86")
+        message(STATUS "Skipping PCH for ${GGML_CPU_NAME}: GCC PCH is only enabled for x86 (arch: ${GGML_SYSTEM_ARCH})")
+    else()
+        target_precompile_headers(${GGML_CPU_NAME} PRIVATE ggml-impl.h)
+    endif()
+
     if (EMSCRIPTEN)
         set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()

ggml/src/ggml-cpu/ops.h

@@ -18,7 +18,15 @@
 #endif
 #endif
 
+// -Winterference-size was introduced in GCC 12
+#if defined(__cplusplus) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Winterference-size"
+#endif
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
+#if defined(__cplusplus) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12
+#pragma GCC diagnostic pop
+#endif
 
 // Work buffer size for im2col operations in CONV2D
 #define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)

ggml/src/ggml-cuda/fattn-common.cuh

@@ -1133,12 +1133,21 @@ void launch_fattn(
 
     dim3 blocks_num;
     if (stream_k) {
-        // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
-        const int max_blocks = max_blocks_per_sm*nsm;
-        const int tiles_nwaves = (ntiles_dst + max_blocks - 1) / max_blocks;
-        const int tiles_efficiency_percent = 100 * ntiles_dst / (max_blocks*tiles_nwaves);
+        auto should_use_stream_k = [](const int cc, const int ntiles_dst, const int max_blocks, const int DKQ) {
+            const int tiles_nwaves             = (ntiles_dst + max_blocks - 1) / max_blocks;
+            const int tiles_efficiency_percent = 100 * ntiles_dst / (max_blocks*tiles_nwaves);
 
-        const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || amd_wmma_available(cc) || tiles_efficiency_percent < 75;
+            if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                return true;
+            }
+            if (amd_wmma_available(cc) && DKQ == 64) {
+                return true; // TODO better configuration
+            }
+            return tiles_efficiency_percent < 75;
+        };
+
+        const int  max_blocks   = max_blocks_per_sm*nsm;
+        const bool use_stream_k = should_use_stream_k(cc, ntiles_dst, max_blocks, Q->ne[0]);
 
         blocks_num.x = ntiles_dst;
         blocks_num.y = 1;

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -158,8 +158,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
 
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256,  8,  64, 2,  32, 128, 128, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 16,  64, 2,  32, 128, 128, 128, 1, true);
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 1, true);
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 256, 2,  64, 128, 128,  64, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 256, 2,  64, 128, 128,  64, 1, true);
 
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 32, 128, 2,  32, 160, 128, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 64, 128, 2,  32, 160, 128, 128, 1, true);
@@ -1826,7 +1826,7 @@ static __global__ void flash_attn_ext_f16(
 #endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
 
 #if defined(AMD_WMMA_AVAILABLE)
-    if (ncols1*ncols2 < 16 || ncols2 == 1 || DKQ > 128) {
+    if (ncols1*ncols2 < 16 || ncols2 == 1 || DKQ > 256) {
         NO_DEVICE_CODE;
         return;
     }

ggml/src/ggml-cuda/fattn.cu

@@ -221,6 +221,24 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
         }
     }
 
+    // On RDNA it is preferable to minimize wasted compute vs. duplicate I/O for the mask.
+    if (amd_wmma_available(cc)) {
+        if (use_gqa_opt && gqa_ratio % 8 == 0) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
+            return;
+        }
+
+        if (use_gqa_opt && gqa_ratio % 4 == 0) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
+            return;
+        }
+
+        if (use_gqa_opt && gqa_ratio % 2 == 0) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
+            return;
+        }
+    }
+
     if (use_gqa_opt && gqa_ratio > 4) {
         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
         return;
@@ -646,8 +664,9 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         }
     }
 
-    // AMD WMMA is always faster than the tile kernel if the full tile width of 16 can be utilized.
-    if ((amd_wmma_available(cc) && gqa_opt_applies && Q->ne[0] <= 128) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[1] * gqa_ratio_eff > 8) {
+    // AMD WMMA is faster than the tile kernel if the wide tiles with high arithmetic intensity can be utilized.
+    if ((amd_wmma_available(cc) && gqa_opt_applies && Q->ne[0] <= 256) && Q->ne[0] != 40 && Q->ne[0] != 72 &&
+            Q->ne[1] * gqa_ratio_eff > (Q->ne[0] <= 128 ? 8 : 16)) {
         return BEST_FATTN_KERNEL_MMA_F16;
     }
 

ggml/src/ggml-cuda/mmq-config-gcn.cuh

@@ -0,0 +1,281 @@
+static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_gcn(ggml_type type, int J, bool fallback) {
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_1, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 512, 2,  64,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_Q2_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 512, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q3_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 3,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_IQ1_S, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_XS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_S, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ3_S, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ4_XS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ4_NL, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_MXFP4, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 512, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2,  64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 512, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+}

ggml/src/ggml-cuda/mmq.cuh

@@ -218,6 +218,7 @@ struct ggml_cuda_mmq_config {
 #include "mmq-config-ampere.cuh"
 #include "mmq-config-blackwell.cuh"
 
+#include "mmq-config-gcn.cuh"
 #include "mmq-config-cdna.cuh"
 #include "mmq-config-rdna2.cuh"
 #include "mmq-config-rdna3.cuh"
@@ -228,6 +229,9 @@ struct ggml_cuda_mmq_config {
 
 static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type type, const int J, const bool fallback, const int cc) {
     if (GGML_CUDA_CC_IS_AMD(cc)) {
+        if (GGML_CUDA_CC_IS_GCN(cc)) {
+            return ggml_cuda_mmq_get_config_gcn(type, J, fallback);
+        }
         if (GGML_CUDA_CC_IS_CDNA(cc)) {
             return ggml_cuda_mmq_get_config_cdna(type, J, fallback);
         }
@@ -256,7 +260,9 @@ static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type ty
 
 static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_type type, int J, bool fallback) {
 #ifdef GGML_USE_HIP
-#ifdef CDNA
+#ifdef GCN
+    return ggml_cuda_mmq_get_config_gcn(type, J, fallback);
+#elif defined(CDNA)
     return ggml_cuda_mmq_get_config_cdna(type, J, fallback);
 #elif defined(RDNA4)
     return ggml_cuda_mmq_get_config_rdna4(type, J, fallback);

ggml/src/ggml-hexagon/htp-opnode.h

@@ -344,6 +344,12 @@ struct htp_opformat {
         } else if (htp_op_is_unary(node.opcode)) {
             const auto * kparams = (const struct htp_unary_kernel_params *) node.kernel_params;
             snprintf(str, max_size, "%s vtcm %d", kparams->col_tile ? "wide-row" : "row-block", (int) kparams->vtcm_size);
+        } else if (node.opcode == HTP_OP_MDEV_GROUP && node.node) {
+            snprintf(str, max_size, "idx %d count %d", (int) node.node->op_params[0], (int) node.dst()->ne[1]);
+        } else if ((node.opcode == HTP_OP_FENCE || node.opcode == HTP_OP_CPY_FENCE) && node.node) {
+            snprintf(str, max_size, "seq 0x%x", (uint32_t) node.node->op_params[0]);
+        } else if (node.opcode == HTP_OP_ALLREDUCE && node.node) {
+            snprintf(str, max_size, "seq 0x%x -> 0x%x", (uint32_t) node.node->op_params[0], (uint32_t) node.node->op_params[1]);
         } else {
             snprintf(str, max_size, "----");
         }

ggml/src/ggml-hexagon/htp/act-ops.c

@@ -3,7 +3,6 @@
 #pragma clang diagnostic ignored "-Wunused-but-set-variable"
 
 #include <HAP_farf.h>
-#include <HAP_perf.h>
 
 #include <math.h>
 #include <string.h>
@@ -15,7 +14,7 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
+#include "hex-common.h"
 #include "htp-tensor.h"
 #include "htp-vtcm.h"
 
@@ -80,6 +79,7 @@ struct htp_act_context {
     uint32_t                 block;
     uint32_t                 src0_nrows;
     uint32_t                 src0_nrows_per_thread;
+    uint32_t                 row_start;
     int                      nc;
 
     uint8_t *                vtcm_src0;
@@ -329,104 +329,104 @@ static void geglu_f32(const float * restrict src0,
     }
 }
 
-#define DEFINE_GLU_PER_THREAD(NAME, OP_STR, CORE_EXPR)                                                                 \
-    static void glu_##NAME##_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {                         \
-        struct htp_act_context * actx = (struct htp_act_context *) data;                                               \
-        htp_act_preamble;                                                                                              \
-                                                                                                                       \
-        struct htp_thread_trace * tr = actx->octx->ctx ? &actx->octx->ctx->trace[ith] : NULL;                          \
-                                                                                                                       \
-        size_t src0_row_size = actx->src0_row_size;                                                                    \
-        size_t src1_row_size = actx->src1_row_size;                                                                    \
-        size_t dst_row_size  = actx->dst_row_size;                                                                     \
-                                                                                                                       \
-        size_t src0_row_stride = actx->src0_row_stride;                                                                \
-        size_t src1_row_stride = actx->src1_row_stride;                                                                \
-                                                                                                                       \
-        const uint32_t src0_nrows            = actx->src0_nrows;                                                       \
-        const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;                                            \
-                                                                                                                       \
-        const uint32_t src0_start_row = src0_nrows_per_thread * ith;                                                   \
-        const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);                       \
-                                                                                                                       \
-        /* no work for this thread */                                                                                  \
-        if (src0_start_row >= src0_end_row) {                                                                          \
-            return;                                                                                                    \
-        }                                                                                                              \
-                                                                                                                       \
-        const uint8_t * restrict data_src0 = actx->data_src0;                                                          \
-        const uint8_t * restrict data_src1 = actx->data_src1;                                                          \
-        uint8_t * restrict data_dst        = actx->data_dst;                                                           \
-                                                                                                                       \
-        const size_t src0_row_size_aligned = actx->src0_row_size_aligned;                                              \
-        const size_t src1_row_size_aligned = actx->src1_row_size_aligned;                                              \
-        const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;                                               \
-                                                                                                                       \
-        uint8_t * restrict src0_spad_data = actx->vtcm_src0 + (ith * actx->vtcm_src0_size_per_thread);                 \
-        uint8_t * restrict src1_spad_data = actx->vtcm_src1 + (ith * actx->vtcm_src1_size_per_thread);                 \
-        uint8_t * restrict dst_spad_data  = actx->vtcm_dst  + (ith * actx->vtcm_dst_size_per_thread);                  \
-                                                                                                                       \
-        size_t src0_spad_half_size = actx->src0_spad_half_size;                                                        \
-        size_t src1_spad_half_size = actx->src1_spad_half_size;                                                        \
-        size_t dst_spad_half_size  = actx->dst_spad_half_size;                                                         \
-                                                                                                                       \
-        const int BLOCK = actx->block;                                                                                 \
-        if (BLOCK == 0) {                                                                                              \
-            FARF(ERROR,                                                                                                \
-                 OP_STR                                                                                                \
-                 " : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",      \
-                 actx->vtcm_src0_size_per_thread, src0_row_size_aligned);                                              \
-            return;                                                                                                    \
-        }                                                                                                              \
-                                                                                                                       \
-        dma_queue * dma_queue = actx->octx->ctx->dma[ith];                                                             \
-                                                                                                                       \
-        /* See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379 */                 \
-        for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) { \
-            const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);                                                 \
-                                                                                                                       \
-            /* Dummy DMA transation for sequencing (interleaving dst,src,dst,...) */                                   \
-            dma_queue_push_vtcm_to_ddr(dma_queue,                                                                      \
-                                       dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),        \
-                                       dst_row_size, dst_row_size_aligned, 0);                                         \
-                                                                                                                       \
-            dma_queue_push(                                                                                            \
-                dma_queue,                                                                                             \
-                dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_stride)),   \
-                src0_row_size_aligned, src0_row_stride, src0_row_size, block_size);                                    \
-            dma_queue_push(                                                                                            \
-                dma_queue,                                                                                             \
-                dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_stride)),   \
-                src1_row_size_aligned, src1_row_stride, src1_row_size, block_size);                                    \
-        }                                                                                                              \
-                                                                                                                       \
-        for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {                                           \
-            const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);                                                 \
-                                                                                                                       \
-            float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;                                                \
-            float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;                                                \
-            float * src1_spad = (float *) dma_queue_pop(dma_queue).dst;                                                \
-                                                                                                                       \
-            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir);                                                     \
-            CORE_EXPR;                                                                                                 \
-            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir);                                                      \
-                                                                                                                       \
-            dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad),              \
-                                       dst_row_size, dst_row_size_aligned, block_size);                                \
-                                                                                                                       \
-            /* prefetch N+2 loop iteration if any */                                                                   \
-            const uint32_t pref_block = (ir + BLOCK * 2);                                                              \
-            if (pref_block < src0_end_row) {                                                                           \
-                const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);                                \
-                dma_queue_push(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_stride)),         \
-                               src0_row_size_aligned, src0_row_stride, src0_row_size, pref_block_size);                \
-                dma_queue_push(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_stride)),         \
-                               src1_row_size_aligned, src1_row_stride, src1_row_size, pref_block_size);                \
-            }                                                                                                          \
-        }                                                                                                              \
-                                                                                                                       \
-        dma_queue_flush(dma_queue);                                                                                    \
-                                                                                                                       \
+#define DEFINE_GLU_PER_THREAD(NAME, OP_STR, CORE_EXPR)                                                                   \
+    static void glu_##NAME##_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {                           \
+        struct htp_act_context * actx = (struct htp_act_context *) data;                                                 \
+        htp_act_preamble;                                                                                                \
+                                                                                                                         \
+        struct htp_thread_trace * tr = actx->octx->ctx ? &actx->octx->ctx->trace[ith] : NULL;                            \
+                                                                                                                         \
+        size_t src0_row_size = actx->src0_row_size;                                                                      \
+        size_t src1_row_size = actx->src1_row_size;                                                                      \
+        size_t dst_row_size  = actx->dst_row_size;                                                                       \
+                                                                                                                         \
+        size_t src0_row_stride = actx->src0_row_stride;                                                                  \
+        size_t src1_row_stride = actx->src1_row_stride;                                                                  \
+                                                                                                                         \
+        const uint32_t src0_nrows            = actx->src0_nrows;                                                         \
+        const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;                                              \
+                                                                                                                         \
+        const uint32_t src0_start_row = actx->row_start + src0_nrows_per_thread * ith;                                   \
+        const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, actx->row_start + src0_nrows);       \
+                                                                                                                         \
+        /* no work for this thread */                                                                                    \
+        if (src0_start_row >= src0_end_row) {                                                                            \
+            return;                                                                                                      \
+        }                                                                                                                \
+                                                                                                                         \
+        const uint8_t * restrict data_src0 = actx->data_src0;                                                            \
+        const uint8_t * restrict data_src1 = actx->data_src1;                                                            \
+        uint8_t * restrict data_dst        = actx->data_dst;                                                             \
+                                                                                                                         \
+        const size_t src0_row_size_aligned = actx->src0_row_size_aligned;                                                \
+        const size_t src1_row_size_aligned = actx->src1_row_size_aligned;                                                \
+        const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;                                                 \
+                                                                                                                         \
+        uint8_t * restrict src0_spad_data = actx->vtcm_src0 + (ith * actx->vtcm_src0_size_per_thread);                   \
+        uint8_t * restrict src1_spad_data = actx->vtcm_src1 + (ith * actx->vtcm_src1_size_per_thread);                   \
+        uint8_t * restrict dst_spad_data  = actx->vtcm_dst  + (ith * actx->vtcm_dst_size_per_thread);                    \
+                                                                                                                         \
+        size_t src0_spad_half_size = actx->src0_spad_half_size;                                                          \
+        size_t src1_spad_half_size = actx->src1_spad_half_size;                                                          \
+        size_t dst_spad_half_size  = actx->dst_spad_half_size;                                                           \
+                                                                                                                         \
+        const int BLOCK = actx->block;                                                                                   \
+        if (BLOCK == 0) {                                                                                                \
+            FARF(ERROR,                                                                                                  \
+                 OP_STR                                                                                                  \
+                 " : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",        \
+                 actx->vtcm_src0_size_per_thread, src0_row_size_aligned);                                                \
+            return;                                                                                                      \
+        }                                                                                                                \
+                                                                                                                         \
+        dma_queue * dma_queue = actx->octx->ctx->dma[ith];                                                               \
+                                                                                                                         \
+        /* See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379 */                   \
+        for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {   \
+            const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);                                                   \
+                                                                                                                         \
+            /* Dummy DMA transation for sequencing (interleaving dst,src,dst,...) */                                     \
+            dma_queue_push_vtcm_to_ddr(dma_queue,                                                                        \
+                                       dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),          \
+                                       dst_row_size, dst_row_size_aligned, 0);                                           \
+                                                                                                                         \
+            dma_queue_push(                                                                                              \
+                dma_queue,                                                                                               \
+                dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_stride)),     \
+                src0_row_size_aligned, src0_row_stride, src0_row_size, block_size);                                      \
+            dma_queue_push(                                                                                              \
+                dma_queue,                                                                                               \
+                dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_stride)),     \
+                src1_row_size_aligned, src1_row_stride, src1_row_size, block_size);                                      \
+        }                                                                                                                \
+                                                                                                                         \
+        for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {                                             \
+            const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);                                                   \
+                                                                                                                         \
+            float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;                                                  \
+            float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;                                                  \
+            float * src1_spad = (float *) dma_queue_pop(dma_queue).dst;                                                  \
+                                                                                                                         \
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir);                                                       \
+            CORE_EXPR;                                                                                                   \
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir);                                                        \
+                                                                                                                         \
+            dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad),                \
+                                       dst_row_size, dst_row_size_aligned, block_size);                                  \
+                                                                                                                         \
+            /* prefetch N+2 loop iteration if any */                                                                     \
+            const uint32_t pref_block = (ir + BLOCK * 2);                                                                \
+            if (pref_block < src0_end_row) {                                                                             \
+                const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);                                  \
+                dma_queue_push(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_stride)),           \
+                               src0_row_size_aligned, src0_row_stride, src0_row_size, pref_block_size);                  \
+                dma_queue_push(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_stride)),           \
+                               src1_row_size_aligned, src1_row_stride, src1_row_size, pref_block_size);                  \
+            }                                                                                                            \
+        }                                                                                                                \
+                                                                                                                         \
+        dma_queue_flush(dma_queue);                                                                                      \
+                                                                                                                         \
     }
 
 DEFINE_GLU_PER_THREAD(swiglu, "swiglu-f32", swiglu_f32(src0_spad, src1_spad, dst_spad, block_size, actx))
@@ -473,14 +473,30 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     }
 
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
-    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
+    const size_t dst_row_size = dst->ne[0] * SIZEOF_FP32;
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = src0_nrows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, sizeof(float), (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(src0_nrows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
 
     // row_size   = bytes of useful data per row (what the kernel touches / what DMA copies).
     // row_stride = bytes between successive rows in DDR (may exceed row_size for non-contig src).
-    const size_t nc_bytes    = dst->ne[0] * SIZEOF_FP32;
-    const size_t src0_row_size = nc_bytes;
-    const size_t src1_row_size = nc_bytes;
-    const size_t dst_row_size  = nc_bytes;
+    const size_t nc_bytes        = dst_row_size;
+    const size_t src0_row_size   = nc_bytes;
+    const size_t src1_row_size   = nc_bytes;
     const size_t src0_row_stride = src0->nb[1];
     const size_t src1_row_stride = src1 ? src1->nb[1] : src0->nb[1];
 
@@ -518,7 +534,7 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     struct htp_act_context actx;
     actx.octx = octx;
 
-    actx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
+    actx.src0_nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
 
     actx.src0_row_size = src0_row_size;
     actx.src1_row_size = src1_row_size;
@@ -545,7 +561,8 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     actx.dst_spad_half_size  = L.dst_bytes_per_thread / 2;
 
     actx.block = actx.src0_spad_half_size / actx.src0_row_size_aligned;
-    actx.src0_nrows = src0_nrows;
+    actx.src0_nrows = nrows;
+    actx.row_start  = row_start;
 
     actx.nc = dst->ne[0];
 
@@ -570,7 +587,7 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     actx.data_src1 = data_src1;
     actx.data_dst  = (uint8_t *) dst->data;
 
-    worker_pool_run_func(octx->ctx->worker_pool, act_op_func, &actx, n_threads);
+    work_queue_run(octx->ctx->work_queue, act_op_func, &actx, n_threads);
     return HTP_STATUS_OK;
 }
 

ggml/src/ggml-hexagon/htp/allreduce-ops.c

@@ -17,6 +17,7 @@
 #include "hex-dma.h"
 #include "hex-profile.h"
 #include "allreduce-ops.h"
+#include "htp-fence.h"
 
 struct htp_allreduce_context {
     struct htp_ops_context * octx;
@@ -242,7 +243,42 @@ DEFINE_ALLREDUCE_THREAD_DMA_2D(add_f32,       float,  hvx_add_f32_aaa, 1, 0)
 DEFINE_ALLREDUCE_THREAD_DMA_2D(add_bcast_f16, __fp16, hvx_add_f16_aaa, 1, 1)
 DEFINE_ALLREDUCE_THREAD_DMA_2D(add_bcast_f32, float,  hvx_add_f32_aaa, 1, 1)
 
+static int validate_allreduce(
+    struct htp_ops_context * octx,
+    const struct htp_allreduce_kernel_params * kparams,
+    uint32_t n_ranks
+) {
+    if (!htp_ops_context_set_n_threads(octx, (uint32_t) kparams->n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    if (kparams->vtcm_size_per_thread <= 0 || kparams->vtcm_size <= 0) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    const bool has_add = (octx->op == HTP_OP_ALLREDUCE_ADD);
+    const size_t n_vtcm_buffers = htp_allreduce_vtcm_buffer_count(
+        n_ranks, octx->n_threads, has_add, kparams->is_row_bcast != 0);
+    const size_t vtcm_size = n_vtcm_buffers * (size_t) kparams->vtcm_size_per_thread;
+    if (vtcm_size != (size_t) kparams->vtcm_size) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+    if (vtcm_size > octx->ctx->vtcm_size) {
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    if (octx->dst->type != HTP_TYPE_F16 && octx->dst->type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    return HTP_STATUS_OK;
+}
+
 int op_allreduce(struct htp_ops_context * octx) {
+    if (octx->ctx->mdev.count > 1 && octx->ctx->mdev.idx > 0) {
+        return HTP_STATUS_OK;
+    }
+
     const struct htp_allreduce_kernel_params * kparams = (const struct htp_allreduce_kernel_params *) octx->kernel_params;
     const struct htp_tensor * dst = octx->dst;
 
@@ -253,38 +289,53 @@ int op_allreduce(struct htp_ops_context * octx) {
         return HTP_STATUS_INVAL_PARAMS;
     }
 
-    if (dst->type != HTP_TYPE_F16 && dst->type != HTP_TYPE_F32) {
-        return HTP_STATUS_NO_SUPPORT;
-    }
-
-    const uint32_t nelem = dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3];
     const uint32_t fence_seq_entry = (uint32_t) octx->op_params[0];
     const uint32_t fence_seq_exit  = (uint32_t) octx->op_params[1];
 
+    const struct htp_tensor * my_sync = octx->src[n_ranks + rank];
+    atomic_uint * my_fence = (atomic_uint *) (uintptr_t) my_sync->data;
+
+    const int status = validate_allreduce(octx, kparams, n_ranks);
+    if (status != HTP_STATUS_OK) {
+        if (status == HTP_STATUS_NO_SUPPORT) {
+            FARF(ERROR, "ggml-hex: allreduce unsupported type %d : rank %u\n", dst->type, rank);
+        }
+        htp_fence_write(my_fence, fence_seq_exit, status);
+        return status;
+    }
+
+    const bool has_add = (octx->op == HTP_OP_ALLREDUCE_ADD);
+    const uint32_t nelem = dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3];
+
     // 1. Entry Barrier: Synchronize all ranks before reading
     struct htp_thread_trace * tr0 = &octx->ctx->trace[0];
     htp_trace_event_start(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_entry);
 
-    const struct htp_tensor * my_sync = octx->src[n_ranks + rank];
-    atomic_uint * my_fence = (atomic_uint *) my_sync->data;
-
-    atomic_store(&my_fence[0], fence_seq_entry);
-    asm volatile ("syncht" : : : "memory");
-    Q6_dccleaninva_A((void *) my_fence);
+    htp_fence_write(my_fence, fence_seq_entry, octx->status);
 
     for (uint32_t j = 0; j < n_ranks; j++) {
         if (j == rank) continue;
         const struct htp_tensor * peer_sync = octx->src[n_ranks + j];
-        atomic_uint * peer_fence = (atomic_uint *) peer_sync->data;
+        atomic_uint * peer_fence = (atomic_uint *) (uintptr_t) peer_sync->data;
         uint64_t spins = 0;
         while (1) {
-            Q6_dccleaninva_A((void *) peer_fence);
-            uint32_t val = atomic_load(&peer_fence[0]);
-            if (val == fence_seq_entry || val == fence_seq_exit) {
+            uint32_t peer_seq;
+            uint32_t peer_status;
+            htp_fence_read(peer_fence, &peer_seq, &peer_status);
+            if ((int32_t)(peer_seq - fence_seq_entry) >= 0) {
+                if (peer_status > HTP_STATUS_OK) {
+                    FARF(ERROR, "ggml-hex: allreduce entry peer %u failed with status %u\n", j, peer_status);
+                    htp_fence_write(my_fence, fence_seq_exit, peer_status);
+                    htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_entry);
+                    return peer_status;
+                }
                 break;
             }
             if (++spins > HTP_FENCE_TIMEOUT) {
-                FARF(ERROR, "ggml-hex: allreduce entry fence-wait TIMEOUT: rank %u waiting on %u (fence %p seq %u)\n", rank, j, peer_fence, fence_seq_entry);
+                FARF(ERROR, "ggml-hex: allreduce entry fence-wait TIMEOUT : rank %u waiting on %u fence %p seq 0x%x peer-seq 0x%x\n",
+                     rank, j, peer_fence, fence_seq_entry, peer_seq);
+                htp_fence_write(my_fence, fence_seq_exit, HTP_STATUS_INTERNAL_ERR);
+                htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_entry);
                 return HTP_STATUS_INTERNAL_ERR;
             }
             hex_pause();
@@ -301,8 +352,6 @@ int op_allreduce(struct htp_ops_context * octx) {
         const uint32_t elems_per_thread     = (uint32_t) kparams->elems_per_thread;
         const uint32_t vtcm_size_per_thread = (uint32_t) kparams->vtcm_size_per_thread;
 
-        const bool has_add = (octx->op == HTP_OP_ALLREDUCE_ADD);
-
         struct htp_allreduce_context actx;
         actx.octx                 = octx;
         actx.n_ranks              = n_ranks;
@@ -339,6 +388,8 @@ int op_allreduce(struct htp_ops_context * octx) {
                 }
                 break;
             default:
+                FARF(ERROR, "ggml-hex: allreduce unsupported kernel %d : rank %u\n", kparams->kernel_type, rank);
+                htp_fence_write(my_fence, fence_seq_exit, HTP_STATUS_NO_SUPPORT);
                 return HTP_STATUS_NO_SUPPORT;
         }
 
@@ -368,23 +419,31 @@ int op_allreduce(struct htp_ops_context * octx) {
     // 4. Exit Barrier: Synchronize all ranks after writing
     htp_trace_event_start(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_exit);
 
-    atomic_store(&my_fence[0], fence_seq_exit);
-    asm volatile ("syncht" : : : "memory");
-    Q6_dccleaninva_A((void *) my_fence);
+    htp_fence_write(my_fence, fence_seq_exit, octx->status);
 
     for (uint32_t j = 0; j < n_ranks; j++) {
         if (j == rank) continue;
         const struct htp_tensor * peer_sync = octx->src[n_ranks + j];
-        atomic_uint * peer_fence = (atomic_uint *) peer_sync->data;
+        atomic_uint * peer_fence = (atomic_uint *) (uintptr_t) peer_sync->data;
         uint64_t spins = 0;
         while (1) {
-            Q6_dccleaninva_A((void *) peer_fence);
-            uint32_t val = atomic_load(&peer_fence[0]);
-            if (val == fence_seq_exit) {
+            uint32_t peer_seq;
+            uint32_t peer_status;
+            htp_fence_read(peer_fence, &peer_seq, &peer_status);
+            if ((int32_t)(peer_seq - fence_seq_exit) >= 0) {
+                if (peer_status > HTP_STATUS_OK) {
+                    FARF(ERROR, "ggml-hex: allreduce exit peer %u failed with status %u\n", j, peer_status);
+                    htp_fence_write(my_fence, fence_seq_exit, peer_status);
+                    htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_exit);
+                    return peer_status;
+                }
                 break;
             }
             if (++spins > HTP_FENCE_TIMEOUT) {
-                FARF(ERROR, "ggml-hex: allreduce exit fence-wait TIMEOUT: rank %u waiting on %u (fence %p seq %u)\n", rank, j, peer_fence, fence_seq_exit);
+                FARF(ERROR, "ggml-hex: allreduce exit fence-wait TIMEOUT : rank %u waiting on %u fence %p seq 0x%x peer-seq 0x%x\n",
+                     rank, j, peer_fence, fence_seq_exit, peer_seq);
+                htp_fence_write(my_fence, fence_seq_exit, HTP_STATUS_INTERNAL_ERR);
+                htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_exit);
                 return HTP_STATUS_INTERNAL_ERR;
             }
             hex_pause();
@@ -394,5 +453,5 @@ int op_allreduce(struct htp_ops_context * octx) {
 
     htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_exit);
 
-    return HTP_STATUS_OK;
+    return octx->status;
 }

ggml/src/ggml-hexagon/htp/allreduce-ops.h

@@ -2,6 +2,8 @@
 #define ALLREDUCE_OPS_H
 
 #include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
 
 #define HTP_ALLREDUCE_MAX_RANKS 4
 
@@ -15,6 +17,15 @@ enum htp_allreduce_kernel_type {
     HTP_ALLREDUCE_KERNEL_DMA_2D,
 };
 
+static inline size_t htp_allreduce_vtcm_buffer_count(
+    uint32_t n_ranks,
+    uint32_t n_threads,
+    bool has_add,
+    bool is_row_bcast
+) {
+    return (size_t) (n_ranks + 1) * n_threads + (has_add ? (is_row_bcast ? 1 : n_threads) : 0);
+}
+
 struct htp_allreduce_kernel_params {
     int32_t rank;
     int32_t n_ranks;

ggml/src/ggml-hexagon/htp/argsort-ops.c

@@ -11,9 +11,10 @@
 #include "hvx-utils.h"
 #include "hex-dma.h"
 
+#include "hex-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
+#include "htp-tensor.h"
 
 #ifndef MIN
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -22,6 +23,9 @@
 struct htp_argsort_context {
     struct htp_ops_context * octx;
     uint32_t                 nrows_per_thread;
+    uint32_t                 total_rows;
+    uint32_t                 row_start;
+    uint32_t                 row_end;
     uint8_t *                vtcm_base;
     size_t                   vtcm_per_thread;
 };
@@ -336,10 +340,9 @@ static void htp_argsort_f32_##ne00##_##order_name(unsigned int n, unsigned int i
     const struct htp_tensor * src0 = octx->src[0];                                                             \
     const struct htp_tensor * dst = octx->dst;                                                                 \
     uint8_t * spad = actx->vtcm_base + actx->vtcm_per_thread * i;                                              \
-    uint32_t total_rows = src0->ne[1] * src0->ne[2] * src0->ne[3];                                             \
     uint32_t rows_per_thread = actx->nrows_per_thread;                                                         \
-    uint32_t start_row = rows_per_thread * i;                                                                  \
-    uint32_t end_row = MIN(start_row + rows_per_thread, total_rows);                                           \
+    uint32_t start_row = actx->row_start + rows_per_thread * i;                                                \
+    uint32_t end_row = MIN(start_row + rows_per_thread, actx->row_end);                                        \
     size_t values_size = hex_round_up(ne00 * sizeof(float), 128);                                              \
     float * values_buf = (float *) spad;                                                                       \
     int32_t * indices_buf = (int32_t *) (spad + values_size);                                                  \
@@ -386,9 +389,6 @@ static void htp_argsort_f32_fallback(unsigned int n, unsigned int i, void * data
 
     // Dimensions
     uint32_t ne00 = src0->ne[0];
-    uint32_t ne01 = src0->ne[1];
-    uint32_t ne02 = src0->ne[2];
-    uint32_t ne03 = src0->ne[3];
 
     uint32_t nb01 = src0->nb[1];
 
@@ -398,10 +398,9 @@ static void htp_argsort_f32_fallback(unsigned int n, unsigned int i, void * data
     enum ggml_sort_order order = (enum ggml_sort_order) octx->op_params[0];
 
     // Rows to process
-    uint32_t total_rows = ne01 * ne02 * ne03;
     uint32_t rows_per_thread = actx->nrows_per_thread;
-    uint32_t start_row = rows_per_thread * i;
-    uint32_t end_row = MIN(start_row + rows_per_thread, total_rows);
+    uint32_t start_row = actx->row_start + rows_per_thread * i;
+    uint32_t end_row = MIN(start_row + rows_per_thread, actx->row_end);
 
     size_t values_size = hex_round_up(ne00 * sizeof(float), 128);
     uint32_t num_vec_ind_values = hmx_ceil_div(ne00, VLEN/(sizeof(int32_t)));
@@ -451,8 +450,28 @@ int op_argsort(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t total_rows = octx->src[0]->ne[1] * octx->src[0]->ne[2] * octx->src[0]->ne[3];
-    const uint32_t n_threads = MIN(total_rows, octx->n_threads);
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const uint32_t total_rows  = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const size_t dst_row_size  = dst->ne[0]  * sizeof(int32_t);
+
+    uint32_t row_start = 0;
+    uint32_t row_end   = total_rows;
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, sizeof(int32_t), (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        row_end   = range.start + range.count;
+    }
+
+    const uint32_t nrows = row_end - row_start;
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
 
     // Allocate scratchpad
     // We need 1 row of float + 1 row of int32 per thread.
@@ -478,7 +497,10 @@ int op_argsort(struct htp_ops_context * octx) {
 
     struct htp_argsort_context actx;
     actx.octx = octx;
-    actx.nrows_per_thread = (total_rows + n_threads - 1) / n_threads;
+    actx.nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
+    actx.total_rows       = nrows;
+    actx.row_start        = row_start;
+    actx.row_end          = row_end;
     actx.vtcm_base = (uint8_t *) octx->ctx->vtcm_base;
     actx.vtcm_per_thread = spad_per_thread;
 
@@ -508,7 +530,7 @@ int op_argsort(struct htp_ops_context * octx) {
     }
 
     // Run jobs
-    worker_pool_run_func(octx->ctx->worker_pool, job_func, &actx, n_threads);
+    work_queue_run(octx->ctx->work_queue, job_func, &actx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/binary-ops.c

@@ -13,9 +13,10 @@
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
 #include "htp-tensor.h"
 
 #ifndef MIN
@@ -36,6 +37,8 @@ struct htp_binary_context {
 
     uint32_t block_max;
     uint32_t nrows_per_thread;
+    uint32_t total_rows;
+    uint32_t row_start;
     size_t   src0_row_size_aligned;
     size_t   src1_row_size_aligned;
     size_t   dst_row_size_aligned;
@@ -48,27 +51,27 @@ struct htp_binary_context {
     const struct htp_tensor * src0 = octx->src[0]; \
     const struct htp_tensor * src1 = octx->src[1]; \
     const struct htp_tensor * dst  = octx->dst;    \
-                                       \
-    const uint32_t ne00 = src0->ne[0]; \
-    const uint32_t ne01 = src0->ne[1]; \
-    const uint32_t ne02 = src0->ne[2]; \
-    const uint32_t ne03 = src0->ne[3]; \
-                                       \
-    const uint32_t ne10 = src1->ne[0]; \
-    const uint32_t ne11 = src1->ne[1]; \
-    const uint32_t ne12 = src1->ne[2]; \
-    const uint32_t ne13 = src1->ne[3]; \
-                                       \
-    const uint32_t nb01 = src0->nb[1]; \
-    const uint32_t nb02 = src0->nb[2]; \
-    const uint32_t nb03 = src0->nb[3]; \
-                                       \
-    const uint32_t nb11 = src1->nb[1]; \
-    const uint32_t nb12 = src1->nb[2]; \
-    const uint32_t nb13 = src1->nb[3]; \
-                                       \
-    const uint32_t nb1 = dst->nb[1];   \
-    const uint32_t nb2 = dst->nb[2];   \
+                                                   \
+    const uint32_t ne00 = src0->ne[0];             \
+    const uint32_t ne01 = src0->ne[1];             \
+    const uint32_t ne02 = src0->ne[2];             \
+    const uint32_t ne03 = src0->ne[3];             \
+                                                   \
+    const uint32_t ne10 = src1->ne[0];             \
+    const uint32_t ne11 = src1->ne[1];             \
+    const uint32_t ne12 = src1->ne[2];             \
+    const uint32_t ne13 = src1->ne[3];             \
+                                                   \
+    const uint32_t nb01 = src0->nb[1];             \
+    const uint32_t nb02 = src0->nb[2];             \
+    const uint32_t nb03 = src0->nb[3];             \
+                                                   \
+    const uint32_t nb11 = src1->nb[1];             \
+    const uint32_t nb12 = src1->nb[2];             \
+    const uint32_t nb13 = src1->nb[3];             \
+                                                   \
+    const uint32_t nb1 = dst->nb[1];               \
+    const uint32_t nb2 = dst->nb[2];               \
     const uint32_t nb3 = dst->nb[3];
 
 static inline uint32_t calc_block_size(struct htp_binary_context * bctx, uint32_t ir, uint32_t end_row, uint32_t ne01, uint32_t ne02) {
@@ -93,87 +96,87 @@ static inline uint32_t calc_block_size(struct htp_binary_context * bctx, uint32_
 }
 
 // Macro for scalar op switch
-#define COMPUTE_SCALAR_OP(DST, SRC, VAL, TYPE, N) \
-    if(TYPE == HTP_TYPE_F32) { \
-        switch (octx->op) { \
-            case HTP_OP_ADD: hvx_add_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
-            case HTP_OP_SUB: hvx_sub_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
-            case HTP_OP_MUL: hvx_mul_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+#define COMPUTE_SCALAR_OP(DST, SRC, VAL, TYPE, N)                                               \
+    if(TYPE == HTP_TYPE_F32) {                                                                  \
+        switch (octx->op) {                                                                     \
+            case HTP_OP_ADD: hvx_add_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break;          \
+            case HTP_OP_SUB: hvx_sub_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break;          \
+            case HTP_OP_MUL: hvx_mul_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break;          \
             case HTP_OP_DIV: hvx_mul_scalar_f32_aa(DST, SRC, 1.0f / (*(float *)VAL), N); break; \
-            default: break; \
-        } \
-    } \
-    else { \
-        switch (octx->op) { \
-            case HTP_OP_ADD: hvx_add_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
-            case HTP_OP_SUB: hvx_sub_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
-            case HTP_OP_MUL: hvx_mul_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
-            case HTP_OP_DIV: hvx_div_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
-            default: break; \
-        } \
+            default: break;                                                                     \
+        }                                                                                       \
+    }                                                                                           \
+    else {                                                                                      \
+        switch (octx->op) {                                                                     \
+            case HTP_OP_ADD: hvx_add_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break;       \
+            case HTP_OP_SUB: hvx_sub_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break;       \
+            case HTP_OP_MUL: hvx_mul_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break;       \
+            case HTP_OP_DIV: hvx_div_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break;       \
+            default: break;                                                                     \
+        }                                                                                       \
     }
 
 // Macro for vector op switch (All Aligned)
-#define COMPUTE_VECTOR_OP_AAA(DST, SRC0, SRC1, TYPE, N) \
-    if(TYPE == HTP_TYPE_F32) { \
-        switch (octx->op) { \
+#define COMPUTE_VECTOR_OP_AAA(DST, SRC0, SRC1, TYPE, N)                  \
+    if(TYPE == HTP_TYPE_F32) {                                           \
+        switch (octx->op) {                                              \
             case HTP_OP_ADD: hvx_add_f32_aaa(DST, SRC0, SRC1, N); break; \
             case HTP_OP_SUB: hvx_sub_f32_aaa(DST, SRC0, SRC1, N); break; \
             case HTP_OP_MUL: hvx_mul_f32_aaa(DST, SRC0, SRC1, N); break; \
             case HTP_OP_DIV: hvx_div_f32_aaa(DST, SRC0, SRC1, N); break; \
-            default: break; \
-        } \
-    } \
-    else { \
-        switch (octx->op) { \
+            default: break;                                              \
+        }                                                                \
+    }                                                                    \
+    else {                                                               \
+        switch (octx->op) {                                              \
             case HTP_OP_ADD: hvx_add_f16_aaa(DST, SRC0, SRC1, N); break; \
             case HTP_OP_SUB: hvx_sub_f16_aaa(DST, SRC0, SRC1, N); break; \
             case HTP_OP_MUL: hvx_mul_f16_aaa(DST, SRC0, SRC1, N); break; \
             case HTP_OP_DIV: hvx_div_f16_aaa(DST, SRC0, SRC1, N); break; \
-            default: break; \
-        } \
+            default: break;                                              \
+        }                                                                \
     }
 
 // Macro for vector op switch (Dst Aligned, Src0 Aligned, Src1 Unaligned)
-#define COMPUTE_VECTOR_OP_AAU(DST, SRC0, SRC1, TYPE, N) \
-    if(TYPE == HTP_TYPE_F32) { \
-        switch (octx->op) { \
+#define COMPUTE_VECTOR_OP_AAU(DST, SRC0, SRC1, TYPE, N)                  \
+    if(TYPE == HTP_TYPE_F32) {                                           \
+        switch (octx->op) {                                              \
             case HTP_OP_ADD: hvx_add_f32_aau(DST, SRC0, SRC1, N); break; \
             case HTP_OP_SUB: hvx_sub_f32_aau(DST, SRC0, SRC1, N); break; \
             case HTP_OP_MUL: hvx_mul_f32_aau(DST, SRC0, SRC1, N); break; \
             case HTP_OP_DIV: hvx_div_f32_aau(DST, SRC0, SRC1, N); break; \
-            default: break; \
-        } \
-    } \
-    else { \
-        switch (octx->op) { \
+            default: break;                                              \
+        }                                                                \
+    }                                                                    \
+    else {                                                               \
+        switch (octx->op) {                                              \
             case HTP_OP_ADD: hvx_add_f16_aau(DST, SRC0, SRC1, N); break; \
             case HTP_OP_SUB: hvx_sub_f16_aau(DST, SRC0, SRC1, N); break; \
             case HTP_OP_MUL: hvx_mul_f16_aau(DST, SRC0, SRC1, N); break; \
             case HTP_OP_DIV: hvx_div_f16_aau(DST, SRC0, SRC1, N); break; \
-            default: break; \
-        } \
+            default: break;                                              \
+        }                                                                \
     }
 
 // Macro for vector op switch (All Unaligned - generic loop used in element repeat)
-#define COMPUTE_VECTOR_OP_UUU(DST, SRC0, SRC1, TYPE, N) \
-    if(TYPE == HTP_TYPE_F32) { \
-        switch (octx->op) { \
+#define COMPUTE_VECTOR_OP_UUU(DST, SRC0, SRC1, TYPE, N)                  \
+    if(TYPE == HTP_TYPE_F32) {                                           \
+        switch (octx->op) {                                              \
             case HTP_OP_ADD: hvx_add_f32_uuu(DST, SRC0, SRC1, N); break; \
             case HTP_OP_SUB: hvx_sub_f32_uuu(DST, SRC0, SRC1, N); break; \
             case HTP_OP_MUL: hvx_mul_f32_uuu(DST, SRC0, SRC1, N); break; \
             case HTP_OP_DIV: hvx_div_f32_uuu(DST, SRC0, SRC1, N); break; \
-            default: break; \
-        } \
-    } \
-    else { \
-        switch (octx->op) { \
+            default: break;                                              \
+        }                                                                \
+    }                                                                    \
+    else {                                                               \
+        switch (octx->op) {                                              \
             case HTP_OP_ADD: hvx_add_f16_uuu(DST, SRC0, SRC1, N); break; \
             case HTP_OP_SUB: hvx_sub_f16_uuu(DST, SRC0, SRC1, N); break; \
             case HTP_OP_MUL: hvx_mul_f16_uuu(DST, SRC0, SRC1, N); break; \
             case HTP_OP_DIV: hvx_div_f16_uuu(DST, SRC0, SRC1, N); break; \
-            default: break; \
-        } \
+            default: break;                                              \
+        }                                                                \
     }
 
 // 1. Scalar src1 (ne10 == 1)
@@ -184,9 +187,8 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
 
     const uint32_t src0_type = octx->src[0]->type;
     const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
-    const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row = bctx->nrows_per_thread * ith;
-    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row = bctx->row_start + bctx->nrows_per_thread * ith;
+    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, bctx->row_start + bctx->total_rows);
     if (start_row >= end_row) return;
 
     FARF(HIGH, "binary-scalar: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
@@ -222,6 +224,8 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
     }
 
     // Main loop
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
 
@@ -242,12 +246,14 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * src1_ptr = (uint8_t *)src1->data + i13 * nb13 + i12 * nb12 + i11 * nb11;
         uint32_t s1_stride = (ne11 == 1) ? 0 : nb11;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
             COMPUTE_SCALAR_OP(r_dst, r_src0, src1_ptr, src0_type, ne00);
             src1_ptr += s1_stride;
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
         dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
@@ -266,6 +272,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         }
         ir += current_block_size;
     }
+
     dma_queue_flush(q);
 }
 
@@ -277,9 +284,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
 
     const uint32_t src0_type = octx->src[0]->type;
     const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
-    const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row = bctx->nrows_per_thread * ith;
-    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row = bctx->row_start + bctx->nrows_per_thread * ith;
+    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, bctx->row_start + bctx->total_rows);
     if (start_row >= end_row) return;
 
     FARF(HIGH, "binary-same-shape: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
@@ -323,18 +329,22 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
         uint8_t * d_spad  = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
         uint8_t * s1_spad = (uint8_t *) dma_queue_pop(q).dst;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_src1 = s1_spad + r * bctx->src1_row_size_aligned;
             uint8_t * r_dst  = d_spad  + r * bctx->dst_row_size_aligned;
             COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         uint32_t i03, i02, i01, rem;
         i03 = fastdiv(ir, &bctx->src0_dim12_div);
@@ -366,6 +376,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         }
         ir += current_block_size;
     }
+
     dma_queue_flush(q);
 }
 
@@ -377,9 +388,8 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
 
     const uint32_t src0_type  = octx->src[0]->type;
     const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
-    const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row  = bctx->nrows_per_thread * ith;
-    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row  = bctx->row_start + bctx->nrows_per_thread * ith;
+    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, bctx->row_start + bctx->total_rows);
     if (start_row >= end_row) return;
 
     FARF(HIGH, "binary-row-bcast: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
@@ -416,17 +426,21 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
         uint8_t * d_spad  = (uint8_t *) dma_queue_pop(q).src;
         uint8_t * s0_spad = (uint8_t *) dma_queue_pop(q).dst;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_src1 = (uint8_t *)s1_ptr; // Constant
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
             COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         uint32_t i03 = fastdiv(ir, &bctx->src0_dim12_div);
         uint32_t rem = ir - i03 * (ne02 * ne01);
@@ -447,6 +461,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         }
         ir += current_block_size;
     }
+
     dma_queue_flush(q);
 }
 
@@ -458,9 +473,8 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
 
     const uint32_t src0_type = octx->src[0]->type;
     const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
-    const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row  = bctx->nrows_per_thread * ith;
-    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row  = bctx->row_start + bctx->nrows_per_thread * ith;
+    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, bctx->row_start + bctx->total_rows);
     if (start_row >= end_row) return;
 
     FARF(HIGH, "binary-complex: %d/%d (%u:%u) row-size %u (%u)", ith, nth, start_row, end_row, nb01, bctx->dst_row_size_aligned);
@@ -493,6 +507,8 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
         uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
@@ -503,6 +519,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
         uint32_t i01 = rem - i02 * ne01;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint32_t r_i01 = i01 + r;
             uint32_t i13 = fastmodulo(i03, ne13, &bctx->src1_dim3_div);
@@ -516,6 +533,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
             // Read src1 from DDR (unaligned)
             COMPUTE_VECTOR_OP_AAU(r_dst, r_src0, r_src1, src0_type, ne00);
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
         dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
@@ -532,6 +550,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         }
         ir += current_block_size;
     }
+
     dma_queue_flush(q);
 }
 
@@ -544,9 +563,8 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
     const uint32_t src0_type = octx->src[0]->type;
     const uint32_t elem_size_bytes = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
     const uint32_t row_size_bytes = ne00 * elem_size_bytes;;
-    const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row  = bctx->nrows_per_thread * ith;
-    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row  = bctx->row_start + bctx->nrows_per_thread * ith;
+    const uint32_t end_row    = MIN(start_row + bctx->nrows_per_thread, bctx->row_start + bctx->total_rows);
     if (start_row >= end_row) return;
 
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
@@ -579,6 +597,8 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
         uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
@@ -589,6 +609,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
         uint32_t i01 = rem - i02 * ne01;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint32_t r_i01 = i01 + r;
             uint32_t i13 = fastmodulo(i03, ne13, &bctx->src1_dim3_div);
@@ -606,6 +627,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
                 COMPUTE_VECTOR_OP_UUU(r_dst + c * elem_size_bytes, r_src0 + c * elem_size_bytes, r_src1_row, src0_type, len);
             }
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
         dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
@@ -622,6 +644,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         }
         ir += current_block_size;
     }
+
     dma_queue_flush(q);
 }
 
@@ -650,9 +673,8 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t nb2 = dst->nb[2];
     const uint32_t nb3 = dst->nb[3];
 
-    const uint32_t total_rows = ne01 * ne02 * ne03;
-    const uint32_t start_row = bctx->nrows_per_thread * ith;
-    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
+    const uint32_t start_row = bctx->row_start + bctx->nrows_per_thread * ith;
+    const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, bctx->row_start + bctx->total_rows);
     if (start_row >= end_row) return;
 
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
@@ -683,6 +705,8 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = start_row; ir < end_row; ) {
         uint32_t current_block_size = calc_block_size(bctx, ir, end_row, ne01, ne02);
         uint8_t * d_spad = (uint8_t *) dma_queue_pop(q).src;
@@ -693,6 +717,7 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
         uint32_t i02 = fastdiv(rem, &bctx->src0_dim1_div);
         uint32_t i01 = rem - i02 * ne01;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint32_t r_i01 = i01 + r; // linear within block since we split at ne01
 
@@ -704,6 +729,7 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
 
             hvx_add_f32_aau(r_dst, r_src0, r_src1, ne00);
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
         dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
@@ -720,6 +746,7 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
         }
         ir += current_block_size;
     }
+
     dma_queue_flush(q);
 }
 
@@ -729,15 +756,31 @@ static int execute_op_binary(struct htp_ops_context * octx) {
     const struct htp_tensor * dst  = octx->dst;
 
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
-    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
-    // Use packed row sizes for VTCM allocation
+    // Use packed row sizes for VTCM allocation and alignment
     const uint32_t src0_type = octx->src[0]->type;
     const size_t elem_size = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
     const size_t src0_row_size = src0->ne[0] * elem_size;
     const size_t src1_row_size = src1->ne[0] * elem_size;
     const size_t dst_row_size  = dst->ne[0]  * elem_size;
 
+    uint32_t row_start = 0;
+    uint32_t nrows     = src0_nrows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, (uint32_t) elem_size, (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(src0_nrows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
     size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
     size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
     size_t dst_row_size_aligned  = hex_round_up(dst_row_size,  VLEN);
@@ -815,7 +858,9 @@ static int execute_op_binary(struct htp_ops_context * octx) {
 
     struct htp_binary_context bctx;
     bctx.octx                  = octx;
-    bctx.nrows_per_thread      = (src0_nrows + n_threads - 1) / n_threads;
+    bctx.nrows_per_thread      = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
+    bctx.total_rows            = nrows;
+    bctx.row_start             = row_start;
     bctx.block_max             = rows_per_buffer;
     bctx.src0_row_size_aligned = src0_row_size_aligned;
     bctx.src1_row_size_aligned = src1_row_size_aligned;
@@ -850,7 +895,7 @@ static int execute_op_binary(struct htp_ops_context * octx) {
         dma_queue_pop(q);
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &bctx, n_threads);
+    work_queue_run(octx->ctx->work_queue, worker_func, &bctx, n_threads);
 
     return HTP_STATUS_OK;
 }
@@ -870,4 +915,3 @@ int op_binary(struct htp_ops_context * octx) {
 
     return HTP_STATUS_NO_SUPPORT;
 }
-

ggml/src/ggml-hexagon/htp/concat-ops.c

@@ -1,5 +1,8 @@
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
+#include "htp-tensor.h"
 #include "hexagon_types.h"
 #include "hexagon_protos.h"
 #include "hvx_hexagon_protos.h"
@@ -13,6 +16,10 @@ struct htp_concat_context {
     struct htp_ops_context * octx;
     uint32_t dim;
     uint32_t nrows_per_thread;
+    uint32_t row_start;
+    uint32_t nrows;
+    uint32_t elem_start;
+    uint32_t nelems;
     struct fastdiv_values div_ne0;
     struct fastdiv_values div_ne1;
     struct fastdiv_values div_ne2;
@@ -28,10 +35,10 @@ static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void *
 
     const uint32_t src0_ne0 = src0->ne[0];
     const uint32_t src1_ne0 = src1->ne[0];
-    const uint32_t ne1      = dst->ne[1];
 
-    const uint32_t start_i = ith * cctx->nrows_per_thread;
-    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < ne1) ? (start_i + cctx->nrows_per_thread) : ne1;
+    const uint32_t row_end = cctx->row_start + cctx->nrows;
+    const uint32_t start_i = cctx->row_start + ith * cctx->nrows_per_thread;
+    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < row_end) ? (start_i + cctx->nrows_per_thread) : row_end;
     if (start_i >= end_i) return;
 
     dma_queue * q = octx->ctx->dma[ith];
@@ -51,6 +58,8 @@ static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void *
     const uint32_t spad0_row_bytes = hex_round_up((src0_ne0 + src1_ne0_padded) * sizeof(float), VLEN);
     uint32_t mu = src1_ne0_padded * spad1_stride;
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t i = start_i; i < end_i; i += block_i) {
         uint32_t current_block_i = (end_i - i < block_i) ? (end_i - i) : block_i;
 
@@ -66,6 +75,7 @@ static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void *
 
         HVX_Vector * vtcm_tmp = (HVX_Vector *)(spad1_base + src1_ne0_padded * spad1_stride);
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
         for (uint32_t j = 0; j < src1_ne0_padded; j += 32) {
             #pragma unroll(4)
             for (uint32_t ii = 0; ii < current_block_i; ii++) {
@@ -75,6 +85,7 @@ static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void *
                 hvx_vmemu(dst_ptr) = vtcm_tmp[ii];
             }
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
 
         dma_queue_pop(q); // src0
 
@@ -95,10 +106,10 @@ static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void *
 
     const uint32_t src0_ne0 = src0->ne[0];
     const uint32_t src1_ne0 = src1->ne[0];
-    const uint32_t ne1      = dst->ne[1];
 
-    const uint32_t start_i = ith * cctx->nrows_per_thread;
-    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < ne1) ? (start_i + cctx->nrows_per_thread) : ne1;
+    const uint32_t row_end = cctx->row_start + cctx->nrows;
+    const uint32_t start_i = cctx->row_start + ith * cctx->nrows_per_thread;
+    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < row_end) ? (start_i + cctx->nrows_per_thread) : row_end;
     if (start_i >= end_i) return;
 
     dma_queue * q = octx->ctx->dma[ith];
@@ -118,6 +129,8 @@ static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void *
     const uint32_t spad0_row_bytes = hex_round_up((src0_ne0 + src1_ne0_padded) * sizeof(__fp16), VLEN);
     uint32_t mu = src1_ne0_padded * spad1_stride;
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t i = start_i; i < end_i; i += block_i) {
         uint32_t current_block_i = (end_i - i < block_i) ? (end_i - i) : block_i;
 
@@ -133,6 +146,7 @@ static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void *
 
         HVX_Vector * vtcm_tmp = (HVX_Vector *)(spad1_base + src1_ne0_padded * spad1_stride);
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
         for (uint32_t j = 0; j < src1_ne0_padded; j += 64) {
             #pragma unroll(4)
             for (uint32_t ii = 0; ii < current_block_i; ii++) {
@@ -142,6 +156,7 @@ static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void *
                 hvx_vmemu(dst_ptr) = vtcm_tmp[ii];
             }
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
 
         dma_queue_pop(q); // src0
 
@@ -164,11 +179,14 @@ static void concat_generic(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t type_size = (dst->type == HTP_TYPE_F32 || dst->type == HTP_TYPE_I32) ? 4 : 2;
 
     const uint32_t ne[4] = {dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]};
-    const uint32_t total_elements = ne[0] * ne[1] * ne[2] * ne[3];
-    const uint32_t chunk_size = (total_elements + nth - 1) / nth;
 
-    const uint32_t start_idx = MIN(ith * chunk_size, total_elements);
-    const uint32_t end_idx   = MIN(start_idx + chunk_size, total_elements);
+    // Per-device element range aligned to prevent false sharing
+    const uint32_t elem_start = cctx->elem_start;
+    const uint32_t nelems     = cctx->nelems;
+    const uint32_t chunk_size = (nelems + nth - 1) / nth;
+
+    const uint32_t start_idx = MIN(elem_start + ith * chunk_size, elem_start + nelems);
+    const uint32_t end_idx   = MIN(start_idx + chunk_size, elem_start + nelems);
 
     // Naive scalar element-wise copy
     for (uint32_t idx = start_idx; idx < end_idx; idx++) {
@@ -236,13 +254,28 @@ int op_concat(struct htp_ops_context * octx) {
     void (*worker_func)(unsigned int, unsigned int, void *) = concat_generic;
 
     if (dim == 0 && is_2d && is_src1_transposed && !is_src0_transposed) {
-        n_threads = MIN(dst->ne[1], n_threads);
-        if (n_threads < 1) {
-            n_threads = 1;
+        const uint32_t total_rows = dst->ne[1];
+        const size_t dst_data_row_size = dst->ne[0] * type_size;
+        uint32_t row_start = 0;
+        uint32_t nrows     = total_rows;
+        if (octx->ctx->mdev.count > 1) {
+            uint32_t rows_per_chunk = 0;
+            htp_tensor_mdev_rows_per_chunk(dst, type_size, (uint32_t) dst_data_row_size, &rows_per_chunk);
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            row_start = range.start;
+            nrows     = range.count;
         }
+
+        if (nrows == 0) {
+            return HTP_STATUS_OK;
+        }
+
+        cctx.row_start = row_start;
+        cctx.nrows     = nrows;
+
         uint32_t block_i = (type_size == 4) ? 32 : 64;
 
-        cctx.nrows_per_thread = hmx_ceil_div(dst->ne[1], n_threads);
+        cctx.nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
 
         // Allocate VTCM
         uint32_t spad1_stride = block_i * type_size;
@@ -270,8 +303,26 @@ int op_concat(struct htp_ops_context * octx) {
         } else {
             worker_func = concat_2d_f16_transposed;
         }
+    } else {
+        const uint32_t total_elements = dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3];
+        uint32_t elem_start = 0;
+        uint32_t nelems     = total_elements;
+        if (octx->ctx->mdev.count > 1) {
+            const uint32_t elems_per_chunk = HEX_L2_LINE_SIZE / type_size;
+            const bool can_split = htp_tensor_mdev_data_aligned(dst) && htp_tensor_is_contiguous(dst, type_size) && !htp_tensor_is_permuted(dst);
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_elements, can_split ? elems_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            elem_start = range.start;
+            nelems     = range.count;
+        }
+
+        if (nelems == 0) {
+            return HTP_STATUS_OK;
+        }
+
+        cctx.elem_start = elem_start;
+        cctx.nelems     = nelems;
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &cctx, n_threads);
+    work_queue_run(octx->ctx->work_queue, worker_func, &cctx, n_threads);
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/cpy-ops.c

@@ -16,6 +16,7 @@
 #include "htp-ops.h"
 #include "hvx-utils.h"
 #include "htp-tensor.h"
+#include "htp-fence.h"
 
 struct htp_copy_context {
     struct htp_ops_context * octx;
@@ -29,7 +30,23 @@ struct htp_copy_context {
     uint32_t          src0_blocks_per_row;
     uint32_t          dst_blocks_per_row;
 
+    uint32_t          elem_start;
+    uint32_t          nelem;
+    uint32_t          elem_per_thread;
+
     uint32_t          src0_nrows_per_thread;
+    uint32_t          row_start;
+    uint32_t          nrows;
+
+    struct fastdiv_values div_ne01;
+    struct fastdiv_values div_ne02_ne01;
+
+    struct fastdiv_values div_ne0;
+    struct fastdiv_values div_ne1_ne0;
+    struct fastdiv_values div_ne2_ne1_ne0;
+    struct fastdiv_values div_ne00;
+    struct fastdiv_values div_ne01_ne00;
+    struct fastdiv_values div_ne02_ne01_ne00;
 };
 
 #define cpy_preamble                              \
@@ -54,131 +71,113 @@ struct htp_copy_context {
     const uint32_t  nb0 = dst->nb[0];             \
     const uint32_t  nb1 = dst->nb[1];             \
     const uint32_t  nb2 = dst->nb[2];             \
-    const uint32_t  nb3 = dst->nb[3];             \
-                                                  \
-    const uint32_t   nr = ne01;
+    const uint32_t  nb3 = dst->nb[3];
 
-#define DEFINE_CPY_SAMESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                       \
-static void cpy_thread_##NAME##_sameshape(unsigned int nth, unsigned int ith, void * data) {                   \
-    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                           \
-    struct htp_ops_context * octx = ct->octx;                                                                  \
-    cpy_preamble;                                                                                              \
-    const uint32_t dr  = ct->src0_nrows_per_thread;                                                            \
-    const uint32_t ir0 = dr * ith;                                                                             \
-    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;                                                    \
-    if (ir0 >= nr) return;                                                                                     \
-    for (uint32_t i03 = 0; i03 < ne03; i03++) {                                                                \
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {                                                            \
-            _Pragma("unroll(4)")                                                                               \
-            for (uint32_t i01 = ir0; i01 < ir1; i01++) {                                                       \
-                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;                     \
-                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;                    \
-                hex_l2fetch(src0_ptr, ne00 * ELEM_SIZE, nb01, 2);                                              \
-                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                               \
-            }                                                                                                  \
-        }                                                                                                      \
-    }                                                                                                          \
+#define DEFINE_CPY_SAMESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                       \
+static void cpy_thread_##NAME##_sameshape(unsigned int nth, unsigned int ith, void * data) {   \
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;                           \
+    struct htp_ops_context * octx = ct->octx;                                                  \
+    cpy_preamble;                                                                              \
+    const uint32_t dr  = ct->src0_nrows_per_thread;                                            \
+    const uint32_t ir0 = ct->row_start + dr * ith;                                             \
+    const uint32_t ir1 = MIN(ir0 + dr, ct->row_start + ct->nrows);                             \
+    if (ir0 >= ir1) return;                                                                    \
+    const bool contiguous = (nb01 == ne00 * ELEM_SIZE) && (nb1 == nb01) &&                     \
+                            (nb02 == ne01 * nb01)      && (nb2 == nb02) &&                     \
+                            (nb03 == ne02 * nb02)      && (nb3 == nb03);                       \
+    const uint32_t ne02_ne01 = ne02 * ne01;                                                    \
+    uint32_t i03 = fastdiv(ir0, &ct->div_ne02_ne01);                                           \
+    uint32_t rem = ir0 - i03 * ne02_ne01;                                                      \
+    uint32_t i02 = fastdiv(rem, &ct->div_ne01);                                                \
+    uint32_t i01 = rem - i02 * ne01;                                                           \
+    uint8_t * dst_ptr  = (uint8_t *) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;               \
+    uint8_t * src0_ptr = (uint8_t *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;              \
+    if (contiguous) {                                                                          \
+        hvx_copy_uu(dst_ptr, src0_ptr, (ir1 - ir0) * ne00, ELEM_SIZE);                         \
+        return;                                                                                \
+    }                                                                                          \
+    for (uint32_t r = ir0; r < ir1; r++) {                                                     \
+        hex_l2fetch(src0_ptr, ne00 * ELEM_SIZE, nb01, 2);                                      \
+        hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                       \
+        dst_ptr  += nb1;                                                                       \
+        src0_ptr += nb01;                                                                      \
+        if (++i01 == ne01) {                                                                   \
+            i01 = 0;                                                                           \
+            if (++i02 == ne02) {                                                               \
+                i02 = 0;                                                                       \
+                i03++;                                                                         \
+            }                                                                                  \
+            dst_ptr  = (uint8_t *) dst->data  + i02*nb2  + i03*nb3;                            \
+            src0_ptr = (uint8_t *) src0->data + i02*nb02 + i03*nb03;                           \
+        }                                                                                      \
+    }                                                                                          \
 }
 
 DEFINE_CPY_SAMESHAPE(f32,  float, 4)
 DEFINE_CPY_SAMESHAPE(f16, __fp16, 2)
 
-#define DEFINE_CPY_RESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                         \
-static void cpy_thread_##NAME##_reshape(unsigned int nth, unsigned int ith, void * data) {                     \
-    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                           \
-    struct htp_ops_context * octx = ct->octx;                                                                  \
-    cpy_preamble;                                                                                              \
-    const uint32_t dr  = ct->src0_nrows_per_thread;                                                            \
-    const uint32_t ir0 = dr * ith;                                                                             \
-    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;                                                    \
-    if (ir0 >= nr) return;                                                                                     \
-    const bool src0_contig = (nb00 == ELEM_SIZE)   &&                                                          \
-                             (nb01 == ne00 * nb00) &&                                                          \
-                             (nb02 == ne01 * nb01) &&                                                          \
-                             (nb03 == ne02 * nb02);                                                            \
-    const bool dst_contig  = (nb0  == ELEM_SIZE)   &&                                                          \
-                             (nb1  == ne0  * nb0)  &&                                                          \
-                             (nb2  == ne1  * nb1)  &&                                                          \
-                             (nb3  == ne2  * nb2);                                                             \
-    if (src0_contig && dst_contig) {                                                                           \
-        for (int64_t i03 = 0; i03 < ne03; i03++) {                                                             \
-            for (int64_t i02 = 0; i02 < ne02; i02++) {                                                         \
-                uint8_t * src_ptr = (uint8_t *) src0->data + i03*nb03 + i02*nb02 + ir0*nb01;                   \
-                uint32_t  flat    = ((i03*ne02 + i02)*ne01 + ir0) * ne00;                                      \
-                uint8_t * dst_ptr = (uint8_t *) dst->data  + flat * ELEM_SIZE;                                 \
-                hvx_copy_uu(dst_ptr, src_ptr, (ir1 - ir0) * ne00, ELEM_SIZE);                                  \
-            }                                                                                                  \
-        }                                                                                                      \
-        return;                                                                                                \
-    }                                                                                                          \
-    const bool reshape_flat_fast = (ne03 == 1 && ne2 == 1 && ne3 == 1) &&                                      \
-                                   (ne0 == ne00 * ne01) && (ne1 == ne02) &&                                    \
-                                   (nb00 == ELEM_SIZE) && (nb0 == ELEM_SIZE);                                  \
-    if (reshape_flat_fast) {                                                                                   \
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {                                                            \
-            for (uint32_t i01 = ir0; i01 < ir1; i01++) {                                                       \
-                uint8_t * src0_ptr = (uint8_t *) src0->data + i01 * nb01 + i02 * nb02;                         \
-                uint8_t * dst_ptr  = (uint8_t *) dst->data  + i01 * ne00 * ELEM_SIZE + i02 * nb1;              \
-                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                               \
-            }                                                                                                  \
-        }                                                                                                      \
-        return;                                                                                                \
-    }                                                                                                          \
-    int64_t k10 = 0;                                                                                           \
-    int64_t i11 = 0;                                                                                           \
-    int64_t i12 = 0;                                                                                           \
-    int64_t i13 = 0;                                                                                           \
-    const int64_t nk00 = ct->src0_blocks_per_row;                                                              \
-    const int64_t nk0  = ct->dst_blocks_per_row;                                                               \
-    for (int64_t i03 = 0; i03 < ne03; i03++) {                                                                 \
-        for (int64_t i02 = 0; i02 < ne02; i02++) {                                                             \
-            k10 += nk00 * ir0;                                                                                 \
-            while (k10 >= nk0) {                                                                               \
-                k10 -= nk0;                                                                                    \
-                if (++i11 == ne1) {                                                                            \
-                    i11 = 0;                                                                                   \
-                    if (++i12 == ne2) {                                                                        \
-                        i12 = 0;                                                                               \
-                        if (++i13 == ne3) {                                                                    \
-                            i13 = 0;                                                                           \
-                        }                                                                                      \
-                    }                                                                                          \
-                }                                                                                              \
-            }                                                                                                  \
-            for (int64_t i01 = ir0; i01 < ir1; i01++) {                                                        \
-                for (int64_t k00 = 0; k00 < nk00; k00++) {                                                     \
-                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); \
-                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);  \
-                    memcpy(dst_ptr, src0_ptr, ELEM_SIZE);                                                      \
-                    if (++k10 == nk0) {                                                                        \
-                        k10 = 0;                                                                               \
-                        if (++i11 == ne1) {                                                                    \
-                            i11 = 0;                                                                           \
-                            if (++i12 == ne2) {                                                                \
-                                i12 = 0;                                                                       \
-                                if (++i13 == ne3) {                                                            \
-                                    i13 = 0;                                                                   \
-                                }                                                                              \
-                            }                                                                                  \
-                        }                                                                                      \
-                    }                                                                                          \
-                }                                                                                              \
-            }                                                                                                  \
-            k10 += nk00 * (ne01 - ir1);                                                                        \
-            while (k10 >= nk0) {                                                                               \
-                k10 -= nk0;                                                                                    \
-                if (++i11 == ne1) {                                                                            \
-                    i11 = 0;                                                                                   \
-                    if (++i12 == ne2) {                                                                        \
-                        i12 = 0;                                                                               \
-                        if (++i13 == ne3) {                                                                    \
-                            i13 = 0;                                                                           \
-                        }                                                                                      \
-                    }                                                                                          \
-                }                                                                                              \
-            }                                                                                                  \
-        }                                                                                                      \
-    }                                                                                                          \
+#define DEFINE_CPY_RESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                               \
+static void cpy_thread_##NAME##_reshape(unsigned int nth, unsigned int ith, void * data) {           \
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                 \
+    struct htp_ops_context * octx = ct->octx;                                                        \
+    cpy_preamble;                                                                                    \
+    const uint32_t th_nelem = ct->elem_per_thread;                                                   \
+    const uint32_t th_start = ct->elem_start + ith * th_nelem;                                       \
+    const uint32_t th_end   = MIN(th_start + th_nelem, ct->elem_start + ct->nelem);                  \
+    if (th_start >= th_end) return;                                                                  \
+                                                                                                     \
+    const uint32_t ne01_ne00      = ne01 * ne00;                                                     \
+    const uint32_t ne02_ne01_ne00 = ne02 * ne01_ne00;                                                \
+    const uint32_t ne1_ne0        = ne1 * ne0;                                                       \
+    const uint32_t ne2_ne1_ne0    = ne2 * ne1_ne0;                                                   \
+                                                                                                     \
+    uint32_t e = th_start;                                                                           \
+    uint32_t i13 = fastdiv(e, &ct->div_ne2_ne1_ne0);                                                 \
+    uint32_t rem = e - i13 * ne2_ne1_ne0;                                                            \
+    uint32_t i12 = fastdiv(rem, &ct->div_ne1_ne0);                                                   \
+    uint32_t rem2 = rem - i12 * ne1_ne0;                                                             \
+    uint32_t i11 = fastdiv(rem2, &ct->div_ne0);                                                      \
+    uint32_t i10 = rem2 - i11 * ne0;                                                                 \
+                                                                                                     \
+    uint32_t i03 = fastdiv(e, &ct->div_ne02_ne01_ne00);                                              \
+    uint32_t rem_s = e - i03 * ne02_ne01_ne00;                                                       \
+    uint32_t i02 = fastdiv(rem_s, &ct->div_ne01_ne00);                                               \
+    uint32_t rem2_s = rem_s - i02 * ne01_ne00;                                                       \
+    uint32_t i01 = fastdiv(rem2_s, &ct->div_ne00);                                                   \
+    uint32_t i00 = rem2_s - i01 * ne00;                                                              \
+                                                                                                     \
+    char * dst_ptr        = (char *)       dst->data  + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3;    \
+    const char * src0_ptr = (const char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03;   \
+                                                                                                     \
+    for (; e < th_end; e++) {                                                                        \
+        *((ELEM_TYPE *) dst_ptr) = *((const ELEM_TYPE *) src0_ptr);                                  \
+                                                                                                     \
+        dst_ptr += nb0;                                                                              \
+        if (++i10 == ne0) {                                                                          \
+            i10 = 0;                                                                                 \
+            if (++i11 == ne1) {                                                                      \
+                i11 = 0;                                                                             \
+                if (++i12 == ne2) {                                                                  \
+                    i12 = 0;                                                                         \
+                    i13++;                                                                           \
+                }                                                                                    \
+            }                                                                                        \
+            dst_ptr = (char *) dst->data + i11*nb1 + i12*nb2 + i13*nb3;                              \
+        }                                                                                            \
+                                                                                                     \
+        src0_ptr += nb00;                                                                            \
+        if (++i00 == ne00) {                                                                         \
+            i00 = 0;                                                                                 \
+            if (++i01 == ne01) {                                                                     \
+                i01 = 0;                                                                             \
+                if (++i02 == ne02) {                                                                 \
+                    i02 = 0;                                                                         \
+                    i03++;                                                                           \
+                }                                                                                    \
+            }                                                                                        \
+            src0_ptr = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;                   \
+        }                                                                                            \
+    }                                                                                                \
 }
 
 DEFINE_CPY_RESHAPE(f32,  float, 4)
@@ -189,22 +188,33 @@ static void cpy_thread_f16_f32_sameshape(unsigned int nth, unsigned int ith, voi
     struct htp_ops_context * octx = ct->octx;
     cpy_preamble;
 
-    // parallelize by src0 rows
     const uint32_t dr  = ct->src0_nrows_per_thread;
-    const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
-    if (ir0 >= nr) return;
+    const uint32_t ir0 = ct->row_start + dr * ith;
+    const uint32_t ir1 = MIN(ir0 + dr, ct->row_start + ct->nrows);
+    if (ir0 >= ir1) return;
 
-    // copy by rows
-    for (uint32_t i03 = 0; i03 < ne03; i03++) {
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {
-            #pragma unroll(2)
-            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
-                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
-                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
-                hex_l2fetch(src0_ptr, ne00 * sizeof(float), nb01, 2);
-                hvx_copy_f16_f32_uu(dst_ptr, src0_ptr, ne00);
+    const uint32_t ne02_ne01 = ne02 * ne01;
+    uint32_t i03 = fastdiv(ir0, &ct->div_ne02_ne01);
+    uint32_t rem = ir0 - i03 * ne02_ne01;
+    uint32_t i02 = fastdiv(rem, &ct->div_ne01);
+    uint32_t i01 = rem - i02 * ne01;
+
+    uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+    uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+
+    for (uint32_t r = ir0; r < ir1; r++) {
+        hex_l2fetch(src0_ptr, ne00 * sizeof(float), nb01, 2);
+        hvx_copy_f16_f32_uu(dst_ptr, src0_ptr, ne00);
+        dst_ptr  += nb1;
+        src0_ptr += nb01;
+        if (++i01 == ne01) {
+            i01 = 0;
+            if (++i02 == ne02) {
+                i02 = 0;
+                i03++;
             }
+            dst_ptr  = (uint8_t*) dst->data  + i02*nb2  + i03*nb3;
+            src0_ptr = (uint8_t*) src0->data + i02*nb02 + i03*nb03;
         }
     }
 }
@@ -214,22 +224,33 @@ static void cpy_thread_f32_f16_sameshape(unsigned int nth, unsigned int ith, voi
     struct htp_ops_context * octx = ct->octx;
     cpy_preamble;
 
-    // parallelize by src0 rows
     const uint32_t dr  = ct->src0_nrows_per_thread;
-    const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
-    if (ir0 >= nr) return;
+    const uint32_t ir0 = ct->row_start + dr * ith;
+    const uint32_t ir1 = MIN(ir0 + dr, ct->row_start + ct->nrows);
+    if (ir0 >= ir1) return;
 
-    // copy by rows
-    for (uint32_t i03 = 0; i03 < ne03; i03++) {
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {
-            #pragma unroll(2)
-            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
-                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
-                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
-                hex_l2fetch(src0_ptr, ne00 * sizeof(__fp16), nb01, 2);
-                hvx_copy_f32_f16_uu(dst_ptr, src0_ptr, ne00);
+    const uint32_t ne02_ne01 = ne02 * ne01;
+    uint32_t i03 = fastdiv(ir0, &ct->div_ne02_ne01);
+    uint32_t rem = ir0 - i03 * ne02_ne01;
+    uint32_t i02 = fastdiv(rem, &ct->div_ne01);
+    uint32_t i01 = rem - i02 * ne01;
+
+    uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+    uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+
+    for (uint32_t r = ir0; r < ir1; r++) {
+        hex_l2fetch(src0_ptr, ne00 * sizeof(__fp16), nb01, 2);
+        hvx_copy_f32_f16_uu(dst_ptr, src0_ptr, ne00);
+        dst_ptr  += nb1;
+        src0_ptr += nb01;
+        if (++i01 == ne01) {
+            i01 = 0;
+            if (++i02 == ne02) {
+                i02 = 0;
+                i03++;
             }
+            dst_ptr  = (uint8_t*) dst->data  + i02*nb2  + i03*nb3;
+            src0_ptr = (uint8_t*) src0->data + i02*nb02 + i03*nb03;
         }
     }
 }
@@ -250,15 +271,19 @@ static inline void cpy_dma_sametype_sameshape(
     dma_queue * q = octx->ctx->dma[0];
 
     if (contiguous_outer) {
-        dma_queue_push(q, dma_make_ptr((void *) dst->data, (const void *) src0->data), nb1, nb01, ne00 * elem_size, ne01 * ne02 * ne03);
-        dma_queue_pop(q);
+        if (!dma_queue_push(q, dma_make_ptr((void *) dst->data, (const void *) src0->data), nb1, nb01, ne00 * elem_size, ne01 * ne02 * ne03)) {
+            dma_queue_flush(q);
+            dma_queue_push(q, dma_make_ptr((void *) dst->data, (const void *) src0->data), nb1, nb01, ne00 * elem_size, ne01 * ne02 * ne03);
+        }
+        dma_queue_flush(q);
         return;
     }
 
     for (uint32_t i03 = 0; i03 < ne03; i03++) {
         for (uint32_t i02 = 0; i02 < ne02; i02++) {
-            uint8_t* dst_ptr  = (uint8_t*) dst->data  + i02*nb2  + i03*nb3;
-            uint8_t* src0_ptr = (uint8_t*) src0->data + i02*nb02 + i03*nb03;
+            uint8_t * dst_ptr  = (uint8_t *) dst->data  + i02 * nb2  + i03 * nb3;
+            uint8_t * src0_ptr = (uint8_t *) src0->data + i02 * nb02 + i03 * nb03;
+
             if (!dma_queue_push(q, dma_make_ptr(dst_ptr, src0_ptr), nb1, nb01, ne00 * elem_size, ne01)) {
                 dma_queue_flush(q);
                 dma_queue_push(q, dma_make_ptr(dst_ptr, src0_ptr), nb1, nb01, ne00 * elem_size, ne01);
@@ -269,10 +294,9 @@ static inline void cpy_dma_sametype_sameshape(
     dma_queue_flush(q);
 }
 
-int op_cpy(struct htp_ops_context * octx) {
+static int exec_cpy(struct htp_ops_context * octx, bool * use_dma) {
     cpy_preamble;
-
-    const uint32_t n_threads = MIN(nr, octx->n_threads);
+    *use_dma = false;
 
     struct htp_copy_context ct;
     ct.octx = octx;
@@ -296,59 +320,117 @@ int op_cpy(struct htp_ops_context * octx) {
     }
 
     const bool sametype   = (src0->type == dst->type);
-    const bool transposed = (nb00 > nb01) || (nb0 > nb1);
+    const bool transposed = (nb00 > nb01) || (nb0 > nb1) ||
+                            (nb00 != ct.src0_type_size) || (nb0 != ct.dst_type_size) ||
+                            (nb01 < ne00 * ct.src0_type_size) || (nb1 < ne0 * ct.dst_type_size);
     const bool sameshape  = !transposed && (ne00 == ne0 && ne01 == ne1 && ne02 == ne2 && ne03 == ne3);
 
-    ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
+    const uint32_t n_threads = octx->n_threads;
 
-    worker_callback_t copy_fun = NULL;
-    bool use_dma = false;
+    const bool dst_is_contiguous = htp_tensor_is_contiguous(dst, ct.dst_type_size);
 
-    if (sametype && sameshape) {
-        use_dma = true;
-    } else if (sameshape) {
-        /**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
-            copy_fun = cpy_thread_f16_f32_sameshape;
-        else if (dst->type == HTP_TYPE_F32 && src0->type == HTP_TYPE_F16)
-            copy_fun = cpy_thread_f32_f16_sameshape;
-        else
-            return HTP_STATUS_NO_SUPPORT;
-    } else if (sametype) {
-        if (src0->type == HTP_TYPE_F32) {
-            copy_fun = cpy_thread_f32_reshape;
-        } else {
-            copy_fun = cpy_thread_f16_reshape;
+    if (sameshape) {
+        const uint32_t total_rows = ne01 * ne02 * ne03;
+        const uint32_t row_size   = ne00 * ct.dst_type_size;
+
+        ct.div_ne01      = init_fastdiv_values(ne01);
+        ct.div_ne02_ne01 = init_fastdiv_values(ne02 * ne01);
+
+        uint32_t row_start = 0;
+        uint32_t nrows     = total_rows;
+
+        if (octx->ctx->mdev.count > 1) {
+            const uint32_t rows_per_chunk = (row_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(row_size, HEX_L2_LINE_SIZE)) : 1;
+            const bool can_split = htp_tensor_mdev_data_aligned(dst) && dst_is_contiguous;
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, can_split ? rows_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            row_start = range.start;
+            nrows     = range.count;
         }
+
+        if (nrows == 0) {
+            return HTP_STATUS_OK;
+        }
+
+        ct.row_start = row_start;
+        ct.nrows     = nrows;
+        ct.src0_nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
+
+        if (sametype && octx->ctx->mdev.count <= 1) {
+            *use_dma = true;
+            cpy_dma_sametype_sameshape(octx, dst, src0, ct.src0_type_size, ne00, ne01, ne02, ne03, nb01, nb02, nb03, nb1, nb2, nb3);
+        } else {
+            work_queue_func_t copy_fun = NULL;
+            if (sametype) {
+                copy_fun = (src0->type == HTP_TYPE_F32) ? cpy_thread_f32_sameshape : cpy_thread_f16_sameshape;
+            } else if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32) {
+                copy_fun = cpy_thread_f16_f32_sameshape;
+            } else if (dst->type == HTP_TYPE_F32 && src0->type == HTP_TYPE_F16) {
+                copy_fun = cpy_thread_f32_f16_sameshape;
+            } else {
+                return HTP_STATUS_NO_SUPPORT;
+            }
+            work_queue_run(octx->ctx->work_queue, copy_fun, &ct, n_threads);
+        }
+    } else if (sametype) {
+        const uint32_t total_elems = ne0 * ne1 * ne2 * ne3;
+        const uint32_t elems_per_line = (ct.dst_type_size == 4) ? 32 : 64;
+
+        ct.div_ne0            = init_fastdiv_values(ne0);
+        ct.div_ne1_ne0        = init_fastdiv_values(ne1 * ne0);
+        ct.div_ne2_ne1_ne0    = init_fastdiv_values(ne2 * ne1 * ne0);
+        ct.div_ne00           = init_fastdiv_values(ne00);
+        ct.div_ne01_ne00      = init_fastdiv_values(ne01 * ne00);
+        ct.div_ne02_ne01_ne00 = init_fastdiv_values(ne02 * ne01 * ne00);
+
+        uint32_t elem_start = 0;
+        uint32_t nelem      = total_elems;
+
+        if (octx->ctx->mdev.count > 1) {
+            const bool can_split = htp_tensor_mdev_data_aligned(dst) && dst_is_contiguous;
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_elems, can_split ? elems_per_line : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            elem_start = range.start;
+            nelem      = range.count;
+        }
+
+        if (nelem == 0) {
+            return HTP_STATUS_OK;
+        }
+
+        ct.elem_start      = elem_start;
+        ct.nelem           = nelem;
+        ct.elem_per_thread = fastdiv(nelem + n_threads - 1, &octx->n_threads_div);
+
+        work_queue_func_t copy_fun = (src0->type == HTP_TYPE_F32) ? cpy_thread_f32_reshape : cpy_thread_f16_reshape;
+        work_queue_run(octx->ctx->work_queue, copy_fun, &ct, n_threads);
     } else {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    FARF(HIGH, "cpy-%s-%s: (%ux%ux%ux%u) -> (%ux%ux%ux%u) : use_dma=%d n_threads %u\n",
-         src0->type == HTP_TYPE_F32 ? "f32" : "f16", dst->type == HTP_TYPE_F32 ? "f32" : "f16",
-         ne00, ne01, ne02, ne03, ne0, ne1, ne2, ne3, use_dma, n_threads);
-
-    if (use_dma) {
-        cpy_dma_sametype_sameshape(octx, dst, src0, ct.src0_type_size, ne00, ne01, ne02, ne03, nb01, nb02, nb03, nb1, nb2, nb3);
-    } else {
-        worker_pool_run_func(octx->ctx->worker_pool, copy_fun, &ct, n_threads);
-    }
-
-    const struct htp_tensor *sync = octx->src[1];
-    if (sync && (sync->flags & HTP_TENSOR_FENCE)) {
-        if (!use_dma) {
-            // htp_tensor_flush_all(octx->ctx, octx->dsts, 1);
-            qurt_mem_cache_clean((qurt_addr_t) 0, 0, QURT_MEM_CACHE_FLUSH_INVALIDATE_ALL, QURT_MEM_DCACHE);
-        }
-
-        atomic_uint * sync_fence = (atomic_uint *) sync->data;
-        const uint32_t seq = (uint32_t) octx->op_params[0];
-
-        atomic_store(&sync_fence[0], seq);
-        asm volatile ("syncht" : : : "memory");
-        Q6_dccleaninva_A((void *) sync_fence);
-
-        FARF(HIGH, "ggml-hex: sync-release : fence %p seq %u\n", sync_fence, seq);
-    }
-
     return HTP_STATUS_OK;
 }
+
+int op_cpy(struct htp_ops_context * octx) {
+    bool use_dma = false;
+    int status = exec_cpy(octx, &use_dma);
+
+    htp_ops_context_set_status(octx, status);
+
+    if (octx->op == HTP_OP_CPY_FENCE) {
+        if (!use_dma) {
+            htp_flush_dirty_ranges(octx->ctx);
+        }
+
+        htp_mdev_group_barrier(octx);
+
+        if (octx->ctx->mdev.idx == 0) {
+            const struct htp_tensor * sync = octx->src[1];
+            const uint32_t seq = (uint32_t) octx->op_params[0];
+            atomic_uint * sync_fence = (atomic_uint *) (uintptr_t) sync->data;
+            htp_fence_write(sync_fence, seq, octx->status);
+
+            FARF(HIGH, "ggml-hex: sync-release : fence %p seq 0x%x status %d\n", sync_fence, seq, octx->status);
+        }
+    }
+
+    return octx->status;
+}

ggml/src/ggml-hexagon/htp/cumsum-ops.c

@@ -7,6 +7,8 @@
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
 #include "htp-tensor.h"
@@ -17,25 +19,25 @@
 #define htp_cumsum_tensors_preamble                         \
     const struct htp_tensor * restrict src0 = octx->src[0]; \
     const struct htp_tensor * restrict dst  = octx->dst;    \
-                                                     \
-    const uint32_t ne00 = src0->ne[0];               \
-    const uint32_t ne01 = src0->ne[1];               \
-    const uint32_t ne02 = src0->ne[2];               \
-    const uint32_t ne03 = src0->ne[3];               \
-                                                     \
-    const uint32_t ne0 = dst->ne[0];                 \
-    const uint32_t ne1 = dst->ne[1];                 \
-    const uint32_t ne2 = dst->ne[2];                 \
-    const uint32_t ne3 = dst->ne[3];                 \
-                                                     \
-    const uint32_t nb00 = src0->nb[0];               \
-    const uint32_t nb01 = src0->nb[1];               \
-    const uint32_t nb02 = src0->nb[2];               \
-    const uint32_t nb03 = src0->nb[3];               \
-                                                     \
-    const uint32_t nb0 = dst->nb[0];                 \
-    const uint32_t nb1 = dst->nb[1];                 \
-    const uint32_t nb2 = dst->nb[2];                 \
+                                                            \
+    const uint32_t ne00 = src0->ne[0];                      \
+    const uint32_t ne01 = src0->ne[1];                      \
+    const uint32_t ne02 = src0->ne[2];                      \
+    const uint32_t ne03 = src0->ne[3];                      \
+                                                            \
+    const uint32_t ne0 = dst->ne[0];                        \
+    const uint32_t ne1 = dst->ne[1];                        \
+    const uint32_t ne2 = dst->ne[2];                        \
+    const uint32_t ne3 = dst->ne[3];                        \
+                                                            \
+    const uint32_t nb00 = src0->nb[0];                      \
+    const uint32_t nb01 = src0->nb[1];                      \
+    const uint32_t nb02 = src0->nb[2];                      \
+    const uint32_t nb03 = src0->nb[3];                      \
+                                                            \
+    const uint32_t nb0 = dst->nb[0];                        \
+    const uint32_t nb1 = dst->nb[1];                        \
+    const uint32_t nb2 = dst->nb[2];                        \
     const uint32_t nb3 = dst->nb[3];
 
 struct htp_cumsum_context {
@@ -46,6 +48,7 @@ struct htp_cumsum_context {
     size_t          dst_row_size_aligned;
     uint32_t        rows_per_thread;
     uint32_t        total_rows;
+    uint32_t        row_start;
 };
 
 #define htp_cumsum_preamble                                                \
@@ -116,11 +119,8 @@ static inline void hvx_cumsum_row_f32(const float * restrict src, float * restri
 static void cumsum_thread_f32_dma(unsigned int nth, unsigned int ith, void * data) {
     htp_cumsum_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
-    const uint32_t ir0 = cctx->rows_per_thread * ith;
-    const uint32_t ir1 = MIN(ir0 + cctx->rows_per_thread, cctx->total_rows);
+    const uint32_t ir0 = cctx->row_start + cctx->rows_per_thread * ith;
+    const uint32_t ir1 = MIN(ir0 + cctx->rows_per_thread, cctx->row_start + cctx->total_rows);
 
     if (ir0 >= ir1) {
         return;
@@ -149,11 +149,15 @@ static void cumsum_thread_f32_dma(unsigned int nth, unsigned int ith, void * dat
                                    src_row_size_aligned, src_row_size, 1);
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = ir0; ir < ir1; ir++) {
         float * dst_spad_row = (float *) dma_queue_pop(dma_queue).src;
         float * src_spad_row = (float *) dma_queue_pop(dma_queue).dst;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         hvx_cumsum_row_f32(src_spad_row, dst_spad_row, ne00);
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         dma_queue_push_vtcm_to_ddr(dma_queue,
                                    dma_make_ptr(dst_data + (ir * dst_row_size), (uint8_t *) dst_spad_row),
@@ -168,12 +172,10 @@ static void cumsum_thread_f32_dma(unsigned int nth, unsigned int ith, void * dat
     }
 
     dma_queue_flush(dma_queue);
-    t2 = HAP_perf_get_qtimer_count();
 
-    FARF(HIGH, "cumsum-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n",
+    FARF(HIGH, "cumsum-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u\n",
          ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ir0, ir1,
-         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
 }
 
 // ---------------------------------------------------------------------------
@@ -183,14 +185,14 @@ static void cumsum_thread_f32_dma(unsigned int nth, unsigned int ith, void * dat
 static void cumsum_thread_f32(unsigned int nth, unsigned int ith, void * data) {
     htp_cumsum_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
     const uint8_t * src_data = (const uint8_t *) src0->data;
     uint8_t *       dst_data = (uint8_t *) dst->data;
 
-    const uint32_t ir0 = cctx->rows_per_thread * ith;
-    const uint32_t ir1 = MIN(ir0 + cctx->rows_per_thread, cctx->total_rows);
+    const uint32_t ir0 = cctx->row_start + cctx->rows_per_thread * ith;
+    const uint32_t ir1 = MIN(ir0 + cctx->rows_per_thread, cctx->row_start + cctx->total_rows);
+
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir0);
 
     for (uint32_t ir = ir0; ir < ir1; ir++) {
         const float * restrict src_row = (const float *) (src_data + ir * cctx->src_row_size);
@@ -198,12 +200,11 @@ static void cumsum_thread_f32(unsigned int nth, unsigned int ith, void * data) {
         hvx_cumsum_row_f32(src_row, dst_row, ne00);
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir0);
 
-    FARF(HIGH, "cumsum-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n",
+    FARF(HIGH, "cumsum-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u\n",
          ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ir0, ir1,
-         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
 }
 
 int op_cumsum_f32(struct htp_ops_context * octx) {
@@ -214,8 +215,25 @@ int op_cumsum_f32(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    const uint32_t total_rows = src0->ne[1] * src0->ne[2] * src0->ne[3];
-    const uint32_t n_threads  = MIN(octx->n_threads, total_rows);
+    const uint32_t total_rows      = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const size_t dst_data_row_size = dst->ne[0] * sizeof(float);
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = total_rows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, sizeof(float), (uint32_t) dst_data_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
 
     const size_t src_row_size         = src0->nb[1];
     const size_t dst_row_size         = dst->nb[1];
@@ -240,14 +258,15 @@ int op_cumsum_f32(struct htp_ops_context * octx) {
         .dst_row_size         = dst_row_size,
         .src_row_size_aligned = src_row_size_aligned,
         .dst_row_size_aligned = dst_row_size_aligned,
-        .rows_per_thread      = (total_rows + n_threads - 1) / n_threads,
-        .total_rows           = total_rows,
+        .rows_per_thread      = fastdiv(nrows + n_threads - 1, &octx->n_threads_div),
+        .total_rows           = nrows,
+        .row_start            = row_start,
     };
 
     if (octx->ctx->vtcm_size < spad_per_thread * n_threads) {
-        worker_pool_run_func(octx->ctx->worker_pool, cumsum_thread_f32, &cctx, n_threads);
+        work_queue_run(octx->ctx->work_queue, cumsum_thread_f32, &cctx, n_threads);
     } else {
-        worker_pool_run_func(octx->ctx->worker_pool, cumsum_thread_f32_dma, &cctx, n_threads);
+        work_queue_run(octx->ctx->work_queue, cumsum_thread_f32_dma, &cctx, n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/diag-ops.c

@@ -5,8 +5,11 @@
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
+#include "htp-tensor.h"
 #include "hvx-types.h"
 #include "hex-utils.h"
 #include "hvx-copy.h"
@@ -15,17 +18,17 @@
 #define htp_diag_tensors_preamble                           \
     const struct htp_tensor * restrict src0 = octx->src[0]; \
     const struct htp_tensor * restrict dst  = octx->dst;    \
-                                                     \
-    const uint32_t ne02 = src0->ne[2];               \
-                                                     \
-    const uint32_t ne0 = dst->ne[0];                 \
-    const uint32_t ne1 = dst->ne[1];                 \
-                                                     \
-    const uint32_t nb02 = src0->nb[2];               \
-    const uint32_t nb03 = src0->nb[3];               \
-                                                     \
-    const uint32_t nb1 = dst->nb[1];                 \
-    const uint32_t nb2 = dst->nb[2];                 \
+                                                            \
+    const uint32_t ne02 = src0->ne[2];                      \
+                                                            \
+    const uint32_t ne0 = dst->ne[0];                        \
+    const uint32_t ne1 = dst->ne[1];                        \
+                                                            \
+    const uint32_t nb02 = src0->nb[2];                      \
+    const uint32_t nb03 = src0->nb[3];                      \
+                                                            \
+    const uint32_t nb1 = dst->nb[1];                        \
+    const uint32_t nb2 = dst->nb[2];                        \
     const uint32_t nb3 = dst->nb[3];
 
 struct htp_diag_context {
@@ -36,6 +39,7 @@ struct htp_diag_context {
     size_t          dst_row_size_aligned;
     uint32_t        batches_per_thread;
     uint32_t        total_batches;
+    uint32_t        batch_start;
 };
 
 #define htp_diag_preamble                                              \
@@ -57,11 +61,8 @@ static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data)
     htp_diag_preamble;
     dma_queue * dma_queue = octx->ctx->dma[ith];
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
-    const uint32_t ib0 = dctx->batches_per_thread * ith;
-    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->total_batches);
+    const uint32_t ib0 = dctx->batch_start + dctx->batches_per_thread * ith;
+    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->batch_start + dctx->total_batches);
 
     if (ib0 >= ib1) {
         return;
@@ -79,6 +80,8 @@ static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data)
     uint8_t * src_spad = octx->src0_spad.data + (ith * src_batch_size_aligned);
     uint8_t * dst_spad = octx->dst_spad.data  + (ith * dst_row_size_aligned);
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ib = ib0; ib < ib1; ib++) {
         const uint32_t i3 = ib / ne02;
         const uint32_t i2 = ib % ne02;
@@ -96,7 +99,9 @@ static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data)
 
         for (uint32_t i1 = 0; i1 < ne1; i1++) {
             // Compute row in VTCM
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (ib * ne1 + i1));
             hvx_diag_row_f32(src_spad_f32, dst_spad_f32, i1, ne0);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (ib * ne1 + i1));
 
             // Write completed row back to DDR
             uint8_t * dst_row = dst_data + i3 * nb3 + i2 * nb2 + i1 * nb1;
@@ -107,12 +112,9 @@ static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data)
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
-
-    FARF(HIGH, "diag-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n",
+    FARF(HIGH, "diag-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u\n",
          ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ib0, ib1,
-         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
 }
 
 // ---------------------------------------------------------------------------
@@ -122,14 +124,14 @@ static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data)
 static void diag_thread_f32(unsigned int nth, unsigned int ith, void * data) {
     htp_diag_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
     const uint8_t * src_data = (const uint8_t *) src0->data;
     uint8_t *       dst_data = (uint8_t *) dst->data;
 
-    const uint32_t ib0 = dctx->batches_per_thread * ith;
-    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->total_batches);
+    const uint32_t ib0 = dctx->batch_start + dctx->batches_per_thread * ith;
+    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->batch_start + dctx->total_batches);
+
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ib0);
 
     for (uint32_t ib = ib0; ib < ib1; ib++) {
         const uint32_t i3 = ib / ne02;
@@ -143,12 +145,11 @@ static void diag_thread_f32(unsigned int nth, unsigned int ith, void * data) {
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ib0);
 
-    FARF(HIGH, "diag-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n",
+    FARF(HIGH, "diag-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u\n",
          ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ib0, ib1,
-         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
 }
 
 int op_diag_f32(struct htp_ops_context * octx) {
@@ -160,7 +161,36 @@ int op_diag_f32(struct htp_ops_context * octx) {
     }
 
     const uint32_t total_batches = src0->ne[2] * src0->ne[3];
-    const uint32_t n_threads     = MIN(octx->n_threads, total_batches);
+    const size_t dst_batch_size  = dst->ne[1] * dst->nb[1];
+
+    uint32_t batch_start = 0;
+    uint32_t nbatches    = total_batches;
+
+    if (octx->ctx->mdev.count > 1) {
+        bool can_split = htp_tensor_mdev_data_aligned(dst) && (dst->ne[0] == 1 || dst->nb[0] == sizeof(float)) && !htp_tensor_is_permuted(dst);
+        uint32_t batches_per_chunk = 1;
+        if (can_split) {
+            if (dst->ne[2] > 1 && (dst->nb[2] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0 &&
+                (dst->ne[3] <= 1 || (dst->nb[3] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0)) {
+                batches_per_chunk = 1;
+            } else if (dst->nb[2] == dst_batch_size &&
+                       (dst->ne[3] <= 1 || dst->nb[3] == dst->nb[2] * dst->ne[2])) {
+                batches_per_chunk = (dst_batch_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(dst_batch_size, HEX_L2_LINE_SIZE)) : 1;
+            } else {
+                can_split = false;
+            }
+        }
+
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_batches, can_split ? batches_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        batch_start = range.start;
+        nbatches    = range.count;
+    }
+
+    if (nbatches == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads     = octx->n_threads;
 
     const size_t src_batch_size         = src0->ne[0] * sizeof(float);
     const size_t dst_row_size           = dst->ne[0] * sizeof(float);
@@ -185,14 +215,15 @@ int op_diag_f32(struct htp_ops_context * octx) {
         .dst_row_size           = dst_row_size,
         .src_batch_size_aligned = src_batch_size_aligned,
         .dst_row_size_aligned   = dst_row_size_aligned,
-        .batches_per_thread     = (total_batches + n_threads - 1) / n_threads,
-        .total_batches          = total_batches,
+        .batches_per_thread     = fastdiv(nbatches + n_threads - 1, &octx->n_threads_div),
+        .total_batches          = nbatches,
+        .batch_start            = batch_start,
     };
 
     if (octx->ctx->vtcm_size < spad_per_thread * n_threads) {
-        worker_pool_run_func(octx->ctx->worker_pool, diag_thread_f32, &dctx, n_threads);
+        work_queue_run(octx->ctx->work_queue, diag_thread_f32, &dctx, n_threads);
     } else {
-        worker_pool_run_func(octx->ctx->worker_pool, diag_thread_f32_dma, &dctx, n_threads);
+        work_queue_run(octx->ctx->work_queue, diag_thread_f32_dma, &dctx, n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/fill-ops.c

@@ -3,10 +3,11 @@
 #pragma clang diagnostic ignored "-Wunused-but-set-variable"
 
 #include <HAP_farf.h>
-#include <HAP_perf.h>
-
 #include <string.h>
 
+#include "hex-common.h"
+#include "hex-profile.h"
+
 #include "hvx-copy.h"
 #include "hvx-utils.h"
 
@@ -14,28 +15,30 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
+#include "htp-tensor.h"
 
 // ggml op_params layout for FILL:
 //   op_params[0] (as float) - the scalar fill value
 
-#define fill_preamble \
+#define fill_preamble                          \
     const struct htp_tensor * dst = octx->dst; \
-    \
-    const uint32_t ne0 = dst->ne[0]; \
-    const uint32_t ne1 = dst->ne[1]; \
-    const uint32_t ne2 = dst->ne[2]; \
-    const uint32_t ne3 = dst->ne[3]; \
-    \
-    const uint32_t nb1 = dst->nb[1]; \
-    const uint32_t nb2 = dst->nb[2]; \
-    const uint32_t nb3 = dst->nb[3]; \
-    \
+                                               \
+    const uint32_t ne0 = dst->ne[0];           \
+    const uint32_t ne1 = dst->ne[1];           \
+    const uint32_t ne2 = dst->ne[2];           \
+    const uint32_t ne3 = dst->ne[3];           \
+                                               \
+    const uint32_t nb1 = dst->nb[1];           \
+    const uint32_t nb2 = dst->nb[2];           \
+    const uint32_t nb3 = dst->nb[3];           \
+                                               \
     const uint32_t nr = ne1 * ne2 * ne3;
 
 struct htp_fill_context {
     struct htp_ops_context * octx;
     uint32_t nrows_per_thread;
     uint32_t total_rows;  // ne1 * ne2 * ne3
+    uint32_t row_start;
     bool     opt_path;
     HVX_Vector splat_vec;
     uint32_t   elem_size;
@@ -47,10 +50,15 @@ static void fill_thread(unsigned int nth, unsigned int ith, void * data) {
     fill_preamble;
 
     // Parallelise over the flat row index spanning ne1*ne2*ne3
-    const uint32_t ir0 = fctx->nrows_per_thread * ith;
-    const uint32_t ir1 = MIN(ir0 + fctx->nrows_per_thread, fctx->total_rows);
+    const uint32_t ir0 = fctx->row_start + fctx->nrows_per_thread * ith;
+    const uint32_t ir1 = MIN(ir0 + fctx->nrows_per_thread, fctx->row_start + fctx->total_rows);
 
-    uint64_t t1 = HAP_perf_get_qtimer_count();
+    if (ir0 >= ir1) {
+        return;
+    }
+
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir0);
 
     if (fctx->opt_path) {
         // Opt path: tensor is fully contiguous, treat as flat array
@@ -69,9 +77,8 @@ static void fill_thread(unsigned int nth, unsigned int ith, void * data) {
         }
     }
 
-    uint64_t t2 = HAP_perf_get_qtimer_count();
-    FARF(HIGH, "fill %u/%u: rows %u:%u usec %u\n",
-         ith, nth, ir0, ir1, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir1);
+    FARF(HIGH, "fill %u/%u: rows %u:%u\n", ith, nth, ir0, ir1);
 }
 
 int op_fill(struct htp_ops_context * octx) {
@@ -85,8 +92,23 @@ int op_fill(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
+    uint32_t row_start = 0;
+    uint32_t nrows     = nr;
+
+    if (octx->ctx->mdev.count > 1) {
+        const uint32_t row_size = nb1;
+        const uint32_t rows_per_chunk = (row_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(row_size, HEX_L2_LINE_SIZE)) : 1;
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(nr, htp_tensor_mdev_data_aligned(dst) ? rows_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
     // nr = ne1*ne2*ne3 (flat row count across all outer dims); parallelise over it.
-    const uint32_t n_threads = MIN(nr, octx->n_threads);
+    const uint32_t n_threads = octx->n_threads;
 
     // Optimize if fully contiguous: skip stride arithmetic, treat as flat array
     const bool opt_path = (nb2 == nb1 * ne1) && (nb3 == nb2 * ne2);
@@ -99,8 +121,9 @@ int op_fill(struct htp_ops_context * octx) {
 
     struct htp_fill_context fctx = {
         .octx             = octx,
-        .nrows_per_thread = (nr + n_threads - 1) / n_threads,
-        .total_rows       = nr,
+        .nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div),
+        .total_rows       = nrows,
+        .row_start        = row_start,
         .opt_path         = opt_path,
     };
 
@@ -117,7 +140,7 @@ int op_fill(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, fill_thread, &fctx, n_threads);
+    work_queue_run(octx->ctx->work_queue, fill_thread, &fctx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -5,7 +5,6 @@
 #include <assert.h>
 #include <HAP_compute_res.h>
 #include <HAP_farf.h>
-#include <HAP_perf.h>
 #include <math.h>
 #include <stdbool.h>
 #include <stdatomic.h>
@@ -75,6 +74,7 @@ struct htp_fa_context {
 
     uint32_t qrows;
     uint32_t qrows_per_thread;
+    uint32_t qrow_start;
 
     bool is_q_fp32;
 
@@ -89,8 +89,6 @@ struct htp_fa_context {
 
     const struct htp_tensor * k;
     const struct htp_tensor * v;
-
-    uint64_t t_start;
 };
 
 struct hmx_fa_context {
@@ -206,10 +204,9 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
     const uint32_t nb3 = dst->nb[3];
 
     // total rows in q
-    const uint32_t nr = factx->qrows;
-    const uint32_t dr = factx->qrows_per_thread;
-    const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = MIN(ir0 + dr, nr);
+    const uint32_t dr  = factx->qrows_per_thread;
+    const uint32_t ir0 = factx->qrow_start + dr * ith;
+    const uint32_t ir1 = MIN(ir0 + dr, factx->qrow_start + factx->qrows);
 
     if (ir0 >= ir1) return;
 
@@ -1888,6 +1885,24 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     const uint32_t n_threads = factx.n_threads;
     const uint32_t G = factx.G;
 
+    // Multi-device: split Q blocks across devices
+    const uint32_t n_q_blocks = (neq1 + Br - 1) / Br;
+    uint32_t q_start_min = 0;
+    uint32_t q_start_max = neq1;
+
+    if (octx->ctx->mdev.count > 1) {
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(n_q_blocks, htp_tensor_mdev_data_aligned(dst) ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        const uint32_t block_start = range.start;
+        const uint32_t block_end   = range.start + range.count;
+
+        if (block_start >= block_end) {
+            return HTP_STATUS_OK;
+        }
+
+        q_start_min = block_start * Br;
+        q_start_max = MIN(block_end * Br, neq1);
+    }
+
     // ======== VTCM allocation (GQA-aware) ========
     // K/V row sizes drive the DMA descriptors (not the VTCM layout) and are used
     // throughout the KV loop below.
@@ -1977,7 +1992,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     // ======== Main loop ========
     for (uint32_t ib3 = 0; ib3 < neq3; ++ib3) {
         const uint32_t im3 = mask ? fastmodulo(ib3, mask->ne[3], &factx.src3_div3) : 0;
-        for (uint32_t q_start = 0; q_start < neq1; q_start += Br) {
+        for (uint32_t q_start = q_start_min; q_start < q_start_max; q_start += Br) {
             const uint32_t n_rows_q    = hex_smin(Br, neq1 - q_start);
             const size_t   n_rows_g    = n_rows_q * G;
             const size_t   g_br_actual = hex_align_up(n_rows_g, HMX_FP16_TILE_N_ROWS);
@@ -1991,8 +2006,9 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
 
                 // 1. Push Q and KV DMAs for the very first iteration.
                 // Subsequent iterations are enqueued early at the end of the previous iteration.
-                if (ib3 == 0 && q_start == 0 && kv_head == 0) {
-                    const uint8_t * q_ptr = (const uint8_t *) q->data;
+                if (ib3 == 0 && q_start == q_start_min && kv_head == 0) {
+                    const uint8_t * q_ptr = (const uint8_t *) q->data + q_start * q->nb[1] +
+                                            (kv_head * factx.G) * q->nb[2] + ib3 * q->nb[3];
                     const size_t q_row_bytes = q_transposed ? n_rows_q * q_row_bytes_trans_factor : q_row_bytes_untransposed;
                     const size_t n_rows      = q_transposed ? factx.G : n_rows_q;
                     dma_queue_push(dma, dma_make_ptr(factx.vtcm_q_dma, q_ptr), q_row_bytes, hex_smax(q_src_stride, q_row_bytes), q_row_bytes, n_rows);
@@ -2311,8 +2327,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                 if (next_kv_head >= n_kv_heads) {
                     next_kv_head = 0;
                     next_q_start = q_start + Br;
-                    if (next_q_start >= neq1) {
-                        next_q_start = 0;
+                    if (next_q_start >= q_start_max) {
+                        next_q_start = q_start_min;
                         next_ib3     = ib3 + 1;
                     }
                 }
@@ -2398,6 +2414,10 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
+    if (!htp_ops_context_set_n_threads(octx, kparams->n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
     if (kparams->kernel_type == HTP_FA_KERNEL_HMX) {
         return hmx_flash_attn_ext(octx);
     }
@@ -2407,8 +2427,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     factx.k = k;
     factx.v = v;
 
-    factx.t_start = HAP_perf_get_qtimer_count();
-
     factx.src0_div21 = kparams->u.hvx.src0_div21;
     factx.src0_div1  = kparams->u.hvx.src0_div1;
 
@@ -2451,8 +2469,30 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     }
 
     // total rows in q
-    factx.qrows = kparams->qrows;
-    factx.qrows_per_thread = kparams->qrows_per_thread;
+    const uint32_t neq1 = q->ne[1];
+    const uint32_t neq2 = q->ne[2];
+    const uint32_t neq3 = q->ne[3];
+    const uint32_t total_qrows = neq1 * neq2 * neq3;
+
+    uint32_t qrow_start = 0;
+    uint32_t qrows      = total_qrows;
+
+    if (octx->ctx->mdev.count > 1) {
+        const bool can_split = htp_tensor_mdev_data_aligned(dst) && ((dst->nb[1] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_qrows, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        qrow_start = range.start;
+        qrows      = range.count;
+    }
+
+    if (qrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
+    factx.qrows            = qrows;
+    factx.qrow_start       = qrow_start;
+    factx.qrows_per_thread = fastdiv(qrows + n_threads - 1, &octx->n_threads_div);
 
     size_t size_vkq_acc = hex_round_up(v->ne[0] * sizeof(float), 128); // VKQ32
 
@@ -2461,18 +2501,18 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
 
     uint8_t * vtcm_cur = octx->ctx->vtcm_base;
 
-    factx.spad_q = vtcm_seq_alloc(&vtcm_cur, size_q_block * octx->n_threads);
-    factx.spad_k = vtcm_seq_alloc(&vtcm_cur, factx.size_k_block * 2 * octx->n_threads);
-    factx.spad_v = vtcm_seq_alloc(&vtcm_cur, factx.size_v_block * 2 * octx->n_threads);
-    factx.spad_m = vtcm_seq_alloc(&vtcm_cur, (mask ? factx.size_m_block * HVX_FA_DMA_CACHE_SIZE : 0) * octx->n_threads);
-    factx.spad_a = vtcm_seq_alloc(&vtcm_cur, size_vkq_acc * octx->n_threads);
+    factx.spad_q = vtcm_seq_alloc(&vtcm_cur, size_q_block * n_threads);
+    factx.spad_k = vtcm_seq_alloc(&vtcm_cur, factx.size_k_block * 2 * n_threads);
+    factx.spad_v = vtcm_seq_alloc(&vtcm_cur, factx.size_v_block * 2 * n_threads);
+    factx.spad_m = vtcm_seq_alloc(&vtcm_cur, (mask ? factx.size_m_block * HVX_FA_DMA_CACHE_SIZE : 0) * n_threads);
+    factx.spad_a = vtcm_seq_alloc(&vtcm_cur, size_vkq_acc * n_threads);
 
     if ((size_t) (vtcm_cur - octx->ctx->vtcm_base) > octx->ctx->vtcm_size) {
         return HTP_STATUS_VTCM_TOO_SMALL;
     }
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        work_queue_run(octx->ctx->work_queue, flash_attn_ext_f16_thread, &factx, octx->n_threads);
+        work_queue_run(octx->ctx->work_queue, flash_attn_ext_f16_thread, &factx, n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/flash-attn-ops.h

@@ -51,6 +51,7 @@ struct htp_fa_kernel_params {
 
     uint32_t qrows;
     uint32_t qrows_per_thread;
+    uint32_t qrow_start;
     float    m0;
     float    m1;
     uint32_t n_head_log2;

ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c

@@ -4,10 +4,13 @@
 
 #include "hvx-utils.h"
 #include "hex-fastdiv.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
+#include "htp-tensor.h"
 
 #ifndef MIN
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -22,6 +25,8 @@ struct htp_gdn_context {
     size_t   state_bytes;
     uint8_t * vtcm_base;
     size_t   vtcm_per_thread;
+    uint32_t row_start;
+    uint32_t nrows;
 };
 
 static inline HVX_Vector gdn_mul_dot_f32(float * restrict dst, const float * restrict mul, const float * restrict dot, uint32_t n) {
@@ -586,8 +591,9 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
     const uint32_t n_seqs   = v->ne[3];
     const uint32_t K        = octx->op_params[0];
 
-    const uint32_t total_rows = H * n_seqs;
-    if (ith >= total_rows) {
+    const uint32_t row_end = gctx->row_start + gctx->nrows;
+
+    if (ith >= gctx->nrows) {
         return;
     }
 
@@ -621,11 +627,11 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
     const uint64_t state_seq_stride = state->nb[3] / sizeof(float);
     const uint64_t state_size_per_snap = (uint64_t) S_v * S_v * H * n_seqs;
 
-    uint32_t ir_prefetch = ith;
+    uint32_t ir_prefetch = gctx->row_start + ith;
     int spad_idx = 0;
 
     // Prefetch preamble (up to 2 steps)
-    for (int k = 0; k < 2 && ir_prefetch < total_rows; k++) {
+    for (int k = 0; k < 2 && ir_prefetch < row_end; k++) {
         const uint32_t piv1 = fastmodulo(ir_prefetch, H, &fd_H);
         const uint32_t piv3 = fastdiv(ir_prefetch, &fd_H);
         const float * ps_in = state_in_base + (uint64_t) piv3 * state_seq_stride + (uint64_t) piv1 * S_v * S_v;
@@ -646,8 +652,11 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (gctx->row_start + ith));
+
     int curr_spad_idx = 0;
-    for (uint32_t ir = ith; ir < total_rows; ir += nth) {
+    for (uint32_t ir = gctx->row_start + ith; ir < row_end; ir += nth) {
         dma_queue_pop(dma);
         dma_queue_pop(dma);
 
@@ -812,7 +821,7 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
                        S_v * sizeof(float), S_v);
 
         // Prefetch next block (if any)
-        if (ir_prefetch < total_rows) {
+        if (ir_prefetch < row_end) {
             const uint32_t piv1 = fastmodulo(ir_prefetch, H, &fd_H);
             const uint32_t piv3 = fastdiv(ir_prefetch, &fd_H);
             const float * ps_in = state_in_base + (uint64_t) piv3 * state_seq_stride + (uint64_t) piv1 * S_v * S_v;
@@ -828,6 +837,7 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
         curr_spad_idx ^= 1;
     }
     dma_queue_flush(dma);
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) row_end);
 }
 
 
@@ -847,8 +857,9 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
     const uint32_t H        = v->ne[1];
     const uint32_t n_seqs   = v->ne[3];
 
-    const uint32_t total_rows = H * n_seqs;
-    if (ith >= total_rows) {
+    const uint32_t row_end = gctx->row_start + gctx->nrows;
+
+    if (ith >= gctx->nrows) {
         return;
     }
 
@@ -881,11 +892,11 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
 
     const uint64_t state_seq_stride = state->nb[3] / sizeof(float);
 
-    uint32_t ir_prefetch = ith;
+    uint32_t ir_prefetch = gctx->row_start + ith;
     int spad_idx = 0;
 
     // Prefetch preamble (up to 2 steps)
-    for (int k = 0; k < 2 && ir_prefetch < total_rows; k++) {
+    for (int k = 0; k < 2 && ir_prefetch < row_end; k++) {
         const uint32_t piv1 = fastmodulo(ir_prefetch, H, &fd_H);
         const uint32_t piv3 = fastdiv(ir_prefetch, &fd_H);
         const float * ps_in = state_in_base + (uint64_t) piv3 * state_seq_stride + (uint64_t) piv1 * S_v * S_v;
@@ -906,8 +917,11 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
         spad_idx ^= 1;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (gctx->row_start + ith));
+
     int curr_spad_idx = 0;
-    for (uint32_t ir = ith; ir < total_rows; ir += nth) {
+    for (uint32_t ir = gctx->row_start + ith; ir < row_end; ir += nth) {
         dma_queue_pop(dma);
         dma_queue_pop(dma);
 
@@ -1057,7 +1071,7 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
                        S_v * sizeof(float), S_v);
 
         // Prefetch next block (if any)
-        if (ir_prefetch < total_rows) {
+        if (ir_prefetch < row_end) {
             const uint32_t piv1 = fastmodulo(ir_prefetch, H, &fd_H);
             const uint32_t piv3 = fastdiv(ir_prefetch, &fd_H);
             const float * ps_in = state_in_base + (uint64_t) piv3 * state_seq_stride + (uint64_t) piv1 * S_v * S_v;
@@ -1073,6 +1087,7 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
         curr_spad_idx ^= 1;
     }
     dma_queue_flush(dma);
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) row_end);
 }
 
 
@@ -1085,10 +1100,6 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
     const struct htp_tensor * state = octx->src[5];
     const struct htp_tensor * dst   = octx->dst;
 
-    if (!q || !k || !v || !g || !beta || !state || !dst) {
-        return HTP_STATUS_INVAL_PARAMS;
-    }
-
     if (q->type != HTP_TYPE_F32 || k->type != HTP_TYPE_F32 || v->type != HTP_TYPE_F32 ||
         g->type != HTP_TYPE_F32 || beta->type != HTP_TYPE_F32 || state->type != HTP_TYPE_F32 ||
         dst->type != HTP_TYPE_F32) {
@@ -1124,16 +1135,37 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
+    const uint32_t total_rows = H * n_seqs;
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = total_rows;
+
+    if (octx->ctx->mdev.count > 1) {
+        const uint32_t head_bytes = S_v * sizeof(float);
+        const uint32_t rows_per_chunk = (head_bytes > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(head_bytes, HEX_L2_LINE_SIZE)) : 1;
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, htp_tensor_mdev_data_aligned(dst) ? rows_per_chunk : 0,
+                                                                             octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
     struct htp_gdn_context gctx;
     gctx.octx = octx;
-    gctx.rows_per_thread = (H * n_seqs + octx->n_threads - 1) / octx->n_threads;
+    gctx.row_start       = row_start;
+    gctx.nrows           = nrows;
+    gctx.rows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
     gctx.state_bytes = (size_t) S_v * S_v * sizeof(float);
 
     size_t state_aligned = (size_t) S_v * S_v * sizeof(float);
     state_aligned = (state_aligned + 127) & ~(size_t)127;
 
-    assert(octx->ctx->vtcm_base != NULL);
-    assert(octx->ctx->vtcm_size >= 2 * state_aligned * octx->n_threads);
+    assert(octx->ctx->vtcm_size >= 2 * state_aligned * n_threads);
 
     gctx.vtcm_base = octx->ctx->vtcm_base;
     gctx.vtcm_per_thread = 2 * state_aligned;
@@ -1148,9 +1180,9 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
          gctx.vtcm_per_thread * octx->n_threads, octx->n_threads);
 
     if (n_tokens == 1) {
-        worker_pool_run_func(octx->ctx->worker_pool, gated_delta_net_f32_tg_thread, &gctx, octx->n_threads);
+        work_queue_run(octx->ctx->work_queue, gated_delta_net_f32_tg_thread, &gctx, n_threads);
     } else {
-        worker_pool_run_func(octx->ctx->worker_pool, gated_delta_net_f32_pp_thread, &gctx, octx->n_threads);
+        work_queue_run(octx->ctx->work_queue, gated_delta_net_f32_pp_thread, &gctx, n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/get-rows-ops.c

@@ -10,6 +10,7 @@
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
+#include "hex-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
 #include "htp-tensor.h"
@@ -23,9 +24,12 @@ struct get_rows_context {
     const struct htp_get_rows_kernel_params * kparams;
     struct htp_get_rows_vtcm_layout vtcm_layout;
     uint8_t * vtcm_base;
+    uint32_t task_start;
+    uint32_t tasks;
+    uint32_t tasks_per_thread;
 };
 
-#define get_rows_preamble \
+#define get_rows_preamble                      \
     const uint32_t ne00 = octx->src[0]->ne[0]; \
     const uint32_t ne01 = octx->src[0]->ne[1]; \
     const uint32_t ne02 = octx->src[0]->ne[2]; \
@@ -61,12 +65,12 @@ static void get_rows_thread_st_##IDX_TYPE(unsigned int nth, unsigned int ith, vo
     struct htp_ops_context * octx = grctx->octx;                                                                       \
     const struct htp_get_rows_kernel_params * kparams = grctx->kparams;                                                \
     get_rows_preamble;                                                                                                 \
-    const uint32_t dr  = kparams->tasks_per_thread;                                                                    \
-    const uint32_t ir0 = dr * ith;                                                                                     \
-    if (ir0 >= kparams->total_tasks) {                                                                                 \
+    const uint32_t dr  = grctx->tasks_per_thread;                                                                      \
+    const uint32_t ir0 = grctx->task_start + dr * ith;                                                                 \
+    if (ir0 >= grctx->task_start + grctx->tasks) {                                                                     \
         return;                                                                                                        \
     }                                                                                                                  \
-    const uint32_t ir1 = MIN(ir0 + dr, kparams->total_tasks);                                                          \
+    const uint32_t ir1 = MIN(ir0 + dr, grctx->task_start + grctx->tasks);                                              \
     const uint32_t row_size_bytes = htp_tensor_get_row_size(octx->src[0]->type, ne00);                                 \
     dma_queue * dma_queue = octx->ctx->dma[ith];                                                                       \
     for (uint32_t i = ir0; i < ir1; ++i) {                                                                             \
@@ -101,12 +105,12 @@ static void get_rows_thread_##TYPE_NAME##_##IDX_TYPE(unsigned int nth, unsigned
     const struct htp_get_rows_kernel_params * kparams = grctx->kparams;                                                \
     get_rows_preamble;                                                                                                 \
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                             \
-    const uint32_t dr  = kparams->tasks_per_thread;                                                                    \
-    const uint32_t ir0 = dr * ith;                                                                                     \
-    if (ir0 >= kparams->total_tasks) {                                                                                 \
+    const uint32_t dr  = grctx->tasks_per_thread;                                                                      \
+    const uint32_t ir0 = grctx->task_start + dr * ith;                                                                 \
+    if (ir0 >= grctx->task_start + grctx->tasks) {                                                                     \
         return;                                                                                                        \
     }                                                                                                                  \
-    const uint32_t ir1 = MIN(ir0 + dr, kparams->total_tasks);                                                          \
+    const uint32_t ir1 = MIN(ir0 + dr, grctx->task_start + grctx->tasks);                                              \
     const uint32_t chunks_per_row = kparams->chunks_per_row;                                                           \
     const uint32_t chunk_size     = kparams->chunk_size;                                                               \
     dma_queue * dma_queue = octx->ctx->dma[ith];                                                                       \
@@ -225,13 +229,41 @@ int op_get_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
+    const struct htp_tensor * dst = octx->dst;
+    const uint32_t total_tasks    = kparams->total_tasks;
+    const size_t dst_row_size     = htp_tensor_get_row_size(dst->type, dst->ne[0]);
+
+    uint32_t task_start = 0;
+    uint32_t tasks      = total_tasks;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t tasks_per_chunk = 1;
+        htp_tensor_mdev_rows_per_chunk(dst, dst_row_size / dst->ne[0], (uint32_t) dst_row_size, &tasks_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_tasks, tasks_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        task_start = range.start;
+        tasks      = range.count;
+    }
+
+    if (tasks == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    if (!htp_ops_context_set_n_threads(octx, (uint32_t) kparams->n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
     struct get_rows_context grctx;
     grctx.octx = octx;
     grctx.kparams = kparams;
     grctx.vtcm_base = (uint8_t *)octx->ctx->vtcm_base;
+    grctx.task_start = task_start;
+    grctx.tasks = tasks;
+    grctx.tasks_per_thread = fastdiv(tasks + n_threads - 1, &octx->n_threads_div);
 
     const uint32_t ne00 = octx->src[0]->ne[0];
-    htp_get_rows_vtcm_layout_build(&grctx.vtcm_layout, octx->src[0]->type, ne00, kparams->n_threads);
+    htp_get_rows_vtcm_layout_build(&grctx.vtcm_layout, octx->src[0]->type, ne00, n_threads);
 
     const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
 
@@ -247,14 +279,14 @@ int op_get_rows(struct htp_ops_context * octx) {
         }
     }
 
-    FARF(HIGH, "get-rows: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-vtcm-size %zu dst-vtcm-size %zu use_dma=%d n_threads %d\n",
+    FARF(HIGH, "get-rows: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-vtcm-size %zu dst-vtcm-size %zu use-dma %d n-threads %d\n",
          octx->src[0]->ne[0], octx->src[0]->ne[1], octx->src[0]->ne[2], octx->src[0]->ne[3],
          octx->src[1]->ne[0], octx->src[1]->ne[1], octx->src[1]->ne[2], octx->src[1]->ne[3],
          octx->dst->ne[0], octx->dst->ne[1], octx->dst->ne[2], octx->dst->ne[3],
-         grctx.vtcm_layout.src0_bytes_per_thread * kparams->n_threads,
-         grctx.vtcm_layout.dst_bytes_per_thread  * kparams->n_threads,
-         kparams->use_dma, kparams->n_threads);
+         grctx.vtcm_layout.src0_bytes_per_thread * n_threads,
+         grctx.vtcm_layout.dst_bytes_per_thread  * n_threads,
+         kparams->use_dma, n_threads);
 
-    work_queue_run(octx->ctx->work_queue, q_func, &grctx, kparams->n_threads);
+    work_queue_run(octx->ctx->work_queue, q_func, &grctx, n_threads);
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/hex-common.h

@@ -77,4 +77,13 @@ static inline bool hex_add_overflow(size_t a, size_t b, size_t *out) {
     return false;
 }
 
+static inline uint32_t hex_gcd_u32(uint32_t a, uint32_t b) {
+    while (b != 0) {
+        uint32_t t = b;
+        b = a % b;
+        a = t;
+    }
+    return a;
+}
+
 #endif // HEX_COMMON_H

ggml/src/ggml-hexagon/htp/hex-utils.h

@@ -39,7 +39,6 @@ static inline void hex_l2fetch_block(const void * addr, size_t size) {
 
 #define HEX_L2_LINE_SIZE           128
 #define HEX_L2_BLOCK_SIZE          (HEX_L2_LINE_SIZE * 4) // flush granularity (lines per loop iteration)
-#define HEX_L2_FLUSH_IL_THRESHOLD  1024                   // inline flush threshold
 #define HEX_L2_FLUSH_WQ_THRESHOLD  (4 * 1024)
 #define HEX_L2_FLUSH_ALL_THRESHOLD (4 * 1024 * 1024)
 

ggml/src/ggml-hexagon/htp/hmx-utils.h

@@ -27,7 +27,7 @@ static inline void hmx_init_column_scales(void *out_scales, HVX_Vector v_scale)
 // vscatter offsets for fused dequant+transpose: write K-values directly to [K][N] tile.
 // word[i] = i*128 maps K-row-pair i to byte offset i*128.
 // Column offset (n*4) is added at runtime.  Entries 0..15 cover one tile (region 2047);
-// entries 16..31 cover the next adjacent tile (region 4095) — pick region size at the
+// entries 16..31 cover the next adjacent tile (region 4095) - pick region size at the
 // call site to scatter into one tile (masked) or two contiguous tiles (unmasked).
 static const int32_t hmx_transpose_scatter_offsets[32] __attribute__((aligned(VLEN))) = {
     0 * 128,  1 * 128,  2 * 128,  3 * 128,  4 * 128,  5 * 128,  6 * 128,  7 * 128,  8 * 128,  9 * 128,  10 * 128,
@@ -198,16 +198,16 @@ static inline void hmx_interleave_cols_to_tiles(__fp16 * restrict tiles_out,
 }
 
 // --- HMX inline asm macros for load-store packetization ---
-#define HMX_LOAD_MPY_F16(act, wt, range) \
-    "{\n" \
+#define HMX_LOAD_MPY_F16(act, wt, range)              \
+    "{\n"                                             \
     "    activation.hf = mxmem(" act ", " range ")\n" \
-    "    weight.hf = mxmem(" wt ", " range ")\n" \
+    "    weight.hf = mxmem(" wt ", " range ")\n"      \
     "}\n"
 
-#define HMX_LOAD_MPY_DEEP_F16(act, wt, range) \
-    "{\n" \
+#define HMX_LOAD_MPY_DEEP_F16(act, wt, range)              \
+    "{\n"                                                  \
     "    activation.hf = mxmem(" act ", " range "):deep\n" \
-    "    weight.hf = mxmem(" wt ", " range ")\n" \
+    "    weight.hf = mxmem(" wt ", " range ")\n"           \
     "}\n"
 
 #define HMX_STORE_AFTER_F16(out, scale_reg) \

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -19,7 +19,7 @@
 #endif
 #define HTP_MAX_MMAPS    16
 
-#define HTP_MAX_DIRTY_RANGES 16
+#define HTP_MAX_DIRTY_RANGES 32
 
 // Memory mapping
 struct htp_mmap {
@@ -29,6 +29,11 @@ struct htp_mmap {
     uint32_t reserved;
 };
 
+struct htp_dirty_range {
+    uint32_t start;
+    uint32_t end;
+};
+
 // Scratchpad state
 struct htp_spad {
     const struct htp_tensor * src;             // original src of the data (for reuse)
@@ -38,6 +43,14 @@ struct htp_spad {
     uint32_t                  size_per_thread; // size per thread
 };
 
+struct htp_mdev_group {
+    uint16_t              idx;
+    uint16_t              count;
+    struct fastdiv_values count_div;
+    uint8_t *             fence_base;
+    uint32_t              fence_seq;
+};
+
 struct htp_context;
 
 // Context while processing an Op
@@ -65,8 +78,10 @@ struct htp_ops_context {
     struct htp_spad src3_spad;
     struct htp_spad dst_spad;
 
-    uint32_t n_threads;
-    uint32_t flags;
+    uint32_t              flags;
+    uint32_t              n_threads;
+    struct fastdiv_values n_threads_div;
+    int                   status;
 };
 
 // Main context for htp DSP backend
@@ -76,6 +91,7 @@ struct htp_context {
     struct htp_mmap        mmap[HTP_MAX_MMAPS];
     dma_queue_t            dma[HTP_MAX_NTHREADS];
     dma_queue_t            dma_cached[HTP_MAX_NTHREADS];
+    struct htp_thread_trace trace[HTP_MAX_NTHREADS + 1];
     work_queue_t           work_queue;
     hmx_queue_t            hmx_queue;
 
@@ -88,7 +104,6 @@ struct htp_context {
     bool                   hmx_enabled;
     bool                   etm;
     uint32_t               profiler;
-    struct htp_thread_trace trace[HTP_MAX_NTHREADS + 1];
 
     uint8_t *              vtcm_base;
     size_t                 vtcm_size;
@@ -97,16 +112,13 @@ struct htp_context {
     atomic_bool            vtcm_needs_release;
 
     uint64_t               max_vmem;
-    struct htp_dirty_range {
-        uint32_t start;
-        uint32_t end;
-        uint32_t bi;
-    } dirty_ranges[HTP_MAX_DIRTY_RANGES];
+    struct htp_dirty_range dirty_ranges[HTP_MAX_DIRTY_RANGES];
 
     // Persistent DDR scratchpad for MUL_MAT_ID mappings
     void *                 ddr_spad_base;
     size_t                 ddr_spad_size;
 
+    struct htp_mdev_group  mdev;
     struct htp_ops_context octx;
 
     qurt_thread_t          main_thread;
@@ -115,6 +127,27 @@ struct htp_context {
     size_t                 footprint;
 };
 
+static inline bool htp_ops_context_set_n_threads(struct htp_ops_context * octx, uint32_t n_threads) {
+    if (n_threads == 0 || n_threads > octx->ctx->n_threads) {
+        return false;
+    }
+
+    if (n_threads != octx->n_threads) {
+        octx->n_threads = n_threads;
+        octx->n_threads_div = n_threads == octx->ctx->n_threads
+            ? octx->ctx->n_threads_div
+            : init_fastdiv_values(n_threads);
+    }
+
+    return true;
+}
+
+static inline void htp_ops_context_set_status(struct htp_ops_context * octx, int status) {
+    if (status > HTP_STATUS_OK && octx->status == HTP_STATUS_OK) {
+        octx->status = status;
+    }
+}
+
 int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
 int op_matmul_nx(struct htp_ops_context * octx);

ggml/src/ggml-hexagon/htp/htp-fence.h

@@ -0,0 +1,89 @@
+#ifndef HTP_FENCE_H
+#define HTP_FENCE_H
+
+#include <stdatomic.h>
+#include <stdint.h>
+
+#include <HAP_farf.h>
+
+#include "hex-utils.h"
+#include "htp-ops.h"
+#include "htp-ctx.h"
+
+static inline atomic_uint * htp_mdev_fence_slot(const void * fence_base, uint32_t idx) {
+    return (atomic_uint *) ((const uint8_t *) fence_base + (size_t) idx * HTP_FENCE_SLOT_SIZE);
+}
+
+static inline void htp_fence_write(void * fence_ptr, uint32_t seq, uint32_t status) {
+    atomic_uint * fence = (atomic_uint *) fence_ptr;
+    atomic_store(&fence[1], status);
+    atomic_store(&fence[0], seq);
+    asm volatile ("syncht" : : : "memory");
+    Q6_dccleaninva_A((void *) fence);
+}
+
+static inline void htp_fence_read(const void * fence_ptr, uint32_t * seq, uint32_t * status) {
+    const atomic_uint * fence = (const atomic_uint *) fence_ptr;
+    Q6_dccleaninva_A((void *) fence);
+    asm volatile ("syncht" : : : "memory");
+    *seq = atomic_load(&fence[0]);
+    *status = atomic_load(&fence[1]);
+}
+
+static inline void htp_mdev_group_barrier(struct htp_ops_context * octx) {
+    struct htp_context * ctx = octx->ctx;
+    if (ctx->mdev.count <= 1) {
+        return;
+    }
+
+    const uint32_t seq = ++ctx->mdev.fence_seq;
+
+    struct htp_thread_trace * tr = &ctx->trace[0];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
+
+    const uint32_t mdev_idx   = ctx->mdev.idx;
+    const uint32_t mdev_count = ctx->mdev.count;
+
+    uint8_t * fence_base   = ctx->mdev.fence_base;
+    atomic_uint * my_fence = htp_mdev_fence_slot(fence_base, mdev_idx);
+    htp_fence_write(my_fence, seq, octx->status);
+
+    for (uint32_t d = 0; d < mdev_count; d++) {
+        if (d == mdev_idx) continue;
+        atomic_uint * peer_fence = htp_mdev_fence_slot(fence_base, d);
+        uint64_t spins = 0;
+        while (1) {
+            uint32_t peer_seq;
+            uint32_t peer_status;
+            htp_fence_read(peer_fence, &peer_seq, &peer_status);
+            if ((int32_t)(peer_seq - seq) >= 0) {
+                if (peer_status > HTP_STATUS_OK) {
+                    FARF(ERROR, "ggml-hex: mdev %u peer %u failed with status %u : seq 0x%08x\n",
+                         mdev_idx, d, peer_status, seq);
+                    htp_ops_context_set_status(octx, peer_status);
+                }
+                break;
+            }
+            if (++spins == 10000) {
+                FARF(ALWAYS, "ggml-hex: mdev %u waiting for mdev %u : seq 0x%08x (b %u op %u) my-fence %p peer-fence %p peer-seq 0x%08x (diff %d)\n",
+                     mdev_idx, d, seq, seq >> 12, seq & 0xfff, my_fence, peer_fence, peer_seq, (int32_t)(peer_seq - seq));
+            }
+            if (spins > HTP_FENCE_TIMEOUT) {
+                FARF(ERROR, "ggml-hex: mdev %u timeout waiting for mdev %u : seq 0x%08x (b %u op %u) peer-fence %p peer-seq 0x%08x\n",
+                     mdev_idx, d, seq, seq >> 12, seq & 0xfff, peer_fence, peer_seq);
+                htp_ops_context_set_status(octx, HTP_STATUS_INTERNAL_ERR);
+                break;
+            }
+            hex_pause();
+        }
+    }
+    asm volatile ("syncht" : : : "memory");
+
+    if (octx->status > HTP_STATUS_OK) {
+        htp_fence_write(my_fence, seq, octx->status);
+    }
+
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
+}
+
+#endif // HTP_FENCE_H

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -77,6 +77,7 @@ enum htp_op_code {
     HTP_OP_GET_ROWS,
     HTP_OP_SCALE,
     HTP_OP_CPY,
+    HTP_OP_CPY_FENCE,
     HTP_OP_ARGSORT,
     HTP_OP_SQR,
     HTP_OP_SQRT,
@@ -100,6 +101,7 @@ enum htp_op_code {
     HTP_OP_ALLREDUCE,
     HTP_OP_ALLREDUCE_ADD,
     HTP_OP_GLU_SWIGLU_CLAMP,
+    HTP_OP_MDEV_GROUP,
 
     HTP_OP_INVALID
 };
@@ -114,6 +116,7 @@ enum htp_op_code {
 #define HTP_OP_MAX_TENSORS 8192 // must stay under 64K (uint16)
 
 #define HTP_FENCE_TIMEOUT  (1000000000ULL)
+#define HTP_FENCE_SLOT_SIZE 128
 
 #define HTP_OP_MAX_VMEM_DEFAULT (3355443200u)
 
@@ -214,30 +217,26 @@ struct htp_prof_desc {
 };
 
 struct htp_opbatch_req {
-    uint32_t id;          // Batch id
+    uint64_t seq;         // Sequence number
     uint32_t n_bufs;      // Number of buffers
     uint32_t n_tensors;   // Number of tensors
     uint32_t n_ops;       // Number of ops
     uint32_t n_traces;    // Number of trace descriptors per thread
-    uint32_t pad;         // unused
-    uint64_t seq;         // Sequence number
     // struct htp_buf_desc  bufs[];    -- dspqueue buf 0
     // struct htp_tensor    tensors[]; -- dspqueue buf 0
     // struct htp_op_desc   ops[];     -- dspqueue buf 0
 };
 
 struct htp_opbatch_rsp {
-    uint32_t id;         // Batch id
-    uint32_t status;     // HTP_STATUS_...
-    uint32_t n_bufs;     // Number of buffers
-    uint32_t n_tensors;  // Number of tensors
-    uint32_t n_ops;      // Number of op profile descriptors
-    uint32_t n_traces[HTP_MAX_NTHREADS + 1];
-    uint32_t usecs;          // Number of usec
-    uint32_t pad;            // align to 8 bytes
+    uint64_t seq;            // Sequence number
     uint64_t cycles_start;   // Start cycle counter
     uint64_t cycles_stop;    // Stop cycle counter
-    uint64_t seq;            // Sequence number
+    uint32_t status;         // HTP_STATUS_...
+    uint32_t n_bufs;         // Number of buffers
+    uint32_t n_tensors;      // Number of tensors
+    uint32_t n_ops;          // Number of op profile descriptors
+    uint32_t usecs;          // Number of usec
+    uint32_t n_traces[HTP_MAX_NTHREADS + 1];
     // struct htp_prof_desc profs[];  -- dspqueue buf 0
 };
 

ggml/src/ggml-hexagon/htp/htp-tensor.c

@@ -20,7 +20,7 @@ struct l2flush_range {
 
 struct l2flush_multi_task {
     struct htp_thread_trace * trace;
-    struct l2flush_range      ranges[HTP_OP_MAX_INPUTS];
+    struct l2flush_range      ranges[HTP_MAX_DIRTY_RANGES];
     uint32_t                  n_ranges;
     uint32_t                  total_blocks;
     uint32_t                  blocks_per_thread;
@@ -73,6 +73,27 @@ static void l2flush_multi_worker(unsigned int n, unsigned int i, void * data) {
     htp_trace_event_stop(tr, HTP_TRACE_EVT_L2FLUSH, gb_first);
 }
 
+static void merge_dirty_ranges(struct htp_context * ctx) {
+    for (uint32_t i = 0; i < HTP_MAX_DIRTY_RANGES; i++) {
+        struct htp_dirty_range * r = &ctx->dirty_ranges[i];
+        if (!r->start) continue;
+
+        for (uint32_t j = 0; j < HTP_MAX_DIRTY_RANGES;) {
+            struct htp_dirty_range * s = &ctx->dirty_ranges[j];
+            if (i == j || !s->start || r->end < s->start || s->end < r->start) {
+                j++;
+                continue;
+            }
+
+            r->start = MIN(r->start, s->start);
+            r->end   = MAX(r->end, s->end);
+            s->start = 0;
+            s->end   = 0;
+            j = 0;
+        }
+    }
+}
+
 void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * const * tensors, uint32_t n) {
     const struct htp_tensor * pending[HTP_OP_MAX_OUTPUTS];
     uint32_t n_pending = 0;
@@ -83,11 +104,6 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
             continue;
         }
 
-        if (t->size <= HEX_L2_FLUSH_IL_THRESHOLD) {
-            hex_l2flush((void *) (uintptr_t) t->data, t->size);
-            continue;
-        }
-
         uint32_t t_start = t->data;
         uint32_t t_end   = t_start + t->size;
 
@@ -110,6 +126,8 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
         }
     }
 
+    merge_dirty_ranges(ctx);
+
     if (n_pending == 0) {
         return;
     }
@@ -132,8 +150,8 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
             struct htp_dirty_range * r = &ctx->dirty_ranges[idx];
             r->start = pending[i]->data;
             r->end   = pending[i]->data + pending[i]->size;
-            r->bi    = pending[i]->bi;
         }
+        merge_dirty_ranges(ctx);
         return;
     }
 
@@ -151,12 +169,12 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
             struct htp_dirty_range * r = &ctx->dirty_ranges[i];
             r->start = pending[i]->data;
             r->end   = pending[i]->data + pending[i]->size;
-            r->bi    = pending[i]->bi;
         }
+        merge_dirty_ranges(ctx);
         return;
     }
 
-    if (total_evict_size > HEX_L2_FLUSH_WQ_THRESHOLD && ctx->n_threads > 1 && n_evict <= HTP_OP_MAX_INPUTS) {
+    if (total_evict_size > HEX_L2_FLUSH_WQ_THRESHOLD && ctx->n_threads > 1 && n_evict <= HTP_MAX_DIRTY_RANGES) {
         struct l2flush_multi_task task;
         task.trace    = ctx->trace;
         task.n_ranges = n_evict;
@@ -195,7 +213,6 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
         struct htp_dirty_range * r = &ctx->dirty_ranges[idx];
         r->start = pending[i]->data;
         r->end   = pending[i]->data + pending[i]->size;
-        r->bi    = pending[i]->bi;
     }
 
     for (uint32_t i = 0; i < n_empty; i++) {
@@ -203,8 +220,9 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
         struct htp_dirty_range * r = &ctx->dirty_ranges[idx];
         r->start = pending[n_evict + i]->data;
         r->end   = pending[n_evict + i]->data + pending[n_evict + i]->size;
-        r->bi    = pending[n_evict + i]->bi;
     }
+
+    merge_dirty_ranges(ctx);
 }
 
 static void make_tensor_clean(struct htp_context * ctx, const struct htp_tensor * t) {
@@ -242,15 +260,77 @@ static inline bool is_tensor_dirty(struct htp_context * ctx, const struct htp_te
     return false;
 }
 
+static void flush_dirty_ranges(struct htp_context * ctx, const struct htp_dirty_range * ranges, uint32_t n_ranges, uint64_t total_dirty) {
+    if (total_dirty >= HEX_L2_FLUSH_WQ_THRESHOLD && ctx->n_threads > 1) {
+        struct l2flush_multi_task task;
+        task.trace    = ctx->trace;
+        task.n_ranges = n_ranges;
+
+        uint32_t block_acc = 0;
+        for (uint32_t i = 0; i < n_ranges; i++) {
+            const struct htp_dirty_range * r = &ranges[i];
+            struct l2flush_range * rg = &task.ranges[i];
+            rg->start = hex_align_down((size_t) r->start, HEX_L2_LINE_SIZE);
+            rg->end   = hex_align_up((size_t) r->end, HEX_L2_LINE_SIZE);
+            rg->block_first = block_acc;
+            rg->n_blocks = (rg->end - rg->start + HEX_L2_BLOCK_SIZE - 1) / HEX_L2_BLOCK_SIZE;
+            block_acc += rg->n_blocks;
+        }
+
+        task.total_blocks      = block_acc;
+        task.blocks_per_thread = fastdiv(block_acc + ctx->n_threads - 1, &ctx->n_threads_div);
+
+        work_queue_run(ctx->work_queue, l2flush_multi_worker, &task, ctx->n_threads);
+    } else {
+        struct htp_thread_trace * tr = &ctx->trace[0];
+        htp_trace_event_start(tr, HTP_TRACE_EVT_L2FLUSH, 0);
+        for (uint32_t i = 0; i < n_ranges; i++) {
+            const struct htp_dirty_range * r = &ranges[i];
+            hex_l2flush((void *) (uintptr_t) r->start, r->end - r->start);
+        }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_L2FLUSH, 0);
+    }
+}
+
+void htp_flush_dirty_ranges(struct htp_context * ctx) {
+    struct htp_dirty_range ranges[HTP_MAX_DIRTY_RANGES];
+    uint32_t n_ranges = 0;
+    uint64_t total_dirty = 0;
+
+    for (uint32_t i = 0; i < HTP_MAX_DIRTY_RANGES; i++) {
+        const struct htp_dirty_range * r = &ctx->dirty_ranges[i];
+        if (!r->start) {
+            continue;
+        }
+        ranges[n_ranges++] = *r;
+        total_dirty += r->end - r->start;
+    }
+
+    if (total_dirty == 0) {
+        return;
+    }
+
+    if (total_dirty > HEX_L2_FLUSH_ALL_THRESHOLD) {
+        flush_all_dcache(ctx);
+        return;
+    }
+
+    flush_dirty_ranges(ctx, ranges, n_ranges, total_dirty);
+    memset(ctx->dirty_ranges, 0, sizeof(ctx->dirty_ranges));
+}
+
 void htp_tensor_flush_all(struct htp_context * ctx, const struct htp_tensor * const * tensors, uint32_t n) {
     const struct htp_tensor * dirty_tensors[HTP_OP_MAX_INPUTS];
+    struct htp_dirty_range ranges[HTP_OP_MAX_INPUTS];
     uint32_t n_dirty = 0;
     uint64_t total_dirty = 0;
 
     for (uint32_t i = 0; i < n; i++) {
         const struct htp_tensor * t = tensors[i];
-        if (t && !(t->flags & (HTP_TENSOR_WEIGHT | HTP_TENSOR_FENCE)) && is_tensor_dirty(ctx, t)) {
+        if (t && is_tensor_dirty(ctx, t)) {
             dirty_tensors[n_dirty++] = t;
+            ranges[n_dirty - 1].start = t->data;
+            ranges[n_dirty - 1].end   = t->data + t->size;
             total_dirty += t->size;
         }
     }
@@ -264,37 +344,8 @@ void htp_tensor_flush_all(struct htp_context * ctx, const struct htp_tensor * co
         return;
     }
 
-    if (total_dirty >= HEX_L2_FLUSH_WQ_THRESHOLD && ctx->n_threads > 1) {
-        struct l2flush_multi_task task;
-        task.trace    = ctx->trace;
-        task.n_ranges = 0;
-
-        uint32_t block_acc = 0;
-        for (uint32_t i = 0; i < n_dirty; i++) {
-            const struct htp_tensor * t = dirty_tensors[i];
-            make_tensor_clean(ctx, t);
-
-            struct l2flush_range * rg = &task.ranges[task.n_ranges++];
-            rg->start = hex_align_down((size_t) t->data, HEX_L2_LINE_SIZE);
-            rg->end   = hex_align_up((size_t) t->data + t->size, HEX_L2_LINE_SIZE);
-            rg->block_first = block_acc;
-            rg->n_blocks = (rg->end - rg->start + HEX_L2_BLOCK_SIZE - 1) / HEX_L2_BLOCK_SIZE;
-            block_acc += rg->n_blocks;
-        }
-
-        task.total_blocks      = block_acc;
-        task.blocks_per_thread = fastdiv(block_acc + ctx->n_threads - 1, &ctx->n_threads_div);
-
-        work_queue_run(ctx->work_queue, l2flush_multi_worker, &task, ctx->n_threads);
-        return;
-    }
-
-    struct htp_thread_trace * tr = &ctx->trace[0];
+    flush_dirty_ranges(ctx, ranges, n_dirty, total_dirty);
     for (uint32_t i = 0; i < n_dirty; i++) {
-        const struct htp_tensor * t = dirty_tensors[i];
-        htp_trace_event_start(tr, HTP_TRACE_EVT_L2FLUSH, t->ti);
-        hex_l2flush((void *) (uintptr_t) t->data, t->size);
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_L2FLUSH, t->ti);
-        make_tensor_clean(ctx, t);
+        make_tensor_clean(ctx, dirty_tensors[i]);
     }
 }

ggml/src/ggml-hexagon/htp/htp-tensor.h

@@ -2,8 +2,20 @@
 #define HTP_TENSOR_H
 
 #include <stdint.h>
+#include <stdbool.h>
 #include "htp-ops.h"
 #include "hex-bitmap.h"
+#include "hex-common.h"
+#include "hex-fastdiv.h"
+
+enum {
+    HTP_TENSOR_MDEV_LINE_SIZE = 128,
+};
+
+struct htp_tensor_mdev_range {
+    uint32_t start;
+    uint32_t count;
+};
 
 static inline void * htp_tensor_data(const struct htp_tensor * t) {
     return (void *) (uintptr_t) t->data;
@@ -13,6 +25,102 @@ static inline uint32_t * htp_tensor_flags(const struct htp_tensor * t) {
     return (uint32_t *) &t->flags;
 }
 
+static inline bool htp_tensor_is_contiguous(const struct htp_tensor * t, uint32_t type_size) {
+    uint32_t next_nb = type_size;
+    if (t->ne[0] != 1 && t->nb[0] != next_nb) {
+        return false;
+    }
+    next_nb *= t->ne[0];
+    for (int i = 1; i < HTP_OP_MAX_DIMS; i++) {
+        if (t->ne[i] != 1 && t->nb[i] != next_nb) {
+            return false;
+        }
+        next_nb *= t->ne[i];
+    }
+    return true;
+}
+
+static inline bool htp_tensor_is_permuted(const struct htp_tensor * t) {
+    return t->nb[0] > t->nb[1] || t->nb[1] > t->nb[2] || t->nb[2] > t->nb[3];
+}
+
+static inline bool htp_tensor_mdev_data_aligned(const struct htp_tensor * t) {
+    return ((uintptr_t) t->data & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0;
+}
+
+static inline bool htp_tensor_can_row_partition(const struct htp_tensor * t, uint32_t elem_size) {
+    if (!htp_tensor_mdev_data_aligned(t)) {
+        return false;
+    }
+    if (t->ne[0] != 1 && t->nb[0] != elem_size) {
+        return false;
+    }
+    if (htp_tensor_is_permuted(t)) {
+        return false;
+    }
+    if (t->ne[1] > 1 && (t->nb[1] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) != 0) return false;
+    if (t->ne[2] > 1 && (t->nb[2] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) != 0) return false;
+    if (t->ne[3] > 1 && (t->nb[3] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) != 0) return false;
+    return true;
+}
+
+static inline bool htp_tensor_mdev_rows_per_chunk(const struct htp_tensor * t, uint32_t elem_size, uint32_t row_size, uint32_t * rows_per_chunk) {
+    *rows_per_chunk = 0;
+
+    if (!htp_tensor_mdev_data_aligned(t)) {
+        return false;
+    }
+    if (t->ne[0] != 1 && t->nb[0] != elem_size) {
+        return false;
+    }
+    if (htp_tensor_is_permuted(t)) {
+        return false;
+    }
+    if (t->ne[1] > 1 && (t->nb[1] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0 &&
+        (t->ne[2] <= 1 || (t->nb[2] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0) &&
+        (t->ne[3] <= 1 || (t->nb[3] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0)) {
+        *rows_per_chunk = 1;
+        return true;
+    }
+    if (t->nb[1] == row_size &&
+        (t->ne[2] <= 1 || t->nb[2] == t->nb[1] * t->ne[1]) &&
+        (t->ne[3] <= 1 || t->nb[3] == t->nb[2] * t->ne[2])) {
+        *rows_per_chunk = (row_size > 0) ? (HTP_TENSOR_MDEV_LINE_SIZE / hex_gcd_u32(row_size, HTP_TENSOR_MDEV_LINE_SIZE)) : 1;
+        return true;
+    }
+    return false;
+}
+
+static inline struct htp_tensor_mdev_range htp_tensor_mdev_partition(uint32_t total_units, uint32_t units_per_chunk, uint32_t mdev_idx, uint32_t mdev_count, const struct fastdiv_values * mdev_count_div) {
+    struct htp_tensor_mdev_range range = { 0, total_units };
+
+    if (mdev_count <= 1) {
+        return range;
+    }
+
+    if (units_per_chunk == 0) {
+        range.start = (mdev_idx == 0) ? 0 : total_units;
+        range.count = (mdev_idx == 0) ? total_units : 0;
+        return range;
+    }
+
+    const uint32_t total_chunks = total_units / units_per_chunk;
+    if (total_chunks < mdev_count) {
+        range.start = (mdev_idx == 0) ? 0 : total_units;
+        range.count = (mdev_idx == 0) ? total_units : 0;
+        return range;
+    }
+
+    const uint32_t chunks_per_mdev = fastdiv(total_chunks + mdev_count - 1, mdev_count_div);
+    range.start = MIN(mdev_idx * chunks_per_mdev * units_per_chunk, total_units);
+    if (mdev_idx == mdev_count - 1) {
+        range.count = total_units - range.start;
+    } else {
+        range.count = MIN(chunks_per_mdev * units_per_chunk, total_units - range.start);
+    }
+    return range;
+}
+
 static inline uint32_t htp_tensor_get_row_size(int type, uint32_t ne00) {
     switch (type) {
         case HTP_TYPE_F32:  return ne00 * 4;
@@ -23,6 +131,7 @@ static inline uint32_t htp_tensor_get_row_size(int type, uint32_t ne00) {
 }
 
 struct htp_context;
+void htp_flush_dirty_ranges(struct htp_context * ctx);
 void htp_tensor_flush_all(struct htp_context * ctx, const struct htp_tensor * const * tensors, uint32_t n);
 void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * const * tensors, uint32_t n);
 

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -16,25 +16,25 @@
 #define UNUSED(x) (void)(x)
 
 #define hvx_arith_loop_body(dst_type, src0_type, src1_type, elem_size, vec_store, vec_op) \
-    do {                                                                       \
-        dst_type * vdst  = (dst_type *) dst;                                   \
-        src0_type * vsrc0 = (src0_type *) src0;                                \
-        src1_type * vsrc1 = (src1_type *) src1;                                \
-                                                                               \
-        const uint32_t epv  = 128 / (elem_size);                               \
-        const uint32_t nvec = n / epv;                                         \
-        const uint32_t nloe = n % epv;                                         \
-                                                                               \
-        uint32_t i = 0;                                                        \
-                                                                               \
-        _Pragma("unroll(4)")                                                   \
-        for (; i < nvec; i++) {                                                \
-            vdst[i] = vec_op(vsrc0[i], vsrc1[i]);                              \
-        }                                                                      \
-        if (nloe) {                                                            \
-            HVX_Vector v = vec_op(vsrc0[i], vsrc1[i]);                         \
-            vec_store((void *) &vdst[i], nloe * (elem_size), v);               \
-        }                                                                      \
+    do {                                                                                  \
+        dst_type * vdst  = (dst_type *) dst;                                              \
+        src0_type * vsrc0 = (src0_type *) src0;                                           \
+        src1_type * vsrc1 = (src1_type *) src1;                                           \
+                                                                                          \
+        const uint32_t epv  = 128 / (elem_size);                                          \
+        const uint32_t nvec = n / epv;                                                    \
+        const uint32_t nloe = n % epv;                                                    \
+                                                                                          \
+        uint32_t i = 0;                                                                   \
+                                                                                          \
+        _Pragma("unroll(4)")                                                              \
+        for (; i < nvec; i++) {                                                           \
+            vdst[i] = vec_op(vsrc0[i], vsrc1[i]);                                         \
+        }                                                                                 \
+        if (nloe) {                                                                       \
+            HVX_Vector v = vec_op(vsrc0[i], vsrc1[i]);                                    \
+            vec_store((void *) &vdst[i], nloe * (elem_size), v);                          \
+        }                                                                                 \
     } while(0)
 
 #if __HVX_ARCH__ < 79
@@ -56,43 +56,43 @@
 #define HVX_OP_MUL_F16(a, b) hvx_vec_mul_f16_f16(a, b)
 
 // Generic macro to define alignment permutations for an op
-#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO, ELEM_TYPE) \
-static inline void OP_NAME##_aaa(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src0 % 128 == 0); \
-    assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
-} \
-static inline void OP_NAME##_aau(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src0 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
-} \
-static inline void OP_NAME##_aua(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
-} \
-static inline void OP_NAME##_auu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
-} \
-static inline void OP_NAME##_uaa(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) src0 % 128 == 0); \
-    assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
-} \
-static inline void OP_NAME##_uau(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) src0 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
-} \
-static inline void OP_NAME##_uua(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
-    assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
-} \
-static inline void OP_NAME##_uuu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO, ELEM_TYPE)                                           \
+static inline void OP_NAME##_aaa(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) dst % 128 == 0);                                                                       \
+    assert((uintptr_t) src0 % 128 == 0);                                                                      \
+    assert((uintptr_t) src1 % 128 == 0);                                                                      \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO);    \
+}                                                                                                             \
+static inline void OP_NAME##_aau(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) dst % 128 == 0);                                                                       \
+    assert((uintptr_t) src0 % 128 == 0);                                                                      \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO);   \
+}                                                                                                             \
+static inline void OP_NAME##_aua(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) dst % 128 == 0);                                                                       \
+    assert((uintptr_t) src1 % 128 == 0);                                                                      \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO);   \
+}                                                                                                             \
+static inline void OP_NAME##_auu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) dst % 128 == 0);                                                                       \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO);  \
+}                                                                                                             \
+static inline void OP_NAME##_uaa(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) src0 % 128 == 0);                                                                      \
+    assert((uintptr_t) src1 % 128 == 0);                                                                      \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO);   \
+}                                                                                                             \
+static inline void OP_NAME##_uau(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) src0 % 128 == 0);                                                                      \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO);  \
+}                                                                                                             \
+static inline void OP_NAME##_uua(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
+    assert((uintptr_t) src1 % 128 == 0);                                                                      \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO);  \
+}                                                                                                             \
+static inline void OP_NAME##_uuu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) {     \
     hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
-} \
+}                                                                                                             \
 
 DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD_F32, float)
 DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB_F32, float)
@@ -103,25 +103,25 @@ DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f16, HVX_OP_SUB_F16, _Float16)
 DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f16, HVX_OP_MUL_F16, _Float16)
 
 // Dispatcher logic
-#define HVX_BINARY_DISPATCHER(OP_NAME) \
+#define HVX_BINARY_DISPATCHER(OP_NAME)                                                                                                       \
 static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
-    if (hex_is_aligned((void *) dst, 128)) { \
-        if (hex_is_aligned((void *) src0, 128)) { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
-        } else { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
-        } \
-    } else { \
-        if (hex_is_aligned((void *) src0, 128)) { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
-        } else { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
-        } \
-    } \
+    if (hex_is_aligned((void *) dst, 128)) {                                                                                                 \
+        if (hex_is_aligned((void *) src0, 128)) {                                                                                            \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems);                                               \
+        } else {                                                                                                                             \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems);                                               \
+        }                                                                                                                                    \
+    } else {                                                                                                                                 \
+        if (hex_is_aligned((void *) src0, 128)) {                                                                                            \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems);                                               \
+        } else {                                                                                                                             \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems);                                               \
+        }                                                                                                                                    \
+    }                                                                                                                                        \
 }
 
 HVX_BINARY_DISPATCHER(hvx_add_f32)
@@ -166,44 +166,44 @@ static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * re
 
 // Scalar Operations
 
-#define hvx_scalar_loop_body(dst_type, src_type, elem_size, vec_store, scalar_op_macro)   \
-    do {                                                                       \
-        dst_type * restrict vdst = (dst_type *) dst;                           \
-        src_type * restrict vsrc = (src_type *) src;                           \
-                                                                               \
-        const uint32_t epv  = 128 / (elem_size);                               \
-        const uint32_t nvec = n / epv;                                         \
-        const uint32_t nloe = n % epv;                                         \
-                                                                               \
-        uint32_t i = 0;                                                        \
-                                                                               \
-        _Pragma("unroll(4)")                                                   \
-        for (; i < nvec; i++) {                                                \
-            HVX_Vector v = vsrc[i];                                            \
-            vdst[i] = scalar_op_macro(v);                                      \
-        }                                                                      \
-        if (nloe) {                                                            \
-            HVX_Vector v = vsrc[i];                                            \
-            v = scalar_op_macro(v);                                            \
-            vec_store((void *) &vdst[i], nloe * (elem_size), v);               \
-        }                                                                      \
+#define hvx_scalar_loop_body(dst_type, src_type, elem_size, vec_store, scalar_op_macro) \
+    do {                                                                                \
+        dst_type * restrict vdst = (dst_type *) dst;                                    \
+        src_type * restrict vsrc = (src_type *) src;                                    \
+                                                                                        \
+        const uint32_t epv  = 128 / (elem_size);                                        \
+        const uint32_t nvec = n / epv;                                                  \
+        const uint32_t nloe = n % epv;                                                  \
+                                                                                        \
+        uint32_t i = 0;                                                                 \
+                                                                                        \
+        _Pragma("unroll(4)")                                                            \
+        for (; i < nvec; i++) {                                                         \
+            HVX_Vector v = vsrc[i];                                                     \
+            vdst[i] = scalar_op_macro(v);                                               \
+        }                                                                               \
+        if (nloe) {                                                                     \
+            HVX_Vector v = vsrc[i];                                                     \
+            v = scalar_op_macro(v);                                                     \
+            vec_store((void *) &vdst[i], nloe * (elem_size), v);                        \
+        }                                                                               \
     } while(0)
 
-#define HVX_OP_ADD_SCALAR_F32(v) \
-    ({ \
+#define HVX_OP_ADD_SCALAR_F32(v)                                   \
+    ({                                                             \
         const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, v); \
-        HVX_Vector out = HVX_OP_ADD_F32(v, val_vec); \
-        Q6_V_vmux_QVV(pred_inf, inf, out); \
+        HVX_Vector out = HVX_OP_ADD_F32(v, val_vec);               \
+        Q6_V_vmux_QVV(pred_inf, inf, out);                         \
     })
 
 #define HVX_OP_MUL_SCALAR_F32(v) HVX_OP_MUL_F32(v, val_vec)
 #define HVX_OP_SUB_SCALAR_F32(v) HVX_OP_SUB_F32(v, val_vec)
 
-#define HVX_OP_ADD_SCALAR_F16(v) \
-    ({ \
+#define HVX_OP_ADD_SCALAR_F16(v)                                   \
+    ({                                                             \
         const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VhVh(inf, v); \
-        HVX_Vector out = HVX_OP_ADD_F16(v, val_vec); \
-        Q6_V_vmux_QVV(pred_inf, inf, out); \
+        HVX_Vector out = HVX_OP_ADD_F16(v, val_vec);               \
+        Q6_V_vmux_QVV(pred_inf, inf, out);                         \
     })
 
 #define HVX_OP_MUL_SCALAR_F16(v) HVX_OP_MUL_F16(v, val_vec)
@@ -212,31 +212,31 @@ static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * re
 // Scalar Variants
 
 // Generic macro to define alignment permutations for an op
-#define DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(OP_NAME, OP_MACRO, SPLAT_MACRO, ELEM_TYPE) \
+#define DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(OP_NAME, OP_MACRO, SPLAT_MACRO, ELEM_TYPE)                                  \
 static inline void OP_NAME##_aa(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
-    const HVX_Vector val_vec = SPLAT_MACRO(val); \
-    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src % 128 == 0); \
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
-} \
+    const HVX_Vector val_vec = SPLAT_MACRO(val);                                                                         \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf);                                                \
+    assert((uintptr_t) dst % 128 == 0);                                                                                  \
+    assert((uintptr_t) src % 128 == 0);                                                                                  \
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO);                          \
+}                                                                                                                        \
 static inline void OP_NAME##_au(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
-    const HVX_Vector val_vec = SPLAT_MACRO(val); \
-    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
-    assert((uintptr_t) dst % 128 == 0); \
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
-} \
+    const HVX_Vector val_vec = SPLAT_MACRO(val);                                                                         \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf);                                                \
+    assert((uintptr_t) dst % 128 == 0);                                                                                  \
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO);                         \
+}                                                                                                                        \
 static inline void OP_NAME##_ua(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
-    const HVX_Vector val_vec = SPLAT_MACRO(val); \
-    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
-    assert((uintptr_t) src % 128 == 0); \
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
-} \
+    const HVX_Vector val_vec = SPLAT_MACRO(val);                                                                         \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf);                                                \
+    assert((uintptr_t) src % 128 == 0);                                                                                  \
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO);                         \
+}                                                                                                                        \
 static inline void OP_NAME##_uu(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
-    const HVX_Vector val_vec = SPLAT_MACRO(val); \
-    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
-} \
+    const HVX_Vector val_vec = SPLAT_MACRO(val);                                                                         \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf);                                                \
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO);                        \
+}                                                                                                                        \
 
 DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_add_scalar_f32, HVX_OP_ADD_SCALAR_F32, hvx_vec_splat_f32, float)
 DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_sub_scalar_f32, HVX_OP_SUB_SCALAR_F32, hvx_vec_splat_f32, float)
@@ -247,17 +247,17 @@ DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_sub_scalar_f16, HVX_OP_SUB_SCALAR_F16,
 DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_mul_scalar_f16, HVX_OP_MUL_SCALAR_F16, hvx_vec_splat_f16, _Float16)
 
 // Dispatcher logic
-#define HVX_BINARY_SCALAR_DISPATCHER(OP_NAME, ELEM_TYPE) \
+#define HVX_BINARY_SCALAR_DISPATCHER(OP_NAME, ELEM_TYPE)                                                                          \
 static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, const uint32_t num_elems) { \
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) { \
-        OP_NAME##_aa(dst, src, val, num_elems); \
-    } else if (hex_is_aligned((void *) dst, 128)) { \
-        OP_NAME##_au(dst, src, val, num_elems); \
-    } else if (hex_is_aligned((void *) src, 128)) { \
-        OP_NAME##_ua(dst, src, val, num_elems); \
-    } else { \
-        OP_NAME##_uu(dst, src, val, num_elems); \
-    } \
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {                                                 \
+        OP_NAME##_aa(dst, src, val, num_elems);                                                                                   \
+    } else if (hex_is_aligned((void *) dst, 128)) {                                                                               \
+        OP_NAME##_au(dst, src, val, num_elems);                                                                                   \
+    } else if (hex_is_aligned((void *) src, 128)) {                                                                               \
+        OP_NAME##_ua(dst, src, val, num_elems);                                                                                   \
+    } else {                                                                                                                      \
+        OP_NAME##_uu(dst, src, val, num_elems);                                                                                   \
+    }                                                                                                                             \
 }
 
 HVX_BINARY_SCALAR_DISPATCHER(hvx_add_scalar_f32, float)
@@ -350,12 +350,12 @@ static inline void hvx_max_scalar_f32(uint8_t * restrict dst, const uint8_t * re
 
 // CLAMP Scalar variants
 
-#define HVX_OP_CLAMP_SCALAR(v) \
-    ({ \
+#define HVX_OP_CLAMP_SCALAR(v)                                           \
+    ({                                                                   \
         HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(v, max_vec); \
         HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(min_vec, v); \
-        HVX_Vector tmp = Q6_V_vmux_QVV(pred_cap_right, max_vec, v); \
-        Q6_V_vmux_QVV(pred_cap_left, min_vec, tmp); \
+        HVX_Vector tmp = Q6_V_vmux_QVV(pred_cap_right, max_vec, v);      \
+        Q6_V_vmux_QVV(pred_cap_left, min_vec, tmp);                      \
     })
 
 static inline void hvx_clamp_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {

ggml/src/ggml-hexagon/htp/hvx-div.h

@@ -219,64 +219,64 @@ static inline HVX_Vector hvx_vec_hybrid_div_f16(HVX_Vector vec1, HVX_Vector vec2
     } while(0)
 
 // Generic macro to define alignment permutations for an op
-#define DEFINE_HVX_DIV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY) \
+#define DEFINE_HVX_DIV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY)                                                                            \
 static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src0 % 128 == 0); \
-    assert((uintptr_t) src1 % 128 == 0); \
-    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a); \
-} \
+    assert((uintptr_t) dst % 128 == 0);                                                                                              \
+    assert((uintptr_t) src0 % 128 == 0);                                                                                             \
+    assert((uintptr_t) src1 % 128 == 0);                                                                                             \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);                                                               \
+}                                                                                                                                    \
 static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src0 % 128 == 0); \
-    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a); \
-} \
+    assert((uintptr_t) dst % 128 == 0);                                                                                              \
+    assert((uintptr_t) src0 % 128 == 0);                                                                                             \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);                                                              \
+}                                                                                                                                    \
 static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src1 % 128 == 0); \
-    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a); \
-} \
+    assert((uintptr_t) dst % 128 == 0);                                                                                              \
+    assert((uintptr_t) src1 % 128 == 0);                                                                                             \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);                                                              \
+}                                                                                                                                    \
 static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a); \
-} \
+    assert((uintptr_t) dst % 128 == 0);                                                                                              \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);                                                             \
+}                                                                                                                                    \
 static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) src0 % 128 == 0); \
-    assert((uintptr_t) src1 % 128 == 0); \
-    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u); \
-} \
+    assert((uintptr_t) src0 % 128 == 0);                                                                                             \
+    assert((uintptr_t) src1 % 128 == 0);                                                                                             \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);                                                              \
+}                                                                                                                                    \
 static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) src0 % 128 == 0); \
-    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u); \
-} \
+    assert((uintptr_t) src0 % 128 == 0);                                                                                             \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);                                                             \
+}                                                                                                                                    \
 static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    assert((uintptr_t) src1 % 128 == 0); \
-    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u); \
-} \
+    assert((uintptr_t) src1 % 128 == 0);                                                                                             \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);                                                             \
+}                                                                                                                                    \
 static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u); \
-} \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);                                                            \
+}                                                                                                                                    \
 
 // Dispatcher logic
-#define HVX_DIV_DISPATCHER(OP_NAME) \
+#define HVX_DIV_DISPATCHER(OP_NAME)                                                                                                          \
 static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
-    if (hex_is_aligned((void *) dst, 128)) { \
-        if (hex_is_aligned((void *) src0, 128)) { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
-        } else { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
-        } \
-    } else { \
-        if (hex_is_aligned((void *) src0, 128)) { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
-        } else { \
-            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
-            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
-        } \
-    } \
+    if (hex_is_aligned((void *) dst, 128)) {                                                                                                 \
+        if (hex_is_aligned((void *) src0, 128)) {                                                                                            \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems);                                               \
+        } else {                                                                                                                             \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems);                                               \
+        }                                                                                                                                    \
+    } else {                                                                                                                                 \
+        if (hex_is_aligned((void *) src0, 128)) {                                                                                            \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems);                                               \
+        } else {                                                                                                                             \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems);                                               \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems);                                               \
+        }                                                                                                                                    \
+    }                                                                                                                                        \
 }
 
 DEFINE_HVX_DIV_OP_VARIANTS(hvx_div_f32, hvx_div_f32_loop_body)

ggml/src/ggml-hexagon/htp/hvx-inverse.h

@@ -169,36 +169,36 @@ static inline HVX_Vector hvx_vec_inverse_f16_guard(HVX_Vector v_sf, HVX_Vector n
     } while(0)
 
 // Generic macro to define alignment permutations for an op
-#define DEFINE_HVX_INV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY) \
+#define DEFINE_HVX_INV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY)                                           \
 static inline void OP_NAME##_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    assert((uintptr_t) src % 128 == 0); \
-    OP_LOOP_BODY(HVX_Vector, HVX_Vector, hvx_vec_store_a); \
-} \
+    assert((uintptr_t) dst % 128 == 0);                                                             \
+    assert((uintptr_t) src % 128 == 0);                                                             \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, hvx_vec_store_a);                                          \
+}                                                                                                   \
 static inline void OP_NAME##_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
-    assert((uintptr_t) dst % 128 == 0); \
-    OP_LOOP_BODY(HVX_Vector, HVX_UVector, hvx_vec_store_a); \
-} \
+    assert((uintptr_t) dst % 128 == 0);                                                             \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, hvx_vec_store_a);                                         \
+}                                                                                                   \
 static inline void OP_NAME##_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
-    assert((uintptr_t) src % 128 == 0); \
-    OP_LOOP_BODY(HVX_UVector, HVX_Vector, hvx_vec_store_u); \
-} \
+    assert((uintptr_t) src % 128 == 0);                                                             \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, hvx_vec_store_u);                                         \
+}                                                                                                   \
 static inline void OP_NAME##_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
-    OP_LOOP_BODY(HVX_UVector, HVX_UVector, hvx_vec_store_u); \
-} \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, hvx_vec_store_u);                                        \
+}                                                                                                   \
 
 // Dispatcher logic
-#define HVX_INV_DISPATCHER(OP_NAME) \
+#define HVX_INV_DISPATCHER(OP_NAME)                                                                          \
 static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) { \
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) { \
-        OP_NAME##_aa(dst, src, num_elems); \
-    } else if (hex_is_aligned((void *) dst, 128)) { \
-        OP_NAME##_au(dst, src, num_elems); \
-    } else if (hex_is_aligned((void *) src, 128)) { \
-        OP_NAME##_ua(dst, src, num_elems); \
-    } else { \
-        OP_NAME##_uu(dst, src, num_elems); \
-    } \
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {                            \
+        OP_NAME##_aa(dst, src, num_elems);                                                                   \
+    } else if (hex_is_aligned((void *) dst, 128)) {                                                          \
+        OP_NAME##_au(dst, src, num_elems);                                                                   \
+    } else if (hex_is_aligned((void *) src, 128)) {                                                          \
+        OP_NAME##_ua(dst, src, num_elems);                                                                   \
+    } else {                                                                                                 \
+        OP_NAME##_uu(dst, src, num_elems);                                                                   \
+    }                                                                                                        \
 }
 
 DEFINE_HVX_INV_OP_VARIANTS(hvx_inverse_f32, hvx_inverse_f32_loop_body)

ggml/src/ggml-hexagon/htp/hvx-scale.h

@@ -68,30 +68,30 @@ static inline void hvx_scale_f32(uint8_t * restrict dst, const uint8_t * restric
     }
 }
 
-#define hvx_scale_offset_f32_loop_body(dst_type, src_type, vec_store)                \
-    do {                                                                             \
-        dst_type * restrict vdst = (dst_type *) dst;                                 \
-        src_type * restrict vsrc = (src_type *) src;                                 \
-                                                                                     \
-        HVX_Vector vs = hvx_vec_splat_f32(scale);                                    \
-        HVX_Vector vo = hvx_vec_splat_f32(offset);                                   \
-                                                                                     \
-        const uint32_t elem_size = sizeof(float);                                    \
-        const uint32_t epv = 128 / elem_size;                                        \
-        const uint32_t nvec = n / epv;                                               \
-        const uint32_t nloe = n % epv;                                               \
-                                                                                     \
-        uint32_t i = 0;                                                              \
-                                                                                     \
-        _Pragma("unroll(4)")                                                         \
-        for (; i < nvec; ++i) {                                                      \
+#define hvx_scale_offset_f32_loop_body(dst_type, src_type, vec_store)                     \
+    do {                                                                                  \
+        dst_type * restrict vdst = (dst_type *) dst;                                      \
+        src_type * restrict vsrc = (src_type *) src;                                      \
+                                                                                          \
+        HVX_Vector vs = hvx_vec_splat_f32(scale);                                         \
+        HVX_Vector vo = hvx_vec_splat_f32(offset);                                        \
+                                                                                          \
+        const uint32_t elem_size = sizeof(float);                                         \
+        const uint32_t epv = 128 / elem_size;                                             \
+        const uint32_t nvec = n / epv;                                                    \
+        const uint32_t nloe = n % epv;                                                    \
+                                                                                          \
+        uint32_t i = 0;                                                                   \
+                                                                                          \
+        _Pragma("unroll(4)")                                                              \
+        for (; i < nvec; ++i) {                                                           \
             HVX_Vector v = Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs), vo); \
-            vdst[i] = Q6_Vsf_equals_Vqf32(v);                                        \
-        }                                                                            \
-        if (nloe) {                                                                  \
+            vdst[i] = Q6_Vsf_equals_Vqf32(v);                                             \
+        }                                                                                 \
+        if (nloe) {                                                                       \
             HVX_Vector v = Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs), vo); \
-            vec_store((void *) &vdst[i], nloe * elem_size, Q6_Vsf_equals_Vqf32(v));  \
-        }                                                                            \
+            vec_store((void *) &vdst[i], nloe * elem_size, Q6_Vsf_equals_Vqf32(v));       \
+        }                                                                                 \
     } while(0)
 
 static inline void hvx_scale_offset_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {

ggml/src/ggml-hexagon/htp/hvx-sigmoid.h

@@ -68,50 +68,50 @@ static inline HVX_Vector hvx_vec_tanh_f32(HVX_Vector x) {
     return Q6_Vsf_equals_Vqf32(res);
 }
 
-#define hvx_sigmoid_loop_body(dst_type, src_type, vec_store)    \
-    do {                                                        \
-        dst_type * restrict vdst = (dst_type *) dst;            \
-        src_type * restrict vsrc = (src_type *) src;            \
-                                                                \
-        const HVX_Vector one     = hvx_vec_splat_f32(1.f);      \
-        const HVX_Vector max_exp = hvx_vec_splat_f32(87.f);     \
-        const HVX_Vector min_exp = hvx_vec_splat_f32(-87.f);    \
-                                                                \
-        const uint32_t epv  = 128 / sizeof(float);              \
-        const uint32_t nvec = n / epv;                          \
-        const uint32_t nloe = n % epv;                          \
-                                                                \
-        uint32_t i = 0;                                         \
-                                                                \
-        _Pragma("unroll(4)")                                    \
-        for (; i < nvec; i++) {                                 \
-             vdst[i] = hvx_vec_fast_sigmoid_f32_guard(vsrc[i], one, max_exp, min_exp); \
-        }                                                       \
-        if (nloe) {                                             \
+#define hvx_sigmoid_loop_body(dst_type, src_type, vec_store)                                  \
+    do {                                                                                      \
+        dst_type * restrict vdst = (dst_type *) dst;                                          \
+        src_type * restrict vsrc = (src_type *) src;                                          \
+                                                                                              \
+        const HVX_Vector one     = hvx_vec_splat_f32(1.f);                                    \
+        const HVX_Vector max_exp = hvx_vec_splat_f32(87.f);                                   \
+        const HVX_Vector min_exp = hvx_vec_splat_f32(-87.f);                                  \
+                                                                                              \
+        const uint32_t epv  = 128 / sizeof(float);                                            \
+        const uint32_t nvec = n / epv;                                                        \
+        const uint32_t nloe = n % epv;                                                        \
+                                                                                              \
+        uint32_t i = 0;                                                                       \
+                                                                                              \
+        _Pragma("unroll(4)")                                                                  \
+        for (; i < nvec; i++) {                                                               \
+             vdst[i] = hvx_vec_fast_sigmoid_f32_guard(vsrc[i], one, max_exp, min_exp);        \
+        }                                                                                     \
+        if (nloe) {                                                                           \
              HVX_Vector tmp = hvx_vec_fast_sigmoid_f32_guard(vsrc[i], one, max_exp, min_exp); \
-             vec_store((void *) &vdst[i], nloe * sizeof(float), tmp); \
-        }                                                       \
+             vec_store((void *) &vdst[i], nloe * sizeof(float), tmp);                         \
+        }                                                                                     \
     } while(0)
 
-#define hvx_tanh_loop_body(dst_type, src_type, vec_store)       \
-    do {                                                        \
-        dst_type * restrict vdst = (dst_type *) dst;            \
-        src_type * restrict vsrc = (src_type *) src;            \
-                                                                \
-        const uint32_t epv  = 128 / sizeof(float);              \
-        const uint32_t nvec = n / epv;                          \
-        const uint32_t nloe = n % epv;                          \
-                                                                \
-        uint32_t i = 0;                                         \
-                                                                \
-        _Pragma("unroll(4)")                                    \
-        for (; i < nvec; i++) {                                 \
-             vdst[i] = hvx_vec_tanh_f32(vsrc[i]);               \
-        }                                                       \
-        if (nloe) {                                             \
-             HVX_Vector tmp = hvx_vec_tanh_f32(vsrc[i]);        \
+#define hvx_tanh_loop_body(dst_type, src_type, vec_store)             \
+    do {                                                              \
+        dst_type * restrict vdst = (dst_type *) dst;                  \
+        src_type * restrict vsrc = (src_type *) src;                  \
+                                                                      \
+        const uint32_t epv  = 128 / sizeof(float);                    \
+        const uint32_t nvec = n / epv;                                \
+        const uint32_t nloe = n % epv;                                \
+                                                                      \
+        uint32_t i = 0;                                               \
+                                                                      \
+        _Pragma("unroll(4)")                                          \
+        for (; i < nvec; i++) {                                       \
+             vdst[i] = hvx_vec_tanh_f32(vsrc[i]);                     \
+        }                                                             \
+        if (nloe) {                                                   \
+             HVX_Vector tmp = hvx_vec_tanh_f32(vsrc[i]);              \
              vec_store((void *) &vdst[i], nloe * sizeof(float), tmp); \
-        }                                                       \
+        }                                                             \
     } while(0)
 
 static inline void hvx_sigmoid_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {

ggml/src/ggml-hexagon/htp/im2col-ops.c

@@ -3,11 +3,12 @@
 #pragma clang diagnostic ignored "-Wunused-but-set-variable"
 
 #include <HAP_farf.h>
-#include <HAP_perf.h>
 #include <hexagon_protos.h>
 #include <hexagon_types.h>
 #include <string.h>
 
+#include "hex-common.h"
+
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
@@ -16,14 +17,19 @@
 #include "hex-dma.h"
 #include "hex-profile.h"
 #include "htp-vtcm.h"
+#include "htp-tensor.h"
 
 struct htp_im2col_context {
     struct htp_ops_context * octx;
+    uint32_t                 patch_base;           // first patch index assigned to this dev
+    uint32_t                 npatches;             // number of patches assigned to this dev
     uint32_t                 npatches_per_thread;  // patches = N*OH*OW (pure-DDR kernel)
 
-    uint32_t pe_rows_per_thread;                   // N*OH rows per worker
-    uint32_t pe_src_row_bytes;                     // one output row's source: IC*KH*IW*4, rounded 256
-    uint32_t pe_dst_row_bytes;                     // one output row's dst: OW*patch_stride*2, rounded 256
+    uint32_t pe_row_base;              // first N*OH row index assigned to this dev (DMA path)
+    uint32_t pe_nrows;                 // number of N*OH rows assigned to this dev (DMA path)
+    uint32_t pe_rows_per_thread;       // N*OH rows per worker
+    uint32_t pe_src_row_bytes;         // one output row's source: IC*KH*IW*4, rounded 256
+    uint32_t pe_dst_row_bytes;         // one output row's dst: OW*patch_stride*2, rounded 256
 
     // Patch-embed DMA path VTCM ping-pong.
     uint8_t * pe_vtcm_src;                         // base of the 2x src buffers region
@@ -58,33 +64,27 @@ static inline void htp_im2col_vtcm_layout_build(struct htp_im2col_vtcm_layout *
         struct htp_im2col_context * ictx        = (struct htp_im2col_context *) data;                     \
         struct htp_ops_context *    octx        = ictx->octx;                                             \
         struct htp_thread_trace * restrict tr   = &octx->ctx->trace[ith];                                 \
+        const struct htp_tensor * restrict src0 = octx->src[0];                                           \
         const struct htp_tensor * restrict src1 = octx->src[1];                                           \
         const struct htp_tensor * restrict dst  = octx->dst;                                              \
-        const int32_t  s0                       = octx->op_params[0];                                     \
-        const int32_t  s1                       = octx->op_params[1];                                     \
-        const int32_t  p0                       = octx->op_params[2];                                     \
-        const int32_t  p1                       = octx->op_params[3];                                     \
-        const int32_t  d0                       = octx->op_params[4];                                     \
-        const int32_t  d1                       = octx->op_params[5];                                     \
-        const uint32_t N                        = src1->ne[3];                                            \
-        const uint32_t IC                       = src1->ne[2];                                            \
-        const uint32_t IH                       = src1->ne[1];                                            \
-        const uint32_t IW                       = src1->ne[0];                                            \
-        const uint32_t KH                       = octx->src[0]->ne[1];                                    \
-        const uint32_t KW                       = octx->src[0]->ne[0];                                    \
+        const int32_t s0 = octx->op_params[0], s1 = octx->op_params[1];                                   \
+        const int32_t p0 = octx->op_params[2], p1 = octx->op_params[3];                                   \
+        const int32_t d0 = octx->op_params[4], d1 = octx->op_params[5];                                   \
+        const uint32_t N = src1->ne[3], IC = src1->ne[2], IH = src1->ne[1], IW = src1->ne[0];             \
+        const uint32_t KH                       = src0->ne[1], KW = src0->ne[0];                          \
         const uint32_t OH                       = dst->ne[2];                                             \
         const uint32_t OW                       = dst->ne[1];                                             \
         const uint32_t patch_stride             = IC * KH * KW;                                           \
         const float * restrict src_data         = (const float *) src1->data;                             \
         DST_CTYPE * restrict dst_data           = (DST_CTYPE *) dst->data;                                \
-        const uint32_t npatches                 = N * OH * OW;                                            \
-        const uint32_t patch_start              = ictx->npatches_per_thread * ith;                        \
-        const uint32_t patch_end                = MIN(patch_start + ictx->npatches_per_thread, npatches); \
-        if (patch_start >= patch_end) {                                                                   \
+        const uint32_t patch_end                = ictx->patch_base + ictx->npatches;                      \
+        const uint32_t patch_start              = ictx->patch_base + ictx->npatches_per_thread * ith;     \
+        const uint32_t patch_stop               = MIN(patch_start + ictx->npatches_per_thread, patch_end);\
+        if (patch_start >= patch_stop) {                                                                  \
             return;                                                                                       \
         }                                                                                                 \
         htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, patch_start);                                   \
-        for (uint32_t p = patch_start; p < patch_end; p++) {                                              \
+        for (uint32_t p = patch_start; p < patch_stop; p++) {                                             \
             const uint32_t iow             = p % OW;                                                      \
             const uint32_t ioh             = (p / OW) % OH;                                               \
             const uint32_t in              = p / (OW * OH);                                               \
@@ -154,10 +154,10 @@ IM2COL_PATCHEMBED_BODY(im2col_patchembed_f32_thread, float, hvx_copy_f32_uu, hvx
         uint8_t *      dst_base         = ictx->pe_vtcm_dst + ith * ictx->pe_dst_size_per_thread;                    \
         float *        srcb             = (float *) src_base;                                                        \
         DST_CTYPE *    dstb             = (DST_CTYPE *) dst_base;                                                    \
-        const uint32_t nrows            = N * OH;                                                                    \
+        const uint32_t row_end_max      = ictx->pe_row_base + ictx->pe_nrows;                                        \
         const uint32_t per_thread       = ictx->pe_rows_per_thread;                                                  \
-        const uint32_t row_start        = per_thread * ith;                                                          \
-        const uint32_t row_end          = MIN(row_start + per_thread, nrows);                                        \
+        const uint32_t row_start        = ictx->pe_row_base + per_thread * ith;                                      \
+        const uint32_t row_end          = MIN(row_start + per_thread, row_end_max);                                  \
         if (row_start >= row_end)                                                                                    \
             return;                                                                                                  \
         for (uint32_t r = row_start; r < row_end; r++) {                                                             \
@@ -266,26 +266,55 @@ int op_im2col(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t N         = src1->ne[3];
-    const uint32_t OH        = dst->ne[2];
-    const uint32_t OW        = dst->ne[1];
-    const uint32_t npatches  = N * OH * OW;
-    const uint32_t n_threads = MIN(octx->n_threads, npatches);
-
-    if ((octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) || n_threads == 0) {
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
         return HTP_STATUS_OK;
     }
 
+    const uint32_t N             = src1->ne[3];
+    const uint32_t OH            = dst->ne[2];
+    const uint32_t OW            = dst->ne[1];
+    const uint32_t total_patches = N * OH * OW;
+    const uint32_t total_rows    = N * OH;
+
+    uint32_t patch_base = 0;
+    uint32_t npatches   = total_patches;
+    if (octx->ctx->mdev.count > 1) {
+        const uint32_t patch_size = dst->nb[1];
+        const uint32_t patches_per_chunk = (patch_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(patch_size, HEX_L2_LINE_SIZE)) : 1;
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_patches, htp_tensor_mdev_data_aligned(dst) ? patches_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        patch_base = range.start;
+        npatches   = range.count;
+    }
+
+    uint32_t row_base = 0;
+    uint32_t nrows    = total_rows;
+    if (octx->ctx->mdev.count > 1) {
+        const uint32_t row_size = dst->nb[2];
+        const uint32_t rows_per_chunk = (row_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(row_size, HEX_L2_LINE_SIZE)) : 1;
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_rows, htp_tensor_mdev_data_aligned(dst) ? rows_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_base = range.start;
+        nrows    = range.count;
+    }
+
+    if (npatches == 0 && nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = MIN(octx->n_threads, MAX(npatches, 1));
+
     struct htp_im2col_context ictx = { 0 };
-    ictx.octx                      = octx;
-    ictx.npatches_per_thread       = (npatches + n_threads - 1) / n_threads;
+    ictx.octx                = octx;
+    ictx.patch_base          = patch_base;
+    ictx.npatches            = npatches;
+    ictx.npatches_per_thread = (npatches + n_threads - 1) / n_threads;
 
     // Clean non-overlapping patch-embed -> DMA kernel (if it fits VTCM);
     // everything else (padding/dilation/stride edges) -> pure-DDR kernel.
-    if (im2col_use_patchembed_dma(octx)) {
-        const uint32_t nrows = N * OH;
-        const uint32_t pth   = MIN(octx->n_threads, nrows);
+    if (im2col_use_patchembed_dma(octx) && nrows > 0) {
+        const uint32_t pth = MIN(octx->n_threads, nrows);
         if (pth > 0 && im2col_patchembed_dma_fits(octx, &ictx, pth)) {
+            ictx.pe_row_base        = row_base;
+            ictx.pe_nrows           = nrows;
             ictx.pe_rows_per_thread = (nrows + pth - 1) / pth;
             if (dst->type == HTP_TYPE_F16) {
                 work_queue_run(octx->ctx->work_queue, im2col_patchembed_dma_thread, &ictx, pth);
@@ -297,6 +326,10 @@ int op_im2col(struct htp_ops_context * octx) {
         // else: doesn't fit -> fall through to the pure-DDR kernel below.
     }
 
+    if (npatches == 0) {
+        return HTP_STATUS_OK;
+    }
+
     if (dst->type == HTP_TYPE_F16) {
         work_queue_run(octx->ctx->work_queue, im2col_patchembed_thread, &ictx, n_threads);
     } else {

ggml/src/ggml-hexagon/htp/main.c

@@ -34,6 +34,7 @@
 #include "work-queue.h"
 #include "hex-profile.h"
 #include "allreduce-ops.h"
+#include "htp-fence.h"
 
 #define HMX_QUEUE_CAPACITY     16
 #define HMX_QUEUE_STACK_SIZE   16384
@@ -710,22 +711,43 @@ static inline void profile_stop(uint32_t mode, struct profile_data * d) {
 static int op_fence(struct htp_ops_context * octx) {
     struct htp_context *ctx = octx->ctx;
     struct htp_thread_trace * tr = &ctx->trace[0];
-    const uint32_t seq = (uint32_t) octx->op_params[0];
+    const uint32_t seq  = (uint32_t) octx->op_params[0];
+    const uint32_t mode = (uint32_t) octx->op_params[1];
 
     htp_trace_event_start(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
 
     const struct htp_tensor * sync = octx->src[0];
-    atomic_uint * sync_fence = (atomic_uint *) sync->data;
+    atomic_uint * sync_fence = (atomic_uint *) (uintptr_t) sync->data;
+
+    if (mode == 1) {
+        htp_flush_dirty_ranges(ctx);
+
+        htp_mdev_group_barrier(octx);
+
+        if (ctx->mdev.idx == 0) {
+            htp_fence_write(sync_fence, seq, octx->status);
+        }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
+        FARF(HIGH, "ggml-hex: sync-signal : fence %p seq 0x%x status %d\n", sync_fence, seq, octx->status);
+        return octx->status;
+    }
+
+    int status = HTP_STATUS_OK;
     uint64_t spins = 0;
     while (1) {
-        Q6_dccleaninva_A((void *) sync_fence);
-        asm volatile ("syncht" : : : "memory");
-        uint32_t val = atomic_load(&sync_fence[0]);
-        if ((int32_t)(val - seq) >= 0) {
+        uint32_t sync_seq;
+        uint32_t sync_status;
+        htp_fence_read(sync_fence, &sync_seq, &sync_status);
+        if ((int32_t)(sync_seq - seq) >= 0) {
+            if (sync_status > HTP_STATUS_OK) {
+                FARF(ERROR, "ggml-hex: sync-wait peer failed with status %u : fence %p seq 0x%x\n", sync_status, sync_fence, seq);
+                status = sync_status;
+            }
             break;
         }
         if (++spins > HTP_FENCE_TIMEOUT) {
-            FARF(ERROR, "ggml-hex: sync-wait TIMEOUT : fence %p spins %llu seq %u\n", sync_fence, spins, seq);
+            FARF(ERROR, "ggml-hex: sync-wait TIMEOUT : fence %p spins %llu seq 0x%x\n", sync_fence, spins, seq);
+            status = HTP_STATUS_INTERNAL_ERR;
             break;
         }
         hex_pause();
@@ -733,12 +755,27 @@ static int op_fence(struct htp_ops_context * octx) {
 
     htp_trace_event_stop(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
 
-    FARF(HIGH, "ggml-hex: sync-done : fence %p spins %llu seq %u\n", sync_fence, spins, seq);
+    FARF(HIGH, "ggml-hex: sync-done : fence %p spins %llu seq 0x%x\n", sync_fence, spins, seq);
+    return status;
+}
+
+static int op_mdev_group(struct htp_ops_context * octx) {
+    struct htp_context * ctx = octx->ctx;
+    const struct htp_tensor * sync = octx->src[0];
+    ctx->mdev.idx   = (uint16_t) octx->op_params[0];
+    ctx->mdev.count = (uint16_t) sync->ne[1];
+    if (ctx->mdev.count > 1) {
+        ctx->mdev.count_div = init_fastdiv_values(ctx->mdev.count);
+        ctx->mdev.fence_base = (uint8_t *) sync->data;
+    }
     return HTP_STATUS_OK;
 }
 
 static int execute_op(struct htp_ops_context * octx) {
     switch (octx->op) {
+        case HTP_OP_MDEV_GROUP:
+            return op_mdev_group(octx);
+
         case HTP_OP_FENCE:
             return op_fence(octx);
 
@@ -812,6 +849,7 @@ static int execute_op(struct htp_ops_context * octx) {
             return op_sum_rows(octx);
 
         case HTP_OP_CPY:
+        case HTP_OP_CPY_FENCE:
             return op_cpy(octx);
 
         case HTP_OP_REPEAT:
@@ -855,7 +893,7 @@ static int execute_op(struct htp_ops_context * octx) {
     }
 
     FARF(ERROR, "Unknown Op %u", octx->op);
-    return -1;
+    return HTP_STATUS_NO_SUPPORT;
 }
 
 static inline bool reuse_buf(struct htp_context *ctx, uint32_t *m_reuse, struct htp_buf_desc *b) {
@@ -984,11 +1022,19 @@ static void prep_tensors(struct htp_context *ctx, struct htp_buf_desc *bufs, str
     }
 }
 
-static int proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, uint32_t idx, struct htp_op_desc * op) {
-    memcpy(octx->op_params, op->params, sizeof(octx->op_params));
+static void mdev_group_init(struct htp_context * ctx, const struct htp_opbatch_req * req) {
+    memset(&ctx->mdev, 0, sizeof(ctx->mdev));
+    ctx->mdev.fence_seq = (uint32_t)((req->seq & 0xfffff) << 12);
+}
+
+static int proc_op_req(struct htp_ops_context * octx, struct htp_buf_desc * bufs, uint32_t n_bufs,
+                       struct htp_tensor * tens, uint32_t idx, struct htp_op_desc * op) {
+    memcpy(octx->op_params,     op->params, sizeof(octx->op_params));
     memcpy(octx->kernel_params, op->kernel_params, sizeof(octx->kernel_params));
-    octx->flags = op->flags;
-    octx->op    = op->opcode;
+    octx->flags         = op->flags;
+    octx->op            = op->opcode;
+    octx->n_threads     = octx->ctx->n_threads;
+    octx->n_threads_div = octx->ctx->n_threads_div;
 
     FARF(HIGH, "proc-op #%u: opcode %u flags 0x%x", idx, octx->op, octx->flags);
 
@@ -1027,9 +1073,13 @@ static int proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, u
             dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
     }
 
+    htp_tensor_dirty_all(octx->ctx, octx->dsts, HTP_OP_MAX_OUTPUTS);
+
+    htp_mdev_group_barrier(octx);
+
     int status = execute_op(octx);
 
-    htp_tensor_dirty_all(octx->ctx, octx->dsts, HTP_OP_MAX_OUTPUTS);
+    htp_ops_context_set_status(octx, status);
 
     octx->src0_spad.src = NULL;
     octx->src1_spad.src = NULL;
@@ -1037,7 +1087,7 @@ static int proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, u
     octx->src3_spad.src = NULL;
     octx->dst_spad.src  = NULL;
 
-    return status;
+    return octx->status;
 }
 
 static void process_opbatch(struct htp_context * ctx, const struct htp_opbatch_req * req, const struct dspqueue_buffer * dbuf) {
@@ -1059,7 +1109,7 @@ static void process_opbatch(struct htp_context * ctx, const struct htp_opbatch_r
         return;
     }
 
-    FARF(HIGH, "processing opbatch #%u: n-bufs %u n-tensors %u n-ops %u n-traces %u : m-size %u b-size %u t-size %u o-size %u", req->id,
+    FARF(HIGH, "processing opbatch #%llu: n-bufs %u n-tensors %u n-ops %u n-traces %u : m-size %u b-size %u t-size %u o-size %u", (unsigned long long) req->seq,
             n_bufs, n_tens, n_ops, req->n_traces, dbuf->size, b_size, t_size, o_size);
 
     // Setup descriptor pointers
@@ -1096,8 +1146,11 @@ static void process_opbatch(struct htp_context * ctx, const struct htp_opbatch_r
 
     struct htp_ops_context *octx = &ctx->octx;
     memset(octx, 0, sizeof(*octx));
-    octx->n_threads = ctx->n_threads;
-    octx->ctx       = ctx;
+    octx->n_threads     = ctx->n_threads;
+    octx->n_threads_div = ctx->n_threads_div;
+    octx->ctx           = ctx;
+
+    mdev_group_init(ctx, req);
 
     work_queue_wakeup(ctx->work_queue);
     if (ctx->hmx_queue) {
@@ -1105,15 +1158,18 @@ static void process_opbatch(struct htp_context * ctx, const struct htp_opbatch_r
     }
 
     int op_status = HTP_STATUS_OK;
-    for (uint32_t i = 0; i < n_ops && op_status == HTP_STATUS_OK; i++) {
+    octx->status  = HTP_STATUS_OK;
+    for (uint32_t i = 0; i < n_ops; i++) {
         struct profile_data prof;
 
         profile_start(ctx->profiler, &prof);
 
-        op_status = proc_op_req(octx, tens, i, &ops[i]);
+        op_status = proc_op_req(octx, bufs, n_bufs, tens, i, &ops[i]);
 
         profile_stop(ctx->profiler, &prof);
 
+        htp_ops_context_set_status(octx, op_status);
+
         if (ctx->profiler) {
             pds[i].opcode = ops[i].opcode;
             pds[i].usecs  = prof.usecs;
@@ -1136,19 +1192,20 @@ static void process_opbatch(struct htp_context * ctx, const struct htp_opbatch_r
     qurt_mem_cache_clean((qurt_addr_t) 0, 0, QURT_MEM_CACHE_FLUSH_INVALIDATE_ALL, QURT_MEM_DCACHE);
     htp_trace_event_stop(&ctx->trace[0], HTP_TRACE_EVT_L2FLUSH, 0);
 
+    htp_mdev_group_barrier(octx);
+
     profile_stop(HTP_PROF_BASIC, &batch_prof);
 
     struct htp_opbatch_rsp rsp;
     memset(&rsp, 0, sizeof(rsp));
-    rsp.id           = req->id;
-    rsp.status       = op_status;
+    rsp.seq          = req->seq;
+    rsp.status       = octx->status;
     rsp.n_bufs       = n_bufs;
     rsp.n_tensors    = n_tens;
     rsp.n_ops        = n_ops;
     rsp.usecs        = batch_prof.usecs;
     rsp.cycles_start = batch_prof.cycles_start;
     rsp.cycles_stop  = batch_prof.cycles_stop;
-    rsp.seq          = req->seq;
 
     if (ctx->profiler == HTP_PROF_TRACE) {
         for (int t = 0; t <= HTP_MAX_NTHREADS; t++) {

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -21,6 +21,7 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
+#include "htp-tensor.h"
 #include "matmul-ops.h"
 #include "htp-vtcm.h"
 
@@ -89,6 +90,8 @@ struct htp_mm_context {
 
     // Precomputed values
     uint32_t src0_nrows_per_thread;
+    uint32_t src0_row_start;
+    uint32_t src0_row_end;
     uint32_t src0_row_size_padded;
     uint32_t src1_nrows;
 
@@ -135,6 +138,23 @@ struct htp_mm_context {
     uint32_t vtcm_dst_size_per_thread;
 };
 
+static int htp_mm_init_context(
+    struct htp_ops_context * octx,
+    const struct htp_mm_kernel_params * kparams
+) {
+    if (!htp_ops_context_set_n_threads(octx, (uint32_t) kparams->n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    if (kparams->n_hmx) {
+        if (kparams->n_act_threads <= 0 || kparams->n_act_threads > (int32_t) octx->n_threads) {
+            return HTP_STATUS_INVAL_PARAMS;
+        }
+    }
+
+    return HTP_STATUS_OK;
+}
+
 // vdelta control to expand first 32 e8m0 values into 32 uint32 elements
 static const uint8_t __attribute__((aligned(128))) expand_x32_e8m0[128] = {
     0x00, 0x00, 0x00, 0x00, 0x01, 0x04, 0x00, 0x00, 0x02, 0x00, 0x08, 0x08, 0x01, 0x02, 0x00, 0x04, 0x04, 0x00, 0x00,
@@ -238,22 +258,24 @@ static void hvx_mm_4d(unsigned int nth, unsigned int ith, void * data) {
     // This is the size of the rest of the dimensions of the result
     const uint32_t nr1 = ne1 * ne2 * ne3;
 
+    const uint32_t src0_nrows = mmctx->src0_row_end - mmctx->src0_row_start;
+
     // distribute the thread work across the inner or outer loop based on which one is larger
     uint32_t dr0, dr1, ith0, ith1;
     if (nr0 > nr1) {
-        dr0  = fastdiv(nr0 + nth - 1, &octx->ctx->n_threads_div);
+        dr0  = fastdiv(src0_nrows + nth - 1, &octx->n_threads_div);
         dr1  = nr1;
         ith0 = ith;
         ith1 = 0;
     } else {
-        dr0  = nr0;
-        dr1  = fastdiv(nr1 + nth - 1, &octx->ctx->n_threads_div);
+        dr0  = src0_nrows;
+        dr1  = fastdiv(nr1 + nth - 1, &octx->n_threads_div);
         ith0 = 0;
         ith1 = ith;
     }
 
-    const uint32_t ir0_start = dr0 * ith0;
-    const uint32_t ir0_end   = MIN(ir0_start + dr0, nr0);
+    const uint32_t ir0_start = mmctx->src0_row_start + dr0 * ith0;
+    const uint32_t ir0_end   = MIN(ir0_start + dr0, mmctx->src0_row_end);
 
     const uint32_t ir1_start = dr1 * ith1;
     const uint32_t ir1_end   = MIN(ir1_start + dr1, nr1);
@@ -312,11 +334,11 @@ static void hvx_mm_4d(unsigned int nth, unsigned int ith, void * data) {
 static void hvx_mm_2d_repacked_##SUFFIX(unsigned int nth, unsigned int ith, void * data) {                                        \
     htp_matmul_preamble;                                                                                                          \
                                                                                                                                   \
-    const uint32_t src0_nrows = ne01 * ne02 * ne03;                                                                               \
+    const uint32_t src0_nrows = mmctx->src0_row_end - mmctx->src0_row_start;                                                      \
     const uint32_t src1_nrows = ne11 * ne12 * ne13;                                                                               \
                                                                                                                                   \
-    const uint32_t src0_start_row  = src0_nrows_per_thread * ith;                                                                 \
-    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);                                     \
+    const uint32_t src0_start_row  = mmctx->src0_row_start + src0_nrows_per_thread * ith;                                         \
+    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, mmctx->src0_row_end);                            \
                                                                                                                                   \
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                                        \
                                                                                                                                   \
@@ -414,10 +436,10 @@ static void hvx_mm_2d_repacked_##SUFFIX(unsigned int nth, unsigned int ith, void
 static void hvx_mv_2d_repacked_##SUFFIX(unsigned int nth, unsigned int ith, void * data) {                                        \
     htp_matmul_preamble;                                                                                                          \
                                                                                                                                   \
-    const uint32_t src0_nrows = ne01;                                                                                             \
+    const uint32_t src0_nrows = mmctx->src0_row_end - mmctx->src0_row_start;                                                      \
                                                                                                                                   \
-    const uint32_t src0_start_row  = src0_nrows_per_thread * ith;                                                                 \
-    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);                                     \
+    const uint32_t src0_start_row  = mmctx->src0_row_start + src0_nrows_per_thread * ith;                                         \
+    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, mmctx->src0_row_end);                            \
                                                                                                                                   \
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                                        \
                                                                                                                                   \
@@ -549,12 +571,22 @@ static void hvx_mm_nx_2d_repacked_##SUFFIX(unsigned int nth, unsigned int ith, v
         uint32_t n_k_tiles_w = ne00 / 32;                                                                                         \
         uint32_t tile_row_stride = n_k_tiles_w * tile_size;                                                                       \
                                                                                                                                   \
-        const uint32_t src0_nrows = ne01 * src_w->ne[2] * src_w->ne[3];                                                           \
-        uint32_t src0_nrows_per_thread = fastdiv(src0_nrows + nth - 1, &octx->ctx->n_threads_div);                                \
+        uint32_t src0_start_row = 0;                                                                                              \
+        uint32_t src0_end_row   = ne01;                                                                                           \
+        if (octx->ctx->mdev.count > 1) {                                                                                          \
+            const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));                                              \
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(ne01, can_split ? 32 : 0,                        \
+                                                         octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div); \
+            src0_start_row = range.start;                                                                                         \
+            src0_end_row   = range.start + range.count;                                                                           \
+        }                                                                                                                         \
+                                                                                                                                  \
+        const uint32_t nrows = src0_end_row - src0_start_row;                                                                     \
+        uint32_t src0_nrows_per_thread = fastdiv(nrows + nth - 1, &octx->n_threads_div);                                          \
         src0_nrows_per_thread = hex_round_up(src0_nrows_per_thread, 32);                                                          \
                                                                                                                                   \
-        const uint32_t start_row = src0_nrows_per_thread * ith;                                                                   \
-        const uint32_t end_row   = MIN(start_row + src0_nrows_per_thread, src0_nrows);                                            \
+        const uint32_t start_row = src0_start_row + src0_nrows_per_thread * ith;                                                  \
+        const uint32_t end_row   = MIN(start_row + src0_nrows_per_thread, src0_end_row);                                          \
         if (start_row >= end_row) continue;                                                                                       \
                                                                                                                                   \
         uint32_t ct_start = start_row / 32;                                                                                       \
@@ -735,11 +767,11 @@ static void hvx_mm_2d(unsigned int nth, unsigned int ith, void * data) {
     assert(n_prefetch >= 2 && n_prefetch <= HTP_MM_MAX_PREFETCH && (n_prefetch & (n_prefetch - 1)) == 0);
     const uint32_t prefetch_mask = n_prefetch - 1;
 
-    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
-    const uint32_t src1_nrows = ne11 * ne12 * ne13;  // src1 rows
+    const uint32_t src0_nrows = mmctx->src0_row_end - mmctx->src0_row_start;  // src0 rows
+    const uint32_t src1_nrows = ne11 * ne12 * ne13;                          // src1 rows
 
-    const uint32_t src0_start_row  = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t src0_start_row  = mmctx->src0_row_start + src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, mmctx->src0_row_end);
     const uint32_t src0_end_row_x2 = src0_start_row + ((src0_end_row - src0_start_row) & ~1U);
 
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];
@@ -781,7 +813,7 @@ static void hvx_mm_2d(unsigned int nth, unsigned int ith, void * data) {
         const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
         htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
-        // Process src1 columns in pairs (2×2 tiling)
+        // Process src1 columns in pairs (2x2 tiling)
         uint32_t ir1 = 0;
         for (; ir1 + 1 < src1_nrows; ir1 += 2) {
             const uint8_t * restrict src1_col0 = (const uint8_t *) (src1_data + (ir1+0) * src1_stride);
@@ -791,7 +823,7 @@ static void hvx_mm_2d(unsigned int nth, unsigned int ith, void * data) {
             mmctx->vec_dot_2x2(ne00, &dst_row0[ir0], &dst_row1[ir0], ss0, ss0 + src0_stride, src1_col0, src1_col1);
         }
 
-        // Handle remaining src1 rows (fallback to 2×1)
+        // Handle remaining src1 rows (fallback to 2x1)
         for (; ir1 < src1_nrows; ++ir1) {
             const uint8_t * restrict src1_col = (const uint8_t *) (src1_data + ir1 * src1_stride);
             float * restrict dst_row          = (float *) (dst->data + (ir1 * dst_row_size));
@@ -833,10 +865,10 @@ static void hvx_mm_2d(unsigned int nth, unsigned int ith, void * data) {
 static void hvx_mv_2d(unsigned int nth, unsigned int ith, void * data) {
     htp_matmul_preamble;
 
-    const uint32_t src0_nrows = ne01;
+    const uint32_t src0_nrows = mmctx->src0_row_end - mmctx->src0_row_start;
 
-    const uint32_t src0_start_row  = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t src0_start_row  = mmctx->src0_row_start + src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, mmctx->src0_row_end);
 
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];
 
@@ -943,13 +975,10 @@ static void hvx_mm_id(unsigned int nth, unsigned int ith, void * data) {
 
     const struct htp_tensor * restrict ids = octx->src[2];
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
-    const uint32_t src0_nrows      = ne01;  // src0 rows per expert
+    const uint32_t src0_nrows      = mmctx->src0_row_end - mmctx->src0_row_start;  // src0 rows per expert
     const uint32_t src1_nrows      = ne11;
-    const uint32_t src0_start_row  = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t src0_start_row  = mmctx->src0_row_start + src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, mmctx->src0_row_end);
 
     hvx_mm_run_quant_task(mmctx, ith);
 
@@ -1036,9 +1065,9 @@ static void hvx_mv_id(unsigned int nth, unsigned int ith, void * data) {
 
     const struct htp_tensor * restrict ids = octx->src[2];
 
-    const uint32_t src0_nrows      = ne01;  // src0 rows per expert
-    const uint32_t src0_start_row  = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t src0_nrows      = mmctx->src0_row_end - mmctx->src0_row_start;  // src0 rows per expert
+    const uint32_t src0_start_row  = mmctx->src0_row_start + src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, mmctx->src0_row_end);
 
     hvx_mm_run_quant_task(mmctx, ith);
 
@@ -1143,12 +1172,22 @@ static void hvx_mv_id_nx(unsigned int nth, unsigned int ith, void * data) {
             const struct htp_tensor * restrict dst   = octx->dsts[p];
             if (!src_w || !dst) continue;
 
-            const uint32_t src0_nrows = src_w->ne[1];
-            uint32_t src0_nrows_per_thread = fastdiv(src0_nrows + nth - 1, &octx->ctx->n_threads_div);
+            const uint32_t ne01 = src_w->ne[1];
+            uint32_t start_row = 0;
+            uint32_t end_row   = ne01;
+            if (octx->ctx->mdev.count > 1) {
+                const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));
+                const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(ne01, can_split ? 32 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+                start_row = range.start;
+                end_row   = range.start + range.count;
+            }
+
+            const uint32_t nrows = end_row - start_row;
+            uint32_t src0_nrows_per_thread = fastdiv(nrows + nth - 1, &octx->n_threads_div);
             src0_nrows_per_thread = hex_round_up(src0_nrows_per_thread, 32);
 
-            const uint32_t src0_start_row = src0_nrows_per_thread * ith;
-            const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+            const uint32_t src0_start_row = start_row + src0_nrows_per_thread * ith;
+            const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, end_row);
             if (src0_start_row >= src0_end_row) continue;
 
             const uint8_t * restrict src0_row = (const uint8_t *) src_w->data + eid * src_w->nb[2];
@@ -1227,12 +1266,22 @@ static void hvx_mm_id_nx(unsigned int nth, unsigned int ith, void * data) {
             const struct htp_tensor * restrict dst   = octx->dsts[p];
             if (!src_w || !dst) continue;
 
-            const uint32_t src0_nrows = src_w->ne[1];
-            uint32_t src0_nrows_per_thread = fastdiv(src0_nrows + nth - 1, &octx->ctx->n_threads_div);
+            const uint32_t ne01 = src_w->ne[1];
+            uint32_t start_row = 0;
+            uint32_t end_row   = ne01;
+            if (octx->ctx->mdev.count > 1) {
+                const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));
+                const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(ne01, can_split ? 32 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+                start_row = range.start;
+                end_row   = range.start + range.count;
+            }
+
+            const uint32_t nrows = end_row - start_row;
+            uint32_t src0_nrows_per_thread = fastdiv(nrows + nth - 1, &octx->n_threads_div);
             src0_nrows_per_thread = hex_round_up(src0_nrows_per_thread, 32);
 
-            const uint32_t src0_start_row = src0_nrows_per_thread * ith;
-            const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+            const uint32_t src0_start_row = start_row + src0_nrows_per_thread * ith;
+            const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, end_row);
             if (src0_start_row >= src0_end_row) continue;
 
             const uint8_t * src0_row = (const uint8_t *) src_w->data + cur_a * src_w->nb[2];
@@ -1323,15 +1372,33 @@ static int hvx_mm_matmul(struct htp_ops_context * octx) {
 
     const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
 
-    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+    const uint32_t src0_nrows = ne01;
     const uint32_t src1_nrows = ne11 * ne12 * ne13;
 
+    uint32_t src0_row_start = 0;
+    uint32_t src0_row_end   = src0_nrows;
+
+    if (octx->ctx->mdev.count > 1) {
+        const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(src0_nrows, can_split ? 32 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        src0_row_start = range.start;
+        src0_row_end   = range.start + range.count;
+    }
+
+    if (src0_row_start >= src0_row_end) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t nrows = src0_row_end - src0_row_start;
+    mmctx->src0_row_start = src0_row_start;
+    mmctx->src0_row_end   = src0_row_end;
+
     bool is_repacked = (src0->type == HTP_TYPE_Q4_0 || src0->type == HTP_TYPE_Q4_1 ||
                         src0->type == HTP_TYPE_Q8_0 || src0->type == HTP_TYPE_IQ4_NL ||
                         src0->type == HTP_TYPE_MXFP4);
 
     // Compute src0_nrows_per_thread
-    mmctx->src0_nrows_per_thread  = fastdiv(src0_nrows + octx->n_threads - 1, &octx->ctx->n_threads_div);
+    mmctx->src0_nrows_per_thread  = fastdiv(nrows + octx->n_threads - 1, &octx->n_threads_div);
     if (is_repacked) {
         mmctx->src0_nrows_per_thread = hex_round_up(mmctx->src0_nrows_per_thread, 32);
     } else {
@@ -1503,13 +1570,13 @@ static int hvx_mm_matmul(struct htp_ops_context * octx) {
         kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_BLOCK) {
         mmctx->vtcm_src1_size_per_thread = L.src1_bytes;
     } else {
-        mmctx->vtcm_src1_size_per_thread = fastdiv(L.src1_bytes, &octx->ctx->n_threads_div);
+        mmctx->vtcm_src1_size_per_thread = fastdiv(L.src1_bytes, &octx->n_threads_div);
     }
 
-    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->ctx->n_threads_div);
-    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->n_threads_div);
+    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->n_threads_div);
 
-    size_t vtcm_size = kparams->vtcm_size > 0 ? (size_t)kparams->vtcm_size : L.total_bytes;
+    const size_t vtcm_size = L.total_bytes;
 
     FARF(HIGH, "matmul-%s : src0-vtcm-size %zu src1-vtcm-size %zu dst-vtcm-size %zu (%zu)\n", mmctx->type,
          L.src0_bytes, L.src1_bytes, L.dst_bytes, vtcm_size);
@@ -1583,13 +1650,21 @@ static void hvx_mm_nx_2d(unsigned int nth, unsigned int ith, void * data) {
 
         const uint32_t ne00 = src_w->ne[0];
         const uint32_t ne01 = src_w->ne[1];
-        const uint32_t src0_nrows = ne01 * src_w->ne[2] * src_w->ne[3];
+        uint32_t start_row = 0;
+        uint32_t end_row   = ne01;
+        if (octx->ctx->mdev.count > 1) {
+            const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(ne01, can_split ? 32 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            start_row = range.start;
+            end_row   = range.start + range.count;
+        }
 
-        uint32_t src0_nrows_per_thread = fastdiv(src0_nrows + nth - 1, &octx->ctx->n_threads_div);
+        const uint32_t nrows = end_row - start_row;
+        uint32_t src0_nrows_per_thread = fastdiv(nrows + nth - 1, &octx->n_threads_div);
         src0_nrows_per_thread += (src0_nrows_per_thread & 1);
 
-        const uint32_t src0_start_row  = src0_nrows_per_thread * ith;
-        const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+        const uint32_t src0_start_row  = start_row + src0_nrows_per_thread * ith;
+        const uint32_t src0_end_row    = MIN(src0_start_row + src0_nrows_per_thread, end_row);
         const uint32_t src0_end_row_x2 = src0_start_row + ((src0_end_row - src0_start_row) & ~1U);
         if (src0_start_row >= src0_end_row) continue;
 
@@ -2638,10 +2713,6 @@ static int hmx_mm_nx_2d_f32(struct htp_ops_context * octx, const struct htp_mm_k
     const struct htp_tensor * restrict src0 = octx->src[0];
     const struct htp_tensor * restrict act  = octx->src[n_weights];
 
-    if (!src0 || !act) {
-        return HTP_STATUS_INVAL_PARAMS;
-    }
-
     const int weight_type = (int) src0->type;
     const int k           = (int) act->ne[0];
     const int k_valid     = (int) act->ne[0];
@@ -2714,16 +2785,31 @@ static int hmx_mm_nx_2d_f32(struct htp_ops_context * octx, const struct htp_mm_k
 
     hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
-    FARF(HIGH, "hmx-mm-nx-2d: n_weights %u m %d k %d wtype %d mc %d nc %d vtcm %zu/%zu",
-         n_weights, m, k, weight_type, m_chunk_n_rows, n_chunk_n_cols, L.total_bytes, vtcm_budget);
+    int m_start = 0;
+    int m_rows  = m;
+    if (octx->ctx->mdev.count > 1) {
+        const bool can_split = htp_tensor_can_row_partition(octx->dsts[0], sizeof(float));
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition((uint32_t) m, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        m_start = (int) range.start;
+        m_rows  = (int) range.count;
+    }
+
+    if (m_rows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    FARF(HIGH, "hmx-mm-nx-2d: n_weights %u m %d (%d..%d) k %d wtype %d mc %d nc %d vtcm %zu/%zu",
+         n_weights, m, m_start, m_start + m_rows, k, weight_type, m_chunk_n_rows, n_chunk_n_cols, L.total_bytes, vtcm_budget);
 
     htp_trace_event_stop(tr, HTP_TRACE_EVT_INIT, 0);
 
+    const size_t mr_end = (size_t)(m_start + m_rows);
+
     if (pipeline) {
         hmx_matmul_job_t job_slots[2];
 
-        for (size_t mr = 0; mr < (size_t) m; mr += m_chunk_n_rows) {
-            const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+        for (size_t mr = (size_t) m_start; mr < mr_end; mr += m_chunk_n_rows) {
+            const size_t n_rows = hex_smin(mr_end - mr, m_chunk_n_rows);
 
             void *vtcm_weight_bufs[2] = { vtcm_scratch0, vtcm_scratch1 };
             void *vtcm_output_bufs[2] = { vtcm_output,   vtcm_scratch2 };
@@ -2822,8 +2908,8 @@ static int hmx_mm_nx_2d_f32(struct htp_ops_context * octx, const struct htp_mm_k
         }
     } else {
         hmx_matmul_job_t job;
-        for (size_t mr = 0; mr < (size_t) m; mr += m_chunk_n_rows) {
-            const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+        for (size_t mr = (size_t) m_start; mr < mr_end; mr += m_chunk_n_rows) {
+            const size_t n_rows = hex_smin(mr_end - mr, m_chunk_n_rows);
 
             struct activation_transfer_params act_params = {
                 .ctx = ctx,
@@ -3095,7 +3181,7 @@ static int hmx_mm_f16_f32_batched(struct htp_context *ctx, const hmx_mm_f16_f32_
                             int chunk_dst_cols = params->n - (int)nc;
                             if (chunk_dst_cols > 0) {
                                 transfer_output_chunk_threaded(ctx, output, src2_chunk, vtcm_output, (int) n_rows, (int) n_cols,
-                                                               params->dst_stride, params->src2_stride, chunk_dst_cols, ctx->n_threads);
+                                                               params->dst_stride, params->src2_stride, chunk_dst_cols, n_threads);
                             }
                         }
                     }
@@ -3216,7 +3302,10 @@ static int hmx_mm_id_2d_f32(struct htp_context *ctx,
                                          int weight_type,
                                          const struct mmid_row_mapping *matrix_rows,
                                          int cur_a,
-                                         int mapping_stride) {
+                                         int mapping_stride,
+                                         int m_start,
+                                         int m_end,
+                                         int n_threads) {
     struct htp_thread_trace * tr = &ctx->trace[0];
     htp_trace_event_start(tr, HTP_TRACE_EVT_INIT, 0);
 
@@ -3247,7 +3336,6 @@ static int hmx_mm_id_2d_f32(struct htp_context *ctx,
     const int n_k_tiles = k / HTP_MM_HMX_TILE_N_COLS;
     const struct fastdiv_values n_k_tiles_div = init_fastdiv_values(n_k_tiles);
 
-    const int n_threads = ctx->n_threads;
     const bool is_quant   = (weight_type != HTP_TYPE_F16 && weight_type != HTP_TYPE_F32);
 
     const size_t vec_dot_size = k * sizeof(__fp16);
@@ -3303,8 +3391,8 @@ static int hmx_mm_id_2d_f32(struct htp_context *ctx,
 
     hmx_matmul_job_t job;
 
-    for (size_t mr = 0; mr < (size_t) m_padded; mr += m_chunk_n_rows) {
-        const size_t n_rows = hex_smin(m_padded - mr, m_chunk_n_rows);
+    for (size_t mr = (size_t) m_start; mr < (size_t) m_end; mr += m_chunk_n_rows) {
+        const size_t n_rows = hex_smin((size_t) m_end - mr, m_chunk_n_rows);
         const size_t n_row_tiles = hmx_ceil_div(n_rows, HTP_MM_HMX_TILE_N_ROWS);
 
         transfer_activation_chunk_gathered_threaded(
@@ -3368,31 +3456,48 @@ static int hmx_mm_op_matmul(struct htp_ops_context * octx, const struct htp_mm_k
     const int act_stride = (int)(src1->nb[1] / sizeof(float));
     const int wgt_stride = (int)(src0->nb[1] / sizeof(__fp16));
 
+    int m_start = 0;
+    int m_rows  = m_total;
+    if (octx->ctx->mdev.count > 1) {
+        const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition((uint32_t) m_total, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        m_start = (int) range.start;
+        m_rows  = (int) range.count;
+    }
+
+    if (m_rows == 0) {
+        return HTP_STATUS_OK;
+    }
+
     const float * src2_ptr = NULL;
     uint32_t src2_stride = 0;
     size_t src2_nb2 = 0;
     size_t src2_nb3 = 0;
     if (src2) {
-        src2_ptr = (const float *) src2->data;
         src2_stride = (src2->ne[1] == 1) ? 0 : (uint32_t) (src2->nb[1] / sizeof(float));
+        src2_ptr = (const float *) src2->data + m_start * src2_stride;
         src2_nb2 = (src2->ne[2] == 1) ? 0 : src2->nb[2];
         src2_nb3 = (src2->ne[3] == 1) ? 0 : src2->nb[3];
     }
 
+    const int dst_stride = (int)(dst->nb[1] / sizeof(float));
+    float       * dst_ptr = (float *)       dst->data  + m_start * dst_stride;
+    const float * act_ptr = (const float *) src1->data + m_start * act_stride;
+
     int ret = -1;
-    const int n_threads = MIN(kparams->n_threads, (int) octx->n_threads);
+    const int n_threads = kparams->n_threads;
     if (kparams->kernel_type == HTP_MM_KERNEL_HMX_F16_BATCHED) {
         hmx_mm_f16_f32_batched_params_t batch_params = {
-            .dst             = (float *) dst->data,
+            .dst             = dst_ptr,
             .src2            = src2_ptr,
-            .activation      = (float *) src1->data,
+            .activation      = act_ptr,
             .weight          = (const __fp16 *) src0->data,
-            .m               = m_total,
+            .m               = m_rows,
             .k               = k,
             .n               = n,
             .act_stride      = act_stride,
             .weight_stride   = wgt_stride,
-            .dst_stride      = (int) (dst->nb[1] / sizeof(float)),
+            .dst_stride      = dst_stride,
             .src2_stride     = src2_stride,
             .ne02            = ne02,
             .ne03            = ne03,
@@ -3420,9 +3525,9 @@ static int hmx_mm_op_matmul(struct htp_ops_context * octx, const struct htp_mm_k
                                      kparams->vtcm_size);
     } else {
         ret = hmx_mm_2d_f32(
-            octx->ctx, (float*) dst->data, src2_ptr, (float*) src1->data, (const uint8_t *) src0->data,
-            m_total, k, n, act_stride, (int) src0->nb[1], (int) src0->type, (int) src1->ne[0],
-            (int)(dst->nb[1] / sizeof(float)), src2_stride, (int)dst->ne[0],
+            octx->ctx, dst_ptr, src2_ptr, act_ptr, (const uint8_t *) src0->data,
+            m_rows, k, n, act_stride, (int) src0->nb[1], (int) src0->type, (int) src1->ne[0],
+            dst_stride, src2_stride, (int)dst->ne[0],
             kparams->m_chunk, kparams->n_chunk, kparams->pipeline, n_threads,
             kparams->n_act_threads,
             &kparams->div_n_act_threads,
@@ -3441,6 +3546,11 @@ static int hmx_mm_op_matmul(struct htp_ops_context * octx, const struct htp_mm_k
 int op_matmul(struct htp_ops_context * octx) {
     const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
 
+    const int status = htp_mm_init_context(octx, kparams);
+    if (status != HTP_STATUS_OK) {
+        return status;
+    }
+
     if (kparams->n_hmx) {
         return hmx_mm_op_matmul(octx, kparams);
     }
@@ -3463,6 +3573,16 @@ static int hmx_mm_op_matmul_id(
         const int32_t cne1 = matrix_row_counts[cur_a];
         if (cne1 == 0) continue;
 
+        const int m_padded = hex_align_up(cne1, 32);
+        int m_start = 0, m_end = m_padded;
+        if (octx->ctx->mdev.count > 1) {
+            const bool can_split = htp_tensor_mdev_data_aligned(dst) && (uint32_t) cne1 >= octx->ctx->mdev.count;
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition((uint32_t) m_padded, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            m_start = (int) range.start;
+            m_end   = (int) (range.start + range.count);
+        }
+        if (m_start >= m_end) continue;
+
         int ret = hmx_mm_id_2d_f32(octx->ctx, (float*) dst->data, (float*) src1->data,
                                    (const uint8_t *) src0->data + cur_a * nb02,
                                    cne1, ne00, ne01,
@@ -3471,7 +3591,8 @@ static int hmx_mm_op_matmul_id(
                                    nb11, nb12,
                                    nb1, nb2,
                                    (int) src0->nb[1], (int) src0->type,
-                                   matrix_rows, cur_a, mmctx->mapping_stride);
+                                   matrix_rows, cur_a, mmctx->mapping_stride,
+                                   m_start, m_end, (int) octx->n_threads);
         if (ret != 0) {
             FARF(ERROR, "HMX matmul failed for expert %u, error %d\n", cur_a, ret);
             return HTP_STATUS_NO_SUPPORT;
@@ -3524,7 +3645,7 @@ static int hvx_mm_matmul_id(
     htp_mm_hvx_vtcm_layout_build(&L, kparams->kernel_type, src0->type, ne10, src1_nrows, octx->n_threads,
                                  0, src0_row_size, src1_row_size, 0, kparams->n_prefetch, true, false);
 
-    size_t vtcm_size = kparams->vtcm_size > 0 ? (size_t)kparams->vtcm_size : L.total_bytes;
+    const size_t vtcm_size = L.total_bytes;
 
     FARF(HIGH, "matmul-id-%s : src0-spad-size %zu src1-spad-size %zu src2-spad-size 0 dst-spad-size %zu (%zu)\n", mmctx->type,
          L.src0_bytes, L.src1_bytes, L.dst_bytes, vtcm_size);
@@ -3554,10 +3675,10 @@ static int hvx_mm_matmul_id(
     mmctx->vtcm_src0_stride = src0_row_size_padded;
     mmctx->vtcm_src1_stride = src1_row_size;
 
-    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->n_threads_div);
     mmctx->vtcm_src1_size_per_thread = L.src1_bytes;
     mmctx->vtcm_src2_size_per_thread = 0;
-    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->n_threads_div);
 
     mmctx->n_quant_rows_per_thread = (src1_nrows + n_quant_tasks - 1) / n_quant_tasks;
     mmctx->quant_task_func = quant_task_func;
@@ -3587,6 +3708,20 @@ static int hmx_mm_op_matmul_id_nx(
         const int32_t cne1 = matrix_row_counts[cur_a];
         if (cne1 == 0) continue;
 
+        const int m_padded = hex_align_up(cne1, 32);
+        int m_start = 0, m_end = m_padded;
+        if (octx->ctx->mdev.count > 1) {
+            bool can_split = (uint32_t) cne1 >= octx->ctx->mdev.count;
+            for (uint32_t p = 0; p < n_weights && can_split; ++p) {
+                const struct htp_tensor * restrict dst = octx->dsts[p];
+                can_split = !dst || htp_tensor_mdev_data_aligned(dst);
+            }
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition((uint32_t) m_padded, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            m_start = (int) range.start;
+            m_end   = (int) (range.start + range.count);
+        }
+        if (m_start >= m_end) continue;
+
         for (uint32_t p = 0; p < n_weights; ++p) {
             const struct htp_tensor * restrict src_w = octx->src[p];
             const struct htp_tensor * restrict dst   = octx->dsts[p];
@@ -3600,7 +3735,8 @@ static int hmx_mm_op_matmul_id_nx(
                                        act->nb[1], act->nb[2],
                                        dst->nb[1], dst->nb[2],
                                        (int) src_w->nb[1], (int) src_w->type,
-                                       matrix_rows, cur_a, mmctx->mapping_stride);
+                                       matrix_rows, cur_a, mmctx->mapping_stride,
+                                       m_start, m_end, (int) octx->n_threads);
             if (ret != 0) {
                 FARF(ERROR, "HMX matmul ID NX failed for expert %u weight %u, error %d\n", cur_a, p, ret);
                 return HTP_STATUS_NO_SUPPORT;
@@ -3656,7 +3792,7 @@ static int hvx_mm_matmul_id_nx(
     htp_mm_hvx_vtcm_layout_build(&L, kparams->kernel_type, src0->type, act->ne[0], src1_nrows, octx->n_threads,
                                  0, src0_row_size, src1_row_size, 0, kparams->n_prefetch, true, false);
 
-    size_t vtcm_size = kparams->vtcm_size > 0 ? (size_t)kparams->vtcm_size : L.total_bytes;
+    const size_t vtcm_size = L.total_bytes;
 
     if (octx->ctx->vtcm_size < vtcm_size) {
         FARF(ERROR, "matmul-id-nx: current VTCM reservation %zu is too small, needed %zu\n",
@@ -3678,9 +3814,9 @@ static int hvx_mm_matmul_id_nx(
     mmctx->vtcm_src0_stride = 0;
     mmctx->vtcm_src1_stride = src1_row_size;
 
-    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->n_threads_div);
     mmctx->vtcm_src1_size_per_thread = L.src1_bytes;
-    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->n_threads_div);
 
     mmctx->n_quant_rows_per_thread = (src1_nrows + n_quant_tasks - 1) / n_quant_tasks;
     mmctx->quant_task_func         = quant_task_func;
@@ -3769,16 +3905,21 @@ static inline void scan_expert_ids(
 int op_matmul_id(struct htp_ops_context * octx) {
     htp_matmul_tensors_preamble;
 
+    const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
+    struct htp_mm_context mmctx_struct = {0};
+    struct htp_mm_context * mmctx = &mmctx_struct;
+
+    const int status = htp_mm_init_context(octx, kparams);
+    if (status != HTP_STATUS_OK) {
+        return status;
+    }
+
     struct htp_thread_trace * tr = &octx->ctx->trace[0];
     htp_trace_event_start(tr, HTP_TRACE_EVT_INIT, 0);
 
-    struct htp_mm_context mmctx_struct = {0};
-    struct htp_mm_context * mmctx = &mmctx_struct;
     mmctx->octx = octx;
     mmctx->act = src1;
 
-    const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
-
     const struct htp_tensor * restrict ids = octx->src[2];
 
     const size_t src0_row_size = nb01;
@@ -3789,9 +3930,6 @@ int op_matmul_id(struct htp_ops_context * octx) {
     const uint32_t src0_nrows = ne01;  // per expert
     const uint32_t src1_nrows = ne11 * ne12 * ne13;
 
-    mmctx->src0_nrows_per_thread = fastdiv(src0_nrows + octx->n_threads - 1, &octx->ctx->n_threads_div);
-    mmctx->src0_nrows_per_thread = hex_round_up(mmctx->src0_nrows_per_thread, 32);
-
     // row groups
     const int n_ids = ids->ne[0];  // n_expert_used
     const int n_as  = ne02;        // n_expert
@@ -3843,6 +3981,29 @@ int op_matmul_id(struct htp_ops_context * octx) {
     if (kparams->n_hmx) {
         s = hmx_mm_op_matmul_id(octx, mmctx);
     } else {
+        uint32_t src0_row_start = 0;
+        uint32_t src0_row_end   = src0_nrows;
+        if (octx->ctx->mdev.count > 1) {
+            const bool can_split = htp_tensor_can_row_partition(dst, sizeof(float));
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(src0_nrows, can_split ? 32 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            src0_row_start = range.start;
+            src0_row_end   = range.start + range.count;
+        }
+
+        if (src0_row_start >= src0_row_end) {
+            if (mapping_buf != octx->ctx->ddr_spad_base) {
+                free(mapping_buf);
+            }
+            return HTP_STATUS_OK;
+        }
+
+        const uint32_t nrows = src0_row_end - src0_row_start;
+        mmctx->src0_row_start = src0_row_start;
+        mmctx->src0_row_end   = src0_row_end;
+
+        mmctx->src0_nrows_per_thread = fastdiv(nrows + octx->n_threads - 1, &octx->n_threads_div);
+        mmctx->src0_nrows_per_thread = hex_round_up(mmctx->src0_nrows_per_thread, 32);
+
         if (hvx_mm_init_vec_dot(mmctx, src0->type) == 0) {
             s = hvx_mm_matmul_id(octx, mmctx, src1_nrows > 1 ? hvx_mm_id : hvx_mv_id);
         } else {
@@ -3858,29 +4019,31 @@ int op_matmul_id(struct htp_ops_context * octx) {
 }
 
 int op_matmul_id_nx(struct htp_ops_context * octx) {
+    const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
+    struct htp_mm_context mmctx_struct = {0};
+    struct htp_mm_context * mmctx = &mmctx_struct;
+
+    const int status = htp_mm_init_context(octx, kparams);
+    if (status != HTP_STATUS_OK) {
+        return status;
+    }
+
     struct htp_thread_trace * tr = &octx->ctx->trace[0];
     htp_trace_event_start(tr, HTP_TRACE_EVT_INIT, 0);
 
-    const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
+    mmctx->octx = octx;
     const uint32_t n_weights = kparams->n_weights;
     const struct htp_tensor * restrict src0 = octx->src[0];
     const struct htp_tensor * restrict act  = octx->src[n_weights];
     const struct htp_tensor * restrict ids  = octx->src[n_weights + 1];
 
-    struct htp_mm_context mmctx_struct = {0};
-    struct htp_mm_context * mmctx = &mmctx_struct;
-    mmctx->octx = octx;
     mmctx->act = act;
 
     const size_t src0_row_size = src0->nb[1];
     const size_t src0_row_size_padded = hex_round_up(src0_row_size, 128);
 
-    const uint32_t src0_nrows = src0->ne[1];
     const uint32_t src1_nrows = act->ne[1] * act->ne[2] * act->ne[3];
 
-    mmctx->src0_nrows_per_thread = fastdiv(src0_nrows + octx->n_threads - 1, &octx->ctx->n_threads_div);
-    mmctx->src0_nrows_per_thread = hex_round_up(mmctx->src0_nrows_per_thread, 32);
-
     const int n_ids = ids->ne[0];
     const int n_as  = src0->ne[2];
 
@@ -3946,6 +4109,12 @@ int op_matmul_id_nx(struct htp_ops_context * octx) {
 }
 int op_matmul_nx(struct htp_ops_context * octx) {
     const struct htp_mm_kernel_params * kparams = (const struct htp_mm_kernel_params *) octx->kernel_params;
+
+    const int status = htp_mm_init_context(octx, kparams);
+    if (status != HTP_STATUS_OK) {
+        return status;
+    }
+
     if (kparams->n_hmx) {
         return hmx_mm_nx_2d_f32(octx, kparams);
     }
@@ -4012,7 +4181,7 @@ int op_matmul_nx(struct htp_ops_context * octx) {
     htp_mm_hvx_vtcm_layout_build(&L, kparams->kernel_type, src0->type, act->ne[0], src1_nrows, octx->n_threads,
                                  0, src0_row_size, src1_row_size, 0, kparams->n_prefetch, false, true);
 
-    size_t vtcm_size = kparams->vtcm_size > 0 ? (size_t)kparams->vtcm_size : L.total_bytes;
+    const size_t vtcm_size = L.total_bytes;
 
     if (octx->ctx->vtcm_size < vtcm_size) {
         FARF(ERROR, "matmul-nx: current VTCM reservation %zu is too small, needed %zu\n",
@@ -4034,9 +4203,9 @@ int op_matmul_nx(struct htp_ops_context * octx) {
     mmctx->vtcm_src0_stride = is_repacked ? 0 : src0_row_size_padded;
     mmctx->vtcm_src1_stride = src1_row_size;
 
-    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_src0_size_per_thread = fastdiv(L.src0_bytes, &octx->n_threads_div);
     mmctx->vtcm_src1_size_per_thread = L.src1_bytes;
-    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->ctx->n_threads_div);
+    mmctx->vtcm_dst_size_per_thread  = fastdiv(L.dst_bytes, &octx->n_threads_div);
 
     mmctx->n_quant_rows_per_thread = (src1_nrows + n_quant_tasks - 1) / n_quant_tasks;
     mmctx->quant_task_func = quant_task_func;

ggml/src/ggml-hexagon/htp/pad-ops.c

@@ -12,8 +12,11 @@
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
+#include "htp-tensor.h"
 
 /* Circular wrap: maps any integer x into [0, n) */
 static inline uint32_t wrap_around(int32_t x, uint32_t n) {
@@ -68,6 +71,7 @@ struct htp_pad_context {
 
     uint32_t nrows_per_thread;
     uint32_t total_dst_rows;
+    uint32_t row_start;
 
     size_t   type_size;
 
@@ -78,39 +82,39 @@ struct htp_pad_context {
     size_t   dst_row_size_aligned;
 };
 
-#define htp_pad_preamble                            \
-    const struct htp_tensor * src = octx->src[0];   \
-    const struct htp_tensor * dst = octx->dst;      \
-                                                    \
-    const uint32_t ne00 = src->ne[0];               \
-    const uint32_t nb00 = src->nb[0];               \
-                                                    \
-    const uint32_t ne0 = dst->ne[0];                \
-    const uint32_t ne1 = dst->ne[1];                \
-    const uint32_t ne2 = dst->ne[2];                \
-    const uint32_t ne3 = dst->ne[3];                \
-                                                    \
-    const uint32_t nb1 = dst->nb[1];                \
-    const uint32_t nb2 = dst->nb[2];                \
-    const uint32_t nb3 = dst->nb[3];                \
-                                                    \
-    const int32_t lp0 = pctx->lp0, rp0 = pctx->rp0; \
-    const int32_t lp1 = pctx->lp1, rp1 = pctx->rp1; \
-    const int32_t lp2 = pctx->lp2, rp2 = pctx->rp2; \
-    const int32_t lp3 = pctx->lp3, rp3 = pctx->rp3; \
-                                                    \
-    const size_t type_size = pctx->type_size;       \
-                                                    \
-    const uint32_t row_start = pctx->nrows_per_thread * ith;                                 \
-    const uint32_t row_end   = MIN(row_start + pctx->nrows_per_thread, pctx->total_dst_rows);
+#define htp_pad_preamble                                                             \
+    const struct htp_tensor * src = octx->src[0];                                    \
+    const struct htp_tensor * dst = octx->dst;                                       \
+                                                                                     \
+    const uint32_t ne00 = src->ne[0];                                                \
+    const uint32_t nb00 = src->nb[0];                                                \
+                                                                                     \
+    const uint32_t ne0 = dst->ne[0];                                                 \
+    const uint32_t ne1 = dst->ne[1];                                                 \
+    const uint32_t ne2 = dst->ne[2];                                                 \
+    const uint32_t ne3 = dst->ne[3];                                                 \
+                                                                                     \
+    const uint32_t nb1 = dst->nb[1];                                                 \
+    const uint32_t nb2 = dst->nb[2];                                                 \
+    const uint32_t nb3 = dst->nb[3];                                                 \
+                                                                                     \
+    const int32_t lp0 = pctx->lp0, rp0 = pctx->rp0;                                  \
+    const int32_t lp1 = pctx->lp1, rp1 = pctx->rp1;                                  \
+    const int32_t lp2 = pctx->lp2, rp2 = pctx->rp2;                                  \
+    const int32_t lp3 = pctx->lp3, rp3 = pctx->rp3;                                  \
+                                                                                     \
+    const size_t type_size = pctx->type_size;                                        \
+                                                                                     \
+    const uint32_t row_start = pctx->row_start + pctx->nrows_per_thread * ith;       \
+    const uint32_t row_end   = MIN(row_start + pctx->nrows_per_thread, pctx->row_start + pctx->total_dst_rows);
 
 
-#define htp_pad_dma_preamble                                        \
-    const size_t src_row_size         = pctx->src_row_size;         \
-    const size_t src_row_size_aligned = pctx->src_row_size_aligned; \
-    const size_t dst_row_size         = pctx->dst_row_size;         \
-    const size_t dst_row_size_aligned = pctx->dst_row_size_aligned; \
-                                                                    \
+#define htp_pad_dma_preamble                                                                \
+    const size_t src_row_size         = pctx->src_row_size;                                 \
+    const size_t src_row_size_aligned = pctx->src_row_size_aligned;                         \
+    const size_t dst_row_size         = pctx->dst_row_size;                                 \
+    const size_t dst_row_size_aligned = pctx->dst_row_size_aligned;                         \
+                                                                                            \
     uint8_t * src_spad_base = octx->src0_spad.data + ith * octx->src0_spad.size_per_thread; \
     uint8_t * dst_spad_base = octx->dst_spad.data  + ith * octx->dst_spad.size_per_thread;  \
                                                                                             \
@@ -125,8 +129,8 @@ static void pad_job_per_thread_hvx(unsigned int nth, unsigned int ith, void * da
     struct htp_ops_context * octx = pctx->octx;
     htp_pad_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
 
     for (uint32_t dst_row = row_start; dst_row < row_end; dst_row++) {
         uint32_t i1, i2, i3;
@@ -165,18 +169,17 @@ static void pad_job_per_thread_hvx(unsigned int nth, unsigned int ith, void * da
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
 
-    FARF(HIGH, "pad-hvx %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u usec %u\n",
+    FARF(HIGH, "pad-hvx %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u\n",
          ith, nth,
          src->ne[0], src->ne[1], src->ne[2], src->ne[3],
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         row_start, row_end,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         row_start, row_end);
 }
 
 // ---------------------------------------------------------------------------
-// HVX + DMA PAD kernel — aligned, double-buffered
+// HVX + DMA PAD kernel - aligned, double-buffered
 // ---------------------------------------------------------------------------
 
 static void pad_job_per_thread_hvx_dma(unsigned int nth, unsigned int ith, void * data) {
@@ -185,9 +188,6 @@ static void pad_job_per_thread_hvx_dma(unsigned int nth, unsigned int ith, void
     htp_pad_preamble;
     htp_pad_dma_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
     // -----------------------------------------------------------------------
     // Priming phase: push 2 pairs of (dummy_dst_DMA, src_DMA) to seed the
     // double-buffer pipeline before the main loop begins.
@@ -222,6 +222,8 @@ static void pad_job_per_thread_hvx_dma(unsigned int nth, unsigned int ith, void
     // Main loop: pop completed DMAs, compute in VTCM with aligned HVX ops,
     // push dst DMA and prefetch src for the next+1 row.
     // -----------------------------------------------------------------------
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = row_start; ir < row_end; ir++) {
         uint8_t * dst_spad_cur = (uint8_t *) dma_queue_pop(dma).src;
         uint8_t * src_spad_cur = (uint8_t *) dma_queue_pop(dma).dst;
@@ -236,6 +238,7 @@ static void pad_job_per_thread_hvx_dma(unsigned int nth, unsigned int ith, void
                                              lp2, rp2, ne2,
                                              lp3, rp3, ne3);
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         if (!interior) {
             hvx_splat_f32_a(dst_spad_cur, 0.0f, ne0);
         } else {
@@ -249,6 +252,7 @@ static void pad_job_per_thread_hvx_dma(unsigned int nth, unsigned int ith, void
                 hvx_copy_f32_ua(dst_interior, src_spad_cur, ne00);
             }
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         dma_queue_push_vtcm_to_ddr(dma,
             dma_make_ptr(dst_ptr, dst_spad_cur),
@@ -274,14 +278,11 @@ static void pad_job_per_thread_hvx_dma(unsigned int nth, unsigned int ith, void
 
     dma_queue_flush(dma);
 
-    t2 = HAP_perf_get_qtimer_count();
-
-    FARF(HIGH, "pad-hvx-dma %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u usec %u\n",
+    FARF(HIGH, "pad-hvx-dma %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u\n",
          ith, nth,
          src->ne[0], src->ne[1], src->ne[2], src->ne[3],
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         row_start, row_end,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         row_start, row_end);
 }
 
 // ---------------------------------------------------------------------------
@@ -293,8 +294,8 @@ static void pad_job_per_thread_hvx_circular(unsigned int nth, unsigned int ith,
     struct htp_ops_context * octx = pctx->octx;
     htp_pad_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
 
     for (uint32_t dst_row = row_start; dst_row < row_end; dst_row++) {
         uint32_t i1, i2, i3;
@@ -344,18 +345,17 @@ static void pad_job_per_thread_hvx_circular(unsigned int nth, unsigned int ith,
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
 
-    FARF(HIGH, "pad-hvx-circ %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u usec %u\n",
+    FARF(HIGH, "pad-hvx-circ %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u\n",
          ith, nth,
          src->ne[0], src->ne[1], src->ne[2], src->ne[3],
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         row_start, row_end,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         row_start, row_end);
 }
 
 // ---------------------------------------------------------------------------
-// HVX + DMA circular PAD kernel — aligned, double-buffered
+// HVX + DMA circular PAD kernel - aligned, double-buffered
 // ---------------------------------------------------------------------------
 
 static void pad_job_per_thread_hvx_circular_dma(unsigned int nth, unsigned int ith, void * data) {
@@ -364,9 +364,6 @@ static void pad_job_per_thread_hvx_circular_dma(unsigned int nth, unsigned int i
     htp_pad_preamble;
     htp_pad_dma_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
     // -----------------------------------------------------------------------
     // Priming phase: push 2 pairs of (dummy_dst_DMA, src_DMA) to seed the
     // double-buffer pipeline.  Every row is a real src DMA (no null DMAs).
@@ -390,6 +387,8 @@ static void pad_job_per_thread_hvx_circular_dma(unsigned int nth, unsigned int i
     // Main loop: pop completed DMAs, assemble circular row in VTCM with
     // aligned HVX ops, push dst DMA and prefetch src for the next+1 row.
     // -----------------------------------------------------------------------
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+
     for (uint32_t ir = row_start; ir < row_end; ir++) {
         uint8_t * dst_spad_cur = (uint8_t *) dma_queue_pop(dma).src;
         uint8_t * src_spad_cur = (uint8_t *) dma_queue_pop(dma).dst;
@@ -398,7 +397,7 @@ static void pad_job_per_thread_hvx_circular_dma(unsigned int nth, unsigned int i
         pad_decompose_row(ir, ne1, ne2, &i1, &i2, &i3);
         uint8_t * dst_ptr = (uint8_t *) dst->data + i1 * nb1 + i2 * nb2 + i3 * nb3;
 
-
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
         if (lp0 > 0) {
             uint8_t * dst_left       = dst_spad_cur;
             const uint8_t * src_left = src_spad_cur + (size_t)(ne00 - (uint32_t)lp0) * type_size;
@@ -430,6 +429,7 @@ static void pad_job_per_thread_hvx_circular_dma(unsigned int nth, unsigned int i
                 }
             }
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);
 
         dma_queue_push_vtcm_to_ddr(dma,
             dma_make_ptr(dst_ptr, dst_spad_cur),
@@ -448,14 +448,11 @@ static void pad_job_per_thread_hvx_circular_dma(unsigned int nth, unsigned int i
 
     dma_queue_flush(dma);
 
-    t2 = HAP_perf_get_qtimer_count();
-
-    FARF(HIGH, "pad-hvx-circ-dma %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u usec %u\n",
+    FARF(HIGH, "pad-hvx-circ-dma %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u\n",
          ith, nth,
          src->ne[0], src->ne[1], src->ne[2], src->ne[3],
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         row_start, row_end,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         row_start, row_end);
 }
 
 int op_pad(struct htp_ops_context * octx) {
@@ -489,19 +486,33 @@ int op_pad(struct htp_ops_context * octx) {
     const uint32_t ne00 = src0->ne[0];
 
     const uint32_t total_dst_rows = dst->ne[1] * dst->ne[2] * dst->ne[3];
-    const uint32_t n_threads = MIN(octx->n_threads, total_dst_rows > 0 ? total_dst_rows : 1);
+    const size_t dst_row_size     = (size_t)ne0 * type_size;
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = total_dst_rows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, type_size, (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_dst_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
 
     const size_t src_row_size         = (size_t)ne00 * type_size;
-    const size_t dst_row_size         = (size_t)ne0  * type_size;
     const size_t src_row_size_aligned = hex_round_up(src_row_size, VLEN);
     const size_t dst_row_size_aligned = hex_round_up(dst_row_size, VLEN);
 
     // Total VTCM needed: 2 buffers (ping+pong) for src and dst, per thread
     const size_t vtcm_needed = (size_t)n_threads * 2 * (src_row_size_aligned + dst_row_size_aligned);
 
-    const int use_dma = (src0->nb[0] == (uint32_t)type_size) &&
-                        (ne00 >= 512) &&
-                        (octx->ctx->vtcm_base != NULL) &&
+    const int use_dma = (src0->nb[0] == (uint32_t)type_size) && (ne00 >= 512) &&
                         (octx->ctx->vtcm_size >= vtcm_needed);
 
     if (use_dma) {
@@ -521,8 +532,9 @@ int op_pad(struct htp_ops_context * octx) {
         .lp1 = lp1, .rp1 = rp1,
         .lp2 = lp2, .rp2 = rp2,
         .lp3 = lp3, .rp3 = rp3,
-        .nrows_per_thread = (total_dst_rows + n_threads - 1) / n_threads,
-        .total_dst_rows   = total_dst_rows,
+        .nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div),
+        .total_dst_rows   = nrows,
+        .row_start        = row_start,
         .type_size        = type_size,
         .src_row_size         = src_row_size,
         .src_row_size_aligned = src_row_size_aligned,
@@ -537,11 +549,10 @@ int op_pad(struct htp_ops_context * octx) {
          dst->ne[0],  dst->ne[1],  dst->ne[2],  dst->ne[3],
          lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
 
-    if      (circular && use_dma) { worker_pool_run_func(octx->ctx->worker_pool, pad_job_per_thread_hvx_circular_dma, &pctx, n_threads); }
-    else if (circular)            { worker_pool_run_func(octx->ctx->worker_pool, pad_job_per_thread_hvx_circular,     &pctx, n_threads); }
-    else if (use_dma)             { worker_pool_run_func(octx->ctx->worker_pool, pad_job_per_thread_hvx_dma,          &pctx, n_threads); }
-    else                          { worker_pool_run_func(octx->ctx->worker_pool, pad_job_per_thread_hvx,              &pctx, n_threads); }
+    if      (circular && use_dma) { work_queue_run(octx->ctx->work_queue, pad_job_per_thread_hvx_circular_dma, &pctx, n_threads); }
+    else if (circular)            { work_queue_run(octx->ctx->work_queue, pad_job_per_thread_hvx_circular,     &pctx, n_threads); }
+    else if (use_dma)             { work_queue_run(octx->ctx->work_queue, pad_job_per_thread_hvx_dma,          &pctx, n_threads); }
+    else                          { work_queue_run(octx->ctx->work_queue, pad_job_per_thread_hvx,              &pctx, n_threads); }
 
     return HTP_STATUS_OK;
 }
-

ggml/src/ggml-hexagon/htp/repeat-ops.c

@@ -12,8 +12,10 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
+#include "htp-tensor.h"
 
 struct htp_repeat_context {
     struct htp_ops_context * octx;
@@ -25,6 +27,7 @@ struct htp_repeat_context {
 
     uint32_t nrows_per_thread;
     uint32_t total_dst_rows;  // ne1 * ne2 * ne3
+    uint32_t row_start;
 
     size_t   type_size;
 };
@@ -62,11 +65,11 @@ static void repeat_job_per_thread(unsigned int nth, unsigned int ith, void * dat
 
     const size_t row_bytes = ne00 * rctx->type_size;
 
-    const uint32_t row_start = rctx->nrows_per_thread * ith;
-    const uint32_t row_end   = MIN(row_start + rctx->nrows_per_thread, rctx->total_dst_rows);
+    const uint32_t row_start = rctx->row_start + rctx->nrows_per_thread * ith;
+    const uint32_t row_end   = MIN(row_start + rctx->nrows_per_thread, rctx->row_start + rctx->total_dst_rows);
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
 
     for (uint32_t dst_row = row_start; dst_row < row_end; dst_row++) {
         // Decompose flat dst row index into (i1, i2, i3)
@@ -89,12 +92,12 @@ static void repeat_job_per_thread(unsigned int nth, unsigned int ith, void * dat
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, row_start);
 
-    FARF(HIGH, "repeat %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u usec %u\n",
+    FARF(HIGH, "repeat %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u\n",
          ith, nth, src->ne[0], src->ne[1], src->ne[2], src->ne[3],
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-         row_start, row_end, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         row_start, row_end);
 }
 
 int op_repeat(struct htp_ops_context * octx) {
@@ -119,21 +122,39 @@ int op_repeat(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t total_dst_rows = dst->ne[1] * dst->ne[2] * dst->ne[3];
-    const uint32_t n_threads = MIN(octx->n_threads, total_dst_rows);
-
     if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
         return HTP_STATUS_OK;
     }
 
+    const uint32_t total_dst_rows  = dst->ne[1] * dst->ne[2] * dst->ne[3];
+    const size_t dst_row_size = dst->ne[0] * type_size;
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = total_dst_rows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, type_size, (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_dst_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
     struct htp_repeat_context rctx = {
         .octx             = octx,
         .nr0              = dst->ne[0] / src0->ne[0],
         .nr1              = dst->ne[1] / src0->ne[1],
         .nr2              = dst->ne[2] / src0->ne[2],
         .nr3              = dst->ne[3] / src0->ne[3],
-        .nrows_per_thread = (total_dst_rows + n_threads - 1) / n_threads,
-        .total_dst_rows   = total_dst_rows,
+        .nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div),
+        .total_dst_rows   = nrows,
+        .row_start        = row_start,
         .type_size        = type_size,
     };
 
@@ -142,7 +163,7 @@ int op_repeat(struct htp_ops_context * octx) {
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
          rctx.nr0, rctx.nr1, rctx.nr2, rctx.nr3);
 
-    worker_pool_run_func(octx->ctx->worker_pool, repeat_job_per_thread, &rctx, n_threads);
+    work_queue_run(octx->ctx->work_queue, repeat_job_per_thread, &rctx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -80,6 +80,8 @@ struct htp_rope_context {
     size_t dst_row_stride;
     size_t src0_row_size_aligned;
     uint32_t src0_nrows;
+    uint32_t row_start;
+    uint32_t nrows;
 
     struct fastdiv_values div_ne2_ne1;
     struct fastdiv_values div_ne1;
@@ -539,11 +541,11 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
 
     htp_rope_preamble;
 
-    const uint32_t src0_nrows = rctx->src0_nrows;
+    const uint32_t src0_nrows = rctx->nrows;
     const uint32_t src0_nrows_per_thread = rctx->src0_nrows_per_thread;
 
-    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t src0_start_row = rctx->row_start + src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, rctx->row_start + src0_nrows);
 
     // no work for this thread
     if (src0_start_row >= src0_end_row) {
@@ -706,9 +708,32 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     }
 
     const struct htp_rope_kernel_params * kparams = (const struct htp_rope_kernel_params *) octx->kernel_params;
-    assert(kparams->n_threads > 0);
+    if (!htp_ops_context_set_n_threads(octx, kparams->n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
     assert(octx->ctx->vtcm_size >= kparams->vtcm_size);
 
+    const uint32_t total_rows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const size_t dst_data_row_size = dst->ne[0] * sizeof(float);
+
+    uint32_t row_start = 0;
+    uint32_t nrows     = total_rows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, sizeof(float), (uint32_t) dst_data_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(
+            total_rows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
     const uint32_t ne0 = dst->ne[0];
     const size_t src0_row_size   = src0->ne[0] * sizeof(float);
     const size_t src0_row_stride = src0->nb[1];
@@ -752,15 +777,17 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     rctx.dst_row_stride        = dst_row_stride;
     rctx.src0_row_size_aligned = kparams->src0_row_size_aligned;
 
-    rctx.src0_nrows            = kparams->src0_nrows;
-    rctx.src0_nrows_per_thread = kparams->src0_nrows_per_thread;
+    rctx.src0_nrows            = nrows;
+    rctx.nrows                 = nrows;
+    rctx.row_start             = row_start;
+    rctx.src0_nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
     rctx.div_ne2_ne1           = kparams->div_ne2_ne1;
     rctx.div_ne1               = kparams->div_ne1;
 
     FARF(HIGH, "rope-f32 n-rows %u n-dims %d ne0 %u ext-factor %.6f theta-scale %.6f attn-factor %.6f\n", rctx.src0_nrows, rctx.n_dims, ne0,
          rctx.ext_factor, rctx.theta_scale, rctx.attn_factor);
 
-    work_queue_run(octx->ctx->work_queue, rope_job_f32, &rctx, kparams->n_threads);
+    work_queue_run(octx->ctx->work_queue, rope_job_f32, &rctx, n_threads);
 
     return err;
 }

ggml/src/ggml-hexagon/htp/set-rows-ops.c

@@ -18,6 +18,7 @@
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 
+#include "hex-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
 #include "htp-tensor.h"
@@ -58,6 +59,9 @@ struct set_rows_context {
     const struct htp_set_rows_kernel_params * kparams;
     struct htp_set_rows_vtcm_layout vtcm_layout;
     uint8_t * vtcm_base;
+    uint32_t task_start;
+    uint32_t tasks;
+    uint32_t tasks_per_thread;
 };
 
 #define SET_ROWS_THREAD_DMA_FN(TYPE_NAME, IDX_TYPE, COMPUTE_EXPR)                                                \
@@ -67,12 +71,12 @@ static void set_rows_thread_dma_##TYPE_NAME##_##IDX_TYPE(unsigned int nth, unsig
     const struct htp_set_rows_kernel_params * kparams = srctx->kparams;                                          \
     set_rows_preamble;                                                                                           \
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                       \
-    const uint32_t dr  = kparams->tasks_per_thread;                                                              \
-    const uint32_t ir0 = dr * ith;                                                                               \
-    if (ir0 >= kparams->total_tasks) {                                                                           \
+    const uint32_t dr  = srctx->tasks_per_thread;                                                                \
+    const uint32_t ir0 = srctx->task_start + dr * ith;                                                           \
+    if (ir0 >= srctx->task_start + srctx->tasks) {                                                               \
         return;                                                                                                  \
     }                                                                                                            \
-    const uint32_t ir1 = MIN(ir0 + dr, kparams->total_tasks);                                                    \
+    const uint32_t ir1 = MIN(ir0 + dr, srctx->task_start + srctx->tasks);                                        \
     dma_queue * dma_queue = octx->ctx->dma[ith];                                                                 \
     const struct htp_set_rows_vtcm_layout * vtcm_layout = &srctx->vtcm_layout;                                   \
     uint8_t * vtcm_src0 = srctx->vtcm_base + vtcm_layout->off_src0 + ith * vtcm_layout->src0_bytes_per_thread;   \
@@ -192,18 +196,44 @@ int op_set_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    if (octx->src[1]->type != HTP_TYPE_I32 && octx->src[1]->type != HTP_TYPE_I64) {
-        return HTP_STATUS_NO_SUPPORT;
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
     }
 
+    const struct htp_tensor * dst = octx->dst;
+    const uint32_t total_tasks    = kparams->total_tasks;
+
+    uint32_t task_start = 0;
+    uint32_t tasks      = total_tasks;
+
+    if (octx->ctx->mdev.count > 1) {
+        const bool can_split = htp_tensor_mdev_data_aligned(dst) && (dst->nb[1] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0 && !htp_tensor_is_permuted(dst);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_tasks, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        task_start = range.start;
+        tasks      = range.count;
+    }
+
+    if (tasks == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    if (!htp_ops_context_set_n_threads(octx, (uint32_t) kparams->n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
     // l2fetch the src1 (indices) tensor in the main thread
     hex_l2fetch_block((const void *)octx->src[1]->data, octx->src[1]->ne[3] * octx->src[1]->nb[3]);
 
     struct set_rows_context srctx;
     srctx.octx = octx;
     srctx.kparams = kparams;
+    srctx.task_start = task_start;
+    srctx.tasks = tasks;
+    srctx.tasks_per_thread = fastdiv(tasks + n_threads - 1, &octx->n_threads_div);
 
-    htp_set_rows_vtcm_layout_build(&srctx.vtcm_layout, octx->dst->type, ne00, kparams->n_threads);
+    htp_set_rows_vtcm_layout_build(&srctx.vtcm_layout, octx->dst->type, ne00, n_threads);
     srctx.vtcm_base = (uint8_t *)octx->ctx->vtcm_base;
 
     work_queue_func_t q_func = NULL;
@@ -216,15 +246,15 @@ int op_set_rows(struct htp_ops_context * octx) {
         default:            return HTP_STATUS_NO_SUPPORT;
     }
 
-    FARF(HIGH, "set-rows: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-vtcm-size %zu dst-vtcm-size %zu n_threads %d\n",
+    FARF(HIGH, "set-rows: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-vtcm-size %zu dst-vtcm-size %zu n-threads %d\n",
          octx->src[0]->ne[0], octx->src[0]->ne[1], octx->src[0]->ne[2], octx->src[0]->ne[3],
          octx->src[1]->ne[0], octx->src[1]->ne[1], octx->src[1]->ne[2], octx->src[1]->ne[3],
          octx->dst->ne[0], octx->dst->ne[1], octx->dst->ne[2], octx->dst->ne[3],
-         srctx.vtcm_layout.src0_bytes_per_thread * kparams->n_threads,
-         srctx.vtcm_layout.dst_bytes_per_thread  * kparams->n_threads,
-         kparams->n_threads);
+         srctx.vtcm_layout.src0_bytes_per_thread * n_threads,
+         srctx.vtcm_layout.dst_bytes_per_thread  * n_threads,
+         n_threads);
 
-    work_queue_run(octx->ctx->work_queue, q_func, &srctx, kparams->n_threads);
+    work_queue_run(octx->ctx->work_queue, q_func, &srctx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/softmax-ops.c

@@ -14,9 +14,11 @@
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
+#include "hex-common.h"
+#include "hex-profile.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
+#include "htp-tensor.h"
 
 #define htp_softmax_preamble3                     \
     const uint32_t ne00 = src0->ne[0];            \
@@ -69,6 +71,8 @@ struct htp_softmax_context {
     struct fastdiv_values fastdiv_ne13; // For mask broadcasting
 
     uint32_t src0_nrows_per_thread;
+    uint32_t row_start;
+    uint32_t nrows;
 };
 
 static void apply_mask(float * restrict wp0,
@@ -223,19 +227,17 @@ static void softmax_job_f32(unsigned int nth, unsigned int ith, void * data) {
 
     htp_softmax_preamble3;
 
-    const uint32_t src0_nrows            = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t src0_nrows            = smctx->nrows;
     const uint32_t src0_nrows_per_thread = smctx->src0_nrows_per_thread;
 
-    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t src0_start_row = smctx->row_start + src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, smctx->row_start + src0_nrows);
 
     // no work for this thread
     if (src0_start_row >= src0_end_row) {
         return;
     }
 
-    uint64_t qt = HAP_perf_get_qtimer_count();
-
     int is_aligned = 1;
     int opt_path   = 0;
 
@@ -262,6 +264,9 @@ static void softmax_job_f32(unsigned int nth, unsigned int ith, void * data) {
     uint32_t prev_i2 = (uint32_t)-1;
     float slope = 1.0f;
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, src0_start_row);
+
     for (uint32_t r = src0_start_row; r < src0_end_row; ++r) {
         uint32_t i1 = fastmodulo(r, ne01, &smctx->fastdiv_ne01);
         uint32_t r_div_ne01 = fastdiv(r, &smctx->fastdiv_ne01);
@@ -323,10 +328,11 @@ static void softmax_job_f32(unsigned int nth, unsigned int ith, void * data) {
         }
     }
 
-    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
-    FARF(HIGH, "softmax-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u : opt %u f16 %u usec %u\n", ith, nth,
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, src0_start_row);
+
+    FARF(HIGH, "softmax-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u : opt %u f16 %u\n", ith, nth,
          ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13,
-         ne0, ne1, ne2, ne3, opt_path, smctx->use_f16, (unsigned) qt);
+         ne0, ne1, ne2, ne3, opt_path, smctx->use_f16);
 }
 
 static int execute_op_softmax_f32(struct htp_ops_context * octx) {
@@ -342,13 +348,32 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
     init_softmax_ctx(&smctx, octx);
 
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
-    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
+    const size_t elem_size = sizeof(float);
+    const size_t dst_row_size = dst->nb[1];
 
-    smctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
+    uint32_t row_start = 0;
+    uint32_t nrows     = src0_nrows;
+
+    if (octx->ctx->mdev.count > 1) {
+        uint32_t rows_per_chunk = 0;
+        htp_tensor_mdev_rows_per_chunk(dst, (uint32_t) elem_size, (uint32_t) dst_row_size, &rows_per_chunk);
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(src0_nrows, rows_per_chunk, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
+
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = octx->n_threads;
+
+    smctx.src0_nrows_per_thread = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
+    smctx.row_start             = row_start;
+    smctx.nrows                 = nrows;
 
     const size_t src0_row_size = src0->nb[1];
     const size_t src1_row_size = src0_row_size;
-    const size_t dst_row_size  = dst->nb[1];
 
     // VTCM scratchpads for all tensors
     // 4 rows per thread, padded to HVX vector size
@@ -383,9 +408,7 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
     octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size; octx->src1_spad.src = NULL;
     octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size; octx->dst_spad.src  = NULL;
 
-    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) return err;
-
-    worker_pool_run_func(octx->ctx->worker_pool, softmax_job_f32, &smctx, n_threads);
+    work_queue_run(octx->ctx->work_queue, softmax_job_f32, &smctx, n_threads);
 
     return err;
 }

ggml/src/ggml-hexagon/htp/solve-tri-ops.c

@@ -1,13 +1,16 @@
 #pragma clang diagnostic ignored "-Wunused-but-set-variable"
 
 #include <HAP_farf.h>
-#include <HAP_perf.h>
 #include <string.h>
 
+#include "hex-common.h"
+#include "hex-profile.h"
+
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
+#include "htp-tensor.h"
 #include "hvx-types.h"
 #include "hvx-utils.h"
 
@@ -15,6 +18,7 @@ struct htp_solve_tri_context {
     struct htp_ops_context * octx;
     uint32_t                 jobs_per_thread;
     uint32_t                 total_jobs;
+    uint32_t                 job_start;
     uint32_t                 k_chunks;
     uint32_t                 col_block;
 };
@@ -89,11 +93,11 @@ static void solve_tri_batch_thread_f32(unsigned int nth, unsigned int ith, void
     const uint32_t col_block = VLEN_FP32;
     const uint32_t k_full    = (k / col_block) * col_block;
 
-    const uint32_t start_batch = sctx->jobs_per_thread * ith;
-    const uint32_t end_batch   = MIN(start_batch + sctx->jobs_per_thread, sctx->total_jobs);
+    const uint32_t start_batch = sctx->job_start + sctx->jobs_per_thread * ith;
+    const uint32_t end_batch   = MIN(start_batch + sctx->jobs_per_thread, sctx->job_start + sctx->total_jobs);
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) start_batch);
 
     for (uint32_t batch = start_batch; batch < end_batch; ++batch) {
         const uint32_t i03 = batch / ne02;
@@ -127,11 +131,10 @@ static void solve_tri_batch_thread_f32(unsigned int nth, unsigned int ith, void
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) end_batch);
 
-    FARF(HIGH, "solve-tri-batch %d/%d: A=(%ux%u) B=(%ux%u) batch %u:%u usec %u\n",
-         ith, nth, n, n, k, n, start_batch, end_batch,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    FARF(HIGH, "solve-tri-batch %d/%d: A=(%ux%u) B=(%ux%u) batch %u:%u\n",
+         ith, nth, n, n, k, n, start_batch, end_batch);
 }
 
 // Chunk-level thread: each job is one (batch, col_chunk) pair.
@@ -148,11 +151,11 @@ static void solve_tri_chunk_thread_f32(unsigned int nth, unsigned int ith, void
 
     const uint32_t ne02 = src0->ne[2];
 
-    const uint32_t start_job = sctx->jobs_per_thread * ith;
-    const uint32_t end_job   = MIN(start_job + sctx->jobs_per_thread, sctx->total_jobs);
+    const uint32_t start_job = sctx->job_start + sctx->jobs_per_thread * ith;
+    const uint32_t end_job   = MIN(start_job + sctx->jobs_per_thread, sctx->job_start + sctx->total_jobs);
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) start_job);
 
     for (uint32_t job = start_job; job < end_job; ++job) {
         const uint32_t batch = job / sctx->k_chunks;
@@ -161,16 +164,14 @@ static void solve_tri_chunk_thread_f32(unsigned int nth, unsigned int ith, void
         const uint32_t i03 = batch / ne02;
         const uint32_t i02 = batch - i03 * ne02;
 
-        const uint32_t col0 = chunk * sctx->col_block;
-        const uint32_t coln = MIN(sctx->col_block, k - col0);
-
         const float * A_batch =
             (const float *) ((const uint8_t *) (uintptr_t) src0->data + i02 * src0->nb[2] + i03 * src0->nb[3]);
         const float * B_batch =
             (const float *) ((const uint8_t *) (uintptr_t) src1->data + i02 * src1->nb[2] + i03 * src1->nb[3]);
         float * X_batch = (float *) ((uint8_t *) (uintptr_t) dst->data + i02 * dst->nb[2] + i03 * dst->nb[3]);
 
-        const bool use_hvx = (coln >= 8);
+        const uint32_t col0 = chunk * sctx->col_block;
+        const uint32_t coln = MIN(sctx->col_block, k - col0);
 
         for (uint32_t row = 0; row < n; ++row) {
             const float diag     = A_batch[row * n + row];
@@ -179,7 +180,7 @@ static void solve_tri_chunk_thread_f32(unsigned int nth, unsigned int ith, void
             const float * A_row = A_batch + row * n;
             const float * B_row = B_batch + row * k;
 
-            if (use_hvx) {
+            if (coln >= 8) {
                 solve_tri_row_hvx(A_row, B_row, X_batch, row, k, col0, coln, inv_diag);
             } else {
                 solve_tri_row_scalar(A_row, B_row, X_batch, row, k, col0, coln, inv_diag);
@@ -187,11 +188,10 @@ static void solve_tri_chunk_thread_f32(unsigned int nth, unsigned int ith, void
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) end_job);
 
-    FARF(HIGH, "solve-tri-chunk %d/%d: A=(%ux%u) B=(%ux%u) job %u:%u usec %u\n",
-         ith, nth, n, n, k, n, start_job, end_job,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    FARF(HIGH, "solve-tri-chunk %d/%d: A=(%ux%u) B=(%ux%u) jobs %u:%u\n",
+         ith, nth, n, n, k, n, start_job, end_job);
 }
 
 int op_solve_tri(struct htp_ops_context * octx) {
@@ -235,32 +235,64 @@ int op_solve_tri(struct htp_ops_context * octx) {
          dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], batched);
 
     if (batched) {
+        uint32_t job_start = 0;
+        uint32_t njobs     = total_batches;
+
+        if (octx->ctx->mdev.count > 1) {
+            const uint32_t batch_size = dst->nb[2];
+            const uint32_t batches_per_chunk = (batch_size > 0) ? (HEX_L2_LINE_SIZE / hex_gcd_u32(batch_size, HEX_L2_LINE_SIZE)) : 1;
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_batches, htp_tensor_mdev_data_aligned(dst) ? batches_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            job_start = range.start;
+            njobs     = range.count;
+        }
+
+        if (njobs == 0) {
+            return HTP_STATUS_OK;
+        }
+
         // Batch-level parallelism
-        const uint32_t n_threads = MIN((uint32_t) octx->n_threads, total_batches);
+        const uint32_t n_threads = octx->n_threads;
 
         struct htp_solve_tri_context sctx = {
             .octx            = octx,
-            .jobs_per_thread = (total_batches + n_threads - 1) / n_threads,
-            .total_jobs      = total_batches,
+            .jobs_per_thread = fastdiv(njobs + n_threads - 1, &octx->n_threads_div),
+            .total_jobs      = njobs,
+            .job_start       = job_start,
             .k_chunks        = k_chunks,
             .col_block       = col_block,
         };
 
-        worker_pool_run_func(octx->ctx->worker_pool, solve_tri_batch_thread_f32, &sctx, n_threads);
+        work_queue_run(octx->ctx->work_queue, solve_tri_batch_thread_f32, &sctx, n_threads);
     } else {
         // Chunk-level parallelism
         const uint32_t total_jobs = total_batches * k_chunks;
-        const uint32_t n_threads  = MIN((uint32_t) octx->n_threads, MAX(total_jobs, 1));
+
+        uint32_t job_start = 0;
+        uint32_t njobs     = total_jobs;
+
+        if (octx->ctx->mdev.count > 1) {
+            const bool can_split = htp_tensor_mdev_data_aligned(dst) && ((dst->nb[1] & (HTP_TENSOR_MDEV_LINE_SIZE - 1)) == 0);
+            const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_jobs, can_split ? 1 : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+            job_start = range.start;
+            njobs     = range.count;
+        }
+
+        if (njobs == 0) {
+            return HTP_STATUS_OK;
+        }
+
+        const uint32_t n_threads = octx->n_threads;
 
         struct htp_solve_tri_context sctx = {
             .octx            = octx,
-            .jobs_per_thread = (total_jobs + n_threads - 1) / n_threads,
-            .total_jobs      = total_jobs,
+            .jobs_per_thread = fastdiv(njobs + n_threads - 1, &octx->n_threads_div),
+            .total_jobs      = njobs,
+            .job_start       = job_start,
             .k_chunks        = k_chunks,
             .col_block       = col_block,
         };
 
-        worker_pool_run_func(octx->ctx->worker_pool, solve_tri_chunk_thread_f32, &sctx, n_threads);
+        work_queue_run(octx->ctx->work_queue, solve_tri_chunk_thread_f32, &sctx, n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/ssm-conv.c

@@ -4,7 +4,6 @@
 
 #include <HAP_farf.h>
 #include <HAP_mem.h>
-#include <HAP_perf.h>
 #include <HAP_ps.h>
 #include <hexagon_protos.h>
 #include <hexagon_types.h>
@@ -16,8 +15,9 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "hex-dma.h"
+#include "hex-profile.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
+#include "htp-tensor.h"
 #include "hvx-utils.h"
 
 #define htp_ssm_conv_tensors_preamble                           \
@@ -63,6 +63,8 @@ struct htp_ssm_conv_context {
     uint32_t nrows_per_thread;
     uint32_t d_inner_tile;
     uint64_t t_start;
+    uint32_t row_start;
+    uint32_t nrows;
 };
 
 #define htp_ssm_conv_preamble                                                   \
@@ -75,9 +77,6 @@ struct htp_ssm_conv_context {
 static void ssm_conv_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
     htp_ssm_conv_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
     const uint32_t d_conv  = src1->ne[0];
     const uint32_t d_inner = src0->ne[1];
     const uint32_t n_t     = dst->ne[1];
@@ -95,14 +94,17 @@ static void ssm_conv_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
 
     // Calculate row range for this thread
     const uint32_t d_inner_per_thread = scctx->nrows_per_thread;
-    const uint32_t d_inner_start = d_inner_per_thread * ith;
-    const uint32_t d_inner_end   = MIN(d_inner_start + d_inner_per_thread, d_inner);
+    const uint32_t d_inner_start = scctx->row_start + d_inner_per_thread * ith;
+    const uint32_t d_inner_end   = MIN(d_inner_start + d_inner_per_thread, scctx->row_start + scctx->nrows);
 
     // No work for this thread
     if (d_inner_start >= d_inner_end) {
         return;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) d_inner_start);
+
     for (uint32_t i3 = 0; i3 < n_s; ++i3) {
         for (uint32_t i2 = 0; i2 < n_t; ++i2) {
             for (uint32_t i1 = d_inner_start; i1 < d_inner_end; ++i1) {
@@ -121,12 +123,12 @@ static void ssm_conv_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) d_inner_end);
 
-    FARF(HIGH, "ssm-conv-f32 %d/%d: %ux%ux%ux%u (%u:%u) * %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n",
+    FARF(HIGH, "ssm-conv-f32 %d/%d: %ux%ux%ux%u (%u:%u) * %ux%ux%ux%u -> %ux%ux%ux%u\n",
          ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], d_inner_start, d_inner_end,
          src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0], dst->ne[1],
-         dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         dst->ne[2], dst->ne[3]);
 }
 
 
@@ -257,9 +259,6 @@ static inline void transpose_src0_block(const float * src0_block,
 static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void *data) {
     htp_ssm_conv_preamble;
 
-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
-
     const uint32_t d_conv  = src1->ne[0];
     const uint32_t d_inner = src0->ne[1];
     const uint32_t n_t     = dst->ne[1];
@@ -273,13 +272,16 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
     const uint32_t dst_stride_seq    = dst->nb[2]  / sizeof(float);
 
     const uint32_t dr  = scctx->nrows_per_thread;
-    const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = MIN(ir0 + dr, d_inner);
+    const uint32_t ir0 = scctx->row_start + dr * ith;
+    const uint32_t ir1 = MIN(ir0 + dr, scctx->row_start + scctx->nrows);
 
     if (ir0 >= ir1) {
         return;
     }
 
+    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir0);
+
     const uint32_t d_inner_per_thread = ir1 - ir0;
     const uint32_t d_inner_stride     = scctx->nrows_per_thread;
     const uint32_t d_inner_tile       = scctx->d_inner_tile;
@@ -319,97 +321,118 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
                         HVX_Vector w = *(const HVX_Vector *) (src1_T + j * d_inner_stride + tile_off + cb);
                         acc          = Q6_Vqf32_vadd_Vqf32Vqf32(acc, Q6_Vqf32_vmpy_VsfVsf(x, w));
                     }
-                    HVX_Vector res = Q6_Vsf_equals_Vqf32(acc);
 
-                    float * dst_ptr = dst_data + i3 * dst_stride_seq + t * dst_stride_token + (ir0 + tile_off + cb);
+                    HVX_Vector y = Q6_Vsf_equals_Vqf32(acc);
+
+                    float * dst_ptr = dst_data + (ir0 + tile_off + cb) + t * dst_stride_token + i3 * dst_stride_seq;
                     if (cb_n == C_TILE) {
-                        *(HVX_UVector *) dst_ptr = res;
+                        *(HVX_UVector *) dst_ptr = y;
                     } else {
-                        hvx_vec_store_u(dst_ptr, cb_n * sizeof(float), res);
+                        hvx_vec_store_u(dst_ptr, cb_n * sizeof(float), y);
                     }
                 }
             }
         }
     }
 
-    t2 = HAP_perf_get_qtimer_count();
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir1);
 
-    FARF(HIGH, "ssm-conv-f32-hvx %d/%d: %ux%ux%ux%u (%u:%u) tile=%u * %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n",
-         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ir0, ir1, d_inner_tile,
+    FARF(HIGH, "ssm-conv-f32-hvx %d/%d: %ux%ux%ux%u (%u:%u) * %ux%ux%ux%u -> %ux%ux%ux%u\n",
+         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ir0, ir1,
          src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0], dst->ne[1],
-         dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+         dst->ne[2], dst->ne[3]);
 }
 
 int op_ssm_conv_f32(struct htp_ops_context * octx) {
-    htp_ssm_conv_tensors_preamble;
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
 
     if (src0->type != HTP_TYPE_F32 || src1->type != HTP_TYPE_F32 || dst->type != HTP_TYPE_F32) {
-        FARF(ERROR, "ssm_conv: only (F32 x F32 -> F32) OPs supported");
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    struct htp_ssm_conv_context scctx = { 0 };
-    scctx.octx = octx;
-
     const uint32_t d_conv  = src1->ne[0];
     const uint32_t d_inner = src0->ne[1];
     const uint32_t n_t     = dst->ne[1];  // tokens per sequence
     const uint32_t n_s     = dst->ne[2];  // number of sequences in the batch
 
-    const uint32_t n_threads = MIN(octx->n_threads, d_inner);
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
 
-    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t use_hvx = 0;
-        if (d_inner >= VLEN_FP32 && n_t >= VLEN_FP32) {
-            use_hvx = 1;
-        }
+    uint32_t row_start = 0;
+    uint32_t nrows     = d_inner;
 
-        scctx.nrows_per_thread = hex_round_up((d_inner + n_threads - 1) / n_threads, VLEN_FP32);
+    if (octx->ctx->mdev.count > 1) {
+        const uint32_t elems_per_chunk = VLEN_FP32;
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(d_inner, htp_tensor_mdev_data_aligned(dst) ? elems_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        row_start = range.start;
+        nrows     = range.count;
+    }
 
-        const uint32_t d_inner_per_thread = scctx.nrows_per_thread;
-        const uint32_t ncs                = src0->ne[0];
+    if (nrows == 0) {
+        return HTP_STATUS_OK;
+    }
 
-        const uint32_t src1_T_size = hex_round_up(d_conv * d_inner_per_thread * sizeof(float), 256);
-        const uint32_t src0_T_max = HTP_SSM_CONV_VTCM_BUDGET > src1_T_size ? HTP_SSM_CONV_VTCM_BUDGET - src1_T_size : 0;
+    const uint32_t n_threads = octx->n_threads;
 
-        uint32_t d_inner_tile = (src0_T_max / sizeof(float)) / ncs;
-        d_inner_tile -= (d_inner_tile % VLEN_FP32);
-        if (d_inner_tile == 0) {
-            FARF(HIGH, "ssm_conv-f32: inner tile rounds to 0 (ncs=%u), falling back to scalar\n", ncs);
+    struct htp_ssm_conv_context scctx = { 0 };
+    scctx.octx      = octx;
+    scctx.row_start = row_start;
+    scctx.nrows     = nrows;
+
+    uint32_t use_hvx = 0;
+    if (nrows >= VLEN_FP32 && n_t >= VLEN_FP32) {
+        use_hvx = 1;
+    }
+
+    const uint32_t raw_rpt = fastdiv(nrows + n_threads - 1, &octx->n_threads_div);
+    scctx.nrows_per_thread = hex_round_up(raw_rpt, VLEN_FP32);
+
+    const uint32_t d_inner_per_thread = scctx.nrows_per_thread;
+    const uint32_t ncs                = src0->ne[0];
+
+    const uint32_t src1_T_size = hex_round_up(d_conv * d_inner_per_thread * sizeof(float), 256);
+    const uint32_t src0_T_max = HTP_SSM_CONV_VTCM_BUDGET > src1_T_size ? HTP_SSM_CONV_VTCM_BUDGET - src1_T_size : 0;
+
+    uint32_t d_inner_tile = (src0_T_max / sizeof(float)) / ncs;
+    d_inner_tile -= (d_inner_tile % VLEN_FP32);
+    if (d_inner_tile == 0) {
+        FARF(HIGH, "ssm_conv-f32: inner tile rounds to 0 (ncs=%u), falling back to scalar\n", ncs);
+        use_hvx = 0;
+    } else {
+        scctx.d_inner_tile = d_inner_tile;
+
+        octx->src0_spad.size_per_thread = hex_round_up(d_inner_tile * ncs * sizeof(float), 256);
+        octx->src1_spad.size_per_thread = src1_T_size;
+        octx->dst_spad.size_per_thread  = 0;
+
+        octx->src0_spad.size = octx->src0_spad.size_per_thread * n_threads;
+        octx->src1_spad.size = octx->src1_spad.size_per_thread * n_threads;
+        octx->dst_spad.size  = 0;
+
+        octx->src0_spad.data = octx->ctx->vtcm_base;
+        octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+        octx->src0_spad.src  = NULL;
+        octx->src1_spad.src  = NULL;
+
+        const size_t total_spad = octx->src0_spad.size + octx->src1_spad.size;
+        if (total_spad > octx->ctx->vtcm_size) {
+            FARF(HIGH, "ssm_conv-f32: scratchpad %zu exceeds VTCM %zu, falling back to scalar\n",
+                 total_spad, octx->ctx->vtcm_size);
             use_hvx = 0;
-        } else {
-            scctx.d_inner_tile = d_inner_tile;
-
-            octx->src0_spad.size_per_thread = hex_round_up(d_inner_tile * ncs * sizeof(float), 256);
-            octx->src1_spad.size_per_thread = src1_T_size;
-            octx->dst_spad.size_per_thread  = 0;
-
-            octx->src0_spad.size = octx->src0_spad.size_per_thread * n_threads;
-            octx->src1_spad.size = octx->src1_spad.size_per_thread * n_threads;
-            octx->dst_spad.size  = 0;
-
-            octx->src0_spad.data = octx->ctx->vtcm_base;
-            octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
-            octx->src0_spad.src  = NULL;
-            octx->src1_spad.src  = NULL;
-
-            const size_t total_spad = octx->src0_spad.size + octx->src1_spad.size;
-            if (total_spad > octx->ctx->vtcm_size) {
-                FARF(HIGH, "ssm_conv-f32: scratchpad %zu exceeds VTCM %zu, falling back to scalar\n",
-                     total_spad, octx->ctx->vtcm_size);
-                use_hvx = 0;
-            }
         }
+    }
 
-        FARF(HIGH, "ssm-conv-f32: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : use_hvx %d\n", src0->ne[0],
-             src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0],
-             dst->ne[1], dst->ne[2], dst->ne[3], use_hvx);
+    FARF(HIGH, "ssm-conv-f32: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : use_hvx %d\n", src0->ne[0],
+         src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0],
+         dst->ne[1], dst->ne[2], dst->ne[3], use_hvx);
 
-        if (use_hvx) {
-            worker_pool_run_func(octx->ctx->worker_pool, ssm_conv_thread_f32_f32_hvx, &scctx, n_threads);
-        } else {
-            worker_pool_run_func(octx->ctx->worker_pool, ssm_conv_thread_f32_f32, &scctx, n_threads);
-        }
+    if (use_hvx) {
+        work_queue_run(octx->ctx->work_queue, ssm_conv_thread_f32_f32_hvx, &scctx, n_threads);
+    } else {
+        work_queue_run(octx->ctx->work_queue, ssm_conv_thread_f32_f32, &scctx, n_threads);
     }
 
     return HTP_STATUS_OK;