experiments

* [CUDA] Reduce CPU-side stalls due to the CUDA command buffer being full

With pipeline parallelism, during prompt processing, the CPU-side CUDA command buffer gets full, stalling the CPU. Due to this, enough work doesn't get submitted to the GPU, causing bubbles in the GPU timeline.
Fix this by setting the CUDA environment variable CUDA_SCALE_LAUNCH_QUEUES to 4x to increase the command buffer size.

* Set the env variable in the CUDA backend registry allocation

* Add link to PR in code comment

* Remove warning logs and update documentation

.github/workflows/check-vendor.yml

@@ -19,7 +19,7 @@ on:
 
 jobs:
   check-vendor:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
 
     steps:
       - name: Checkout

.github/workflows/close-issue.yml

@@ -10,7 +10,7 @@ permissions:
 
 jobs:
   close-issues:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     permissions:
       issues: write
       pull-requests: write

.github/workflows/editorconfig.yml

@@ -20,7 +20,7 @@ concurrency:
 
 jobs:
   editorconfig:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     steps:
       - uses: actions/checkout@v6
       - uses: editorconfig-checker/action-editorconfig-checker@v2

.github/workflows/gguf-publish.yml

@@ -21,7 +21,7 @@ on:
 jobs:
   deploy:
 
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
 
     steps:
     - uses: actions/checkout@v6

.github/workflows/labeler.yml

@@ -7,7 +7,7 @@ jobs:
     permissions:
       contents: read
       pull-requests: write
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     steps:
     - uses: actions/checkout@v6
       with:

.github/workflows/pre-tokenizer-hashes.yml

@@ -12,7 +12,7 @@ on:
 
 jobs:
     pre-tokenizer-hashes:
-        runs-on: ubuntu-latest
+        runs-on: ubuntu-slim
 
         steps:
         - name: Checkout repository

.github/workflows/python-check-requirements.yml

@@ -20,7 +20,7 @@ concurrency:
 
 jobs:
   python-check-requirements:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     name: check-requirements
     steps:
       - name: Check out source repository

.github/workflows/python-lint.yml

@@ -15,7 +15,7 @@ concurrency:
 
 jobs:
   flake8-lint:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     name: Lint
     steps:
       - name: Check out source repository

.github/workflows/python-type-check.yml

@@ -29,9 +29,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: "3.11"
-      - name: Install Python dependencies
-        # TODO: use a venv
-        run: pip install -r requirements/requirements-all.txt
+          pip-install: -r requirements/requirements-all.txt
       - name: Type-check with Pyright
         uses: jakebailey/pyright-action@v2
         with:

.github/workflows/release.yml

@@ -897,7 +897,7 @@ jobs:
             for (let file of await fs.readdirSync('./release')) {
               if (path.extname(file) === '.zip' || file.endsWith('.tar.gz')) {
                 console.log('uploadReleaseAsset', file);
-                await github.repos.uploadReleaseAsset({
+                await github.rest.repos.uploadReleaseAsset({
                   owner: context.repo.owner,
                   repo: context.repo.repo,
                   release_id: release_id,

.github/workflows/server.yml

@@ -72,7 +72,7 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON
+          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
@@ -108,7 +108,7 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON
+          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup

.github/workflows/update-ops-docs.yml

@@ -14,7 +14,7 @@ on:
 
 jobs:
     update-ops-docs:
-        runs-on: ubuntu-latest
+        runs-on: ubuntu-slim
 
         steps:
         - name: Checkout repository

.github/workflows/winget.yml

@@ -8,7 +8,7 @@ on:
 jobs:
   update:
     name: Update Winget Package
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     if: github.repository_owner == 'ggml-org'
 
     steps:

CODEOWNERS

@@ -18,6 +18,7 @@
 /common/jinja/                          @ngxson @CISC @aldehir
 /common/llguidance.*                    @ggerganov
 /common/log.*                           @ggerganov
+/common/ngram-map.*                     @srogmann
 /common/peg-parser.*                    @aldehir
 /common/sampling.*                      @ggerganov
 /common/speculative.*                   @ggerganov

README.md

@@ -132,6 +132,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
+- [x] [RWKV-7](https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf)
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)

common/CMakeLists.txt

@@ -73,6 +73,8 @@ add_library(${TARGET} STATIC
     log.h
     ngram-cache.cpp
     ngram-cache.h
+    ngram-map.cpp
+    ngram-map.h
     peg-parser.cpp
     peg-parser.h
     preset.cpp

common/arg.cpp

@@ -6,6 +6,7 @@
 #include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "preset.h"
 
 // fix problem with std::min and std::max
@@ -1216,21 +1217,25 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-lcs", "--lookup-cache-static"}, "FNAME",
         "path to static lookup cache to use for lookup decoding (not updated by generation)",
         [](common_params & params, const std::string & value) {
-            params.lookup_cache_static = value;
+            params.speculative.lookup_cache_static = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LOOKUP}));
+    ).set_examples({LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-lcd", "--lookup-cache-dynamic"}, "FNAME",
         "path to dynamic lookup cache to use for lookup decoding (updated by generation)",
         [](common_params & params, const std::string & value) {
-            params.lookup_cache_dynamic = value;
+            params.speculative.lookup_cache_dynamic = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LOOKUP}));
+    ).set_examples({LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-c", "--ctx-size"}, "N",
         string_format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx),
         [](common_params & params, int value) {
             params.n_ctx = value;
+            if (value == 0) {
+                // disable context reduction in llama_params_fit if the user explicitly requests the full context size:
+                params.fit_params_min_ctx = UINT32_MAX;
+            }
         }
     ).set_env("LLAMA_ARG_CTX_SIZE"));
     add_opt(common_arg(
@@ -1573,7 +1578,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--temp"}, "N",
-        string_format("temperature (default: %.1f)", (double)params.sampling.temp),
+        string_format("temperature (default: %.2f)", (double)params.sampling.temp),
         [](common_params & params, const std::string & value) {
             params.sampling.temp = std::stof(value);
             params.sampling.temp = std::max(params.sampling.temp, 0.0f);
@@ -1590,7 +1595,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam().set_env("LLAMA_ARG_TOP_K"));
     add_opt(common_arg(
         {"--top-p"}, "N",
-        string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
+        string_format("top-p sampling (default: %.2f, 1.0 = disabled)", (double)params.sampling.top_p),
         [](common_params & params, const std::string & value) {
             params.sampling.top_p = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P;
@@ -1598,7 +1603,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--min-p"}, "N",
-        string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
+        string_format("min-p sampling (default: %.2f, 0.0 = disabled)", (double)params.sampling.min_p),
         [](common_params & params, const std::string & value) {
             params.sampling.min_p = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P;
@@ -1606,14 +1611,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--top-nsigma"}, "N",
-        string_format("top-n-sigma sampling (default: %.1f, -1.0 = disabled)", params.sampling.top_n_sigma),
+        string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma),
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
-        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
+        string_format("xtc probability (default: %.2f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_probability = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY;
@@ -1621,7 +1626,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-threshold"}, "N",
-        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
+        string_format("xtc threshold (default: %.2f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_threshold = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD;
@@ -1629,7 +1634,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--typical"}, "N",
-        string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sampling.typ_p),
+        string_format("locally typical sampling, parameter p (default: %.2f, 1.0 = disabled)", (double)params.sampling.typ_p),
         [](common_params & params, const std::string & value) {
             params.sampling.typ_p = std::stof(value);
         }
@@ -1648,7 +1653,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--repeat-penalty"}, "N",
-        string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
+        string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_repeat = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
@@ -1656,21 +1661,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--presence-penalty"}, "N",
-        string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_present),
+        string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_present = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
-        string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
+        string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_freq = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-multiplier"}, "N",
-        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
+        string_format("set DRY sampling multiplier (default: %.2f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
         [](common_params & params, const std::string & value) {
             params.sampling.dry_multiplier = std::stof(value);
         }
@@ -1751,14 +1756,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-range"}, "N",
-        string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),
+        string_format("dynamic temperature range (default: %.2f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),
         [](common_params & params, const std::string & value) {
             params.sampling.dynatemp_range = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-exp"}, "N",
-        string_format("dynamic temperature exponent (default: %.1f)", (double)params.sampling.dynatemp_exponent),
+        string_format("dynamic temperature exponent (default: %.2f)", (double)params.sampling.dynatemp_exponent),
         [](common_params & params, const std::string & value) {
             params.sampling.dynatemp_exponent = std::stof(value);
         }
@@ -1774,7 +1779,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-lr"}, "N",
-        string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
+        string_format("Mirostat learning rate, parameter eta (default: %.2f)", (double)params.sampling.mirostat_eta),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_eta = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA;
@@ -1782,7 +1787,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-ent"}, "N",
-        string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
+        string_format("Mirostat target entropy, parameter tau (default: %.2f)", (double)params.sampling.mirostat_tau),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_tau = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU;
@@ -1916,28 +1921,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_YARN_ORIG_CTX"));
     add_opt(common_arg(
         {"--yarn-ext-factor"}, "N",
-        string_format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor),
+        string_format("YaRN: extrapolation mix factor (default: %.2f, 0.0 = full interpolation)", (double)params.yarn_ext_factor),
         [](common_params & params, const std::string & value) {
             params.yarn_ext_factor = std::stof(value);
         }
     ).set_env("LLAMA_ARG_YARN_EXT_FACTOR"));
     add_opt(common_arg(
         {"--yarn-attn-factor"}, "N",
-        string_format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor),
+        string_format("YaRN: scale sqrt(t) or attention magnitude (default: %.2f)", (double)params.yarn_attn_factor),
         [](common_params & params, const std::string & value) {
             params.yarn_attn_factor = std::stof(value);
         }
     ).set_env("LLAMA_ARG_YARN_ATTN_FACTOR"));
     add_opt(common_arg(
         {"--yarn-beta-slow"}, "N",
-        string_format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow),
+        string_format("YaRN: high correction dim or alpha (default: %.2f)", (double)params.yarn_beta_slow),
         [](common_params & params, const std::string & value) {
             params.yarn_beta_slow = std::stof(value);
         }
     ).set_env("LLAMA_ARG_YARN_BETA_SLOW"));
     add_opt(common_arg(
         {"--yarn-beta-fast"}, "N",
-        string_format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast),
+        string_format("YaRN: low correction dim or beta (default: %.2f)", (double)params.yarn_beta_fast),
         [](common_params & params, const std::string & value) {
             params.yarn_beta_fast = std::stof(value);
         }
@@ -3331,14 +3336,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_DRAFT_MIN"));
     add_opt(common_arg(
         {"--draft-p-split"}, "P",
-        string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split),
+        string_format("speculative decoding split probability (default: %.2f)", (double)params.speculative.p_split),
         [](common_params & params, const std::string & value) {
             params.speculative.p_split = std::stof(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT"));
     add_opt(common_arg(
         {"--draft-p-min"}, "P",
-        string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min),
+        string_format("minimum speculative decoding probability (greedy) (default: %.2f)", (double)params.speculative.p_min),
         [](common_params & params, const std::string & value) {
             params.speculative.p_min = std::stof(value);
         }
@@ -3392,6 +3397,68 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.replacements.push_back({ tgt, dft });
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+    add_opt(common_arg(
+        {"--spec-draftless"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-map-mod]",
+        string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n",
+            common_speculative_type_to_str(params.speculative.type).c_str()),
+        [](common_params & params, const std::string & value) {
+            if (value == "none") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NONE;
+            } else if (value == "ngram-cache") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_CACHE;
+            } else if (value == "ngram-simple") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE;
+            } else if (value == "ngram-map-k") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K;
+            } else if (value == "ngram-map-k4v") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V;
+            } else if (value == "ngram-map-mod") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD;
+            } else {
+                throw std::invalid_argument("unknown speculative decoding type without draft model");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-size-n"}, "N",
+        string_format("ngram size N for ngram-simple/ngram-map speculative decoding, length of lookup n-gram (default: %d)", params.speculative.ngram_size_n),
+        [](common_params & params, int value) {
+            if (value < 1 || value > 1024) {
+                throw std::invalid_argument("ngram size N must be between 1 and 1024 inclusive");
+            }
+            params.speculative.ngram_size_n = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-size-m"}, "N",
+        string_format("ngram size M for ngram-simple/ngram-map speculative decoding, length of draft m-gram (default: %d)", params.speculative.ngram_size_m),
+        [](common_params & params, int value) {
+            if (value < 1 || value > 1024) {
+                throw std::invalid_argument("ngram size M must be between 1 and 1024 inclusive");
+            }
+            params.speculative.ngram_size_m = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-check-rate"}, "N",
+        string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
+        [](common_params & params, int value) {
+            if (value < 1) {
+                throw std::invalid_argument("ngram check rate must be at least 1");
+            }
+            params.speculative.ngram_check_rate = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-min-hits"}, "N",
+        string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),
+        [](common_params & params, int value) {
+            if (value < 1) {
+                throw std::invalid_argument("ngram min hits must be at least 1");
+            }
+            params.speculative.ngram_min_hits = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-ctkd", "--cache-type-k-draft"}, "TYPE",
         string_format(

common/chat-parser.cpp

@@ -1630,7 +1630,7 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     }
     auto msg = builder.result();
     if (!is_partial) {
-        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat({msg}).at(0).dump().c_str());
     }
     return msg;
 }
@@ -1663,7 +1663,7 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena & parser, const std
         mapper.from_ast(ctx.ast, result);
     }
     if (!is_partial) {
-        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat({msg}).at(0).dump().c_str());
     }
     return msg;
 }

common/chat.cpp

@@ -7,9 +7,6 @@
 #include "log.h"
 #include "regex-partial.h"
 
-// #include <minja/chat-template.hpp>
-// #include <minja/minja.hpp>
-
 #include "jinja/parser.h"
 #include "jinja/value.h"
 #include "jinja/runtime.h"
@@ -56,39 +53,73 @@ static bool has_content_or_tool_calls(const common_chat_msg & msg) {
     return !msg.content.empty() || !msg.tool_calls.empty();
 }
 
-template <>
-json common_chat_msg::to_json_oaicompat() const
-{
-    json message {
-        {"role", "assistant"},
-    };
-    if (!reasoning_content.empty()) {
-        message["reasoning_content"] = reasoning_content;
+json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
+    if (!content.empty() && !content_parts.empty()) {
+        throw std::runtime_error("Cannot specify both content and content_parts");
     }
-    if (content.empty() && !tool_calls.empty()) {
-        message["content"] = json();
+    json jmsg {
+        {"role", role},
+    };
+    if (!content.empty()) {
+        jmsg["content"] = content;
+    } else if (!content_parts.empty()) {
+        if (concat_typed_text) {
+            std::string text;
+            for (const auto & part : content_parts) {
+                if (part.type != "text") {
+                    LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
+                    continue;
+                }
+                if (!text.empty()) {
+                    text += '\n';
+                }
+                text += part.text;
+            }
+            jmsg["content"] = text;
+        } else {
+            auto & parts = jmsg["content"] = json::array();
+            for (const auto & part : content_parts) {
+                parts.push_back({
+                    {"type", part.type},
+                    {"text", part.text},
+                });
+            }
+        }
     } else {
-        message["content"] = content;
+        jmsg["content"] = "";
+    }
+    if (!reasoning_content.empty()) {
+        jmsg["reasoning_content"] = reasoning_content;
+    }
+    if (!tool_name.empty()) {
+        jmsg["name"] = tool_name;
+    }
+    if (!tool_call_id.empty()) {
+        jmsg["tool_call_id"] = tool_call_id;
     }
     if (!tool_calls.empty()) {
-        auto arr = json::array();
-        for (const auto & tc : tool_calls) {
-            arr.push_back({
+        jmsg["tool_calls"] = json::array();
+        auto & jtool_calls = jmsg["tool_calls"];
+        for (const auto & tool_call : tool_calls) {
+            json tc {
                 {"type", "function"},
                 {"function", {
-                    {"name", tc.name},
-                    {"arguments", tc.arguments},
+                    {"name", tool_call.name},
+                    {"arguments", tool_call.arguments},
                 }},
-                {"id", tc.id},
-                // // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo).
-                // // We only generate a random id for the ones that don't generate one by themselves
-                // // (they also won't get to see it as their template likely doesn't use it, so it's all for the client)
-                // {"id", tc.id.empty() ? gen_tool_call_id() : tc.id},
-            });
+            };
+            if (!tool_call.id.empty()) {
+                tc["id"] = tool_call.id;
+            }
+            // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo).
+            // We only generate a random id for the ones that don't generate one by themselves
+            // (they also won't get to see it as their template likely doesn't use it, so it's all for the client)
+            // {"id", tc.id.empty() ? gen_tool_call_id() : tc.id},
+            jtool_calls.push_back(tc);
         }
-        message["tool_calls"] = arr;
     }
-    return message;
+
+    return jmsg;
 }
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & msg_prv, const common_chat_msg & msg_new) {
@@ -256,7 +287,6 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
     return rendered_no_thinking.prompt != rendered_with_thinking.prompt;
 }
 
-template <>
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messages) {
     std::vector<common_chat_msg> msgs;
 
@@ -350,80 +380,15 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
     return msgs;
 }
 
-template <>
 json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
     json messages = json::array();
     for (const auto & msg : msgs) {
-        if (!msg.content.empty() && !msg.content_parts.empty()) {
-            throw std::runtime_error("Cannot specify both content and content_parts");
-        }
-        json jmsg {
-            {"role", msg.role},
-        };
-        if (!msg.content.empty()) {
-            jmsg["content"] = msg.content;
-        } else if (!msg.content_parts.empty()) {
-            if (concat_typed_text) {
-                std::string text;
-                for (const auto & part : msg.content_parts) {
-                    if (part.type != "text") {
-                        LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
-                        continue;
-                    }
-                    if (!text.empty()) {
-                        text += '\n';
-                    }
-                    text += part.text;
-                }
-                jmsg["content"] = text;
-            } else {
-                auto & parts = jmsg["content"] = json::array();
-                for (const auto & part : msg.content_parts) {
-                    parts.push_back({
-                        {"type", part.type},
-                        {"text", part.text},
-                    });
-                }
-            }
-        } else {
-            jmsg["content"] = "";
-        }
-        if (!msg.reasoning_content.empty()) {
-            jmsg["reasoning_content"] = msg.reasoning_content;
-        }
-        if (!msg.tool_name.empty()) {
-            jmsg["name"] = msg.tool_name;
-        }
-        if (!msg.tool_call_id.empty()) {
-            jmsg["tool_call_id"] = msg.tool_call_id;
-        }
-        if (!msg.tool_calls.empty()) {
-            auto & tool_calls = jmsg["tool_calls"] = json::array();
-            for (const auto & tool_call : msg.tool_calls) {
-                json tc {
-                    {"type", "function"},
-                    {"function", {
-                        {"name", tool_call.name},
-                        {"arguments", tool_call.arguments},
-                    }},
-                };
-                if (!tool_call.id.empty()) {
-                    tc["id"] = tool_call.id;
-                }
-                tool_calls.push_back(tc);
-            }
-        }
+        json jmsg = msg.to_json_oaicompat(concat_typed_text);
         messages.push_back(jmsg);
     }
     return messages;
 }
 
-template <>
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const std::string & messages) {
-    return common_chat_msgs_parse_oaicompat(json::parse(messages));
-}
-
-template <>
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 
@@ -459,12 +424,6 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-template <>
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const std::string & tools) {
-    return common_chat_tools_parse_oaicompat(json::parse(tools));
-}
-
-template <>
 json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {
     if (tools.empty()) {
         return json();
@@ -484,7 +443,7 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
     return result;
 }
 
-template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
+json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
     if (!diff.reasoning_content_delta.empty()) {
         delta["reasoning_content"] = diff.reasoning_content_delta;
@@ -2691,6 +2650,51 @@ static common_chat_params common_chat_params_init_exaone_moe(const common_chat_t
     return data;
 }
 
+static common_chat_params common_chat_params_init_translate_gemma(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // This template does not support tools or reasoning
+    // we just need to transform the messages into the correct schema
+
+    templates_params inputs_new = inputs;
+    json & messages = inputs_new.messages;
+
+    // default to chat_template_kwargs, or en-GB if not specified
+    std::string default_src_lang = inputs.extra_context.value("source_lang_code", "en-GB");
+    std::string default_tgt_lang = inputs.extra_context.value("target_lang_code", "en-GB");
+
+    GGML_ASSERT(messages.is_array());
+    for (auto & message : messages) {
+        if (message.contains("role") && message["role"].get<std::string>() != "user") {
+            continue;
+        }
+        if (!message.contains("content")) {
+            message["content"] = json::array();
+        }
+        if (message.contains("content") && !message["content"].is_array()) {
+            auto content_str = message["content"].get<std::string>();
+            // default to en-GB if not specified (to make common_chat_format_example works)
+            auto src_lang = message.contains("source_lang_code")
+                        ? message["source_lang_code"].get<std::string>() : default_src_lang;
+            auto tgt_lang = message.contains("target_lang_code")
+                        ? message["target_lang_code"].get<std::string>() : default_tgt_lang;
+            message["content"] = json::array({
+                json{
+                    {"type", "text"},
+                    {"text", content_str},
+                    {"source_lang_code", src_lang},
+                    {"target_lang_code", tgt_lang},
+                }
+            });
+        }
+    }
+
+    data.prompt = apply(tmpl, inputs_new, std::nullopt, std::nullopt);
+    data.format = COMMON_CHAT_FORMAT_GENERIC;
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2867,13 +2871,13 @@ static common_chat_params common_chat_templates_apply_jinja(
     const struct common_chat_templates_inputs & inputs)
 {
     templates_params params;
-    params.tools = common_chat_tools_to_json_oaicompat<json>(inputs.tools);
+    params.tools = common_chat_tools_to_json_oaicompat(inputs.tools);
     const auto & tmpl = params.tools.is_array() && tmpls->template_tool_use
         ? *tmpls->template_tool_use
         : *tmpls->template_default;
     const auto & src = tmpl.source();
     const auto & caps = tmpl.original_caps();
-    params.messages = common_chat_msgs_to_json_oaicompat<json>(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
+    params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
     params.add_generation_prompt = inputs.add_generation_prompt;
     params.tool_choice = inputs.tool_choice;
     params.reasoning_format = inputs.reasoning_format;
@@ -2943,6 +2947,10 @@ static common_chat_params common_chat_templates_apply_jinja(
         src.find("<arg_value>") != std::string::npos &&
         params.json_schema.is_null()) {
         workaround::func_args_not_string(params.messages);
+        if (!params.extra_context.contains("clear_thinking")) {
+            // by default, do not clear reasoning_content (added since GLM-4.7)
+            params.extra_context["clear_thinking"] = false;
+        }
         return common_chat_params_init_glm_4_5(tmpl, params);
     }
 
@@ -3082,6 +3090,12 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_solar_open(tmpl, params);
     }
 
+    // TranslateGemma
+    if (src.find("[source_lang_code]") != std::string::npos &&
+        src.find("[target_lang_code]") != std::string::npos) {
+        return common_chat_params_init_translate_gemma(tmpl, params);
+    }
+
     // Plain handler (no tools)
     if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
         return common_chat_params_init_without_tools(tmpl, params);
@@ -3174,3 +3188,9 @@ common_chat_params common_chat_templates_apply(
         ? common_chat_templates_apply_jinja(tmpls, inputs)
         : common_chat_templates_apply_legacy(tmpls, inputs);
 }
+
+std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates) {
+    GGML_ASSERT(chat_templates != nullptr);
+    GGML_ASSERT(chat_templates->template_default != nullptr);
+    return chat_templates->template_default->caps.to_map();
+}

common/chat.h

@@ -10,6 +10,8 @@
 #include <vector>
 #include <map>
 
+#include <nlohmann/json_fwd.hpp>
+
 struct common_chat_templates;
 
 struct common_chat_tool_call {
@@ -26,6 +28,11 @@ struct common_chat_msg_content_part {
     std::string type;
     std::string text;
 
+    // TODO @ngxson : no known chat templates support reasoning_content in content parts yet
+    //                this can be useful for models with interleaved thinking (like Kimi-K2)
+    //                if you see any templates explicitly support this, please ping me
+    // std::string reasoning_content;
+
     bool operator==(const common_chat_msg_content_part & other) const {
         return type == other.type && text == other.text;
     }
@@ -40,7 +47,7 @@ struct common_chat_msg {
     std::string tool_name;
     std::string tool_call_id;
 
-    template <class T> T to_json_oaicompat() const;
+    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     bool empty() const {
         return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
@@ -232,13 +239,13 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-// T can be std::string containing JSON or nlohmann::ordered_json
-template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);
-template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-// Parses a JSON array of tools in OpenAI's chat completion tool call API format.
-// T can be std::string containing JSON or nlohmann::ordered_json
-template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
-template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
-template <class T> T common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
+nlohmann::ordered_json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
+
+// 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/common.cpp

@@ -1097,7 +1097,10 @@ common_init_result::common_init_result(common_params & params) :
     if (params.fit_params) {
         LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
+            params.tensor_split,
+            params.tensor_buft_overrides.data(),
+            params.fit_params_target.data(),
+            params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1208,10 +1211,6 @@ std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
     return pimpl->lora;
 }
 
-void common_init_result::free_context() {
-    pimpl->context.reset();
-}
-
 common_init_result_ptr common_init_from_params(common_params & params) {
     common_init_result_ptr res(new common_init_result(params));
 

common/common.h

@@ -164,6 +164,17 @@ enum common_params_sampling_config : uint64_t {
     COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
 };
 
+enum common_speculative_type {
+    COMMON_SPECULATIVE_TYPE_NONE,          // no speculative decoding
+    COMMON_SPECULATIVE_TYPE_DRAFT,         // draft model
+    COMMON_SPECULATIVE_TYPE_EAGLE3,        // eagle draft model
+    COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD,
+    COMMON_SPECULATIVE_TYPE_NGRAM_CACHE,   // self-speculative decoding with 3-level n-gram cache
+    COMMON_SPECULATIVE_TYPE_COUNT          // number of types, unknown type
+};
 
 // sampling parameters
 struct common_params_sampling {
@@ -251,6 +262,7 @@ struct common_params_speculative {
     int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
+
     std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
@@ -261,6 +273,20 @@ struct common_params_speculative {
     struct cpu_params cpuparams_batch;
 
     struct common_params_model model;
+
+    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding
+
+    uint16_t ngram_size_n     = 12; // ngram size for lookup
+    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
+    uint16_t ngram_check_rate =  1; // check rate for ngram lookup
+    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
+
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
+
+    bool has_dft() const {
+        return !model.path.empty() || !model.hf_repo.empty();
+    }
 };
 
 struct common_params_vocoder {
@@ -378,8 +404,6 @@ struct common_params {
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
     std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
-    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
     // llama-debug specific options
@@ -575,10 +599,6 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
-
-    bool has_speculative() const {
-        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
-    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -714,8 +734,6 @@ struct common_init_result {
 
     std::vector<llama_adapter_lora_ptr> & lora();
 
-    void free_context();
-
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;

common/http.h

@@ -60,10 +60,10 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
 #ifndef CPPHTTPLIB_OPENSSL_SUPPORT
     if (parts.scheme == "https") {
         throw std::runtime_error(
-            "HTTPS is not supported. Please rebuild with:\n"
+            "HTTPS is not supported. Please rebuild with one of:\n"
             "  -DLLAMA_BUILD_BORINGSSL=ON\n"
             "  -DLLAMA_BUILD_LIBRESSL=ON\n"
-            "or ensure dev files of an OpenSSL-compatible library are available when building."
+            "  -DLLAMA_OPENSSL=ON (default, requires OpenSSL dev files installed)"
         );
     }
 #endif

common/jinja/caps.cpp

@@ -61,14 +61,23 @@ static void caps_print_stats(value & v, const std::string & path) {
                 ops.c_str());
 }
 
+std::map<std::string, bool> caps::to_map() const {
+    return {
+        {"requires_typed_content", requires_typed_content},
+        {"supports_tools", supports_tools},
+        {"supports_tool_calls", supports_tool_calls},
+        {"supports_parallel_tool_calls", supports_parallel_tool_calls},
+        {"supports_system_role", supports_system_role},
+        {"supports_preserve_reasoning", supports_preserve_reasoning},
+    };
+}
+
 std::string caps::to_string() const {
     std::ostringstream ss;
     ss << "Caps(\n";
-    ss << "  requires_typed_content=" << requires_typed_content << "\n";
-    ss << "  supports_tools=" << supports_tools << "\n";
-    ss << "  supports_tool_calls=" << supports_tool_calls << "\n";
-    ss << "  supports_parallel_tool_calls=" << supports_parallel_tool_calls << "\n";
-    ss << "  supports_system_role=" << supports_system_role << "\n";
+    for (const auto & [key, value] : to_map()) {
+        ss << "  " << key << "=" << (value ? "true" : "false") << "\n";
+    }
     ss << ")";
     return ss.str();
 }
@@ -229,6 +238,40 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
+    // case: preserve reasoning content in chat history
+    caps_try_execute(
+        prog,
+        [&]() {
+            // messages
+            return json::array({
+                {
+                    {"role", "user"},
+                    {"content", "User message"}
+                },
+                {
+                    {"role", "assistant"},
+                    {"content", "Assistant message"},
+                    {"reasoning_content", "Reasoning content"}
+                },
+                {
+                    {"role", "user"},
+                    {"content", "User message"}
+                },
+            });
+        },
+        [&]() {
+            // tools
+            return json::array();
+        },
+        [&](bool, value & messages, value &) {
+            auto & content = messages->at(1)->at("reasoning_content");
+            caps_print_stats(content, "messages[1].reasoning_content");
+            if (content->stats.used) {
+                result.supports_preserve_reasoning = true;
+            }
+        }
+    );
+
     JJ_DEBUG("%s\n", result.to_string().c_str());
 
     return result;

common/jinja/caps.h

@@ -3,6 +3,7 @@
 #include "runtime.h"
 
 #include <string>
+#include <map>
 
 namespace jinja {
 
@@ -11,14 +12,17 @@ struct caps {
     bool supports_tool_calls = true;
     bool supports_system_role = true;
     bool supports_parallel_tool_calls = true;
+    bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
 
     bool requires_typed_content = false; // default: use string content
 
+    // for reporting on server
+    std::map<std::string, bool> to_map() const;
+
     // for debugging
     std::string to_string() const;
 };
 
 caps caps_get(jinja::program & prog);
-void debug_print_caps(const caps & c);
 
 } // namespace jinja

common/ngram-cache.cpp

@@ -192,12 +192,12 @@ void common_ngram_cache_draft(
             break;
         }
 
-        LOG(" - draft candidate: token=%d\n", drafted_token);
+        LOG_DBG(" - draft candidate: token=%d\n", drafted_token);
         draft.push_back(drafted_token);
     }
 }
 
-void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) {
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, const std::string & filename) {
     std::ofstream file_out(filename, std::ios::binary);
     for (std::pair<common_ngram, common_ngram_cache_part> item : ngram_cache) {
         const common_ngram      ngram        = item.first;
@@ -217,10 +217,9 @@ void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & fil
             file_out.write(reinterpret_cast<const char *>(&count), sizeof(int32_t));
         }
     }
-
 }
 
-common_ngram_cache common_ngram_cache_load(std::string & filename) {
+common_ngram_cache common_ngram_cache_load(const std::string & filename) {
     std::ifstream hashmap_file(filename, std::ios::binary);
     if (!hashmap_file) {
         throw std::ifstream::failure("Unable to open file " + filename);

common/ngram-cache.h

@@ -88,12 +88,12 @@ void common_ngram_cache_draft(
 // Save an ngram cache to a file.
 // ngram_cache: the ngram cache to save.
 // filename:    the path under which to save the ngram cache.
-void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, const std::string & filename);
 
 // Load an ngram cache saved with common_ngram_cache_save.
 // filename: the path from which to load the ngram cache.
 // returns:  an ngram cache containing the information saved to filename.
-common_ngram_cache common_ngram_cache_load(std::string & filename);
+common_ngram_cache common_ngram_cache_load(const std::string & filename);
 
 // Merge two ngram caches.
 // ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.

common/ngram-map.cpp

@@ -0,0 +1,457 @@
+#include "common.h"
+#include "log.h"
+#include "ngram-map.h"
+
+#include <cinttypes>
+#include <cstdint>
+#include <cstdio>
+#include <sstream>
+
+// Print the values of a sublist of `llama_tokens & inp` to a string in the form [v0, v1, v2, ...].
+static std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length) {
+    std::ostringstream oss;
+    oss << '[';
+    for (size_t i = 0; i < length; ++i) {
+        if (i > 0) {
+            oss << ", ";
+        }
+        oss << inp[start + i];
+    }
+    oss << ']';
+    return oss.str();
+}
+
+
+// n-gram simple
+//
+
+/**
+ * Perform speculative generation using the model's own token history.
+ * Searches for a matching pattern in the token history and returns draft tokens.
+ *
+ * @param state     Current state of this implementation
+ * @param tokens    Token history to search in
+ * @param sampled   Last sampled token
+ * @return Vector of draft tokens, empty if no matching pattern is found
+ */
+llama_tokens common_ngram_simple_draft(
+        common_ngram_simple_state & state,
+        const llama_tokens & tokens, llama_token sampled) {
+
+    // Simple implementation of self-speculative decoding without a draft model.
+    //
+    const size_t cur_len = tokens.size();
+    // Only check every check_rate tokens to save compute
+    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
+    if (state.idx_last_check + state.config.check_rate > cur_len && cur_len > state.idx_last_check) {
+        llama_tokens draft_tokens;
+        return draft_tokens;
+    }
+
+    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
+    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+
+    // vector for tokens we want to verify.
+    // return empty vector if there is no match.
+    llama_tokens draft_tokens;
+
+    // We need at least n_draft_min + n_draft_max + 1 tokens.
+    if (cur_len <= static_cast<size_t>(n_draft_min + n_draft_max + 1)) {
+        return draft_tokens;
+    }
+
+    // pattern search
+    llama_tokens pattern;
+    pattern.reserve(n_draft_min);
+    for (size_t j = cur_len - n_draft_min + 1; j < cur_len; ++j) {
+        pattern.push_back(tokens[j]);
+    }
+    pattern.push_back(sampled); // add the last token to the pattern
+
+    // We do a search in the token history.
+    state.idx_last_check = cur_len;
+
+    size_t match_pos = 0; // we ignore position 0, position 0 == no match
+                          // search backwards, but skip the current match (we are currently there)
+    for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
+        bool match = true;
+        for (size_t k = 0; k < pattern.size(); ++k) {
+            if (tokens[j + k] != pattern[k]) {
+                match = false;
+                break;
+            }
+        }
+        if (match) {
+            match_pos = j;
+            break;
+        }
+    }
+    if (match_pos == 0) {
+        return draft_tokens;
+    }
+
+    const size_t copy_max = std::min(
+            n_draft_max,
+            cur_len - (match_pos + n_draft_min)
+            );
+    if (copy_max < n_draft_min) {
+        return draft_tokens;
+    }
+    LOG_DBG("%s: #tokens = %zu: found matching pattern at pos %zu, length %zu, draft length %zu\n",
+            __func__, cur_len,
+            match_pos, pattern.size(), copy_max);
+
+    draft_tokens.reserve(copy_max);
+    for (size_t j = 0; j < copy_max; ++j) {
+        draft_tokens.push_back(tokens[match_pos + n_draft_min + j]);
+    }
+    return draft_tokens;
+}
+
+
+// n-gram map
+//
+
+// maximum number of counted values of a ngram map value.
+#define COMMON_NGRAM_MAX_VALUE_COUNT 16380
+
+void common_ngram_map_draft(common_ngram_map & map,
+        const llama_tokens & inp, llama_token sampled,
+        llama_tokens & draft) {
+    // reset last key and value.
+    map.last_draft_created   = false;
+    map.last_draft_key_idx   = 0;
+    map.last_draft_value_idx = 0;
+
+    const size_t cur_len = inp.size();
+    const uint16_t n = map.size_key;
+    const uint16_t m = map.size_value;
+    if (cur_len < static_cast<size_t>(2 * n + m)) {
+        return;
+    }
+
+    // Only check every check_rate tokens to save compute
+    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
+    if (map.idx_last_check + map.check_rate > cur_len && cur_len > map.idx_last_check) {
+        return;
+    }
+    map.idx_last_check = cur_len;
+
+    // search pattern, the key n-gram
+    std::vector<llama_token> key_tokens;
+    key_tokens.reserve(n);
+    for (size_t j = cur_len - n + 1; j < cur_len; ++j) {
+        key_tokens.push_back(inp[j]);
+    }
+    key_tokens.push_back(sampled);
+
+    // search for the key in the map
+    size_t match_pos = 0;
+    for (size_t j = cur_len - n - m - 1; j > 0; --j) {
+        bool match = true;
+        for (size_t k = 0; k < n; ++k) {
+            if (inp[j + k] != key_tokens[k]) {
+                match = false;
+                break;
+            }
+        }
+        if (match) {
+           match_pos = j;
+           break;
+        }
+    }
+    if (match_pos > 0) {
+        LOG_INF("%s: cur_len = %zu, n = %d, m = %d, sz_tkns = %zu, sampled = %d, match_pos = %zu\n", __func__,
+            cur_len, n, m, key_tokens.size(), sampled, match_pos);
+    }
+
+    if (match_pos == 0) {
+        return;
+    }
+
+    // We have a match, now we look for the statistics of the key.
+    size_t key_offset = map.keys.size(); // offset in the map
+    // We iterate through the std::vector<common_ngram_map_key> map->keys.
+    for (size_t i = 0; i < map.keys.size(); ++i) {
+        bool match = true;
+        for (size_t j = 0; j < n; ++j) {
+            if (inp[map.keys[i].key_idx + j] != key_tokens[j]) {
+                match = false;
+                break;
+            }
+        }
+        if (match) {
+            key_offset = i;
+            break;
+        }
+    }
+    if (key_offset == map.keys.size()) {
+        // We create a new key-entry, it will get offset key_offset.
+        common_ngram_map_key new_key;
+        new_key.key_idx = match_pos;
+        new_key.stat_idx = 0;
+        new_key.key_num = 0;
+        for (int i = 0; i < COMMON_NGRAM_MAX_VALUES; ++i) {
+            new_key.values[i].value_num = 0;
+            new_key.values[i].n_accepted = m;
+        }
+        map.keys.push_back(new_key);
+    }
+
+    // our key n-gram:
+    common_ngram_map_key & curr_key = map.keys[key_offset];
+
+    // update number of key hits
+    curr_key.key_num = (uint16_t) std::min((int) map.keys[key_offset].key_num + 1,
+            (int) COMMON_NGRAM_MAX_VALUE_COUNT);
+
+    if (map.key_only) {
+        // simple mode:
+        // Fill in the draft with the m tokens following the key.
+        // We work with value values[0] only.
+        int n_draft_tokens = std::min((int) m, (int) curr_key.values[0].n_accepted);
+
+        for (int i = 0; i < n_draft_tokens; ++i) {
+            draft.push_back(inp[match_pos + n + i]);
+        }
+
+        LOG_INF("%s: key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
+                key_offset, curr_key.key_num, draft.size());
+
+        map.last_draft_created   = false;
+        map.last_draft_key_idx   = key_offset;
+        map.last_draft_value_idx = 0; // value 0 is used for simple mode
+        return;
+    }
+
+    if (curr_key.key_num < map.min_hits) {
+        // not enough hits to consider this a good draft
+        LOG_DBG("%s: key_offset = %zu, key_num = %d, min_hits = %d, no draft\n", __func__,
+                key_offset, curr_key.key_num, map.min_hits);
+        return;
+    }
+
+    // complex mode: examine the different m-grams after this key n-gram.
+    //
+
+    // determine all (max COMMON_NGRAM_MAX_VALUES) m-grams after the key n-gram.
+    for (size_t i = curr_key.stat_idx; i <= match_pos; ++i) {
+        // begins the key n-gram at index i?
+        bool match_key = true;
+        for (size_t k = 0; k < n; ++k) {
+            if (inp[i + k] != key_tokens[k]) {
+                match_key = false;
+                break;
+            }
+        }
+        if (!match_key) {
+            continue;
+        }
+
+        // Do we haven a existing value m-gram or a new one after the key at index i?
+        size_t idx_begin_value_key = i + n;
+        int idx_value = -1;
+        for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+            size_t idx_begin_value_v = curr_key.values[v].value_idx;
+            if (idx_begin_value_v == 0) {
+                // We found an empty value slot => we found a new value m-gram after the key n-gram.
+                curr_key.values[v].value_idx = idx_begin_value_key;
+                curr_key.values[v].value_num = 0;
+                curr_key.values[v].n_accepted = m;
+                idx_value = v;
+                break;
+            }
+            bool match = true;
+            for (size_t j = 0; j < m; ++j) {
+                if (inp[idx_begin_value_key + j] != inp[idx_begin_value_v + j]) {
+                    match = false;
+                    break;
+                }
+            }
+            if (match) {
+                // We found an existing value m-gram after the key n-gram.
+                idx_value = v;
+                break;
+            }
+        }
+        if (idx_value >= 0) {
+            // We found a value m-gram of the key n-gram.
+            curr_key.values[idx_value].value_num = (uint16_t) std::min((int) curr_key.values[idx_value].value_num + 1,
+                    (int) COMMON_NGRAM_MAX_VALUE_COUNT);
+        }
+    }
+    // the statistics are updated up to match_pos.
+    curr_key.stat_idx = match_pos;
+
+    // Do we have a value we could use for the draft?
+    uint16_t max_occur = 0;
+    int slot_max = 0;
+    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+        uint16_t curr_occur = curr_key.values[v].value_num;
+        if (curr_occur > max_occur) {
+            max_occur = curr_occur;
+            slot_max = v;
+        }
+    }
+    // What is sum of the other occurences?
+    uint32_t sum_occur = 0;
+    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+        if (v == slot_max) {
+            continue;
+        }
+        uint16_t curr_occur = curr_key.values[v].value_num;
+        sum_occur += curr_occur;
+    }
+
+    LOG_INF("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__,
+            key_offset,
+            max_occur, sum_occur, slot_max,
+            curr_key.values[0].value_idx, curr_key.values[0].value_num,
+            curr_key.values[1].value_idx, curr_key.values[1].value_num,
+            curr_key.values[2].value_idx, curr_key.values[2].value_num,
+            curr_key.values[3].value_idx, curr_key.values[3].value_num
+        );
+    // Print the tokens of the four values (if idx != 0), use LOG_INF
+    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+        if (curr_key.values[v].value_idx != 0) {
+            LOG_INF("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str());
+        }
+    }
+
+    if (sum_occur > 0 && max_occur < 3 * sum_occur) {
+        // The most frequent value is not much more frequent than the other values.
+        // We do not use the draft.
+        return;
+    }
+
+    // We use the most frequent value values[slot_max] for the draft.
+    // Fill in the draft with the m tokens following the key.
+    int n_draft_tokens = std::min((int) m, (int) curr_key.values[slot_max].n_accepted);
+
+    for (int i = 0; i < n_draft_tokens; ++i) {
+        draft.push_back(inp[match_pos + n + i]);
+    }
+
+    LOG_INF("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__,
+            key_offset, slot_max,
+            curr_key.key_num, draft.size());
+
+    map.last_draft_created   = true;
+    map.last_draft_key_idx   = key_offset;
+    map.last_draft_value_idx = slot_max; // value used for draft generation.
+}
+
+void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) {
+    if (!map.last_draft_created) {
+        return;
+    }
+
+    // find the key and its chosen value.
+    const size_t key_idx = map.last_draft_key_idx;
+    const size_t val_idx = map.last_draft_value_idx;
+
+    // find key corresponding to key_idx.
+    common_ngram_map_key & curr_key = map.keys[key_idx];
+    // find value corresponding to val_idx.
+    struct common_ngram_map_value & curr_value = curr_key.values[val_idx]; // value used for draft generation.
+
+    // update the value statistics
+    LOG_INF("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
+            n_accepted, curr_value.n_accepted);
+    curr_value.n_accepted = n_accepted;
+}
+
+//
+// n-gram mod
+//
+
+common_ngram_mod::common_ngram_mod(uint16_t m) : m(m) {
+    int64_t n = 1;
+    for (int32_t i = 0; i < N_MODS; ++i) {
+        n *= mods[i];
+    }
+
+    entries.resize(n);
+
+    const size_t size_bytes = entries.size() * sizeof(common_ngram_mod_entry);
+
+    LOG_INF("%s: size = %.3f MB\n", __func__, size_bytes / (1024.0 * 1024.0));
+}
+
+void common_ngram_mod::add(const llama_token * tokens) {
+    const uint64_t i = idx(tokens);
+
+    common_ngram_mod_entry & entry = entries[i];
+
+    if (entry.n_choices < COMMON_NGRAM_MOD_MAX_CHOICES) {
+        entry.n_choices++;
+    }
+
+    entry.choices[entry.head] = tokens[N_MODS];
+    entry.head = (entry.head + 1) % COMMON_NGRAM_MOD_MAX_CHOICES;
+}
+
+llama_token common_ngram_mod::get(const llama_token * tokens, int32_t offs) const {
+    const uint64_t i = idx(tokens);
+
+    const common_ngram_mod_entry & entry = entries[i];
+
+    if (entry.n_choices == 0) {
+        return LLAMA_TOKEN_NULL;
+    }
+
+    const int32_t k = (offs + entry.head) % entry.n_choices;
+
+    return entry.choices[k];
+}
+
+uint64_t common_ngram_mod::idx(const llama_token * tokens) {
+    uint64_t rh = 0;
+    uint64_t res = 0;
+    for (uint64_t i = 0; i < N_MODS; ++i) {
+        rh = rh * 31 + tokens[i];
+        res = res * mods[i] + (rh % mods[i]);
+    }
+    return res;
+}
+
+void common_ngram_mod_draft(
+        common_ngram_mod & mod,
+        const llama_tokens & inp,
+        llama_token sampled,
+        llama_tokens & draft) {
+    const size_t N_MODS = common_ngram_mod::N_MODS;
+
+    const size_t cur_len = inp.size();
+    if (cur_len < N_MODS) {
+        return;
+    }
+
+    if (mod.n_calls++ % 64 == 0) {
+        const size_t n_start = (256*(mod.n_calls/64)) % GGML_PAD(cur_len, 256);
+        for (size_t i = 0; i < 256 && n_start + i < cur_len - N_MODS; ++i) {
+            mod.add(inp.data() + n_start + i);
+        }
+    }
+
+    draft.resize(N_MODS + mod.m);
+    for (size_t i = 0; i < N_MODS - 1; ++i) {
+        draft[i] = inp[cur_len - N_MODS + 1 + i];
+    }
+    draft[N_MODS - 1] = sampled;
+
+    for (size_t i = 0; i < mod.m; ++i) {
+        const llama_token token = mod.get(draft.data() + i, cur_len + i);
+        if (token == LLAMA_TOKEN_NULL) {
+            draft.clear();
+            return;
+        }
+        draft[N_MODS + i] = token;
+    }
+
+    // only return the m tokens that were drafted
+    for (size_t i = 0; i < mod.m; ++i) {
+        draft[i] = draft[N_MODS + i];
+    }
+    draft.resize(mod.m);
+}

common/ngram-map.h

@@ -0,0 +1,143 @@
+#pragma once
+//
+// common/ngram-map.h: structures used to manage a map from n-grams to a list of m-grams
+//
+// These structures are used to do a lookup of n-grams followed by m-grams in token history.
+//
+// There are two algorithms implemented:
+// 1. ngram_simple: lookup of n-grams followed by m-grams in token history.
+// 2. ngram_map: lookup of n-grams followed by m-grams in token history using a map.
+//    The map is a vector of key n-grams, and for each key n-gram there is a list of value m-grams.
+//
+
+#include "llama.h"
+#include "common.h"
+
+#include <vector>
+
+// n-gram simple
+//
+
+// config of n-gram simple.
+struct common_ngram_simple_config {
+    uint16_t   size_ngram;      // size of n-grams to lookup in self-mode
+    uint16_t   size_mgram;      // size of m-grams to draft in self-mode
+    uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
+};
+
+// current state (and config) of n-gram simple.
+struct common_ngram_simple_state {
+    common_ngram_simple_config config;
+
+    size_t idx_last_check = 0; // index of last check in context history (mutable)
+
+    common_ngram_simple_state(const common_ngram_simple_config & config)
+        : config(config) {}
+};
+
+// Searches for a n-gram in the history and checks whether a draft sequence should be generated.
+// state:              the ngram simple state to search in.
+// inp:                the tokens generated so far.
+// sampled:            the token that was just sampled.
+// draft:              vector to store the draft tokens, initially empty.
+llama_tokens common_ngram_simple_draft(
+        common_ngram_simple_state & state,
+        const llama_tokens & tokens, llama_token sampled);
+
+
+// n-gram map
+//
+
+// maximum number of m-gram values stored for each key n-gram.
+#define COMMON_NGRAM_MAX_VALUES 4
+
+// statistics of a m-gram after a known n-gram
+struct common_ngram_map_value {
+    size_t   value_idx = 0;  // index of value m-gram in token-history (0 if unused)
+    uint16_t value_num = 0;  // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
+    int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
+};
+
+// statistics of a n-gram
+struct common_ngram_map_key {
+    size_t   key_idx;   // index of key n-gram in token-history
+    size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)
+
+    uint16_t key_num;   // number of occurences of this key n-gram in token-history
+    common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
+};
+
+// map from n-grams to following m-grams in token-history
+struct common_ngram_map {
+    uint16_t size_key;   // size of key n-grams
+    uint16_t size_value; // size of value m-grams
+
+    bool key_only;       // true if only key n-grams are used, no values.
+
+    // first draft: vector only, no map.
+    std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
+    uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
+    uint16_t min_hits;   // minimum number of key hits to consider a draft
+
+    common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
+                     uint16_t check_rate, uint16_t min_hits)
+        : size_key(sz_key), size_value(sz_value), key_only(only_keys),
+          check_rate(check_rate), min_hits(min_hits) {}
+
+    bool     last_draft_created   = false; // true if a draft was created at last call.
+    size_t   last_draft_key_idx   = 0; // index of last key used for draft generation.
+    uint16_t last_draft_value_idx = 0; // index of last value used for draft generation.
+
+    size_t   idx_last_check       = 0; // index of last check in context history
+};
+
+
+// Searches for the n-gram in the history and checks whether a draft sequence should be generated.
+// map:                the ngram map to search in.
+// inp:                the tokens generated so far.
+// sampled:            the token that was just sampled.
+// draft:              vector to store the draft tokens, initially empty.
+void common_ngram_map_draft(
+    common_ngram_map & map,
+    const llama_tokens & inp, llama_token sampled,
+    llama_tokens & draft);
+
+// Update the statistics of a value after a draft was processed.
+void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted);
+
+//
+// n-gram mod
+//
+
+#define COMMON_NGRAM_MOD_MAX_CHOICES 4
+
+struct common_ngram_mod_entry {
+    uint32_t head = 0;
+    uint32_t n_choices = 0;
+
+    llama_token choices[COMMON_NGRAM_MOD_MAX_CHOICES];
+};
+
+struct common_ngram_mod {
+    common_ngram_mod(uint16_t m);
+
+    void        add(const llama_token * tokens);
+    llama_token get(const llama_token * tokens, int32_t offs) const;
+
+    uint64_t n_calls = 0;
+
+    uint16_t m;
+
+    std::vector<common_ngram_mod_entry> entries;
+
+    static constexpr int32_t N_MODS = 17;
+    static constexpr int32_t mods[N_MODS] = { 2, 1, 1, 1, 8, 1, 1, 1, 16, 1, 1, 1, 32, 1, 1, 1, 64, };
+
+    static uint64_t idx(const llama_token * tokens);
+};
+
+void common_ngram_mod_draft(
+    common_ngram_mod & mod,
+    const llama_tokens & inp,
+    llama_token sampled,
+    llama_tokens & draft);

common/speculative.cpp

@@ -1,97 +1,54 @@
 #include "speculative.h"
 
+#include "common.h"
 #include "ggml.h"
 #include "llama.h"
 #include "log.h"
-#include "common.h"
+#include "ngram-cache.h"
+#include "ngram-map.h"
 #include "sampling.h"
 
-#include <cstring>
 #include <algorithm>
+#include <cstring>
+#include <iomanip>
 #include <map>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
 
-struct common_speculative {
-    struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
-    struct llama_context * ctx_dft;
-    struct common_sampler * smpl;
-
-    llama_batch batch;
-    llama_tokens prompt_dft;
-    bool vocab_dft_compatible = true; // whether retokenization is needed
-    std::map<std::string, std::string> tgt_dft_replacements = {};
+const std::vector<enum common_speculative_type> common_speculative_types = {
+    COMMON_SPECULATIVE_TYPE_NONE,
+    COMMON_SPECULATIVE_TYPE_DRAFT,
+    COMMON_SPECULATIVE_TYPE_EAGLE3,
+    COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD,
+    COMMON_SPECULATIVE_TYPE_NGRAM_CACHE
 };
 
-struct common_speculative * common_speculative_init(
-        struct llama_context * ctx_tgt,
-        struct llama_context * ctx_dft) {
-    auto * result = new common_speculative {
-        /* .ctx_tgt    = */ ctx_tgt,
-        /* .ctx_dft    = */ ctx_dft,
-        /* .smpl       = */ nullptr,
-        /* .batch      = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
-        /* .prompt_dft = */ {},
-        /* .vocab_dft_compatible = */ false,
-    };
+const std::map<std::string, enum common_speculative_type> common_speculative_type_from_name_map = {
+    {"none",          COMMON_SPECULATIVE_TYPE_NONE},
+    {"draft",         COMMON_SPECULATIVE_TYPE_DRAFT},
+    {"eagle3",        COMMON_SPECULATIVE_TYPE_EAGLE3},
+    {"ngram_simple",  COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
+    {"ngram_map_k",   COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
+    {"ngram_map_k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V},
+    {"ngram_map_mod", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD},
+    {"ngram_cache",   COMMON_SPECULATIVE_TYPE_NGRAM_CACHE}
+};
 
-    // TODO: optimize or pass from outside?
-#if 0
-    {
-        common_params_sampling params;
-        params.no_perf = false;
+struct common_speculative_config {
+    common_speculative_type type;
+    common_params_speculative params;
 
-        params.top_k = 40;
-        params.top_p = 0.9;
-
-        params.samplers = {
-            COMMON_SAMPLER_TYPE_TOP_K,
-            COMMON_SAMPLER_TYPE_TOP_P,
-            COMMON_SAMPLER_TYPE_INFILL,
-        };
-
-        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
-    }
-#else
-    {
-        common_params_sampling params;
-        params.no_perf = false;
-
-        params.top_k = 10;
-
-        params.samplers = {
-            COMMON_SAMPLER_TYPE_TOP_K,
-        };
-
-        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
-    }
-#endif
-
-    result->vocab_dft_compatible = common_speculative_are_compatible(ctx_tgt, ctx_dft);
-    LOG_DBG("vocab_dft_compatible = %d\n", result->vocab_dft_compatible);
-
-    return result;
-}
-
-void common_speculative_free(struct common_speculative * spec) {
-    if (spec == nullptr) {
-        return;
-    }
-
-    common_sampler_free(spec->smpl);
-
-    llama_batch_free(spec->batch);
-
-    delete spec;
-}
-
-bool common_speculative_are_compatible(
-    const struct llama_context * ctx_tgt,
-    const struct llama_context * ctx_dft) {
-    const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
-    const struct llama_model * model_dft = llama_get_model(ctx_dft);
+    common_speculative_config(common_speculative_type t,
+            const common_params_speculative & p = common_params_speculative{}) : type(t), params(p) {}
+};
 
+static bool common_speculative_are_compatible(
+    const struct llama_model * model_tgt,
+    const struct llama_model * model_dft) {
     const struct llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
     const struct llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
 
@@ -134,11 +91,12 @@ bool common_speculative_are_compatible(
         for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
             const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
             const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
+
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
                 LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__);
                 LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i,
-                        common_token_to_piece(ctx_tgt, i).c_str(),
-                        common_token_to_piece(ctx_dft, i).c_str());
+                        common_token_to_piece(vocab_tgt, i).c_str(),
+                        common_token_to_piece(vocab_dft, i).c_str());
                 return false;
             }
         }
@@ -147,50 +105,437 @@ bool common_speculative_are_compatible(
     return true;
 }
 
-void common_speculative_add_replacement_tgt_dft(
-        struct common_speculative * spec,
-        const char *source, const char *dest) {
-    spec->tgt_dft_replacements[source] = dest;
+// state of an implementation of speculative decoding
+//
+// each implementation has a unique type and a state that is implementation-specific
+// in a subclass of common_speculative_state
+struct common_speculative_state {
+    const enum common_speculative_type type;
+
+    size_t drafts_call_count       = 0; // number of times this implementation was called.
+    size_t drafts_generated_count  = 0; // number of times a draft or part was generated by this implementation.
+    size_t drafts_accepted_count   = 0; // number of times a draft or part was accepted by the target model.
+    size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
+    size_t drafts_accepted_tokens  = 0; // number of tokens accepted by the target model.
+
+    // TODO: track performance of most recent calls
+    const bool gen_perf = true; // whether to generate performance stats.
+
+    int64_t gen_duration_us = 0; // total time spent in this implementation in microseconds.
+
+    virtual ~common_speculative_state() = default;
+
+    common_speculative_state(enum common_speculative_type type) : type(type) {}
+};
+
+struct common_speculative_state_draft : public common_speculative_state {
+    struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
+    struct llama_context * ctx_dft;
+
+    struct common_sampler * smpl;
+
+    llama_batch  batch;
+    llama_tokens prompt_dft;
+
+    bool vocab_cmpt = true; // whether retokenization is needed
+    std::unordered_map<std::string, std::string> vocab_map;
+
+    common_speculative_state_draft(
+            enum common_speculative_type type,
+            struct llama_context * ctx_tgt,
+            struct llama_context * ctx_dft,
+            const std::vector<std::pair<std::string, std::string>> & replacements)
+        : common_speculative_state(type)
+        , ctx_tgt(ctx_tgt)
+        , ctx_dft(ctx_dft)
+    {
+        batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
+        smpl = nullptr;
+
+        // TODO: optimize or pass from outside?
+        // {
+        //     common_params_sampling params;
+        //     params.no_perf = false;
+        //
+        //     params.top_k = 40;
+        //     params.top_p = 0.9;
+        //
+        //     params.samplers = {
+        //         COMMON_SAMPLER_TYPE_TOP_K,
+        //         COMMON_SAMPLER_TYPE_TOP_P,
+        //         COMMON_SAMPLER_TYPE_INFILL,
+        //     };
+        //
+        //     result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+        // }
+        {
+            common_params_sampling params;
+            params.no_perf = false;
+            params.top_k = 10;
+            params.samplers = {
+                COMMON_SAMPLER_TYPE_TOP_K,
+            };
+
+            smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+        }
+
+        vocab_cmpt = common_speculative_are_compatible(llama_get_model(ctx_tgt), llama_get_model(ctx_dft));
+        LOG_DBG("vocab_cmpt = %d\n", vocab_cmpt);
+
+        if (!vocab_cmpt) {
+            LOG_WRN("the target and draft vocabs are not compatible - tokens will be translated between the two\n");
+
+            for (const auto & pair : replacements) {
+                vocab_map[pair.first] = pair.second;
+            }
+        }
+    }
+
+    ~common_speculative_state_draft() override {
+        llama_perf_context_print(ctx_dft);
+
+        llama_free(ctx_dft);
+
+        common_sampler_free(smpl);
+
+        llama_batch_free(batch);
+    }
+};
+
+struct common_speculative_state_eagle3 : public common_speculative_state {
+    common_speculative_state_eagle3(enum common_speculative_type type) : common_speculative_state(type) {}
+};
+
+// state of self-speculation (simple implementation, not ngram-map)
+struct common_speculative_state_ngram_simple : public common_speculative_state {
+
+    common_ngram_simple_state state;
+
+    common_speculative_state_ngram_simple(
+            enum common_speculative_type type,
+            common_ngram_simple_state state)
+        : common_speculative_state(type), state(state) {}
+};
+
+struct common_speculative_state_ngram_map_k : public common_speculative_state {
+    // draft ngram map for speculative decoding without draft model
+    common_ngram_map map;
+
+    common_speculative_state_ngram_map_k(
+            enum common_speculative_type type,
+            common_ngram_map map)
+        : common_speculative_state(type), map(std::move(map)) {}
+};
+
+struct common_speculative_state_ngram_map_k4v : public common_speculative_state_ngram_map_k {
+    common_speculative_state_ngram_map_k4v(
+            enum common_speculative_type type,
+            common_ngram_map map)
+        : common_speculative_state_ngram_map_k(type, std::move(map)) {}
+};
+
+struct common_speculative_state_ngram_mod : public common_speculative_state {
+    common_ngram_mod mod;
+
+    common_speculative_state_ngram_mod(
+            enum common_speculative_type type,
+            common_ngram_mod mod)
+        : common_speculative_state(type), mod(std::move(mod)) {}
+};
+
+struct common_speculative_state_ngram_cache : public common_speculative_state {
+    uint16_t n_draft;
+    bool save_dynamic;
+    bool save_static;
+
+    common_ngram_cache ngram_cache_context;
+    common_ngram_cache ngram_cache_dynamic;
+    common_ngram_cache ngram_cache_static;
+
+    size_t cache_size = 0; // number of tokens in n-gram cache
+
+    common_speculative_state_ngram_cache(
+            const enum common_speculative_type type,
+            const std::string & path_static,
+            const std::string & path_dynamic,
+            uint16_t            n_draft,
+            bool                save_dynamic,
+            bool                save_static)
+        : common_speculative_state(type)
+        , n_draft(n_draft)
+        , save_dynamic(save_dynamic)
+        , save_static(save_static)
+    {
+        if (!path_static.empty()) {
+            try {
+                ngram_cache_static = common_ngram_cache_load(path_static);
+            } catch (...) {
+                LOG_ERR("failed to open static lookup cache: %s", path_static.c_str());
+                GGML_ABORT("Couldn't read static lookup cache");
+            }
+        }
+
+        if (!path_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = common_ngram_cache_load(path_dynamic);
+            } catch (...) {
+                LOG_ERR("failed to open dynamic lookup cache: %s", path_dynamic.c_str());
+                GGML_ABORT("Couldn't read dynamic lookup cache");
+            }
+        }
+    }
+};
+
+struct common_speculative {
+    std::vector<std::unique_ptr<common_speculative_state>> impls; // list of implementations to use and their states
+    common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats)
+};
+
+static common_ngram_map get_common_ngram_map(const common_speculative_config & config) {
+    uint16_t size_key   = config.params.ngram_size_n;
+    uint16_t size_value = config.params.ngram_size_m;
+    bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+    uint16_t check_rate = config.params.ngram_check_rate;
+    uint16_t min_hits   = config.params.ngram_min_hits;
+
+    return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
+}
+
+static struct common_speculative_state_ngram_cache create_state_ngram_cache(
+        const std::string & path_static, const std::string & path_dynamic,
+        const common_speculative_config & config) {
+    uint16_t n_draft = 8; // TODO get from config?
+
+    // TODO bool param in common/common.h to set save_static/save_dynamic?
+    bool save_static = false;
+    bool save_dynamic = false;
+
+    common_speculative_state_ngram_cache state(config.type, path_static, path_dynamic, n_draft, save_static, save_dynamic);
+
+    return state;
+}
+
+std::string common_speculative_type_name_str() {
+    std::string result;
+    for (size_t i = 0; i < common_speculative_types.size(); i++) {
+        if (i > 0) {
+            result += ", ";
+        }
+        result += common_speculative_type_to_str(common_speculative_types[i]);
+    }
+    return result;
+}
+
+std::string common_speculative_type_to_str(enum common_speculative_type type) {
+    switch (type) {
+        case COMMON_SPECULATIVE_TYPE_NONE:          return "none";
+        case COMMON_SPECULATIVE_TYPE_DRAFT:         return "draft";
+        case COMMON_SPECULATIVE_TYPE_EAGLE3:        return "eagle3";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:  return "ngram_simple";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:   return "ngram_map_k";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram_map_k4v";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD: return "ngram_map_mod";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE:   return "ngram_cache";
+        default:                                    return "unknown";
+    }
+}
+
+enum common_speculative_type common_speculative_type_from_name(const std::string & name) {
+    const auto it = common_speculative_type_from_name_map.find(name);
+    if (it == common_speculative_type_from_name_map.end()) {
+        return COMMON_SPECULATIVE_TYPE_COUNT;
+    }
+    return it->second;
+}
+
+// initialization of the speculative decoding system
+//
+struct common_speculative * common_speculative_init(
+        const struct common_params_speculative & params,
+              struct llama_context             * ctx_tgt,
+        const struct llama_context_params      & cparams_dft,
+              struct llama_model               * model_dft) {
+    llama_context * ctx_dft = nullptr;
+    if (model_dft) {
+        ctx_dft = llama_init_from_model(model_dft, cparams_dft);
+        if (ctx_dft == nullptr) {
+            LOG_ERR("%s", "failed to create draft context\n");
+            return nullptr;
+        }
+    }
+
+    // Compute the implementations to use based on the config and their order of preference
+    std::vector<common_speculative_config> configs = {}; // list of speculative configs to try
+    {
+        bool has_draft = !params.model.path.empty();
+        bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
+
+        bool has_ngram_cache   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
+        bool has_ngram_simple  = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE);
+        bool has_ngram_map_k   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+        bool has_ngram_map_k4v = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+        bool has_ngram_map_mod = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD);
+
+        // In a more complex implementation we could use the same implementation but with different parameters.
+        // This was initially used in PR-18471 but removed to simplify the code.
+        if (has_ngram_simple) {
+            // This implementation can guess a lot of tokens without any draft model.
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, params));
+        }
+        if (has_ngram_map_k) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, params));
+        }
+        if (has_ngram_map_k4v) {
+            // This implementation can guess tokens with high acceptance rate but is more expensive.
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
+        }
+        if (has_ngram_map_mod) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD, params));
+        }
+        if (has_ngram_cache) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
+        }
+        if (has_draft) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT, params));
+        }
+        if (has_draft_eagle3) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_EAGLE3, params));
+        }
+    }
+
+    std::vector<std::unique_ptr<common_speculative_state>> implementations = {};
+
+    for (const common_speculative_config & config : configs) {
+        LOG_DBG("%s: adding implementation %s\n", __func__, common_speculative_type_to_str(config.type).c_str());
+        switch (config.type) {
+            case COMMON_SPECULATIVE_TYPE_NONE:
+                break;
+            case COMMON_SPECULATIVE_TYPE_DRAFT: {
+                implementations.push_back(std::make_unique<common_speculative_state_draft>(config.type,
+                    /* .ctx_tgt      = */ ctx_tgt,
+                    /* .ctx_dft      = */ ctx_dft,
+                    /* .replacements = */ params.replacements
+                ));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_EAGLE3: {
+                implementations.push_back(std::make_unique<common_speculative_state_eagle3>(config.type));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
+                common_ngram_map ngram_map = get_common_ngram_map(config);
+
+                uint16_t ngram_size_key   = ngram_map.size_key;
+                uint16_t mgram_size_value = ngram_map.size_value;
+                uint16_t check_rate       = ngram_map.check_rate;
+
+                auto config_simple = common_ngram_simple_config{
+                    /* .size_ngram      = */ ngram_size_key,
+                    /* .size_mgram      = */ mgram_size_value,
+                    /* .check_rate      = */ check_rate
+                };
+                auto state = std::make_unique<common_speculative_state_ngram_simple>(
+                    /* .type            = */ config.type,
+                    /* .state           = */ common_ngram_simple_state(config_simple)
+                );
+                implementations.push_back(std::move(state));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: {
+                implementations.push_back(std::make_unique<common_speculative_state_ngram_map_k>(
+                    (config.type),
+                    get_common_ngram_map(config)
+                ));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: {
+                implementations.push_back(std::make_unique<common_speculative_state_ngram_map_k4v>(
+                    (config.type),
+                    get_common_ngram_map(config)
+                ));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD: {
+                common_ngram_mod mod(config.params.ngram_size_m);
+                implementations.push_back(std::make_unique<common_speculative_state_ngram_mod>(
+                    (config.type),
+                    std::move(mod)
+                ));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: {
+                auto state = create_state_ngram_cache(
+                        params.lookup_cache_static, params.lookup_cache_dynamic, config);
+                implementations.push_back(std::make_unique<common_speculative_state_ngram_cache>(state));
+
+                break;
+            }
+            default:
+                break;
+        }
+    }
+
+    if (implementations.empty()) {
+        LOG_WRN("%s", "no implementations specified for speculative decoding\n");
+        return nullptr;
+    }
+
+    auto * result = new common_speculative {
+        /* .impls          = */ std::move(implementations)
+    };
+
+    return result;
+}
+
+void common_speculative_free(struct common_speculative * spec) {
+    if (spec == nullptr) {
+        return;
+    }
+
+    delete spec;
 }
 
 static std::string replace_to_dft(
-        struct common_speculative * spec,
-        const std::string& input) {
+        struct common_speculative_state_draft * spec,
+        const std::string & input) {
     std::string result = input;
-    for (const auto & pair : spec->tgt_dft_replacements) {
+
+    for (const auto & pair : spec->vocab_map) {
         size_t pos = result.find(pair.first);
         while (pos != std::string::npos) {
             result.replace(pos, pair.first.length(), pair.second);
             pos = result.find(pair.first, pos + pair.second.length());
         }
     }
+
     return result;
 }
 
 static std::string replace_to_tgt(
-        struct common_speculative * spec,
+        struct common_speculative_state_draft * spec,
         const std::string& input) {
     std::string result = input;
-    for (const auto& pair : spec->tgt_dft_replacements) {
+
+    for (const auto & pair : spec->vocab_map) {
         size_t pos = result.find(pair.second);
         while (pos != std::string::npos) {
             result.replace(pos, pair.second.length(), pair.first);
             pos = result.find(pair.second, pos + pair.first.length());
         }
     }
+
     return result;
 }
 
-
-llama_tokens common_speculative_gen_draft(
-        struct common_speculative * spec,
+static llama_tokens common_speculative_use_draft_model(
+        struct common_speculative_state_draft * spec,
         struct common_speculative_params params,
-        const llama_tokens & prompt_tgt_main_model, // specified in target model vocab
+        const llama_tokens & prompt_tgt, // specified in target model vocab
         llama_token id_last) {
-    auto & batch  = spec->batch;
-    auto & ctx_tgt = spec->ctx_tgt;
-    auto & ctx_dft = spec->ctx_dft;
-    auto & smpl   = spec->smpl;
+    auto & batch      = spec->batch;
+    auto & ctx_tgt    = spec->ctx_tgt;
+    auto & ctx_dft    = spec->ctx_dft;
+    auto & smpl       = spec->smpl;
     auto & prompt_dft = spec->prompt_dft;
 
     auto * mem_dft = llama_get_memory(ctx_dft);
@@ -200,13 +545,16 @@ llama_tokens common_speculative_gen_draft(
 
     const int n_ctx = llama_n_ctx(ctx_dft) - params.n_draft;
 
-    llama_tokens prompt_tgt_draft_model;
-    if (!spec->vocab_dft_compatible) {
+    llama_tokens prompt_cnv;
+    if (!spec->vocab_cmpt) {
         std::string text;
-        text = common_detokenize(ctx_tgt, prompt_tgt_main_model, true);
+
+        text = common_detokenize(ctx_tgt, prompt_tgt, true);
         text = replace_to_dft(spec, text);
+
         LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
-        prompt_tgt_draft_model = common_tokenize(ctx_dft, text, false, true);
+
+        prompt_cnv = common_tokenize(ctx_dft, text, false, true);
 
         // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation
         const auto * model_tgt = llama_get_model(ctx_tgt);
@@ -214,6 +562,7 @@ llama_tokens common_speculative_gen_draft(
 
         int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false);
         GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last");
+
         text.resize(-n_chars);
         llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false);
         text = replace_to_dft(spec, text);
@@ -221,23 +570,22 @@ llama_tokens common_speculative_gen_draft(
         LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
         id_last = common_tokenize(ctx_dft, text, false, true)[0];
     }
-    // prompt_tgt's tokens will always be compatible with ctx_dft
-    const llama_tokens &prompt_tgt =
-        spec->vocab_dft_compatible ? prompt_tgt_main_model : prompt_tgt_draft_model;
 
-    const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
+    const llama_tokens & prompt_cur = spec->vocab_cmpt ? prompt_tgt : prompt_cnv;
+
+    const int i_start = std::max<int>(0, (int) prompt_cur.size() - n_ctx);
 
     // reuse as much as possible from the old draft context
     // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
     for (int i = 0; i < (int) prompt_dft.size(); ++i) {
         int cur = 0;
-        while (i_start + cur < (int) prompt_tgt.size() &&
+        while (i_start + cur < (int) prompt_cur.size() &&
                i       + cur < (int) prompt_dft.size() &&
-               prompt_tgt[i_start + cur] == prompt_dft[i + cur]) {
+               prompt_cur[i_start + cur] == prompt_dft[i + cur]) {
             cur++;
         }
 
-        if ((cur >= params.n_reuse || n_ctx >= (int) prompt_tgt.size()) && cur > reuse_n) {
+        if ((cur >= 256 || n_ctx >= (int) prompt_cur.size()) && cur > reuse_n) {
             reuse_i = i;
             reuse_n = cur;
         }
@@ -282,11 +630,11 @@ llama_tokens common_speculative_gen_draft(
     // prepare a batch to evaluate any new tokens in the prompt
     common_batch_clear(batch);
 
-    for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) {
-        //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
-        common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
+    for (size_t i = i_start + reuse_n; i < prompt_cur.size(); ++i) {
+        //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_cur[i]);
+        common_batch_add(batch, prompt_cur[i], i - i_start, { 0 }, false);
 
-        prompt_dft.push_back(prompt_tgt[i]);
+        prompt_dft.push_back(prompt_cur[i]);
     }
 
     // we should rarely end-up here during normal decoding
@@ -348,7 +696,7 @@ llama_tokens common_speculative_gen_draft(
         prompt_dft.push_back(id);
     }
 
-    if (!spec->vocab_dft_compatible) {
+    if (!spec->vocab_cmpt) {
         std::string detokenized = common_detokenize(ctx_dft, result, true);
         detokenized = replace_to_tgt(spec, detokenized);
         LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
@@ -357,5 +705,211 @@ llama_tokens common_speculative_gen_draft(
             result.resize(params.n_draft);
         }
     }
+
     return result;
 }
+
+/**
+ * Perform speculative generation using a 3-tier n-gram cache.
+ *
+ * @param state     Current state of this implementation
+ * @param tokens    Token history to search in
+ * @param sampled   Last sampled token
+ * @return Vector of draft tokens, empty if draft is found
+ */
+static llama_tokens common_speculative_gen_ngram_cache(
+        common_speculative_state_ngram_cache & state,
+        const llama_tokens & tokens, llama_token sampled) {
+    if (state.cache_size < tokens.size() + 1) {
+        llama_tokens tokens_new;
+        tokens_new.reserve(tokens.size() + 1 - state.cache_size);
+        for (size_t j = state.cache_size; j < tokens.size(); ++j) {
+            tokens_new.push_back(tokens[j]);
+        }
+        tokens_new.push_back(sampled); // add the last token
+
+        // Update context ngram cache with new tokens:
+        common_ngram_cache_update(state.ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX,
+                tokens_new, tokens_new.size(), false);
+        state.cache_size = tokens.size() + 1;
+    }
+
+    llama_tokens inp;
+    inp.reserve(tokens.size() + 1);
+    for (size_t j = 0; j < tokens.size(); ++j) {
+        inp.push_back(tokens[j]);
+    }
+    inp.push_back(sampled);
+
+    llama_tokens draft;
+    draft.push_back(sampled);
+
+    common_ngram_cache_draft(inp, draft, state.n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX,
+            state.ngram_cache_context,
+            state.ngram_cache_dynamic,
+            state.ngram_cache_static);
+
+    if (draft.size() > 0) {
+        // delete first token in draft (which is the sampled token)
+        draft.erase(draft.begin());
+    }
+
+    return draft;
+}
+llama_tokens common_speculative_gen_draft(
+        struct common_speculative * spec,
+        struct common_speculative_params params,
+        const llama_tokens & prompt_tgt, // specified in target model vocab
+        llama_token id_last) {
+    llama_tokens result = {};
+
+    spec->curr_impl = nullptr; // reset current implementation
+
+    // TODO: avoid dynamic casts
+    for (auto & impl : spec->impls) {
+        impl->drafts_call_count++;
+        const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
+
+        switch (impl->type) {
+            case COMMON_SPECULATIVE_TYPE_NONE:
+            {
+            } break;
+            case COMMON_SPECULATIVE_TYPE_DRAFT:
+            {
+                // Create a draft using a draft model.
+                auto * draft_impl = dynamic_cast<struct common_speculative_state_draft *>(impl.get());
+                if (draft_impl) {
+                    result = common_speculative_use_draft_model(draft_impl, params, prompt_tgt, id_last);
+                } else {
+                    GGML_ABORT("unexpected implementation in type %d", impl.get()->type);
+                }
+            } break;
+            case COMMON_SPECULATIVE_TYPE_EAGLE3:
+            {
+                // Work in progress: https://github.com/ggml-org/llama.cpp/pull/18039
+            } break;
+            case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:
+            {
+                // Use common_ngram_map_draft to generate a draft from the current context.
+                auto * state = dynamic_cast<struct common_speculative_state_ngram_simple *>(impl.get());
+                if (state) {
+                    result = common_ngram_simple_draft(state->state, prompt_tgt, id_last);
+                } else {
+                    GGML_ABORT("unexpected implementation in type %d", impl.get()->type);
+                }
+            } break;
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:
+            {
+                // Use common_ngram_map_draft to generate a draft from the current context.
+                auto * state = dynamic_cast<common_speculative_state_ngram_map_k *>(impl.get());
+                if (state) {
+                    common_ngram_map_draft(state->map, prompt_tgt, id_last, result);
+                } else {
+                    GGML_ABORT("unexpected implementation in type %d", impl.get()->type);
+                }
+            } break;
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V:
+            {
+                // Use common_ngram_map_draft to generate a draft from the current context.
+                auto * state = dynamic_cast<common_speculative_state_ngram_map_k *>(impl.get());
+                if (state) {
+                    common_ngram_map_draft(state->map, prompt_tgt, id_last, result);
+                } else {
+                    GGML_ABORT("unexpected implementation in type %d", impl.get()->type);
+                }
+            } break;
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_MOD:
+            {
+                auto * state = dynamic_cast<common_speculative_state_ngram_mod *>(impl.get());
+                if (state) {
+                    common_ngram_mod_draft(state->mod, prompt_tgt, id_last, result);
+                } else {
+                    GGML_ABORT("unexpected implementation in type %d", impl.get()->type);
+                }
+            } break;
+            case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE:
+            {
+                auto * state = dynamic_cast<common_speculative_state_ngram_cache *>(impl.get());
+                if (state) {
+                    result = common_speculative_gen_ngram_cache(*state, prompt_tgt, id_last);
+                } else {
+                    GGML_ABORT("unexpected implementation in type %d", impl.get()->type);
+                }
+            } break;
+            case COMMON_SPECULATIVE_TYPE_COUNT:
+            {
+                GGML_ABORT("invalid speculative type COUNT");
+            }
+        }
+
+        const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
+        impl->gen_duration_us += t_now_us - t_start_us; // accumulate duration for this implementation
+
+        if (!result.empty()) {
+            LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
+                    common_speculative_type_to_str(impl.get()->type).c_str(),
+                    prompt_tgt.size(),
+                    impl.get()->drafts_call_count, result.size());
+            spec->curr_impl = impl.get(); // set current implementation for stats
+            impl->drafts_generated_count++;
+            impl->drafts_generated_tokens += result.size();
+
+            break; // We have a draft, so break out of the loop and return it.
+        }
+    }
+
+    return result;
+}
+
+void common_speculative_accept(struct common_speculative * spec, uint16_t n_accepted) {
+    if (n_accepted == 0) {
+        return;
+    }
+
+    common_speculative_state * impl = spec->curr_impl;
+
+    GGML_ASSERT(impl);
+
+    if (n_accepted > 0) {
+        impl->drafts_accepted_count++;
+        impl->drafts_accepted_tokens += n_accepted;
+    }
+
+    LOG_WRN("XXXXXXXXXXXXX n_accepted = %d\n", n_accepted);
+
+    if (impl->type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K ||
+        impl->type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V) {
+
+        // TODO: add common_speculative_state::accept() to base class and remove this dynamic cast
+        auto * state = dynamic_cast<struct common_speculative_state_ngram_map_k *>(impl);
+        if (state) {
+            common_ngram_map_accept(state->map, n_accepted);
+        }
+    }
+}
+
+void common_speculative_print_stats(const struct common_speculative * spec) {
+    if (spec == nullptr) {
+        return;
+    }
+
+    for (const auto & impl : spec->impls) {
+        std::string str_perf;
+        if (impl->gen_perf) {
+            std::ostringstream oss;
+            oss << std::fixed << std::setprecision(3) << impl->gen_duration_us / 1000.0;
+            str_perf = ", dur = " + oss.str() + " ms";
+        } else {
+            str_perf = "";
+        }
+
+        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
+                common_speculative_type_to_str(impl->type).c_str(),
+                impl->drafts_call_count,
+                impl->drafts_generated_count,
+                impl->drafts_accepted_count,
+                impl->drafts_generated_tokens,
+                impl->drafts_accepted_tokens,
+                str_perf.c_str());
+    }
+}

common/speculative.h

@@ -7,29 +7,36 @@ struct common_speculative;
 
 struct common_speculative_params {
     int n_draft = 16;  // max drafted tokens
-    int n_reuse = 256;
 
     float p_min = 0.75f; // min probability required to accept a token in the draft
 };
 
+// comma separated list of all types
+std::string common_speculative_type_name_str();
+
+// convert string to type
+enum common_speculative_type common_speculative_type_from_name(const std::string & name);
+
+// convert type to string
+std::string common_speculative_type_to_str(enum common_speculative_type type);
+
 struct common_speculative * common_speculative_init(
-        struct llama_context * ctx_tgt,
-        struct llama_context * ctx_dft
-);
+        const struct common_params_speculative & params,
+              struct llama_context             * ctx_tgt,
+        const struct llama_context_params      & cparams_dft,
+              struct llama_model               * model_dft);
 
 void common_speculative_free(struct common_speculative * spec);
 
-bool common_speculative_are_compatible(
-        const struct llama_context * ctx_tgt,
-        const struct llama_context * ctx_dft);
-
-void common_speculative_add_replacement_tgt_dft(
-        struct common_speculative * spec,
-        const char *source, const char *dest);
-
 // sample up to n_draft tokens and add them to the batch using the draft model
 llama_tokens common_speculative_gen_draft(
                struct common_speculative * spec,
         struct common_speculative_params   params,
                       const llama_tokens & prompt,
                              llama_token   id_last);
+
+// informs the speculative decoder that n_accepted tokens were accepted by the target model
+void common_speculative_accept(struct common_speculative * spec, uint16_t n_accepted);
+
+// print statistics about the speculative decoding
+void common_speculative_print_stats(const struct common_speculative * spec);

docs/build.md

@@ -248,6 +248,14 @@ You may set the [cuda environmental variables](https://docs.nvidia.com/cuda/cuda
 CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf
 ```
 
+#### CUDA_SCALE_LAUNCH_QUEUES
+
+The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU.
+
+**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
+
+See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details.
+
 ### Unified Memory
 
 The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.

docs/speculative.md

@@ -0,0 +1,120 @@
+# Speculative Decoding
+
+llama.cpp supports speculative decoding, a technique that can significantly accelerate token generation by predicting multiple tokens ahead of the main model.
+
+[Speculative decoding](https://en.wikipedia.org/wiki/Transformer_(deep_learning)#Speculative_decoding) leverages the fact that computing n tokens in a batch (as in prompt processing) is more efficient than computing n sequentially (as in response generation). By generating draft tokens quickly and then verifying them with the target model in a single batch, this approach can achieve substantial speedups when the draft predictions are frequently correct.
+
+## Implementations
+
+The `llama-server` application supports several implementations of speculative decoding:
+
+### Draft Model (`draft`)
+
+A much smaller model (called the _draft model_) generates drafts.
+A draft model is the most used approach in speculative decoding.
+
+### n-gram Cache (`ngram-cache`)
+
+An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences.
+A draft is computed using probabilities derived from these statistics. External statistics can also be loaded from files for improved accuracy.
+
+See:
+
+- #5479, #6828, #6848
+
+### n-gram Map (`ngram-simple`, `ngram-map-*`)
+
+These implementations search the token history for patterns and use matching sequences as draft candidates.
+They require no additional model but rely on patterns that have already appeared in the generated text.
+An example to use this approach can be the rewriting of source code by a LLM.
+
+#### n-gram Map (`ngram-simple`)
+
+This implementation looks for the last n-gram in history that matches the current n-gram and creates a draft using the m tokens following the matched n-gram. It is the simplest self-speculative approach with minimal overhead.
+
+#### n-gram Map Key (`ngram-map-k`)
+
+This implementation looks for the current n-gram of size n (called the _key_) in the token history. If the key n-gram is followed by the same m tokens (called the _mgram_) multiple times, it creates a draft using these m tokens. This approach requires a minimum number of occurrences (argument `--spec-ngram-min-hits`) before generating drafts.
+
+The number of accepted tokens is stored for each used n-gram.
+
+#### n-gram Map Key-4-Values (`ngram-map-k4v`)
+
+This experimental implementation looks for the current n-gram of size n (called the _key_) in the token history. For each key, up to four _values_ (n-grams of size m, called _mgrams_) are tracked. An internal statistic counts the occurrences of each mgram after the key n-gram. If one mgram is significantly more frequent than the others, it is used as the draft.
+
+The number of accepted tokens is stored for each used n-gram.
+
+**Example:** Server options to be used if there are a lot of longer repetitions.
+```bash
+llama-server [...] --spec-draftless ngram-map-k4v --spec-ngram-size-n 8 --spec-ngram-size-m 8 --spec-ngram-min-hits 2
+```
+
+
+## Command-Line Options (draftless)
+
+If a draft model is combined with a draftless decoding the draftless decoding has higher precedence.
+
+```
+--spec-draftless [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v]
+                                        type of speculative decoding to use when no draft model is provided
+                                        (default: none)
+--spec-ngram-size-n N                   ngram size N for ngram-simple/ngram-map speculative decoding, length
+                                        of lookup n-gram (default: 12)
+--spec-ngram-size-m N                   ngram size M for ngram-simple/ngram-map speculative decoding, length
+                                        of draft m-gram (default: 48)
+--spec-ngram-check-rate N               ngram check rate for ngram-simple/ngram-map speculative decoding
+                                        (default: 1)
+--spec-ngram-min-hits N                 minimum hits for ngram-map speculative decoding (default: 1)
+```
+
+### `--spec-draftless TYPE`
+
+Specifies a type of speculative decoding without draft model.
+
+| Type | Description |
+|------|-------------|
+| `none` | No speculative decoding (default) |
+| `ngram-cache` | Use n-gram cache lookup |
+| `ngram-simple` | Use simple n-gram pattern matching |
+| `ngram-map-k` | Use n-gram pattern matching with n-gram-keys |
+| `ngram-map-k4v` | Use n-gram pattern matching with n-gram-keys and up to four m-gram values (experimental) |
+
+**Example:** Server-instance used to refactor source code.
+```bash
+./llama-server [...] --spec-draftless ngram-simple
+```
+
+### `--spec-ngram-size-n N`
+
+Sets the size N of the lookup n-gram for n-gram map based speculative decoding.
+The n-gram size N determines how many tokens in a row to look back when searching for matching patterns.
+
+### `--spec-ngram-size-m M`
+
+Sets the size M of the draft m-gram for n-gram map based speculative decoding.
+The m-gram size determines how many tokens to draft when a match is found.
+Larger values can provide more speedup but may reduce acceptance rate.
+
+### `--spec-ngram-check-rate R`
+
+This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
+
+### `--spec-ngram-min-hits H`
+
+This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
+
+## Statistics
+Each speculative decoding implementation prints statistics.
+
+```
+draft acceptance rate = 0.57576 (  171 accepted /   297 generated)
+statistics ngram_simple: #calls = 15, #gen drafts = 5, #acc drafts = 5, #gen tokens = 187, #acc tokens = 73
+statistics draft: #calls = 10, #gen drafts = 10, #acc drafts = 10, #gen tokens = 110, #acc tokens = 98
+```
+
+- `#calls`: number of calls of this implementations
+- `#gen drafts`: number of drafts generated by this implementation
+- `#acc drafts`: number of drafts accepted (partially) by the main model
+- `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)
+- `#acc tokens`: number of tokens accepted by the main model
+

examples/lookup/lookup-create.cpp

@@ -32,9 +32,9 @@ int main(int argc, char ** argv){
 
     common_ngram_cache ngram_cache;
     common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
-    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.lookup_cache_static.c_str());
 
-    common_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+    common_ngram_cache_save(ngram_cache, params.speculative.lookup_cache_static);
 
     return 0;
 }

examples/lookup/lookup-stats.cpp

@@ -46,18 +46,18 @@ int main(int argc, char ** argv){
     {
         const int64_t t_start_draft_us = ggml_time_us();
 
-        if (!params.lookup_cache_static.empty()) {
+        if (!params.speculative.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 

examples/lookup/lookup.cpp

@@ -51,18 +51,18 @@ int main(int argc, char ** argv){
         const int64_t t_start_draft_us = ggml_time_us();
         common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
 
-        if (!params.lookup_cache_static.empty()) {
+        if (!params.speculative.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
@@ -210,7 +210,7 @@ int main(int argc, char ** argv){
 
     // Update dynamic ngram cache with context ngram cache and save it to disk:
     common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-    common_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+    common_ngram_cache_save(ngram_cache_dynamic, params.speculative.lookup_cache_dynamic);
 
     LOG("\n\n");
 

examples/model-conversion/scripts/utils/perplexity-gen.sh

@@ -3,6 +3,7 @@
 set -e
 
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -25,9 +26,13 @@ mkdir -p ppl
 OUTPUTFILE="ppl/$(basename $CONVERTED_MODEL).kld"
 echo "Model: $CONVERTED_MODEL"
 
-cmake --build ../../build --target llama-perplexity -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
 
-../.././build/bin/llama-perplexity -m $CONVERTED_MODEL \
+cmake --build $BUILD_DIR --target llama-perplexity -j8
+
+${BUILD_DIR}/bin/llama-perplexity -m $CONVERTED_MODEL \
     -f ppl/wikitext-2-raw/wiki.test.raw \
     --kl-divergence-base $OUTPUTFILE
 

examples/model-conversion/scripts/utils/perplexity-run-simple.sh

@@ -3,6 +3,7 @@
 set -e
 
 QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 if [ -z "$QUANTIZED_MODEL" ]; then
     echo "Error: Model path must be provided either as:" >&2
@@ -20,8 +21,12 @@ if [ ! -d "ppl/wikitext-2-raw" ]; then
     popd
 fi
 
-cmake --build ../../build --target llama-perplexity -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
 
-../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
+cmake --build $BUILD_DIR --target llama-perplexity -j8
+
+${BUILD_DIR}/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
 
 

examples/model-conversion/scripts/utils/perplexity-run.sh

@@ -3,7 +3,8 @@
 set -e
 
 QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
-LOGITS_FILE="${1:-"$LOGITS_FILE"}"
+LOGITS_FILE="${2:-"$LOGITS_FILE"}"
+BUILD_DIR="${3:-"$BUILD_DIR"}"
 
 if [ -z "$QUANTIZED_MODEL" ]; then
     echo "Error: Model path must be provided either as:" >&2
@@ -18,11 +19,15 @@ if [ ! -f ${LOGITS_FILE} ]; then
     exit 1
 fi
 
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
 echo "Model: $QUANTIZED_MODEL"
 echo "Data file: $LOGITS_FILE"
 
-cmake --build ../../build --target llama-perplexity -j8
+cmake --build $BUILD_DIR --target llama-perplexity -j8
 
-../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL \
+${BUILD_DIR}/bin/llama-perplexity -m $QUANTIZED_MODEL \
     --kl-divergence-base $LOGITS_FILE \
     --kl-divergence

examples/model-conversion/scripts/utils/quantize.sh

@@ -6,6 +6,7 @@ CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
 QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
 TOKEN_EMBD_TYPE="${3:-"${TOKEN_EMBD_TYPE}"}"
 OUTPUT_TYPE="${4:-"${OUTPUT_TYPE}"}"
+BUILD_DIR="${5:-"$BUILD_DIR"}"
 QUANTIZED_MODEL=$CONVERTED_MODEL
 
 # Final check if we have a model path
@@ -33,12 +34,16 @@ else
     exit 1
 fi
 
-cmake --build ../../build --target llama-quantize -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
+cmake --build $BUILD_DIR --target llama-quantize -j8
 
 echo $TOKEN_EMBD_TYPE
 echo $OUTPUT_TYPE
 
-CMD_ARGS=("../../build/bin/llama-quantize")
+CMD_ARGS=("${BUILD_DIR}/bin/llama-quantize")
 [[ -n "$TOKEN_EMBD_TYPE" ]] && CMD_ARGS+=("--token-embedding-type" "$TOKEN_EMBD_TYPE")
 [[ -n "$OUTPUT_TYPE" ]]     && CMD_ARGS+=("--output-tensor-type" "$OUTPUT_TYPE")
 CMD_ARGS+=("$CONVERTED_MODEL" "$QUANTIZED_MODEL" "$QUANTIZED_TYPE")

examples/model-conversion/scripts/utils/run-embedding-server.sh

@@ -4,6 +4,7 @@ set -e
 #
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -13,10 +14,14 @@ if [ -z "$CONVERTED_MODEL" ]; then
     exit 1
 fi
 
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
 echo $CONVERTED_MODEL
 
-cmake --build ../../build --target llama-server
+cmake --build $BUILD_DIR --target llama-server
 
-../../build/bin/llama-server -m $CONVERTED_MODEL \
+${BUILD_DIR}/bin/llama-server -m $CONVERTED_MODEL \
     --embedding \
     --pooling none

examples/speculative-simple/speculative-simple.cpp

@@ -34,10 +34,9 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     llama_model * model_tgt = NULL;
-    //llama_model * model_dft = NULL;
+    llama_model * model_dft = NULL;
 
     llama_context * ctx_tgt = NULL;
-    llama_context * ctx_dft = NULL;
 
     // load the target model
     auto llama_init_tgt = common_init_from_params(params);
@@ -63,12 +62,7 @@ int main(int argc, char ** argv) {
 
     auto llama_init_dft = common_init_from_params(params);
 
-    //model_dft = llama_init_dft->model();
-    ctx_dft   = llama_init_dft->context();
-
-    if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
-        LOG_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params.speculative.model.path.c_str(), params.model.path.c_str());
-    }
+    model_dft = llama_init_dft->model();
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -129,13 +123,9 @@ int main(int argc, char ** argv) {
     // init the speculator
     struct common_speculative_params params_spec;
     params_spec.n_draft = n_draft;
-    params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
     params_spec.p_min   = p_min;
 
-    struct common_speculative * spec = common_speculative_init(ctx_tgt, ctx_dft);
-    for (auto &pair : params.speculative.replacements) {
-        common_speculative_add_replacement_tgt_dft(spec, pair.first.c_str(), pair.second.c_str());
-    }
+    struct common_speculative * spec = common_speculative_init(params.speculative, ctx_tgt, common_context_params_to_llama(params), model_dft);
 
     llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
 
@@ -249,8 +239,6 @@ int main(int argc, char ** argv) {
     LOG_INF("\n");
     LOG_INF("draft:\n\n");
 
-    llama_perf_context_print(ctx_dft);
-
     LOG_INF("\n");
     LOG_INF("target:\n\n");
     common_perf_print(ctx_tgt, smpl);

ggml/src/ggml-cpu/arch-fallback.h

@@ -38,9 +38,10 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -48,9 +49,10 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -70,12 +72,14 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
@@ -94,9 +98,10 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -104,9 +109,10 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -126,9 +132,10 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -136,9 +143,10 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -165,18 +173,20 @@
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -202,9 +212,10 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -212,9 +223,10 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -242,9 +254,10 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -252,9 +265,10 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -25,9 +25,8 @@
 #define UNUSED GGML_UNUSED
 
 #if defined(__aarch64__) && defined(__ARM_NEON) && (defined(__ARM_FEATURE_MATMUL_INT8) || defined(__ARM_FEATURE_DOTPROD))
-static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
-                                             int16x8_t *     out_mins,
-                                             int8_t *        out_scales) {
+// Helper for decoding scales and mins of Q4_K and Q5_K block formats
+static inline void decode_q_Kx8_6bit_scales(const uint8_t * scales_in, int16x8_t * out_mins, int8_t * out_scales) {
     constexpr uint32_t kmask1 = 0x3f3f3f3f;
     constexpr uint32_t kmask2 = 0x0f0f0f0f;
     constexpr uint32_t kmask3 = 0x03030303;
@@ -561,7 +560,7 @@ void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 for (int i = 0; i < 2; i++) {
                     int8_t    aux_q4sb[8];
                     const int offset = sb * 24 + i * 12;
-                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
                     q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
                 }
 
@@ -701,7 +700,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
                 for (int i = 0; i < 2; i++) {
                     int8_t    aux_q4sb[8];
                     const int offset = sb * 24 + i * 12;
-                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
                     q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
                 }
 
@@ -786,6 +785,293 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
     ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_q5_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+    const uint8x16_t mone      = vdupq_n_u8(1);
+    const uint8x16_t mtwo      = vdupq_n_u8(2);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[ncols_interleaved / 4];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q5_Kx8 * GGML_RESTRICT q5_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < ncols_interleaved / 4; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q5_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q5_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q5_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q5_d_1, q8_d);
+            float32x4_t q5_dmin_0  = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q5_dmin_1  = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0   = vmulq_f32(q5_dmin_0, q8_d);
+            float32x4_t sb_min_1   = vmulq_f32(q5_dmin_1, q8_d);
+
+            // 2 sb each iteration
+            int32x4_t acc_lo[col_pairs];
+            int32x4_t acc_hi[col_pairs];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+
+            // Load qh once per block and shift after each subblock
+            const uint8_t * qh_base = q5_ptr[b].qh;
+            uint8x16_t      qh[col_pairs][4];
+            for (int cp = 0; cp < col_pairs; cp++) {
+                qh[cp][0] = vld1q_u8(qh_base + 16 * cp);
+                qh[cp][1] = vld1q_u8(qh_base + 16 * cp + 64);
+                qh[cp][2] = vld1q_u8(qh_base + 16 * cp + 128);
+                qh[cp][3] = vld1q_u8(qh_base + 16 * cp + 192);
+            }
+
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_pairs; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q5sb_mins[2];  // int16 as its needed for bias_acc later
+                int16x8_t q5sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q5sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q_Kx8_6bit_scales(&q5_ptr[b].scales[offset], &q5sb_mins[i], aux_q5sb);
+                    q5sb_scales[i] = vmovl_s8(vld1_s8(aux_q5sb));
+                }
+
+                const uint8_t * qs_base = q5_ptr[b].qs + sb * QK_K;
+
+                // Load the 64 quants from q8K duplicated to use vecdots with the interleaved columns
+                const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
+                int8x16_t      q8_qs[8];
+                for (int i = 0; i < 8; i++) {
+                    q8_qs[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base + i * 8));
+                }
+
+                // Q5s column pair loop unrolled
+                {
+                    // Cols 01
+                    uint8x16_t qs_0 = vld1q_u8(qs_base);
+                    uint8x16_t qs_1 = vld1q_u8(qs_base + 64);
+                    uint8x16_t qs_2 = vld1q_u8(qs_base + 128);
+                    uint8x16_t qs_3 = vld1q_u8(qs_base + 192);
+
+                    uint8x16_t hbit_lo_0 = vandq_u8(qh[0][0], mone);
+                    uint8x16_t hbit_lo_1 = vandq_u8(qh[0][1], mone);
+                    uint8x16_t hbit_lo_2 = vandq_u8(qh[0][2], mone);
+                    uint8x16_t hbit_lo_3 = vandq_u8(qh[0][3], mone);
+                    uint8x16_t hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[0][0], mtwo), 3);
+                    uint8x16_t hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[0][1], mtwo), 3);
+                    uint8x16_t hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[0][2], mtwo), 3);
+                    uint8x16_t hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[0][3], mtwo), 3);
+
+                    qh[0][0] = vshrq_n_u8(qh[0][0], 2);
+                    qh[0][1] = vshrq_n_u8(qh[0][1], 2);
+                    qh[0][2] = vshrq_n_u8(qh[0][2], 2);
+                    qh[0][3] = vshrq_n_u8(qh[0][3], 2);
+
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+
+                    // Cols 23
+                    qs_0 = vld1q_u8(qs_base + 16);
+                    qs_1 = vld1q_u8(qs_base + 80);
+                    qs_2 = vld1q_u8(qs_base + 144);
+                    qs_3 = vld1q_u8(qs_base + 208);
+
+                    hbit_lo_0 = vandq_u8(qh[1][0], mone);
+                    hbit_lo_1 = vandq_u8(qh[1][1], mone);
+                    hbit_lo_2 = vandq_u8(qh[1][2], mone);
+                    hbit_lo_3 = vandq_u8(qh[1][3], mone);
+                    hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[1][0], mtwo), 3);
+                    hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[1][1], mtwo), 3);
+                    hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[1][2], mtwo), 3);
+                    hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[1][3], mtwo), 3);
+
+                    qh[1][0] = vshrq_n_u8(qh[1][0], 2);
+                    qh[1][1] = vshrq_n_u8(qh[1][1], 2);
+                    qh[1][2] = vshrq_n_u8(qh[1][2], 2);
+                    qh[1][3] = vshrq_n_u8(qh[1][3], 2);
+
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+
+                    // Cols 45
+                    qs_0 = vld1q_u8(qs_base + 32);
+                    qs_1 = vld1q_u8(qs_base + 96);
+                    qs_2 = vld1q_u8(qs_base + 160);
+                    qs_3 = vld1q_u8(qs_base + 224);
+
+                    hbit_lo_0 = vandq_u8(qh[2][0], mone);
+                    hbit_lo_1 = vandq_u8(qh[2][1], mone);
+                    hbit_lo_2 = vandq_u8(qh[2][2], mone);
+                    hbit_lo_3 = vandq_u8(qh[2][3], mone);
+                    hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[2][0], mtwo), 3);
+                    hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[2][1], mtwo), 3);
+                    hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[2][2], mtwo), 3);
+                    hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[2][3], mtwo), 3);
+
+                    qh[2][0] = vshrq_n_u8(qh[2][0], 2);
+                    qh[2][1] = vshrq_n_u8(qh[2][1], 2);
+                    qh[2][2] = vshrq_n_u8(qh[2][2], 2);
+                    qh[2][3] = vshrq_n_u8(qh[2][3], 2);
+
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+
+                    // Cols 45
+                    qs_0 = vld1q_u8(qs_base + 48);
+                    qs_1 = vld1q_u8(qs_base + 112);
+                    qs_2 = vld1q_u8(qs_base + 176);
+                    qs_3 = vld1q_u8(qs_base + 240);
+
+                    hbit_lo_0 = vandq_u8(qh[3][0], mone);
+                    hbit_lo_1 = vandq_u8(qh[3][1], mone);
+                    hbit_lo_2 = vandq_u8(qh[3][2], mone);
+                    hbit_lo_3 = vandq_u8(qh[3][3], mone);
+                    hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[3][0], mtwo), 3);
+                    hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[3][1], mtwo), 3);
+                    hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[3][2], mtwo), 3);
+                    hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[3][3], mtwo), 3);
+
+                    qh[3][0] = vshrq_n_u8(qh[3][0], 2);
+                    qh[3][1] = vshrq_n_u8(qh[3][1], 2);
+                    qh[3][2] = vshrq_n_u8(qh[3][2], 2);
+                    qh[3][3] = vshrq_n_u8(qh[3][3], 2);
+
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+                }
+
+                // Prepare bsum vectors for bias computation
+                // Each pair of subblocks share the same bsums
+                int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                // Iterates over a pair of column pairs (4 columns) to use a single 128 register
+                // p = 0 -> 0123  p2 -> 4567
+                for (int i = 0, p = 0; p < col_pairs; i++, p += 2) {
+                    int16x4_t   group_scales_lo = p == 0 ? vget_low_s16(q5sb_scales[0]) : vget_high_s16(q5sb_scales[0]);
+                    int16x4_t   group_scales_hi = p == 0 ? vget_low_s16(q5sb_scales[1]) : vget_high_s16(q5sb_scales[1]);
+                    int16x4_t   group_mins_lo   = p == 0 ? vget_low_s16(q5sb_mins[0]) : vget_high_s16(q5sb_mins[0]);
+                    int16x4_t   group_mins_hi   = p == 0 ? vget_low_s16(q5sb_mins[1]) : vget_high_s16(q5sb_mins[1]);
+                    float32x4_t sb_scale        = p == 0 ? sb_scale_0 : sb_scale_1;
+                    float32x4_t sb_min          = p == 0 ? sb_min_0 : sb_min_1;
+
+                    // 0123 or 4567
+                    float32x4_t sumf_0 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_lo), vpaddq_s32(acc_lo[p], acc_lo[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_0);
+
+                    float32x4_t sumf_1 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_hi), vpaddq_s32(acc_hi[p], acc_hi[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_1);
+
+                    // FUSED BIAS: Compute and subtract bias immediately
+                    // bias = (bsums_lo * mins_lo + bsums_hi * mins_hi) * sb_min
+                    int32x4_t bias = vmull_s16(bsums_vec_lo, group_mins_lo);
+                    bias = vmlal_s16(bias, bsums_vec_hi, group_mins_hi);
+                    float32x4_t bias_f32 = vcvtq_f32_s32(bias);
+                    acc_f32[i] = vmlsq_f32(acc_f32[i], sb_min, bias_f32);
+                }
+            }  // for sb
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q8_0_4x4_q8_0(int                        n,
                              float * GGML_RESTRICT      s,
                              size_t                     bs,
@@ -2431,7 +2717,7 @@ void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     for (int i = 0; i < 2; i++) {
                         int8_t    aux_q4sb[8];
                         const int offset = sb * 24 + i * 12;
-                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                        decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
                         q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
                     }
 
@@ -2595,7 +2881,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
                     int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
                     for (int i = 0; i < 2; i++) {
                         const int offset = sb * 24 + i * 12;
-                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
+                        decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
                     }
 
                     // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
@@ -2738,6 +3024,252 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
     ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemm_q5_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    col_pairs     = ncols_interleaved / 2;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+    const uint8x16_t mone          = vdupq_n_u8(1);
+    const uint8x16_t mtwo          = vdupq_n_u8(2);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * GGML_RESTRICT q5_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // bsums pairs belongs to the same q8_k subblock
+                const int16x8_t bsums[4]{
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[4][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                int32x4_t sb_acc[4];    // Aux accumulators to store subblock (partial) results
+                int32x4_t acc[8];       // rows 01 stored in [0][1][2][3] rows 23 stored in [4][5][6][7]
+                int32x4_t bias_acc[8];  // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567, [2]->r1 0123 ...
+                for (int i = 0; i < 8; i++) {
+                    acc[i]      = vdupq_n_s32(0);
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                // Load qh once per block and shift after each subblock
+                const uint8_t * qh_base = q5_ptr[b].qh;
+                uint8x16_t      qh[col_pairs][4];
+                for (int cp = 0; cp < col_pairs; cp++) {
+                    qh[cp][0] = vld1q_u8(qh_base + 16 * cp);
+                    qh[cp][1] = vld1q_u8(qh_base + 16 * cp + 64);
+                    qh[cp][2] = vld1q_u8(qh_base + 16 * cp + 128);
+                    qh[cp][3] = vld1q_u8(qh_base + 16 * cp + 192);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int8_t    q5sb_scales[2][8];
+                    int16x8_t q5sb_mins[2];  // int16 as its needed for bias_acc later
+                    for (int i = 0; i < 2; i++) {
+                        const int offset = sb * 24 + i * 12;
+                        decode_q_Kx8_6bit_scales(&q5_ptr[b].scales[offset], &q5sb_mins[i], q5sb_scales[i]);
+                    }
+
+                    // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
+                    const int8_t * q8_base = q8_ptr[b].qs + sb * 256;
+
+                    int8x16_t q8_qs_01[8];
+                    int8x16_t q8_qs_23[8];
+
+                    // Load 32-byte per row pair, 1 subblock each time
+                    for (int i = 0; i < 8; i++) {
+                        const int offset = i * 32;  // 16 for row 01, 16 for row 23
+                        q8_qs_01[i]      = vld1q_s8(q8_base + offset);
+                        q8_qs_23[i]      = vld1q_s8(q8_base + offset + 16);
+                    }
+
+                    const int8x16_t q8s[2][8] = {
+                        { q8_qs_01[0], q8_qs_01[1], q8_qs_01[2], q8_qs_01[3], q8_qs_01[4], q8_qs_01[5], q8_qs_01[6],
+                         q8_qs_01[7] },
+                        { q8_qs_23[0], q8_qs_23[1], q8_qs_23[2], q8_qs_23[3], q8_qs_23[4], q8_qs_23[5], q8_qs_23[6],
+                         q8_qs_23[7] },
+                    };
+
+                    // Q5s columns iterated in pairs (01, 23, 45, 67)
+                    for (int cp = 0; cp < col_pairs; cp++) {
+                        for (int i = 0; i < 4; i++) {
+                            sb_acc[i] = vdupq_n_s32(0);
+                        }
+
+                        uint8x16_t qs_cp_0 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 0);    // 0 .. 7 & 32..39
+                        uint8x16_t qs_cp_1 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 64);   // 8 ..15 & 40..47
+                        uint8x16_t qs_cp_2 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 128);  // 16..23 & 48..55
+                        uint8x16_t qs_cp_3 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 192);  // 24..31 & 56..63
+
+                        // This is the only part of the algorithm that differs with Q4_K
+                        // Extract High bits and pack into 5 bit weights
+                        uint8x16_t hbit_lo_0    = vandq_u8(qh[cp][0], mone);
+                        uint8x16_t hbit_hi_0    = vshlq_n_u8(vandq_u8(qh[cp][0], mtwo), 3);
+                        qh[cp][0]               = vshrq_n_u8(qh[cp][0], 2);
+                        // Same as Q4_K, i8mm to dequantize the weights.
+                        const int8x16_t qs_lo_0 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_0, m4b), hbit_lo_0, 4));
+                        int32x4_t       acc_0   = sb_acc[0];
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_0, q8s[0][0]);
+                        int32x4_t acc_2 = sb_acc[2];
+                        acc_2           = vmmlaq_s32(acc_2, qs_lo_0, q8s[1][0]);
+                        const int8x16_t qs_hi_0 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_0, 4), hbit_hi_0));
+                        int32x4_t       acc_1   = sb_acc[1];
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_0, q8s[0][4]);
+                        int32x4_t acc_3         = sb_acc[3];
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_0, q8s[1][4]);
+
+                        // Repeat for the other 3 columns (8..15, 16..23, 24..31)
+                        uint8x16_t hbit_hi_1    = vshlq_n_u8(vandq_u8(qh[cp][1], mtwo), 3);
+                        uint8x16_t hbit_lo_1    = vandq_u8(qh[cp][1], mone);
+                        qh[cp][1]               = vshrq_n_u8(qh[cp][1], 2);
+                        const int8x16_t qs_lo_1 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_1, m4b), hbit_lo_1, 4));
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_1, q8s[0][1]);
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_1, q8s[1][1]);
+                        const int8x16_t qs_hi_1 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_1, 4), hbit_hi_1));
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_1, q8s[0][5]);
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_1, q8s[1][5]);
+
+                        uint8x16_t hbit_hi_2    = vshlq_n_u8(vandq_u8(qh[cp][2], mtwo), 3);
+                        uint8x16_t hbit_lo_2    = vandq_u8(qh[cp][2], mone);
+                        qh[cp][2]               = vshrq_n_u8(qh[cp][2], 2);
+                        const int8x16_t qs_lo_2 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_2, m4b), hbit_lo_2, 4));
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_2, q8s[0][2]);
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_2, q8s[1][2]);
+                        const int8x16_t qs_hi_2 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_2, 4), hbit_hi_2));
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_2, q8s[0][6]);
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_2, q8s[1][6]);
+
+                        uint8x16_t hbit_lo_3    = vandq_u8(qh[cp][3], mone);
+                        uint8x16_t hbit_hi_3    = vshlq_n_u8(vandq_u8(qh[cp][3], mtwo), 3);
+                        qh[cp][3]               = vshrq_n_u8(qh[cp][3], 2);
+                        const int8x16_t qs_lo_3 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_3, m4b), hbit_lo_3, 4));
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_3, q8s[0][3]);
+                        sb_acc[0]               = acc_0;
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_3, q8s[1][3]);
+                        sb_acc[2]               = acc_2;
+
+                        // Scales[i] corresponds to column i
+                        const int       scale_offset = cp * 2;
+                        const int32_t   s0           = q5sb_scales[0][scale_offset];
+                        const int32_t   s1           = q5sb_scales[0][scale_offset + 1];
+                        const int32x4_t block_scale  = vcombine_s32(vdup_n_s32(s0), vdup_n_s32(s1));
+                        acc[cp]                      = vmlaq_s32(acc[cp], sb_acc[0], block_scale);
+                        acc[cp + 4]                  = vmlaq_s32(acc[cp + 4], sb_acc[2], block_scale);
+
+                        const int8x16_t qs_hi_3 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_3, 4), hbit_hi_3));
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_3, q8s[0][7]);
+                        sb_acc[1]               = acc_1;
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_3, q8s[1][7]);
+                        sb_acc[3]               = acc_3;
+
+                        const int32_t   s2           = q5sb_scales[1][scale_offset];
+                        const int32_t   s3           = q5sb_scales[1][scale_offset + 1];
+                        const int32x4_t block_scale2 = vcombine_s32(vdup_n_s32(s2), vdup_n_s32(s3));
+                        acc[cp]                      = vmlaq_s32(acc[cp], sb_acc[1], block_scale2);
+                        acc[cp + 4]                  = vmlaq_s32(acc[cp + 4], sb_acc[3], block_scale2);
+                    }
+
+                    // Multiply Acc bsum + mins
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        // Each pair of subblocks share the same bsums
+                        // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                        int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][q8_row * 2]);
+                        int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][q8_row * 2 + 1]);
+
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_lo, vget_low_s16(q5sb_mins[0]));
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_hi, vget_low_s16(q5sb_mins[1]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_lo, vget_high_s16(q5sb_mins[0]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_hi, vget_high_s16(q5sb_mins[1]));
+                    }
+                }  // for sb
+
+                // Reorder of i8mm output with bias and output layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d    = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q5_dmin = vcvt_f32_f16(vld1_f16((const __fp16 *) (q5_ptr[b].dmin + j * 4)));
+                        const float32x4_t dmins   = vmulq_f32(q5_dmin, q8_d);
+
+                        float32x4_t       q5_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q5_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q5_d, q8_d);
+
+                        acc_f32[2 * i + j] = vmlsq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(bias_acc[2 * i + j]), dmins);
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // With the previous reorder, the tile is already in the correct memory layout.
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
 
 void ggml_gemm_q8_0_4x4_q8_0(int                        n,
                              float * GGML_RESTRICT      s,

ggml/src/ggml-cpu/common.h

@@ -6,6 +6,9 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 
+#define GGML_FA_TILE_Q  32
+#define GGML_FA_TILE_KV 16
+
 #ifdef __cplusplus
 
 #include <utility>
@@ -84,4 +87,9 @@ static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_pa
     return {ir0, ir1};
 }
 
+struct ggml_fa_tile_config {
+    static constexpr size_t Q  = GGML_FA_TILE_Q;
+    static constexpr size_t KV = GGML_FA_TILE_KV;
+};
+
 #endif

ggml/src/ggml-cpu/ggml-cpu.c

@@ -14,6 +14,7 @@
 #include "vec.h"
 #include "ops.h"
 #include "ggml.h"
+#include "common.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -2866,10 +2867,12 @@ struct ggml_cplan ggml_graph_plan(
                     } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
-                        const int64_t ne10 = node->src[1]->ne[0]; // DK
-                        const int64_t ne20 = node->src[2]->ne[0]; // DV
+                        const int64_t DK = node->src[1]->ne[0];
+                        const int64_t DV = node->src[2]->ne[0];
 
-                        cur = sizeof(float)*(1*ne10 + 2*ne20)*n_tasks; // 1x head size K + 2x head size V (per thread)
+                        // Tiled flash attention scratch (tile sizes defined in common.h)
+                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
+                        cur = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
                     } break;
                 case GGML_OP_FLASH_ATTN_BACK:
                     {

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -1797,10 +1797,27 @@ class tinyBLAS_Q0_AVX {
       } \
    } \
 
+template<typename T>
+struct mma_instr;
+
+template<>
+struct mma_instr<ggml_bf16_t> {
+    static inline void outer_product(acc_t *acc, vec_t a, vec_t b) {
+        __builtin_mma_xvbf16ger2pp(acc, a, b);
+    }
+};
+
+template<>
+struct mma_instr<ggml_fp16_t> {
+    static inline void outer_product(acc_t *acc, vec_t a, vec_t b) {
+        __builtin_mma_xvf16ger2pp(acc, a, b);
+    }
+};
+
 template <typename TA, typename TB, typename TC>
-class tinyBLAS_BF16_PPC {
+class tinyBLAS_HP16_PPC {
   public:
-    tinyBLAS_BF16_PPC(int64_t k,
+    tinyBLAS_HP16_PPC(int64_t k,
                 const TA *A, int64_t lda,
                 const TB *B, int64_t ldb,
                 TC *C, int64_t ldc,
@@ -2118,8 +2135,8 @@ class tinyBLAS_BF16_PPC {
             packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A);
             packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -2135,8 +2152,8 @@ class tinyBLAS_BF16_PPC {
             packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A);
             packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -2155,10 +2172,10 @@ class tinyBLAS_BF16_PPC {
             packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A);
             packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]);
-                __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_2, vec_A[x+4], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_3, vec_A[x+4], vec_B[x+4]);
             }
         }
 
@@ -2189,7 +2206,7 @@ class tinyBLAS_BF16_PPC {
                 packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A);
                 packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B);
                 for (int x = 0; x<2; x++) {
-                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                    mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
                 }
             }
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
@@ -2224,8 +2241,8 @@ class tinyBLAS_BF16_PPC {
                 packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A);
                 packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B);
                 for (int x = 0; x<4; x++) {
-                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                    __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                    mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                    mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
                 }
             }
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
@@ -3418,16 +3435,19 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             return tb.matmul(m, n);
         }
 #elif defined(__MMA__)
-        if ((k % 8))
-                return false;
-        if(Btype == GGML_TYPE_BF16) {
-           tinyBLAS_BF16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
-            (const ggml_bf16_t *)A, lda,
-            (const ggml_bf16_t *)B, ldb,
-            (float *)C, ldc,
-            params->ith, params->nth};
-        tb.matmul(m, n);
-        return true;
+        if (k % 8) {
+            return false;
+        }
+
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS_HP16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc,
+                params->ith, params->nth };
+
+            tb.matmul(m, n);
+            return true;
         }
 #elif defined(__riscv_zvfbfwma)
         #if LMUL == 1
@@ -3516,6 +3536,21 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
         #endif
             return tb.matmul(m, n);
         }
+#elif defined(__MMA__)
+        if (k % 8) {
+            return false;
+        }
+
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS_HP16_PPC<ggml_fp16_t, ggml_fp16_t, float> tb{ k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc,
+                params->ith, params->nth };
+
+            tb.matmul(m, n);
+            return true;
+        }
 #endif
         return false;
     }

ggml/src/ggml-cpu/ops.cpp

@@ -8164,6 +8164,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         // online softmax / attention
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
+
         for (int64_t ic = 0; ic < nek1; ++ic) {
             const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
@@ -8271,6 +8272,280 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
     }
 }
 
+static void ggml_compute_forward_flash_attn_ext_tiled(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        int ir0, int ir1) {
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    GGML_ASSERT(k->type == v->type);
+    const ggml_type kv_type = k->type;
+
+    const auto * kv_type_traits_cpu = ggml_get_type_traits_cpu(kv_type);
+    const ggml_from_float_t kv_from_float = kv_type_traits_cpu->from_float;
+    const ggml_vec_dot_t    kv_vec_dot    = kv_type_traits_cpu->vec_dot;
+    const size_t kv_type_size = ggml_type_size(kv_type);
+
+    // broadcast factors
+    const int64_t rk2 = neq2/nek2;
+    const int64_t rk3 = neq3/nek3;
+
+    const int64_t rv2 = neq2/nev2;
+    const int64_t rv3 = neq3/nev3;
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) dst->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0) {
+        scale /= logit_softcap;
+    }
+
+    const uint32_t n_head      = neq2;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    int ith = params->ith;
+
+    static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
+    static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;
+
+    GGML_ASSERT(nek1 % KV_TILE_SZ == 0 && "KV sequence length must be divisible by KV_TILE_SZ");
+
+    int ir = ir0;
+    while (ir < ir1) {
+        // q indices for the start of this tile
+        const int iq3 = ir/(neq2*neq1);
+        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
+        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
+
+        // Number of valid rows in this tile:
+        // - limited by tile size (Q_TILE_SZ)
+        // - limited by chunk boundary (ir1 - ir)
+        // - limited by head boundary (neq1 - iq1) to avoid crossing into next head
+        const int tile_rows = MIN(Q_TILE_SZ, MIN((int)(ir1 - ir), (int)(neq1 - iq1)));
+        GGML_ASSERT(tile_rows > 0);
+
+        const uint32_t h = iq2; // head index
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+
+        float S[Q_TILE_SZ];
+        float M[Q_TILE_SZ];
+
+        for (int i = 0 ; i < Q_TILE_SZ; ++i) {
+            S[i] = 0.;
+            M[i] = -INFINITY;
+        }
+
+        // Per-thread scratch layout:
+        // Q_q:    Q_TILE_SZ * DK (converted Q tile in KV type)
+        // KQ:     Q_TILE_SZ * KV_TILE_SZ (attention scores in float)
+        // mask:   Q_TILE_SZ * KV_TILE_SZ (mask in float)
+        // VKQ32:  Q_TILE_SZ * DV (FP32 output accumulator)
+        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile - used for f166 conversion)
+        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + CACHE_LINE_SIZE_F32);
+
+        void  * Q_q    = base;
+        float * KQ     = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float));
+        float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ;
+        float * VKQ32  = mask32 + Q_TILE_SZ * KV_TILE_SZ;
+        float * V32    = VKQ32 + Q_TILE_SZ * DV;  // F32 buffer for V tile
+
+        memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float));
+        memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+
+        // k indices
+        const int ik3 = iq3 / rk3;
+        const int ik2 = iq2 / rk2;
+
+        // v indices
+        const int iv3 = iq3 / rv3;
+        const int iv2 = iq2 / rv2;
+
+        for (int tq = 0; tq < tile_rows; tq++) {
+            const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
+            kv_from_float(pq, (char *)Q_q + tq * DK * kv_type_size, DK);
+        }
+        // Zero-pad remaining rows
+        for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
+            memset((char *)Q_q + tq * DK * kv_type_size, 0, DK * kv_type_size);
+        }
+
+        for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
+
+            // skip the tile entirely if all the masks are -inf
+            if (mask) {
+                bool can_skip = true;
+                for (int tq = 0; tq < tile_rows; tq++) {
+                    const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]);
+                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]);
+                        if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) {
+                            can_skip = false;
+                        }
+                    }
+                }
+
+                if (can_skip) {
+                    continue;
+                }
+            }
+
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                const void * q_row = (const char *)Q_q + tq * DK * kv_type_size;
+                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    const void * k_row = (const char *) k->data + ((ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3);
+                    float s;
+                    kv_vec_dot(DK, &s, 0, k_row, 0, q_row, 0, 1);
+                    KQ[tq * KV_TILE_SZ + tk] = s * scale;
+                }
+            }
+
+            if (logit_softcap != 0.0f) {
+                ggml_vec_tanh_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, KQ);
+                ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, logit_softcap);
+            }
+
+            if (mask) {
+                ggml_vec_add_f32(tile_rows * KV_TILE_SZ, KQ, KQ, mask32);
+            }
+
+            bool skip[Q_TILE_SZ] = {};
+
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                float * kq_row = KQ + tq * KV_TILE_SZ;
+
+                float tile_max;
+                ggml_vec_max_f32(KV_TILE_SZ, &tile_max, kq_row);
+
+                if (tile_max == -INFINITY) {
+                    skip[tq] = true;
+                    continue;
+                }
+
+                const float Mold = M[tq];
+                const float Mnew = fmaxf(Mold, tile_max);
+
+                if (Mnew > Mold) {
+                    const float ms = expf(Mold - Mnew);
+                    ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms);
+                    S[tq] *= ms;
+                }
+                M[tq] = Mnew;
+
+
+                S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew);
+            }
+
+            // Convert V tile to F32 first (if F16), then do MAD
+            // On x86, ggml_vec_mad_f16 internall converts F16<->F32 on every load/store, so pre-converting is faster.
+            // TODO: on ARM, native f16 should be faster
+            if (kv_type == GGML_TYPE_F16) {
+                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    const ggml_fp16_t * v_row = (const ggml_fp16_t *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
+                    ggml_fp16_to_fp32_row(v_row, V32 + tk * DV, DV);
+                }
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    if (skip[tq]) continue;
+                    float * vkq_row = VKQ32 + tq * DV;
+                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                        const float p = KQ[tq * KV_TILE_SZ + tk];
+                        ggml_vec_mad_f32(DV, vkq_row, V32 + tk * DV, p);
+                    }
+                }
+            } else {
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    if (skip[tq]) continue;
+                    float * vkq_row = VKQ32 + tq * DV;
+                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                        const float p = KQ[tq * KV_TILE_SZ + tk];
+                        const float * v_row = (const float *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
+                        ggml_vec_mad_f32(DV, vkq_row, v_row, p);
+                    }
+                }
+            }
+        }
+
+        // sinks (apply only to valid rows in the tile)
+        if (sinks) {
+            const float s = ((float *)((char *) sinks->data))[h];
+
+            for (int tq = 0; tq < tile_rows; tq++) {
+                float ms = 1.0f;
+                float vs = 1.0f;
+
+                if (s > M[tq]) {
+                    ms = expf(M[tq] - s);
+                    ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms);
+                } else {
+                    vs = expf(s - M[tq]);
+                }
+
+                S[tq] = S[tq] * ms + vs;
+            }
+        }
+
+        for (int tq = 0; tq < tile_rows; tq++) {
+            // V /= S
+            const float S_inv = S[tq] == 0.0f ? 0.0f : 1.0f / S[tq];
+            ggml_vec_scale_f32(DV, VKQ32 + tq * DV, S_inv);
+
+            // dst indices
+            const int i1 = iq1 + tq;
+            const int i2 = iq2;
+            const int i3 = iq3;
+
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32 + tq * DV, nb1);
+        }
+
+        ir += tile_rows;
+    }
+}
+
 static void ggml_compute_forward_flash_attn_ext_f16(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8343,6 +8618,15 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     // The number of elements in each chunk
     const int64_t dr = (nr + nchunk - 1) / nchunk;
 
+    static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
+    static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+    const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
+    const bool use_tiled = (q->type == GGML_TYPE_F32 &&
+                            kv_is_f32_or_f16 &&
+                            k->type == v->type &&
+                            nek1 % KV_TILE_SZ == 0 &&
+                            neq1 >= Q_TILE_SZ);  // Only use tiled for batch >= tile size
+
     // The first chunk comes from our thread_id, the rest will get auto-assigned.
     int current_chunk = ith;
 
@@ -8350,7 +8634,11 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         const int64_t ir0 = dr * current_chunk;
         const int64_t ir1 = MIN(ir0 + dr, nr);
 
-        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+        if (use_tiled) {
+            ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+        } else {
+            ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+        }
 
         current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
     }

ggml/src/ggml-cpu/repack.cpp

@@ -474,15 +474,8 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     assert (n % qk == 0);
     assert (nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
     UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
     float sumf[8];
     float sum_minf[8];
@@ -616,6 +609,100 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemv_q5_K_8x8_q8_K_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int             qk                = QK_K;
+    const int             nb                = n / qk;
+    const int             ncols_interleaved = 8;
+    const int             blocklen          = 8;
+    static const uint32_t kmask1            = 0x3f3f3f3f;
+    static const uint32_t kmask2            = 0x0f0f0f0f;
+    static const uint32_t kmask3            = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float    sumf[8];
+    float    sum_minf[8];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j]     = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3]      = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2]      = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
+
+                const int qh_shift = (k / 4) * 2;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi  = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                        const int qh_idx      = (k * 8 + i) % 32;
+                        const int qh_chunk    = qh_idx / 8;
+                        const int qh_pos      = qh_idx % 8;
+                        const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
+
+                        const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                        const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                        const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                        const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                        const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                        const int q8_offset = (k >> 2) * 64 + (k % 4) * blocklen + i;
+
+                        sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                        sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
+                                   GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1212,6 +1299,108 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemm_q5_K_8x8_q8_K_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK_K;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen          = 8;
+
+    constexpr uint32_t kmask1 = 0x3f3f3f3f;
+    constexpr uint32_t kmask2 = 0x0f0f0f0f;
+    constexpr uint32_t kmask3 = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float    sumf[4][8];
+    float    sum_minf[4][8];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j]     = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2]      = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
+                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
+
+                    const int qh_shift = (k / 4) * 2;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi  = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                                const int qh_idx      = (k * 8 + i) % 32;
+                                const int qh_chunk    = qh_idx / 8;
+                                const int qh_pos      = qh_idx % 8;
+                                const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
+
+                                const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                                const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                                const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                                const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                                const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                                const int q8_offset = (k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i;
+
+                                sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                                sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for (int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
+                                              GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
 
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -1622,7 +1811,95 @@ static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_in
         out.scales[i] = in[src1].scales[src2];
     }
     return out;
+}
 
+static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_interleave) {
+    block_q5_Kx8 out;
+    //Delta(scale) and dmin values of the eight Q5_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 8; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 4 / blck_size_interleave;
+
+    // Interleave Q5_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end; ++i) {
+        int src_id     = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    // Repeat for low bits 8 bytes at a time as well, since
+    // the high bits are interleaved in Q5_K and the index is
+    // qh_idx = (qs_idx % 32);
+    // qh_val = qh[qh_idx] >> (qs_idx / 32);
+    for (int i = 0; i < end / 4; ++i) {
+        int src_id     = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qh[src_offset], sizeof(uint64_t));
+        memcpy(&out.qh[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    // The below logic is copied over from Q4_K
+    // The point is to unpack all the scales and mins for each sub block every time we load 12 bytes.
+    // Currently the Q5_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value)
+    // The output Q5_Kx8 structure has 96 bytes
+    // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q5_K structure
+    // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q5_K structures
+    uint8_t s[8], m[8];
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = in[j].scales[i] & 63;
+            m[j] = in[j].scales[i + 4] & 63;
+        }
+
+        out.scales[i * 12]      = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 1]  = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 2]  = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 3]  = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 4]  = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 5]  = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 6]  = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 7]  = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 8]  = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 9]  = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4);
+    }
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i + 8] & 15);
+            m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i + 8] & 240) >> 4);
+        }
+
+        out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4);
+    }
+
+    return out;
 }
 
 static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
@@ -1718,6 +1995,38 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor *       t,
+                                    int                        interleave_block,
+                                    const void * GGML_RESTRICT data,
+                                    size_t                     data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q5_K);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q5_Kx8 *     dst = (block_q5_Kx8 *) t->data;
+    const block_q5_K * src = (const block_q5_K *) data;
+    block_q5_K         dst_tmp[8];
+    int                nrow    = ggml_nrows(t);
+    int                nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q5_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q5_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
 static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
     GGML_ASSERT(interleave_block == 8);
@@ -1936,6 +2245,10 @@ template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
+}
+
 template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
 }
@@ -1973,6 +2286,10 @@ template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1981,8 +2298,8 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
@@ -2013,20 +2330,24 @@ template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
-}
-
 template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
@@ -2432,6 +2753,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
+    // instance for Q5_K
+    static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
+
     // instance for Q2
     static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K;
 
@@ -2482,6 +2806,12 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q2_K_8x8_q8_K;
             }
         }
+    } else if (cur->type == GGML_TYPE_Q5_K) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q5_K_8x8_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {

ggml/src/ggml-cpu/repack.h

@@ -44,6 +44,7 @@ struct block_q4_Kx8 {
 };
 
 static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
 struct block_q2_Kx8 {
     ggml_half d[8];      // super-block scale for quantized scales
     ggml_half dmin[8];   // super-block scale for quantized mins
@@ -52,6 +53,18 @@ struct block_q2_Kx8 {
 };
 
 static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
+
+struct block_q5_Kx8 {
+    ggml_half d[8];              // super-block scale for quantized scales
+    ggml_half dmin[8];           // super-block scale for quantized mins
+    uint8_t   scales[96];        // scales and mins, quantized with 6 bits
+    uint8_t   qh[QK_K * 8 / 8];  // high bits of 5-bit quants
+    uint8_t   qs[QK_K * 8 / 2];  // low bits of 5-bit quants (in groups of 4)
+};
+
+static_assert(sizeof(block_q5_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 5,
+              "wrong q5_K block size/padding");
+
 struct block_q8_Kx4 {
     float d[4];              // delta
     int8_t qs[QK_K * 4];     // quants
@@ -82,20 +95,22 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -111,17 +126,19 @@ void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GG
 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);

ggml/src/ggml-cuda/common.cuh

@@ -1327,10 +1327,44 @@ struct ggml_backend_cuda_context {
     cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
     cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
-    std::unique_ptr<ggml_cuda_graph> cuda_graph;
-
     int curr_stream_no = 0;
 
+#ifdef USE_CUDA_GRAPH
+    // Map from first_node_ptr to cuda_graph - allows multiple graphs per context
+    // when the computation is split across CPU/GPU (e.g., with --n-cpu-moe)
+    std::unordered_map<const void *, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
+
+    ggml_cuda_graph * cuda_graph(const void * first_node_ptr) {
+        auto it = cuda_graphs.find(first_node_ptr);
+        if (it == cuda_graphs.end()) {
+            cuda_graphs[first_node_ptr] = std::make_unique<ggml_cuda_graph>();
+            return cuda_graphs[first_node_ptr].get();
+        }
+        return it->second.get();
+    }
+
+    // Check if any CUDA graph is enabled for this context (used by kernels that need to know
+    // if graphs are in use without having access to the specific graph key)
+    bool any_cuda_graph_enabled() const {
+        for (const auto & [key, graph] : cuda_graphs) {
+            if (graph && graph->is_enabled()) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    // Check if any CUDA graph has an instance for this context
+    bool any_cuda_graph_has_instance() const {
+        for (const auto & [key, graph] : cuda_graphs) {
+            if (graph && graph->instance != nullptr) {
+                return true;
+            }
+        }
+        return false;
+    }
+#endif // USE_CUDA_GRAPH
+
     explicit ggml_backend_cuda_context(int device) :
         device(device),
         name(GGML_CUDA_NAME + std::to_string(device)) {

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

@@ -629,8 +629,8 @@ static __global__ void flash_attn_mask_to_KV_max(
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
-        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11,
-        const int nbatch_fa) {
+        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03,
+        const int ne11, const int ne12, const int nbatch_fa) {
     constexpr int ncols = ncols1*ncols2;
 
     const int bidx0 = blockIdx.x;
@@ -641,11 +641,14 @@ static __global__ void flash_attn_stream_k_fixup(
 
     const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
 
-    const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
-    const int iter_j = (ne01 + (ncols1    - 1)) / ncols1;
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
 
-    const int kbc0      = int64_t(bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
-    const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int iter_k     = (ne11      + (nbatch_fa - 1)) / nbatch_fa;
+    const int iter_j     = (ne01      + (ncols1    - 1)) / ncols1;
+    const int iter_z_gqa = (gqa_ratio + (ncols2    - 1)) / ncols2;
+
+    const int kbc0      = int64_t(bidx0 + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
+    const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
 
     const bool did_not_have_any_data   = kbc0 == kbc0_stop;
     const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -654,15 +657,19 @@ static __global__ void flash_attn_stream_k_fixup(
         return;
     }
 
-    const int sequence = kbc0 / (iter_k*iter_j*(ne02/ncols2));
-    const int head = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
-    const int jt = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
+    // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index
+    const int sequence =  kbc0 /(iter_k*iter_j*iter_z_gqa*ne12);
+    const int z_KV     = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
+    const int zt_gqa   = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
+    const int jt       = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
 
-    if (jt*ncols1 + j >= ne01) {
+    const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
+
+    if (jt*ncols1 + j >= ne01 || zt_gqa*ncols2 + c >= gqa_ratio) {
         return;
     }
 
-    dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + head*(ncols2*D) + (j*ne02 + c)*D + tid;
+    dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + zt_Q*D + (j*ne02 + c)*D + tid;
 
     // Load the partial result that needs a fixup:
     float dst_val = 0.0f;
@@ -681,7 +688,7 @@ static __global__ void flash_attn_stream_k_fixup(
     int bidx = bidx0 - 1;
     int kbc_stop = kbc0;
     while(true) {
-        const int kbc = int64_t(bidx)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+        const int kbc = int64_t(bidx)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
         if (kbc == kbc_stop) { // Did not have any data.
             bidx--;
             kbc_stop = kbc;
@@ -778,13 +785,11 @@ void launch_fattn(
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
-    const bool is_mla = DV == 512; // TODO better parameterization
-
     const ggml_tensor * Q = dst->src[0];
     const ggml_tensor * K = dst->src[1];
     const ggml_tensor * V = dst->src[2];
 
-    GGML_ASSERT(V || is_mla);
+    const bool V_is_K_view = V->view_src && V->view_offs == 0 && (V->view_src == K || V->view_src == K->view_src);
 
     const ggml_tensor * mask  = dst->src[3];
     const ggml_tensor * sinks = dst->src[4];
@@ -794,9 +799,9 @@ void launch_fattn(
     GGML_ASSERT(Q->type == GGML_TYPE_F32);
     GGML_ASSERT(KQV->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(      Q->nb[0] == ggml_element_size(Q));
-    GGML_ASSERT(      K->nb[0] == ggml_element_size(K));
-    GGML_ASSERT(!V || V->nb[0] == ggml_element_size(V));
+    GGML_ASSERT(Q->nb[0] == ggml_element_size(Q));
+    GGML_ASSERT(K->nb[0] == ggml_element_size(K));
+    GGML_ASSERT(V->nb[0] == ggml_element_size(V));
 
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
 
@@ -817,10 +822,10 @@ void launch_fattn(
     size_t nb12 = K->nb[2];
     size_t nb13 = K->nb[3];
 
-    const char * V_data = V ? (const char *) V->data : nullptr;
-    size_t nb21 = V ? V->nb[1] : nb11;
-    size_t nb22 = V ? V->nb[2] : nb12;
-    size_t nb23 = V ? V->nb[3] : nb13;
+    const char * V_data = (const char *) V->data;
+    size_t nb21 = V->nb[1];
+    size_t nb22 = V->nb[2];
+    size_t nb23 = V->nb[3];
 
     if (need_f16_K && K->type != GGML_TYPE_F16) {
         const size_t bs = ggml_blck_size(K->type);
@@ -849,36 +854,45 @@ void launch_fattn(
         K_data = (char *) K_f16.ptr;
     }
 
-    if (V && need_f16_V && V->type != GGML_TYPE_F16) {
-        const size_t bs = ggml_blck_size(V->type);
-        const size_t ts = ggml_type_size(V->type);
-
-        V_f16.alloc(ggml_nelements(V));
-        if (ggml_is_contiguously_allocated(V)) {
-            to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
-            to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
-            V_data = (char *) V_f16.ptr;
-
-            nb21 = nb21*bs*sizeof(half)/ts;
-            nb22 = nb22*bs*sizeof(half)/ts;
-            nb23 = nb23*bs*sizeof(half)/ts;
+    if (need_f16_V && V->type != GGML_TYPE_F16) {
+        if (V_is_K_view) {
+            V_data = K_data;
+            nb21   = nb11;
+            nb22   = nb12;
+            nb23   = nb13;
         } else {
-            GGML_ASSERT(V->nb[0] == ts);
-            to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(V->type);
-            const int64_t s01 = nb21 / ts;
-            const int64_t s02 = nb22 / ts;
-            const int64_t s03 = nb23 / ts;
-            to_fp16(V_data, V_f16.ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3], s01, s02, s03, main_stream);
+            const size_t bs = ggml_blck_size(V->type);
+            const size_t ts = ggml_type_size(V->type);
 
-            nb21 = V->ne[0] * sizeof(half);
-            nb22 = V->ne[1] * nb21;
-            nb23 = V->ne[2] * nb22;
+            V_f16.alloc(ggml_nelements(V));
+            if (ggml_is_contiguously_allocated(V)) {
+                to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
+                to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
+                V_data = (char *) V_f16.ptr;
+
+                nb21 = nb21*bs*sizeof(half)/ts;
+                nb22 = nb22*bs*sizeof(half)/ts;
+                nb23 = nb23*bs*sizeof(half)/ts;
+            } else {
+                GGML_ASSERT(V->nb[0] == ts);
+                to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(V->type);
+                const int64_t s01 = nb21 / ts;
+                const int64_t s02 = nb22 / ts;
+                const int64_t s03 = nb23 / ts;
+                to_fp16(V_data, V_f16.ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3], s01, s02, s03, main_stream);
+
+                nb21 = V->ne[0] * sizeof(half);
+                nb22 = V->ne[1] * nb21;
+                nb23 = V->ne[2] * nb22;
+            }
+            V_data = (char *) V_f16.ptr;
         }
-        V_data = (char *) V_f16.ptr;
     }
 
-    const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
-    const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
+    const int ntiles_x     = ((Q->ne[1] + ncols1 - 1) / ncols1);
+    const int gqa_ratio    = Q->ne[2] / K->ne[2];
+    const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2);
+    const int ntiles_total = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];
 
     // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
     // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
@@ -953,7 +967,7 @@ void launch_fattn(
 
         blocks_num.x = ntiles_x;
         blocks_num.y = parallel_blocks;
-        blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
+        blocks_num.z = ntiles_z_gqa*K->ne[2]*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
@@ -1007,7 +1021,7 @@ void launch_fattn(
 
             flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], nbatch_fa);
+                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], K->ne[2], nbatch_fa);
         }
     } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(DV, 1, 1);

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

@@ -400,7 +400,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps,
-    bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter, bool oob_check,
+    bool use_logit_softcap, bool V_is_K_view, bool needs_fixup, bool is_fixup, bool last_iter, bool oob_check,
     typename T_A_KQ, typename T_B_KQ, typename T_C_KQ, typename T_A_VKQ, typename T_B_VKQ, typename T_C_VKQ>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const float2 * const __restrict__ Q_f2,
@@ -442,8 +442,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int stride_tile_Q = DKQ/2     + 4;
     constexpr int stride_tile_K = nbatch_K2 + 4;
 
-    static_assert(!mla || nbatch_K2 >= nbatch_V2, "bad nbatch_K2, nbatch_V2 for MLA");
-    constexpr int stride_tile_V = mla ? stride_tile_K : nbatch_V2 + 4;
+    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
 
     const int k_VKQ_0 = kb0 * nbatch_fa;
 #if defined(TURING_MMA_AVAILABLE)
@@ -456,7 +455,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 
     if constexpr (nstages > 1) {
         static_assert(!oob_check, "OOB check incompatible with multi-stage pipeline");
-        static_assert(!mla, "multi-stage loading not implemented for MLA");
+        static_assert(!V_is_K_view, "K data reuse not implemented multi-stage loading");
         static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi stage loading");
         constexpr bool use_cp_async = true;
         cp_async_wait_all();
@@ -471,8 +470,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
     }
 
+    // For MLA K and V have the same data.
+    // Therefore, iterate over K in reverse and later re-use the data if possible.
 #pragma unroll
-    for (int k0_start = 0; k0_start < DKQ/2; k0_start += nbatch_K2) {
+    for (int k0_start = (DKQ/2-1) - (DKQ/2-1) % nbatch_K2; k0_start >= 0; k0_start -= nbatch_K2) {
         const int k0_stop = k0_start + nbatch_K2 < DKQ/2 ? k0_start + nbatch_K2 : DKQ/2;
         const int k0_diff = k0_stop - k0_start;
 
@@ -776,6 +777,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
     if constexpr (nstages > 1) {
+        static_assert(!V_is_K_view, "K data reuse not implemented multi-stage loading");
         // Preload K tile for next iteration:
         constexpr bool use_cp_async = true;
         cp_async_wait_all();
@@ -791,10 +793,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 
-    // For MLA K and V have the same data.
-    // Therefore, iterate over V in reverse and re-use the data if possible.
-    static_assert(!mla || nstages <= 1, "combination of MLA and multi-stage loading not implemented");
-    constexpr int reusable_cutoff = mla ? (DKQ - 1) - (DKQ - 1) % (2*nbatch_K2) - (DKQ - DV) : DV;
 #if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
     T_A_VKQ A_identity;
     make_identity_mat(A_identity);
@@ -802,12 +800,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
-    for (int i0_stop = DV; i0_stop > 0; i0_stop -= 2*nbatch_V2) {
-        const int i0_start = i0_stop - 2*nbatch_V2 > 0 ? i0_stop - 2*nbatch_V2 : 0;
-        const int i0_diff  = i0_stop - i0_start;
+    for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
+        static_assert(DV % (2*nbatch_V2) == 0, "bad loop size");
+        const int i0_stop = i0_start + 2*nbatch_V2;
+        const int i0_diff = i0_stop - i0_start;
 
         if constexpr (nstages <= 1) {
-            if (i0_start < reusable_cutoff) {
+            if (!V_is_K_view || i0_stop > 2*nbatch_K2) {
                 constexpr bool use_cp_async = nstages == 1;
                 flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check>
                     (V_h2 + int64_t(k_VKQ_0)*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V, k_VKQ_sup);
@@ -817,7 +816,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 __syncthreads();
             }
         }
-        const half2 * tile_V_i = i0_start < reusable_cutoff ? tile_V : tile_V + (i0_start - reusable_cutoff)/2;
+        const half2 * tile_V_i = !V_is_K_view || i0_stop > 2*nbatch_K2 ? tile_V : tile_V + i0_start/2;
 
 #if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         constexpr int i0_stride = cols_per_warp == 8 ? T_C_VKQ::I : 2*T_C_VKQ::J;
@@ -920,7 +919,7 @@ template<int ncols> struct mma_tile_sizes {
 };
 #endif // defined(TURING_MMA_AVAILABLE)
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, bool use_logit_softcap, bool V_is_K_view, bool needs_fixup, bool is_fixup>
 static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -934,6 +933,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const float logit_softcap,
         const uint3 ne01,
         const int ne02,
+        const int gqa_ratio,
         const int ne11,
         const int stride_Q1,
         const int stride_Q2,
@@ -941,6 +941,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int stride_V,
         const int stride_mask,
         const int jt,
+        const int zt_gqa,
         const int kb0_start,
         const int kb0_stop) {
 #if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
@@ -974,8 +975,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     constexpr int stride_tile_Q = DKQ/2     + 4;
     constexpr int stride_tile_K = nbatch_K2 + 4;
 
-    static_assert(!mla || nbatch_K2 >= nbatch_V2, "bad nbatch_K2, nbatch_V2 for MLA");
-    constexpr int stride_tile_V = mla ? stride_tile_K : nbatch_V2 + 4;
+    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
     constexpr int stride_tile_KV_max = stride_tile_K > stride_tile_V ? stride_tile_K : stride_tile_V;
 
     extern __shared__ half2 tile_Q[];
@@ -1024,7 +1024,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             const int j = jc / ncols2;
             const int c = jc % ncols2;
 
-            if (jt*ncols1 + j < int(ne01.z)) {
+            if ((ncols1 == 1 || jt*ncols1 + j < int(ne01.z)) && (ncols2 == 1 || zt_gqa*ncols2 + c < gqa_ratio)) {
 #pragma unroll
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
@@ -1079,7 +1079,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             constexpr bool last_iter = false;
             constexpr int  k_VKQ_sup = nbatch_fa;
             flash_attn_ext_f16_iter
-                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
                  T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
                 (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
                  ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1088,7 +1088,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         constexpr bool last_iter = true;
         const     int  k_VKQ_sup = ne11 - kb0*nbatch_fa;
         flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
               T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
             (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
              ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1099,7 +1099,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             constexpr bool last_iter = false;
             constexpr int  k_VKQ_sup = nbatch_fa;
             flash_attn_ext_f16_iter
-                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
                  T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
                 (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
                  ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1108,7 +1108,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         constexpr bool last_iter = true;
         constexpr int  k_VKQ_sup = nbatch_fa;
         flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
              T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
             (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
              ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1410,7 +1410,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                     const int j_dst = jc_dst / ncols2;
                     const int c_dst = jc_dst % ncols2;
 
-                    if (!is_fixup && jt*ncols1 + j_dst >= int(ne01.z)) {
+                    if (!is_fixup && ((ncols1 > 1 && jt*ncols1 + j_dst >= int(ne01.z)) || (ncols2 > 1 && zt_gqa*ncols2 + c_dst >= gqa_ratio))) {
                         continue;
                     }
 
@@ -1449,14 +1449,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     }
 #else
     GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dstk_fixup,
-        scale, slope, logit_softcap, ne01, ne02,
+        scale, slope, logit_softcap, ne01, ne02, gqa_ratio,
         stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
         jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
 #endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool mla>
+template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool V_is_K_view>
 __launch_bounds__(ggml_cuda_fattn_mma_get_nthreads(DKQ, DV, ncols1*ncols2), ggml_cuda_fattn_mma_get_occupancy(DKQ, DV, ncols1*ncols2))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
@@ -1508,8 +1508,6 @@ static __global__ void flash_attn_ext_f16(
     }
 #endif // defined(AMD_WMMA_AVAILABLE)
 
-    static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
-
     constexpr int ncols     = ncols1 * ncols2;
     constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int nthreads  = ggml_cuda_fattn_mma_get_nthreads(DKQ, DV, ncols);
@@ -1522,14 +1520,15 @@ static __global__ void flash_attn_ext_f16(
     const int stride_K    = nb11 / sizeof(half2);
     const int stride_mask = nb31 / sizeof(half);
 
-    const int stride_V = mla ? stride_K : nb21 / sizeof(half2);
+    const int stride_V = V_is_K_view ? stride_K : nb21 / sizeof(half2);
 
-    const int iter_k = (ne11   + (nbatch_fa - 1)) / nbatch_fa;
-    const int iter_j = (ne01.z + (ncols1    - 1)) / ncols1;
+    const int iter_k     = (ne11      + (nbatch_fa - 1)) / nbatch_fa;
+    const int iter_j     = (ne01.z    + (ncols1    - 1)) / ncols1;
+    const int iter_z_gqa = (gqa_ratio + (ncols2    - 1)) / ncols2;
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int       kbc      = int64_t(blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
-    const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    int       kbc      = int64_t(blockIdx.x + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
+    const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
 
     // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined.
     // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup).
@@ -1540,22 +1539,24 @@ static __global__ void flash_attn_ext_f16(
     int kb0_stop  = min(iter_k, kb0_start + kbc_stop - kbc);
 
     while (kbc < kbc_stop && kb0_stop == iter_k) {
-        const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
-        const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
-        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+        // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index
+        const int sequence =  kbc /(iter_k*iter_j*iter_z_gqa*ne12);
+        const int z_KV     = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
+        const int zt_gqa   = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
+        const int jt       = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
 
-        const int head0 = zt * ncols2;
+        const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
 
-        const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02* head0);
-        const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
+        const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02*zt_Q);
+        const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*z_KV);
         const half   * mask_h = ncols2 == 1 && !mask ? nullptr :
             (const half *) (mask + nb33*(sequence % ne33));
-        float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2);
+        float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + zt_Q) * (DV/2);
 
-        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
-        const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
+        const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
+        const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
 
-        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
+        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;
 
         if (KV_max) {
             kb0_stop = min(kb0_stop, KV_max[sequence*iter_j + jt] / nbatch_fa);
@@ -1563,14 +1564,14 @@ static __global__ void flash_attn_ext_f16(
         constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         if (kb0_start == 0) {
             constexpr bool needs_fixup = false; // CUDA block is working on an entire tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup>
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
                 (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
-                 ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop);
+                 ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
         } else {
             constexpr bool needs_fixup = true; // CUDA block is missing the beginning of a tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup>
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
                 (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
-                 ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop);
+                 ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
         }
 
         kbc += iter_k;
@@ -1584,22 +1585,24 @@ static __global__ void flash_attn_ext_f16(
         return;
     }
 
-    const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
-    const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
-    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+    // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index.
+    const int sequence =  kbc /(iter_k*iter_j*iter_z_gqa*ne12);
+    const int z_KV     = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
+    const int zt_gqa   = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
+    const int jt       = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
 
-    const int head0 = zt * ncols2;
+    const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
 
-    const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02* head0);
-    const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
+    const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02*zt_Q);
+    const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*z_KV);
     const half   * mask_h = ncols2 == 1 && !mask ? nullptr :
         (const half *) (mask + nb33*(sequence % ne33));
-    float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2);
+    float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + zt_Q) * (DV/2);
 
-    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
-    const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
+    const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
+    const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
 
-    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
+    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;
 
     if (KV_max) {
         kb0_stop = min(kb0_stop, KV_max[sequence*iter_j + jt] / nbatch_fa);
@@ -1607,9 +1610,9 @@ static __global__ void flash_attn_ext_f16(
 
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
-    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup>
+    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
         (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
-         ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop);
+         ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
 #else
     GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
         max_bias, m0, m1, n_head_log2, logit_softcap,
@@ -1643,7 +1646,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     const int cols_per_warp = std::min(ncols, get_cols_per_warp(cc));
     const int nwarps        = nthreads / WARP_SIZE;
 
-    constexpr bool mla = DKQ == 576;
+    constexpr bool V_is_K_view = DKQ == 576; // Guaranteed by the kernel selection logic in fattn.cu
 
     const size_t nbytes_shared_KV_1stage = nbatch_fa            * std::max(nbatch_K2 + 4,  nbatch_V2 + 4) * sizeof(half2);
     const size_t nbytes_shared_KV_2stage = nbatch_fa            *         (nbatch_K2 + 4 + nbatch_V2 + 4) * sizeof(half2);
@@ -1668,7 +1671,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     fattn_kernel_t fattn_kernel;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view>;
 
 #if !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
@@ -1679,7 +1682,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
 #endif // !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view>;
 
 #if !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
@@ -1743,3 +1746,5 @@ extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
 extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  1, 32);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  2, 32);

ggml/src/ggml-cuda/fattn.cu

@@ -18,9 +18,11 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
         }
     }
 
-    if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
-        return;
+    if constexpr (ncols2 <= 16) {
+        if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) {
+            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
+            return;
+        }
     }
 
     if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || amd_wmma_available(cc) || Q->ne[1] <= 32/ncols2) {
@@ -33,6 +35,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
 
 template <int DKQ, int DV>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
@@ -46,7 +49,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     //     are put into the template specialization without GQA optimizations.
     bool use_gqa_opt = mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
     for (const ggml_tensor * t : {Q, K, V, mask}) {
-        if (t == nullptr) {
+        if (t == nullptr || ggml_is_quantized(t->type)) {
             continue;
         }
         for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
@@ -60,17 +63,38 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
 
-    if (use_gqa_opt && gqa_ratio % 8 == 0) {
+    // On Volta the GQA optimizations aren't as impactful vs. minimizing wasted compute:
+    if (cc == GGML_CUDA_CC_VOLTA) {
+        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;
+        }
+
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(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;
     }
 
-    if (use_gqa_opt && gqa_ratio % 4 == 0) {
+    if (use_gqa_opt && gqa_ratio > 2) {
         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio % 2 == 0) {
+    if (use_gqa_opt && gqa_ratio > 1) {
         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
         return;
     }
@@ -79,6 +103,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
 }
 
 static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
@@ -121,8 +146,30 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
 
             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
             const int gqa_ratio = Q->ne[2] / K->ne[2];
-            GGML_ASSERT(gqa_ratio % 4 == 0);
-            if (gqa_ratio % 16 == 0) {
+            if (gqa_ratio == 20) { // GLM 4.7 Flash
+                if (cc >= GGML_CUDA_CC_BLACKWELL) {
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    if (Q->ne[1] <= 4) {
+                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+                        break;
+                    }
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                if (cc >= GGML_CUDA_CC_TURING) {
+                    if (Q->ne[1] <= 4) {
+                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 32>(ctx, dst);
+                        break;
+                    }
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                // Volta:
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+            } else if (gqa_ratio % 16 == 0) {
                 ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
             } else {
                 ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
@@ -234,9 +281,9 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 
     // The effective batch size for the kernel can be increased by gqa_ratio.
     // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
-    bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
     for (const ggml_tensor * t : {Q, K, V, mask}) {
-        if (t == nullptr) {
+        if (t == nullptr || ggml_is_quantized(t->type)) {
             continue;
         }
         for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
@@ -247,6 +294,8 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         }
     }
 
+    const bool V_is_K_view = V->view_src && V->view_offs == 0 && (V->view_src == K || V->view_src == K->view_src);
+
     const int cc = ggml_cuda_info().devices[device].cc;
 
     switch (K->ne[0]) {
@@ -266,7 +315,10 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
+            if (!gqa_opt_applies) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            if (!V_is_K_view) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2969,18 +2969,25 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
     return true;
 }
 
+static const void * ggml_cuda_graph_get_key(ggml_cgraph * cgraph) {
+    return cgraph->nodes[0];
+}
+
 static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
 
     bool res = false;
 
-    if (cuda_ctx->cuda_graph->instance == nullptr) {
+    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+
+    if (graph->instance == nullptr) {
         res = true;
     }
 
     // Check if the graph size has changed
-    if (cuda_ctx->cuda_graph->props.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
+    if (graph->props.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
         res = true;
-        cuda_ctx->cuda_graph->props.resize(cgraph->n_nodes + cgraph->n_leafs);
+        graph->props.resize(cgraph->n_nodes + cgraph->n_leafs);
     }
 
     // Loop over nodes in GGML graph to determine if CUDA graph update is required
@@ -2988,37 +2995,38 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
     for (int i = 0; i < cgraph->n_nodes; i++) {
         bool props_match = true;
         if (!res) {
-            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &cuda_ctx->cuda_graph->props[i]);
+            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &graph->props[i]);
         }
         if (!props_match) {
             res = true;
         }
-        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[i], cgraph->nodes[i]);
+        ggml_cuda_graph_node_set_properties(&graph->props[i], cgraph->nodes[i]);
     }
 
     for (int i = 0; i < cgraph->n_leafs; i++) {
-        bool props_match= true;
+        bool props_match = true;
         if (!res) {
-            props_match = ggml_cuda_graph_node_properties_match(cgraph->leafs[i], &cuda_ctx->cuda_graph->props[cgraph->n_nodes + i]);
+            props_match = ggml_cuda_graph_node_properties_match(cgraph->leafs[i], &graph->props[cgraph->n_nodes + i]);
         }
         if (!props_match) {
             res = true;
         }
-        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[cgraph->n_nodes + i], cgraph->leafs[i]);
+        ggml_cuda_graph_node_set_properties(&graph->props[cgraph->n_nodes + i], cgraph->leafs[i]);
     }
 
     return res;
 }
 
-static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx) {
+static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
 #if CUDART_VERSION >= 12000
     cudaGraphExecUpdateResultInfo result_info;
-    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &result_info);
 #else
     cudaGraphNode_t errorNode;
     cudaGraphExecUpdateResult result_info;
-    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &errorNode, &result_info);
 #endif // CUDART_VERSION >= 12000
 
     if (stat == cudaErrorGraphExecUpdateFailure) {
@@ -3029,14 +3037,14 @@ static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_c
         // The pre-existing graph exec cannot be updated due to violated constraints
         // so instead clear error and re-instantiate
         (void)cudaGetLastError();
-        CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
-        cuda_ctx->cuda_graph->instance = nullptr;
-        CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        CUDA_CHECK(cudaGraphExecDestroy(graph->instance));
+        graph->instance = nullptr;
+        CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
     } else {
         GGML_ASSERT(stat == cudaSuccess);
     }
 }
-#endif
+#endif // USE_CUDA_GRAPH
 
 static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
                                                 const ggml_tensor * view,
@@ -3241,7 +3249,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     return false;
 }
 
-static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required) {
+static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required, const void * graph_key) {
     bool graph_evaluated_or_captured = false;
 
     // flag used to determine whether it is an integrated_gpu
@@ -3695,13 +3703,14 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
         }
 
 #ifdef USE_CUDA_GRAPH
+        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
         if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
-            if (cuda_ctx->cuda_graph->graph != nullptr) {
-                CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
-                cuda_ctx->cuda_graph->graph = nullptr;
+            if (graph->graph != nullptr) {
+                CUDA_CHECK(cudaGraphDestroy(graph->graph));
+                graph->graph = nullptr;
             }
 
-            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
+            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &graph->graph));
             graph_evaluated_or_captured = true; // CUDA graph has been captured
 
             std::lock_guard<std::mutex> lock(ggml_cuda_lock);
@@ -3714,40 +3723,39 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
     }
 
     if (use_cuda_graph) {
-        if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
-            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+        if (graph->instance == nullptr) { // Create executable graph from captured graph.
+            CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
         }
         if (cuda_graph_update_required) { // Update graph executable
-            ggml_cuda_graph_update_executable(cuda_ctx);
+            ggml_cuda_graph_update_executable(cuda_ctx, graph_key);
         }
         // Launch graph
-        CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
+        CUDA_CHECK(cudaGraphLaunch(graph->instance, cuda_ctx->stream()));
 #else
         graph_evaluated_or_captured = true;
 #endif  // USE_CUDA_GRAPH
     }
 }
 
-static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx) {
+static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
 
 #ifdef USE_CUDA_GRAPH
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
-    if (cuda_ctx->cuda_graph == nullptr) {
-        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
-    }
-
-    if (cuda_ctx->cuda_graph->graph == nullptr) {
+    if (graph->graph == nullptr) {
         if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
-            if (!cuda_ctx->cuda_graph->disable_due_to_gpu_arch) {
+            if (!graph->disable_due_to_gpu_arch) {
                 GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
             }
-            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
+            graph->disable_due_to_gpu_arch = true;
         }
     }
 
-    return cuda_ctx->cuda_graph->is_enabled();
+    return graph->is_enabled();
 #else
     GGML_UNUSED(cuda_ctx);
+    GGML_UNUSED(graph_key);
     return false;
 #endif // USE_CUDA_GRAPH
 }
@@ -3759,15 +3767,19 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
 
     bool use_cuda_graph             = false;
     bool cuda_graph_update_required = false;
+    const void * graph_key = nullptr;
 
 #ifdef USE_CUDA_GRAPH
-    use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx);
+    graph_key = ggml_cuda_graph_get_key(cgraph);
 
-    if (cuda_ctx->cuda_graph->is_enabled()) {
+    use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+    if (graph->is_enabled()) {
         cuda_graph_update_required = ggml_cuda_graph_update_required(cuda_ctx, cgraph);
         use_cuda_graph             = ggml_cuda_graph_check_compability(cgraph);
 
-        cuda_ctx->cuda_graph->record_update(use_cuda_graph, cuda_graph_update_required);
+        graph->record_update(use_cuda_graph, cuda_graph_update_required);
     }
 #endif // USE_CUDA_GRAPH
 
@@ -3781,7 +3793,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required);
+    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required, graph_key);
 
     return GGML_STATUS_SUCCESS;
 }
@@ -3814,7 +3826,14 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
 static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
 
-    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx);
+#ifdef USE_CUDA_GRAPH
+    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
+    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+#else
+    const bool use_cuda_graph = false;
+    GGML_UNUSED(cuda_ctx);
+    GGML_UNUSED(cgraph);
+#endif
 
     static bool enable_graph_optimization = [] {
         const char * env     = getenv("GGML_CUDA_GRAPH_OPT");
@@ -4857,6 +4876,16 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
         static std::mutex mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
+            // Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance
+            // PR: https://github.com/ggml-org/llama.cpp/pull/19042
+            if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) {
+#ifdef _WIN32
+                _putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x");
+#else
+                setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set
+#endif // _WIN32
+            }
+
             ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
             const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 

ggml/src/ggml-cuda/mean.cu

@@ -31,14 +31,15 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 #endif // USE_CUDA_GRAPH
     if ((nrows == 1) &&
 #ifdef USE_CUDA_GRAPH
-            // CUDA_GRAPHS_DISABLED
-            ((ncols > 65536) &&
-             ((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
-              ctx.cuda_graph->is_enabled())) ||
-        // CUDA_GRAPHS ENABLED
-        ((ncols > 32768) &&
-         !((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
-            ctx.cuda_graph->is_enabled()))) {
+            // Determine if CUDA graphs are effectively disabled for this context
+            // (no graph instance exists and we're not capturing, OR graphs are explicitly enabled)
+            (((ncols > 65536) &&
+              (((!ctx.any_cuda_graph_has_instance()) && (iscapturing == cudaStreamCaptureStatusNone)) ||
+               ctx.any_cuda_graph_enabled())) ||
+            // CUDA graphs are enabled - use lower threshold
+             ((ncols > 32768) &&
+              !(((!ctx.any_cuda_graph_has_instance()) && (iscapturing == cudaStreamCaptureStatusNone)) ||
+                ctx.any_cuda_graph_enabled())))) {
 #else
         (ncols > 65536)) {
 #endif // USE_CUDA_GRAPH

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 1, 32);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 32);

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -71,7 +71,7 @@ for type_k in TYPES_KV:
             f.write(SOURCE_FATTN_VEC.format(type_k=type_k, type_v=type_v))
 
 for ncols in [8, 16, 32, 64]:
-    for ncols2 in [1, 2, 4, 8, 16]:
+    for ncols2 in [1, 2, 4, 8, 16, 32]:
         if ncols2 > ncols:
             continue
         ncols1 = ncols // ncols2
@@ -83,9 +83,9 @@ for ncols in [8, 16, 32, 64]:
                     continue
                 if head_size_kq == 72:
                     continue
-                if head_size_kq != 576 and ncols2 == 16:
+                if head_size_kq != 576 and ncols2 in (16, 32):
                     continue
-                if head_size_kq == 576 and ncols2 not in (4, 16):
+                if head_size_kq == 576 and ncols2 not in (4, 16, 32):
                     continue
                 head_size_v = head_size_kq if head_size_kq != 576 else 512
                 f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))

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

@@ -2,9 +2,9 @@
 #pragma clang diagnostic ignored "-Wunused-function"
 #pragma clang diagnostic ignored "-Wunused-but-set-variable"
 
+#include <assert.h>
 #include <HAP_farf.h>
 #include <HAP_perf.h>
-
 #include <math.h>
 #include <string.h>
 
@@ -111,7 +111,7 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
     hvx_vec_store_u(r, 4, rsum);
 }
 
-// MAD: y (F32) += x (F16) * v (float)
+// MAD: y (F32) += x (F16) * s (float)
 static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict x, int n, float s) {
     const HVX_Vector * restrict ptr_x = (const HVX_Vector *) x;
     HVX_Vector * restrict ptr_y = (HVX_Vector *) y;
@@ -318,9 +318,12 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
             uint32_t ic = 0;
 
             // Process in blocks of 32 (VLEN_FP32)
-            for (; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32) {
+            static_assert(FLASH_ATTN_BLOCK_SIZE / VLEN_FP32 == 4, "FLASH_ATTN_BLOCK_SIZE changed, fix HVX_Vector_x4 usage");
+            HVX_Vector_x4 scores_x4;
+            HVX_Vector v_max = hvx_vec_splat_f32(-INFINITY);
+            for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
-                float __attribute__((aligned(VLEN))) scores_arr[VLEN_FP32];
+                float __attribute__((aligned(VLEN))) scores_arr[FLASH_ATTN_BLOCK_SIZE];
                 for (int j = 0; j < VLEN_FP32; ++j) {
                     const uint32_t cur_ic = ic + j;
                     const uint8_t * k_ptr = k_base + cur_ic * size_k_row_padded;
@@ -356,36 +359,43 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
-                // 4. Online Softmax Update
-                HVX_Vector v_max = hvx_vec_reduce_max_f32(scores);
-                float m_block = hvx_vec_get_f32(v_max);
+                scores_x4.v[iv] = scores;
+                v_max = Q6_Vsf_vmax_VsfVsf(scores, v_max);
+            }
 
+            {
+                // 4. Online Softmax Update
+                v_max = hvx_vec_reduce_max_f32(v_max);
+                float m_block = hvx_vec_get_f32(v_max);
                 float M_old = M;
                 float M_new = (m_block > M) ? m_block : M;
                 M = M_new;
 
-                float ms = expf(M_old - M_new);
-
+                const float ms = expf(M_old - M_new);
                 hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
-                S = S * ms;
 
                 HVX_Vector M_new_vec = hvx_vec_splat_f32(M_new);
-                HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_new_vec);
-                HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
+                HVX_Vector p_sum_vec = hvx_vec_splat_f32(0.0f);
+                for (uint32_t ic2 = 0, iv = 0; ic2 + VLEN_FP32 <= current_block_size; ic2 += VLEN_FP32, ++iv) {
+                    HVX_Vector scores = scores_x4.v[iv];
+                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_new_vec);
+                    HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
 
-                HVX_Vector p_sum_vec = hvx_vec_reduce_sum_f32(P);
-                float p_sum = hvx_vec_get_f32(p_sum_vec);
-                S += p_sum;
+                    p_sum_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(p_sum_vec, P));
 
-                // 5. Accumulate V
-                float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
-                *(HVX_Vector*)p_arr = P;
+                    // 5. Accumulate V
+                    float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
+                    *(HVX_Vector*)p_arr = P;
 
-                for (int j = 0; j < VLEN_FP32; ++j) {
-                    const uint32_t cur_ic = ic + j;
-                    const uint8_t * v_ptr = v_base + cur_ic * size_v_row_padded;
-                    hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, p_arr[j]);
+                    for (int j = 0; j < VLEN_FP32; ++j) {
+                        const uint32_t cur_ic = ic2 + j;
+                        const uint8_t * v_ptr = v_base + cur_ic * size_v_row_padded;
+                        hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, p_arr[j]);
+                    }
                 }
+
+                p_sum_vec = hvx_vec_reduce_sum_f32(p_sum_vec);
+                S = S * ms + hvx_vec_get_f32(p_sum_vec);
             }
 
             // Leftover

ggml/src/ggml-metal/ggml-metal-device.m

@@ -785,8 +785,12 @@ ggml_metal_device_t ggml_metal_device_init(void) {
             dev->props.op_offload_min_batch_size  = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             dev->props.max_buffer_size            = dev->mtl_device.maxBufferLength;
-            dev->props.max_working_set_size       = dev->mtl_device.recommendedMaxWorkingSetSize;
             dev->props.max_theadgroup_memory_size = dev->mtl_device.maxThreadgroupMemoryLength;
+            if (@available(macOS 10.12, iOS 16.0, *)) {
+                dev->props.max_working_set_size   = dev->mtl_device.recommendedMaxWorkingSetSize;
+            } else {
+                dev->props.max_working_set_size   = dev->mtl_device.maxBufferLength;
+            }
 
             strncpy(dev->props.name, [[dev->mtl_device name] UTF8String], sizeof(dev->props.name) - 1);
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -85,7 +85,8 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_8x_flat
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
-    mul_mv_q6_k
+    mul_mv_q6_k_f32
+    mul_mv_q6_k_f32_flat
     mul_mv_q8_0_f32
     mul_mv_q8_0_f32_flat
     mul_mv_mxfp4_f32

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -398,6 +398,7 @@ struct ggml_backend_opencl_context {
     int adreno_wave_size;
 
     cl_bool non_uniform_workgroups;
+    size_t  image_max_buffer_size;
 
     cl_context context;
     cl_command_queue queue;
@@ -407,6 +408,10 @@ struct ggml_backend_opencl_context {
     ggml_cl_buffer prealloc_scales_trans;
     ggml_cl_buffer prealloc_act_trans;
 
+    // prealloc buffers for src0 and src1
+    ggml_cl_buffer prealloc_src0;
+    ggml_cl_buffer prealloc_src1;
+
     cl_program program_add;
     cl_program program_add_id;
     cl_program program_clamp;
@@ -528,8 +533,10 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_restore_block_q4_0_noshuffle;
+    cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_q6_K_f32_flat;
     cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
     cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat;
     cl_kernel kernel_solve_tri_f32;
@@ -887,6 +894,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1109,14 +1118,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
-    // mul_mv_q6_k
+    // mul_mv_q6_k_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
         const std::string kernel_src {
-            #include "mul_mv_q6_k.cl.h"
+            #include "mul_mv_q6_k_f32.cl.h"
         };
 #else
-        const std::string kernel_src = read_file("mul_mv_q6_k.cl");
+        const std::string kernel_src = read_file("mul_mv_q6_k_f32.cl");
 #endif
         backend_ctx->program_mul_mv_q6_K =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
@@ -1125,6 +1134,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_q6_k_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q6_k_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_q8_0_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2658,6 +2684,9 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL);
     GGML_LOG_INFO("ggml_opencl: max mem alloc size: %zu MB\n", backend_ctx->max_alloc_size/1024/1024);
 
+    clGetDeviceInfo(device, CL_DEVICE_IMAGE_MAX_BUFFER_SIZE, sizeof(size_t), &backend_ctx->image_max_buffer_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: device max image buffer size (pixels): %lu\n", backend_ctx->image_max_buffer_size);
+
     clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), &backend_ctx->max_workgroup_size, NULL);
     GGML_LOG_INFO("ggml_opencl: device max workgroup size: %lu\n", backend_ctx->max_workgroup_size);
 
@@ -2911,6 +2940,50 @@ struct ggml_tensor_extra_cl_q8_0 {
     }
 };
 
+struct ggml_tensor_extra_cl_q6_K {
+    // Lower 4 bits of quantized weights.
+    cl_mem ql = nullptr;
+    // Upper 2 bits of quantized weights.
+    cl_mem qh = nullptr;
+    // Scales for each block.
+    cl_mem s  = nullptr;
+    // Scales for each super block.
+    cl_mem d  = nullptr;
+
+    size_t size_ql = 0;
+    size_t size_qh = 0;
+    size_t size_s  = 0;
+    size_t size_d  = 0;
+
+    ~ggml_tensor_extra_cl_q6_K() {
+        reset();
+    }
+
+    void reset() {
+        if (ql != nullptr) {
+            CL_CHECK(clReleaseMemObject(ql));
+            ql = nullptr;
+        }
+        if (qh != nullptr) {
+            CL_CHECK(clReleaseMemObject(qh));
+            qh = nullptr;
+        }
+        if (s != nullptr) {
+            CL_CHECK(clReleaseMemObject(s));
+            s = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+
+        size_ql = 0;
+        size_qh = 0;
+        size_s  = 0;
+        size_d  = 0;
+    }
+};
+
 //------------------------------------------------------------------------------
 // Backend API
 //------------------------------------------------------------------------------
@@ -3457,6 +3530,12 @@ struct ggml_backend_opencl_buffer_context {
         for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
             delete e;
         }
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
+            delete e;
+        }
     }
 
     ggml_tensor_extra_cl * ggml_opencl_alloc_temp_tensor_extra() {
@@ -3519,6 +3598,21 @@ struct ggml_backend_opencl_buffer_context {
         return extra;
     }
 
+    ggml_tensor_extra_cl_q6_K * ggml_opencl_alloc_temp_tensor_extra_q6_K() {
+        ggml_tensor_extra_cl_q6_K * extra;
+        if (temp_tensor_extras_q6_K.empty()) {
+            extra = new ggml_tensor_extra_cl_q6_K();
+        } else {
+            extra = temp_tensor_extras_q6_K.back();
+            temp_tensor_extras_q6_K.pop_back();
+        }
+
+        temp_tensor_extras_q6_K_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
     void reset() {
         for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) {
             temp_tensor_extras.push_back(e);
@@ -3539,6 +3633,11 @@ struct ggml_backend_opencl_buffer_context {
             temp_tensor_extras_q8_0.push_back(e);
         }
         temp_tensor_extras_q8_0_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
+            temp_tensor_extras_q6_K.push_back(e);
+        }
+        temp_tensor_extras_q6_K_in_use.clear();
     }
 
     // Pools for extras. Available extras are in `temp_tensor_extras`. Extras
@@ -3554,6 +3653,8 @@ struct ggml_backend_opencl_buffer_context {
     std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
+    std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K;
+    std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K_in_use;
 
     // The buffer_context is initially created by ggml_backend_buft_alloc_buffer
     // before any tensor is initialized (at the beginning of alloc_tensor_range).
@@ -4060,6 +4161,92 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
 
         return;
     }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q6_K * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q6_K();
+
+        size_t size_ql = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        size_t size_qh = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/4;
+        size_t size_s  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/16;
+        size_t size_d  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        GGML_ASSERT(size_ql + size_qh + size_s + size_d == ggml_nbytes(tensor) &&
+            "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        cl_buffer_region region;
+
+        // Subbuffer for ql
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_ql;
+        extra->ql = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Subbuffer for qh
+        region.origin = align_to(previous_origin + size_ql, backend_ctx->alignment);
+        region.size = size_qh;
+        extra->qh = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Subbuffer for scales
+        region.origin = align_to(previous_origin + size_qh, backend_ctx->alignment);
+        region.size = size_s;
+        extra->s = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Create subbuffer for d.
+        region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Flatten the weights
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q6_K;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        extra->size_ql = size_ql;
+        extra->size_qh = size_qh;
+        extra->size_s  = size_s;
+        extra->size_d  = size_d;
+
+        tensor->extra  = extra;
+        return;
+    }
 #endif // GGML_OPENCL_SOA_Q
 
     ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
@@ -4269,6 +4456,34 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {1, 1, 1};
 
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
         cl_event evt;
         CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
             global_work_size, local_work_size, 0, NULL, &evt));
@@ -4711,6 +4926,81 @@ static bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct gg
             (ne0 >= 32 && ne1 >= 32 && ne10 >= 32);
 }
 
+// Copy a noncontiguous tensor to contiguous tensor. ne[] remains the same but
+// nb[] is recalculated such that tensor is contiguous.
+static void ggml_cl_copy_to_contiguous(ggml_backend_t backend, const ggml_tensor * src, cl_mem dst,
+                                       cl_ulong &nb0, cl_ulong &nb1, cl_ulong &nb2, cl_ulong &nb3) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int tensor_type_size = ggml_type_size(src->type);
+
+    const int ne00 = src->ne[0];
+    const int ne01 = src->ne[1];
+    const int ne02 = src->ne[2];
+    const int ne03 = src->ne[3];
+
+    const cl_ulong nb00 = src->nb[0];
+    const cl_ulong nb01 = src->nb[1];
+    const cl_ulong nb02 = src->nb[2];
+    const cl_ulong nb03 = src->nb[3];
+
+    const int ne0 = src->ne[0];
+    const int ne1 = src->ne[1];
+    const int ne2 = src->ne[2];
+    const int ne3 = src->ne[3];
+
+    nb0 = tensor_type_size;
+    nb1 = tensor_type_size*ne00;
+    nb2 = tensor_type_size*ne00*ne01;
+    nb3 = tensor_type_size*ne00*ne01*ne02;
+
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *)src->extra;
+
+    cl_ulong offset0 = extra->offset + src->view_offs;
+    cl_ulong offsetd = 0;
+
+    cl_kernel kernel;
+
+    switch (src->type) {
+        case GGML_TYPE_F32:
+            kernel = backend_ctx->kernel_cpy_f32_f32;
+            break;
+        case GGML_TYPE_F16:
+            kernel = backend_ctx->kernel_cpy_f16_f16;
+            break;
+        default:
+            GGML_ASSERT(false && "not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &dst));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb3));
+
+    const int nth = MIN(64, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src);
+}
+
 static void ggml_cl_nop(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     UNUSED(backend);
     UNUSED(src0);
@@ -7682,6 +7972,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
     ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
     ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+    ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
 #endif
 
     const int  ne00 = src0 ? src0->ne[0] : 0;
@@ -7724,9 +8015,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     cl_context context = backend_ctx->context;
 
     if(src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32){
-        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0) {
+        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0  &&
+            // dst is wrapped with image1d_buffer, the size limit applies, also src0
+            (ne0 * ne1 * dst->ne[2] * dst->nb[0] / 4 <= backend_ctx->image_max_buffer_size)) {
             // For KQ
             if (ggml_is_permuted(src0) && ggml_is_permuted(src1) &&
+                ((nb01 * ne01 / 4)/4 <= backend_ctx->image_max_buffer_size) &&
                 nb00 <= nb02 &&
                 nb02 <= nb01 &&
                 nb01 <= nb03 &&
@@ -7737,7 +8031,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 return;
             }
             // For KQV
-            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+                ((nb02 * ne02 / 4)/4 <= backend_ctx->image_max_buffer_size)) {
                 ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
                 return;
             }
@@ -8043,9 +8338,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
     // GEMM using local memory
     // Current BK = 16, so ne00 % 16 == 0
-    if (ggml_is_contiguous(src0) &&
-        ggml_is_contiguous(src1) &&
-        src1t == GGML_TYPE_F32 &&
+    if (src1t == GGML_TYPE_F32 &&
         ne00 % 16 == 0 &&
         ne11 > 1) {
         switch(src0t) {
@@ -8057,10 +8350,42 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 int batch_stride_b = ne10*ne11;
                 int batch_stride_d = ne0*ne1;
 
-                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                cl_mem mem_src0 = extra0->data_device;
+                cl_mem mem_src1 = extra1->data_device;
+
+                cl_ulong nb00_cont = nb00;
+                cl_ulong nb01_cont = nb01;
+                cl_ulong nb02_cont = nb02;
+                cl_ulong nb03_cont = nb03;
+
+                cl_ulong nb10_cont = nb10;
+                cl_ulong nb11_cont = nb11;
+                cl_ulong nb12_cont = nb12;
+                cl_ulong nb13_cont = nb13;
+
+                cl_ulong offset0_cont = offset0;
+                cl_ulong offset1_cont = offset1;
+
+                if (!ggml_is_contiguous(src0)) {
+                    backend_ctx->prealloc_src0.allocate(backend_ctx->context, ggml_nbytes(src0));
+                    ggml_cl_copy_to_contiguous(backend, src0, backend_ctx->prealloc_src0.buffer,
+                        nb00_cont, nb01_cont, nb02_cont, nb03_cont);
+                    mem_src0 = backend_ctx->prealloc_src0.buffer;
+                    offset0_cont = 0;
+                }
+
+                if (!ggml_is_contiguous(src1)) {
+                    backend_ctx->prealloc_src1.allocate(backend_ctx->context, ggml_nbytes(src1));
+                    ggml_cl_copy_to_contiguous(backend, src1, backend_ctx->prealloc_src1.buffer,
+                        nb10_cont, nb11_cont, nb12_cont, nb13_cont);
+                    mem_src1 = backend_ctx->prealloc_src1.buffer;
+                    offset1_cont = 0;
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &mem_src0));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0_cont));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &mem_src1));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1_cont));
                 CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
                 CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
                 CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
@@ -8092,10 +8417,42 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 int batch_stride_b = ne10*ne11;
                 int batch_stride_d = ne0*ne1;
 
-                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                cl_mem mem_src0 = extra0->data_device;
+                cl_mem mem_src1 = extra1->data_device;
+
+                cl_ulong nb00_cont = nb00;
+                cl_ulong nb01_cont = nb01;
+                cl_ulong nb02_cont = nb02;
+                cl_ulong nb03_cont = nb03;
+
+                cl_ulong nb10_cont = nb10;
+                cl_ulong nb11_cont = nb11;
+                cl_ulong nb12_cont = nb12;
+                cl_ulong nb13_cont = nb13;
+
+                cl_ulong offset0_cont = offset0;
+                cl_ulong offset1_cont = offset1;
+
+                if (!ggml_is_contiguous(src0)) {
+                    backend_ctx->prealloc_src0.allocate(backend_ctx->context, ggml_nbytes(src0));
+                    ggml_cl_copy_to_contiguous(backend, src0, backend_ctx->prealloc_src0.buffer,
+                        nb00_cont, nb01_cont, nb02_cont, nb03_cont);
+                    mem_src0 = backend_ctx->prealloc_src0.buffer;
+                    offset0_cont = 0;
+                }
+
+                if (!ggml_is_contiguous(src1)) {
+                    backend_ctx->prealloc_src1.allocate(backend_ctx->context, ggml_nbytes(src1));
+                    ggml_cl_copy_to_contiguous(backend, src1, backend_ctx->prealloc_src1.buffer,
+                            nb10_cont, nb11_cont, nb12_cont, nb13_cont);
+                    mem_src1 = backend_ctx->prealloc_src1.buffer;
+                    offset1_cont = 0;
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &mem_src0));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0_cont));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &mem_src1));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1_cont));
                 CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
                 CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
                 CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
@@ -8123,6 +8480,10 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 if (ne11 < 32) {
                     break;
                 }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
                 kernel = backend_ctx->kernel_mul_mm_q8_0_f32_l4_lm;
                 nth0 = 128; // calculated as (BM*BN)/(TM*TN)
 
@@ -8495,14 +8856,49 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q6_K_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q6_K->ql));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q6_K->qh));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q6_K->s));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q6_K->d));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+#else
             kernel = backend_ctx->kernel_mul_mv_q6_K_f32;
 
             if (backend_ctx->gpu_family == INTEL) {
-                nth0 = 2;
-                nth1 = 16;
+                nth0 = 16;
+                nth1 = 2;
+                ndst = 1;
             } else if (backend_ctx->gpu_family == ADRENO) {
-                nth0 = 2;
-                nth1 = 64;
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 1;
             } else {
                 GGML_ASSERT(false && "TODO: Unknown GPU");
             }
@@ -8522,6 +8918,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
             CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
             break;
         case GGML_TYPE_MXFP4: {
 #ifdef GGML_OPENCL_SOA_Q
@@ -8624,7 +9021,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     } else if (src0t == GGML_TYPE_Q5_K) {
         GGML_ASSERT(false && "not implemented");
     } else if (src0t == GGML_TYPE_Q6_K) {
-        size_t global_work_size[] = {(size_t)(ne01+1)/2*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+        size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
         size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
 
         backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);

ggml/src/ggml-opencl/kernels/cvt.cl

@@ -46,6 +46,16 @@ struct block_q4_0
     uint8_t qs[QK4_0 / 2];
 };
 
+//------------------------------------------------------------------------------
+// block_q6_K
+//------------------------------------------------------------------------------
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;                  // super-block scale
+};
+
 //------------------------------------------------------------------------------
 // kernel_convert_block_q4_0
 // Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
@@ -263,3 +273,63 @@ kernel void kernel_restore_block_q8_0(
         b->qs[i] = q[i];
     }
 }
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q6_K
+// Convert the block_q6_K format to 3 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+// Each thread processes a super block.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q6_K(
+    global struct block_q6_K * src0,
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d
+) {
+    global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        ql[i] = b->ql[i];
+    }
+    for (int i = 0; i < QK_K/4; ++i) {
+        qh[i] = b->qh[i];
+    }
+    for (int i = 0; i < QK_K/16; ++i) {
+        s[i] = b->scales[i];
+    }
+}
+
+// Restore block_q6_K from flattened arrays.
+// Each thread processes a super block.
+kernel void kernel_restore_block_q6_K(
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d,
+    global struct block_q6_K * dst
+) {
+    global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    b->d = *d;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        b->ql[i] = ql[i];
+    }
+    for (int i = 0; i < QK_K/4; ++i) {
+        b->qh[i] = qh[i];
+    }
+    for (int i = 0; i < QK_K/16; ++i) {
+        b->scales[i] = s[i];
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl

@@ -0,0 +1,194 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32_flat
+//------------------------------------------------------------------------------
+#define Q6_K_MASK1 0x03
+#define Q6_K_MASK2 0x0C
+#define Q6_K_MASK3 0x30
+#define Q6_K_MASK4 0xC0
+
+#define QK_K       256
+
+inline float block_q_6_K_dot_y_flat(
+    global uchar * blk_ql,
+    global uchar * blk_qh,
+    global char  * blk_scales,
+    global half  * blk_d,
+    global float * yy,
+    int ib,
+    int ip,
+    int is,
+    int l0
+) {
+    int y_offset   = 128*ip + l0;
+    int q_offset_l =  64*ip + l0;
+    int q_offset_h =  32*ip + l0;
+
+    global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
+    global uchar * q2 = q1         + QK_K/8;
+    global uchar * qh = blk_qh     + ib*64 + q_offset_h;
+    global char  * sc = blk_scales + ib*16 + is;
+
+    global float * y = yy + ib * QK_K + y_offset;
+
+    float dall = blk_d[ib];
+
+    float  sumf = 0;
+    float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+    sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+
+    return sumf;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32_flat(
+        global uchar * src0_ql,
+        global uchar * src0_qh,
+        global char  * src0_s,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;
+
+    ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    ulong offset_src0_ql = offset_src0 * 128;
+    ulong offset_src0_qh = offset_src0 * 64;
+    ulong offset_src0_s  = offset_src0 * 16;
+    ulong offset_src0_d  = offset_src0;
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + offset_src0_ql;
+    global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
+    global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+    global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;
+
+    int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix  = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip  = tid/8;   // first or second half of (super) block (0 or 1)
+    int il  = tid%8;   // each half has 8 parts, one per scale
+    int n   = 4;       // 4 scales at a time (and 4 sums)
+    int l0  = n*il;    // offset into half-block, 0..28
+    int is  = 8*ip + l0/16; // 0, 1, 8, 9
+
+    float4 sumf = 0;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        if (first_row + 0 < ne01) {
+            sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 1 < ne01) {
+            sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 2 < ne01) {
+            sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 3 < ne01) {
+            sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0);
+        }
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -1157,13 +1157,28 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+inline void * aligned_malloc_host(size_t alignment, size_t size) {
+#ifdef _WIN32
+    return _aligned_malloc(size, alignment);
+#else
+    return aligned_alloc(alignment, size);
+#endif
+}
+
+inline void free_aligned_mem_host(void * memblock) {
+#ifdef _WIN32
+    _aligned_free(memblock);
+#else
+    free(memblock);
+#endif
+}
+
 static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_sycl_host_free(buffer->context);
+    free_aligned_mem_host((void *)buffer->context);
 }
 
 static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr = ggml_sycl_host_malloc(size);
-
+    void * ptr = aligned_malloc_host(TENSOR_ALIGNMENT, size);
     if (ptr == nullptr) {
         // fallback to cpu buffer
         return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);

ggml/src/gguf.cpp

@@ -585,6 +585,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
                 break;
             }
 
+            // check that the size of the tensor in bytes is representable
+            if (ok && uint64_t(ggml_nelements(&info.t)/ggml_blck_size(info.t.type)) > SIZE_MAX/ggml_type_size(info.t.type)) {
+                GGML_LOG_ERROR("%s: tensor '%s' with shape (%" PRIi64 ", %" PRIi64 ", %" PRIi64 ", %" PRIi64 ") has a size in bytes > %zu\n",
+                    __func__, info.t.name, info.t.ne[0], info.t.ne[1], info.t.ne[2], info.t.ne[3], SIZE_MAX);
+                ok = false;
+                break;
+            }
+
             // calculate byte offsets given the tensor shape and type
             info.t.nb[0] = type_size;
             info.t.nb[1] = info.t.nb[0]*(info.t.ne[0]/blck_size);

include/llama.h

@@ -489,6 +489,7 @@ extern "C" {
     //   - returns true if the parameters could be successfully modified to fit device memory
     //   - this function is NOT thread safe because it modifies the global llama logger state
     //   - only parameters that have the same value as in llama_default_model_params are modified
+    //     with the exception of the context size which is modified if and only if equal to 0
     LLAMA_API enum llama_params_fit_status llama_params_fit(
                                    const char   * path_model,
                     struct llama_model_params   * mparams,
@@ -1475,12 +1476,12 @@ extern "C" {
     /// @details Build a split GGUF final path for this chunk.
     ///          llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
     //  Returns the split_path length.
-    LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
+    LLAMA_API int32_t llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int32_t split_no, int32_t split_count);
 
     /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
     ///          llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
     //  Returns the split_prefix length.
-    LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
+    LLAMA_API int32_t llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int32_t split_no, int32_t split_count);
 
     // Print system information
     LLAMA_API const char * llama_print_system_info(void);

scripts/compare-llama-bench.py

@@ -29,7 +29,7 @@ LLAMA_BENCH_DB_FIELDS = [
     "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
     "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
     "use_mmap",     "embeddings",   "no_op_offload",  "n_prompt",   "n_gen",        "n_depth",
-    "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",
+    "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",    "n_cpu_moe"
 ]
 
 LLAMA_BENCH_DB_TYPES = [
@@ -38,7 +38,7 @@ LLAMA_BENCH_DB_TYPES = [
     "TEXT",    "INTEGER", "INTEGER", "TEXT",    "TEXT",    "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "INTEGER", "TEXT",    "TEXT",
     "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
-    "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",
+    "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",    "INTEGER",
 ]
 
 # All test-backend-ops SQL fields
@@ -59,7 +59,7 @@ assert len(TEST_BACKEND_OPS_DB_FIELDS) == len(TEST_BACKEND_OPS_DB_TYPES)
 
 # Properties by which to differentiate results per commit for llama-bench:
 LLAMA_BENCH_KEY_PROPERTIES = [
-    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
+    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "n_cpu_moe", "tensor_buft_overrides", "model_filename", "model_type",
     "n_batch", "n_ubatch", "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v",
     "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen", "n_depth"
 ]

src/llama-context.cpp

@@ -793,7 +793,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
-        const uint32_t n_embd_out = model.hparams.get_n_embd_out();
+        const uint32_t n_embd_out = model.hparams.n_embd_out();
         return embd + j*n_embd_out;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
@@ -1279,7 +1279,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
                 {
                     // extract token embeddings
                     GGML_ASSERT(embd != nullptr);
-                    const uint32_t n_embd_out = hparams.get_n_embd_out();
+                    const uint32_t n_embd_out = hparams.n_embd_out();
 
                     GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size);
                     ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float));
@@ -1688,7 +1688,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     {
                         // extract token embeddings
                         GGML_ASSERT(embd != nullptr);
-                        const uint32_t n_embd_out = hparams.get_n_embd_out();
+                        const uint32_t n_embd_out = hparams.n_embd_out();
                         float * embd_out = embd + n_outputs_prev*n_embd_out;
 
                         if (n_outputs) {
@@ -1821,7 +1821,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
 
     const auto n_batch    = cparams.n_batch;
     const auto n_vocab    = vocab.n_tokens();
-    const auto n_embd_out = hparams.get_n_embd_out();
+    const auto n_embd_out = hparams.n_embd_out();
 
     bool has_logits = true;
     bool has_embd   = cparams.embeddings;
@@ -2173,13 +2173,6 @@ llm_graph_cb llama_context::graph_get_cb() const {
             ggml_set_name(cur, name);
         }
 
-        if (!cparams.offload_kqv) {
-            if (strcmp(name, "kqv_merged_cont") == 0) {
-                // all nodes between the KV store and the attention output are run on the CPU
-                ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu);
-            }
-        }
-
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
         const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer;
@@ -2559,6 +2552,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
+    // [TAG_CONTEXT_STATE_LOGITS]
     // write logits
     {
         LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
@@ -2903,7 +2897,7 @@ void llama_context::opt_epoch_iter(
                 };
                 ctx_compute_opt = ggml_init(params);
             }
-            ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits());
+            ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_inp_tokens(), res->get_logits());
             ggml_opt_alloc(opt_ctx, train);
 
             res->set_inputs(&ubatch);

src/llama-graph.cpp

@@ -23,7 +23,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     }
 
     if (ubatch->embd) {
-        const int64_t n_embd   = embd->ne[0];
+        GGML_ASSERT(n_embd == embd->ne[0]);
+
         const int64_t n_tokens = ubatch->n_tokens;
 
         ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd));
@@ -33,8 +34,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
 bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
     bool res = true;
 
-    res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
-    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[1] == params.ubatch.n_tokens);
+    res &= (!params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
+    res &= (!params.ubatch.embd)  || (embd   &&   embd->ne[1] == params.ubatch.n_tokens);
 
     return res;
 }
@@ -406,6 +407,27 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
     return res;
 }
 
+void llm_graph_input_attn_k::set_input(const llama_ubatch * ubatch) {
+    mctx->set_input_k_idxs(self_k_idxs, ubatch);
+
+    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+}
+
+bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+
+    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
+
+    return res;
+}
+
 void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
     mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
@@ -634,7 +656,8 @@ int64_t llm_graph_result::get_max_nodes() const {
 }
 
 void llm_graph_result::reset() {
-    t_tokens      = nullptr;
+    t_inp_tokens  = nullptr;
+    t_inp_embd    = nullptr;
     t_logits      = nullptr;
     t_embd        = nullptr;
     t_embd_pooled = nullptr;
@@ -1338,17 +1361,29 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd_inp();
+    const int64_t n_embd_inp = hparams.n_embd_inp();
+    const int64_t n_embd     = hparams.n_embd;
 
-    auto inp = std::make_unique<llm_graph_input_embd>();
+    assert(n_embd_inp >= n_embd);
 
-    ggml_tensor * cur = nullptr;
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd_inp);
 
-    if (ubatch.token) {
-        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
-        //cb(inp->tokens, "inp_tokens", -1);
-        ggml_set_input(inp->tokens);
-        res->t_tokens = inp->tokens;
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+    cb(inp->tokens, "inp_tokens", -1);
+    ggml_set_input(inp->tokens);
+    res->t_inp_tokens = inp->tokens;
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd_inp, ubatch.n_tokens);
+    cb(inp->embd, "inp_embd", -1);
+    ggml_set_input(inp->embd);
+
+    // select one of the 2 inputs, based on the batch contents
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18550
+    std::array<ggml_tensor *, 2> inps;
+
+    // token embeddings path (ubatch.token != nullptr)
+    {
+        auto & cur = inps[0];
 
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
 
@@ -1369,19 +1404,36 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 
             cur = ggml_add(ctx0, cur, inpL_delta);
         }
-    } else {
-        inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
-        ggml_set_input(inp->embd);
+
+        if (n_embd_inp != n_embd) {
+            cur = ggml_pad(ctx0, cur, hparams.n_embd_inp() - n_embd, 0, 0, 0);
+        }
+    }
+
+    // vector embeddings path (ubatch.embd != nullptr)
+    {
+        auto & cur = inps[1];
 
         cur = inp->embd;
     }
 
+    assert(ggml_are_same_shape (inps[0], inps[1]));
+    assert(ggml_are_same_stride(inps[0], inps[1]));
+
+    ggml_tensor * cur = ggml_build_forward_select(gf, inps.data(), inps.size(), ubatch.token ? 0 : 1);
+
+    if (n_embd_inp != n_embd) {
+        cur = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0);
+    }
+
+    res->t_inp_embd = cur;
+
     // For Granite architecture
     if (hparams.f_embedding_scale != 0.0f) {
         cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);
     }
 
-    cb(cur, "inp_embd", -1);
+    cb(cur, "embd", -1);
 
     res->add_input(std::move(inp));
 
@@ -1480,7 +1532,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //}
 
     const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
-    const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
+    const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc  : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
     ggml_set_input(cur);
@@ -1578,6 +1630,11 @@ ggml_tensor * llm_graph_context::build_attn_mha(
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
         cb(cur, LLAMA_TENSOR_NAME_FATTN, il);
 
+        if (!cparams.offload_kqv) {
+            // all nodes between the KV store and the attention output are run on the CPU
+            ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu);
+        }
+
         ggml_flash_attn_ext_add_sinks(cur, sinks);
         ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
 
@@ -1787,9 +1844,11 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
         ggml_tensor * sinks,
-        ggml_tensor * v_mla,
+        ggml_tensor * v_mla, // TODO: remove
             float     kq_scale,
             int       il) const {
+    GGML_ASSERT(v_mla == nullptr);
+
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
     // expand k later to enable rope fusion which directly writes into k-v cache
@@ -1832,6 +1891,93 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
+static std::unique_ptr<llm_graph_input_attn_k> build_attn_inp_k_impl(
+           ggml_context * ctx0,
+     const llama_ubatch & ubatch,
+    const llama_hparams & hparams,
+    const llama_cparams & cparams,
+    const llama_kv_cache_context * mctx_cur) {
+
+    auto inp = std::make_unique<llm_graph_input_attn_k>(hparams, cparams, mctx_cur);
+
+    {
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
+
+        const auto n_kv     = mctx_cur->get_n_kv();
+        const auto n_tokens = ubatch.n_tokens;
+        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
+        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        ggml_set_input(inp->self_kq_mask);
+
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+    }
+
+    return inp;
+}
+
+llm_graph_input_attn_k * llm_graph_context::build_attn_inp_k() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
+
+    auto inp = build_attn_inp_k_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+
+    return (llm_graph_input_attn_k *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_k * inp,
+        ggml_tensor * wo,
+        ggml_tensor * wo_b,
+        ggml_tensor * q_cur,
+        ggml_tensor * k_cur,
+        ggml_tensor * v_cur,
+        ggml_tensor * kq_b,
+        ggml_tensor * sinks,
+        ggml_tensor * v_mla,
+            float     kq_scale,
+            int       il) const {
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
+    ggml_build_forward_expand(gf, q_cur);
+    ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);
+
+    const auto * mctx_cur = inp->mctx;
+
+    // store to KV cache
+    {
+        const auto & k_idxs = inp->get_k_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+    }
+
+    const auto & kq_mask = inp->get_kq_mask();
+
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
+
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
+    cb(cur, "kqv_out", il);
+
+    if (wo) {
+        cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
+    }
+
+    if (wo_b) {
+        cur = ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_iswa * inp,
         ggml_tensor * wo,

src/llama-graph.h

@@ -106,7 +106,7 @@ using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
 
 class llm_graph_input_embd : public llm_graph_input_i {
 public:
-    llm_graph_input_embd()          = default;
+    llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}
     virtual ~llm_graph_input_embd() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -115,6 +115,8 @@ public:
 
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+
+    const int64_t n_embd = 0;
 };
 
 class llm_graph_input_pos : public llm_graph_input_i {
@@ -315,6 +317,39 @@ public:
     const llama_kv_cache_context * mctx;
 };
 
+// V-less input for the KV cache
+// ref: https://github.com/ggml-org/llama.cpp/pull/19067
+class llm_graph_input_attn_k : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_k(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_context * mctx) :
+        hparams(hparams),
+        cparams(cparams),
+        mctx(mctx) {
+    }
+    ~llm_graph_input_attn_k() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+
+    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+
+    const llama_hparams hparams;
+    const llama_cparams cparams;
+
+    const llama_kv_cache_context * mctx;
+};
+
 class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
     llm_graph_input_attn_kv_iswa(
@@ -566,7 +601,7 @@ public:
 
     virtual ~llm_graph_result() = default;
 
-    ggml_tensor * get_tokens()      const { return t_tokens; }
+    ggml_tensor * get_inp_tokens()  const { return t_inp_tokens; }
     ggml_tensor * get_logits()      const { return t_logits; }
     ggml_tensor * get_embd()        const { return t_embd; }
     ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
@@ -593,7 +628,8 @@ public:
     void set_params(const llm_graph_params & params);
 
     // important graph nodes
-    ggml_tensor * t_tokens      = nullptr;
+    ggml_tensor * t_inp_tokens  = nullptr;
+    ggml_tensor * t_inp_embd    = nullptr; // [n_embd_inp, n_tokens]
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
@@ -830,6 +866,21 @@ struct llm_graph_context {
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
             ggml_tensor * sinks, // [n_head_q]
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] // TODO: remove
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_k  * build_attn_inp_k() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_k * inp,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;

src/llama-hparams.cpp

@@ -72,8 +72,8 @@ uint32_t llama_hparams::n_embd_inp() const {
     return n_embd_inp;
 }
 
-uint32_t llama_hparams::get_n_embd_out() const {
-    return n_embd_out > 0 ? n_embd_out : n_embd;
+uint32_t llama_hparams::n_embd_out() const {
+    return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
 }
 
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
@@ -175,6 +175,21 @@ bool llama_hparams::is_swa(uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+bool llama_hparams::is_mla() const {
+    assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) ||
+           (n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0));
+
+    return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0;
+}
+
+uint32_t llama_hparams::n_embd_head_k_mla() const {
+    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k;
+}
+
+uint32_t llama_hparams::n_embd_head_v_mla() const {
+    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v;
+}
+
 bool llama_hparams::has_kv(uint32_t il) const {
     if (n_layer_kv_from_start >= 0) {
         if (il < (uint32_t) n_layer_kv_from_start) {

src/llama-hparams.h

@@ -53,8 +53,8 @@ struct llama_hparams {
     uint32_t n_rel_attn_bkts = 0;
 
     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
-    uint32_t n_embd_head_k_mla = 0;
-    uint32_t n_embd_head_v_mla = 0;
+    uint32_t n_embd_head_k_mla_impl = 0;
+    uint32_t n_embd_head_v_mla_impl = 0;
 
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
@@ -164,7 +164,7 @@ struct llama_hparams {
     uint32_t n_cls_out = 1;
 
     // output embedding dimension (0 = use n_embd)
-    uint32_t n_embd_out = 0;
+    uint32_t n_embd_out_impl = 0;
 
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
@@ -239,7 +239,7 @@ struct llama_hparams {
     uint32_t n_embd_inp() const;
 
     // dimension of output embeddings
-    uint32_t get_n_embd_out() const;
+    uint32_t n_embd_out() const;
 
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
@@ -269,6 +269,12 @@ struct llama_hparams {
 
     bool is_swa(uint32_t il) const;
 
+    // note: currently only support if either all or none of the layers are MLA
+    bool is_mla() const;
+
+    uint32_t n_embd_head_k_mla() const;
+    uint32_t n_embd_head_v_mla() const;
+
     bool has_kv(uint32_t il) const;
 
     // number of layers for which has_kv() returns true

src/llama-kv-cache.cpp

@@ -97,6 +97,8 @@ llama_kv_cache::llama_kv_cache(
                 __func__, hparams.n_embd_v_gqa_max());
     }
 
+    const bool is_mla = hparams.is_mla();
+
     for (uint32_t il = 0; il < hparams.n_layer; il++) {
         if (!hparams.has_kv(il)) {
             LLAMA_LOG_DEBUG("%s: layer %3d: does not have KV cache\n", __func__, il);
@@ -130,18 +132,21 @@ llama_kv_cache::llama_kv_cache(
             throw std::runtime_error("failed to create ggml context for kv cache");
         }
 
-        ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
-        ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
+        const bool has_k = true;
+        const bool has_v = !is_mla;
 
-        ggml_format_name(k, "cache_k_l%d", il);
-        ggml_format_name(v, "cache_v_l%d", il);
+        ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream) : nullptr;
+        ggml_tensor * v = has_v ? ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream) : nullptr;
+
+        has_k && ggml_format_name(k, "cache_k_l%d", il);
+        has_v && ggml_format_name(v, "cache_v_l%d", il);
 
         std::vector<ggml_tensor *> k_stream;
         std::vector<ggml_tensor *> v_stream;
 
         for (uint32_t s = 0; s < n_stream; ++s) {
-            k_stream.push_back(ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]));
-            v_stream.push_back(ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]));
+            k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]) : nullptr);
+            v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr);
         }
 
         map_layer_ids[il] = layers.size();
@@ -647,7 +652,10 @@ bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_co
                 const auto & layer = layers[il];
 
                 ggml_backend_tensor_copy(layer.k_stream[ssrc], layer.k_stream[sdst]);
-                ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
+
+                if (layer.v_stream[ssrc]) {
+                    ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
+                }
             }
         }
     }
@@ -1516,7 +1524,7 @@ size_t llama_kv_cache::size_v_bytes() const {
     size_t size_v_bytes = 0;
 
     for (const auto & layer : layers) {
-        size_v_bytes += ggml_nbytes(layer.v);
+        size_v_bytes += layer.v ? ggml_nbytes(layer.v) : 0;
     }
 
     return size_v_bytes;
@@ -1594,6 +1602,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
     const auto & n_embd_head_k = hparams.n_embd_head_k;
   //const auto & n_embd_head_v = hparams.n_embd_head_v;
 
+    const auto & n_rot = hparams.n_rot;
+
+    const auto n_embd_nope = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
+
     auto inp = std::make_unique<llm_graph_input_k_shift>(this);
 
     inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
@@ -1614,10 +1626,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
 
         ggml_tensor * k =
             ggml_view_3d(ctx, layer.k,
-                n_embd_head_k, n_head_kv, get_size()*n_stream,
+                n_rot, n_head_kv, get_size()*n_stream,
                 ggml_row_size(layer.k->type, n_embd_head_k),
                 ggml_row_size(layer.k->type, n_embd_k_gqa),
-                0);
+                ggml_row_size(layer.k->type, n_embd_nope));
 
         ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
 
@@ -1794,6 +1806,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[cr.strm];
+            if (!v) {
+                continue;
+            }
 
             // Write value type
             const int32_t v_type_i = (int32_t) v->type;
@@ -1820,6 +1835,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[cr.strm];
+            if (!v) {
+                continue;
+            }
 
             // Write value type
             const int32_t v_type_i = (int32_t) v->type;
@@ -2023,6 +2041,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[strm];
+            if (!v) {
+                continue;
+            }
 
             // Read type of value
             int32_t v_type_i_ref;
@@ -2064,6 +2085,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[strm];
+            if (!v) {
+                continue;
+            }
 
             // Read type of value
             int32_t v_type_i_ref;

src/llama-model-saver.cpp

@@ -146,8 +146,8 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
     add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
     add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
-    if (hparams.n_embd_out > 0) {
-        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out);
+    if (hparams.n_embd_out_impl > 0) {
+        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out_impl);
     }
     add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
     add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);

src/llama-model.cpp

@@ -512,7 +512,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_CONTEXT_LENGTH,          hparams.n_ctx_train);
     ml.get_key(LLM_KV_EMBEDDING_LENGTH,        hparams.n_embd);
-    ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out, false);
+    ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out_impl, false);
     ml.get_key(LLM_KV_BLOCK_COUNT,             hparams.n_layer);
     ml.get_key(LLM_KV_EXPERT_COUNT,            hparams.n_expert,        false);
     ml.get_key(LLM_KV_EXPERT_USED_COUNT,       hparams.n_expert_used,   false);
@@ -1697,15 +1697,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         case LLM_ARCH_DEEPSEEK2:
             {
                 // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-                bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+                const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
                 if (!is_lite) {
                     ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
                 }
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
-                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla, false);
-                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
+                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
+                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
@@ -1736,6 +1737,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
+                    case 47: type = LLM_TYPE_30B_A3B; break;
                     case 60: type = LLM_TYPE_236B; break;
                     case 61: type = LLM_TYPE_671B; break;
                     default: type = LLM_TYPE_UNKNOWN;
@@ -4909,14 +4911,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_DEEPSEEK2:
                 {
-                    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-                    const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
-
-                    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+                    const bool is_mla = hparams.is_mla();
 
                     // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
-                    const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
-                    const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+                    const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+                    const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
 
                     const int64_t n_embd_head_qk_rope = hparams.n_rot;
                     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
@@ -4941,13 +4940,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        if (!is_lite) {
+                        if (q_lora_rank > 0) {
                             layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
                         }
 
                         layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
 
-                        if (!is_lite) {
+                        if (q_lora_rank > 0) {
                             layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
                             layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
                         } else {
@@ -6597,7 +6596,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
 
                     // for LFM2-ColBert-350M
-                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.get_n_embd_out()}, TENSOR_NOT_REQUIRED);
+                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
                 } break;
             case LLM_ARCH_SMALLTHINKER:
                 {
@@ -7316,8 +7315,8 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
         LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
-        LLAMA_LOG_INFO("%s: n_embd_head_k_mla     = %d\n",     __func__, hparams.n_embd_head_k_mla);
-        LLAMA_LOG_INFO("%s: n_embd_head_v_mla     = %d\n",     __func__, hparams.n_embd_head_v_mla);
+        LLAMA_LOG_INFO("%s: n_embd_head_k_mla     = %d\n",     __func__, hparams.n_embd_head_k_mla());
+        LLAMA_LOG_INFO("%s: n_embd_head_v_mla     = %d\n",     __func__, hparams.n_embd_head_v_mla());
         LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_expert_shared       = %d\n",     __func__, hparams.n_expert_shared);
         LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
@@ -8162,7 +8161,7 @@ int32_t llama_model_n_embd_inp(const llama_model * model) {
 }
 
 int32_t llama_model_n_embd_out(const llama_model * model) {
-    return model->hparams.get_n_embd_out();
+    return model->hparams.n_embd_out();
 }
 
 int32_t llama_model_n_layer(const llama_model * model) {

src/llama-quant.cpp

@@ -422,57 +422,6 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         ++qs.i_ffn_up;
     }
 
-    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-    //}
-    // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
-    //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
-    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-    //}
-    // This can be used to reduce the size of the Q5_K_S model.
-    // The associated PPL increase is fully in line with the size reduction
-    //else {
-    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
-    //}
-    bool convert_incompatible_tensor = false;
-    {
-        const int64_t nx = tensor->ne[0];
-        const int64_t ny = tensor->ne[1];
-        const int64_t qk_k = ggml_blck_size(new_type);
-
-        if (nx % qk_k != 0) {
-            LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
-            convert_incompatible_tensor = true;
-        } else {
-            ++qs.n_k_quantized;
-        }
-    }
-
-    if (convert_incompatible_tensor) {
-        switch (new_type) {
-            case GGML_TYPE_TQ1_0:
-            case GGML_TYPE_TQ2_0:  new_type = GGML_TYPE_Q4_0; break;  // TODO: use a symmetric type instead
-            case GGML_TYPE_IQ2_XXS:
-            case GGML_TYPE_IQ2_XS:
-            case GGML_TYPE_IQ2_S:
-            case GGML_TYPE_IQ3_XXS:
-            case GGML_TYPE_IQ3_S:
-            case GGML_TYPE_IQ1_S:
-            case GGML_TYPE_IQ1_M:
-            case GGML_TYPE_Q2_K:
-            case GGML_TYPE_Q3_K:
-            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
-            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
-            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
-            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
-            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
-        }
-        if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
-            new_type = GGML_TYPE_F16;
-        }
-        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
-        ++qs.n_fallback;
-    }
-
     return new_type;
 }
 
@@ -875,21 +824,69 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             // get more optimal quantization type based on the tensor shape, layer, etc.
             if (!params->pure && ggml_is_quantized(default_type)) {
-                int fallback = qs.n_fallback;
-                new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
-                // unless the user specifies a type, and the tensor geometry will not require fallback quantisation
-                if (params->tensor_types && qs.n_fallback - fallback == 0) {
+                // if the user provided tensor types - use those
+                bool manual = false;
+                if (params->tensor_types) {
                     const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
                     const std::string tensor_name(tensor->name);
                     for (const auto & [tname, qtype] : tensor_types) {
                         if (std::regex pattern(tname); std::regex_search(tensor_name, pattern)) {
                             if  (qtype != new_type) {
-                                LLAMA_LOG_DEBUG("(overriding %s) ", ggml_type_name(new_type));
+                                LLAMA_LOG_WARN("(manual override: %s -> %s) ", ggml_type_name(new_type), ggml_type_name(qtype));
                                 new_type = qtype; // if two or more types are specified for the same tensor, the last match wins
+                                manual = true;
+                                break;
                             }
                         }
                     }
                 }
+
+                // if not manual - use the standard logic for choosing the quantization type based on the selected mixture
+                if (!manual) {
+                    new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
+                }
+
+                // incompatible tensor shapes are handled here - fallback to a compatible type
+                {
+                    bool convert_incompatible_tensor = false;
+
+                    const int64_t nx = tensor->ne[0];
+                    const int64_t ny = tensor->ne[1];
+                    const int64_t qk_k = ggml_blck_size(new_type);
+
+                    if (nx % qk_k != 0) {
+                        LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
+                        convert_incompatible_tensor = true;
+                    } else {
+                        ++qs.n_k_quantized;
+                    }
+
+                    if (convert_incompatible_tensor) {
+                        switch (new_type) {
+                            case GGML_TYPE_TQ1_0:
+                            case GGML_TYPE_TQ2_0:  new_type = GGML_TYPE_Q4_0; break;  // TODO: use a symmetric type instead
+                            case GGML_TYPE_IQ2_XXS:
+                            case GGML_TYPE_IQ2_XS:
+                            case GGML_TYPE_IQ2_S:
+                            case GGML_TYPE_IQ3_XXS:
+                            case GGML_TYPE_IQ3_S:
+                            case GGML_TYPE_IQ1_S:
+                            case GGML_TYPE_IQ1_M:
+                            case GGML_TYPE_Q2_K:
+                            case GGML_TYPE_Q3_K:
+                            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
+                            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
+                            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
+                            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
+                            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
+                        }
+                        if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
+                            new_type = GGML_TYPE_F16;
+                        }
+                        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
+                        ++qs.n_fallback;
+                    }
+                }
             }
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;

src/llama.cpp

@@ -311,8 +311,12 @@ static void llama_params_fit_impl(
                             __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
-                    LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                        __func__, hp_nct, n_ctx_min);
+                    if (n_ctx_min == UINT32_MAX) {
+                        LLAMA_LOG_INFO("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
+                    } else {
+                        LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                            __func__, hp_nct, n_ctx_min);
+                    }
                 }
             } else {
                 LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
@@ -1091,25 +1095,55 @@ int32_t llama_chat_apply_template(
 // model split
 //
 
-int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) {
+int32_t llama_split_path(
+    char * split_path,
+    size_t maxlen,
+    const char * path_prefix,
+    int32_t split_no,
+    int32_t split_count) {
+
     static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf";
-    if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) {
-        return strlen(split_path);
+
+    const int written = snprintf(
+        split_path,
+        maxlen,
+        SPLIT_PATH_FORMAT,
+        path_prefix,
+        split_no + 1,
+        split_count
+    );
+
+    if (written < 0 || (size_t) written >= maxlen) {
+        return 0;
     }
-    return 0;
+
+    return (int32_t) written;
 }
 
-int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count) {
-    std::string str_split_path(split_path);
-    char postfix[32];
-    snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count);
-    std::string str_postfix(postfix);
+int32_t llama_split_prefix(
+    char * split_prefix,
+    size_t maxlen,
+    const char * split_path,
+    int32_t split_no,
+    int32_t split_count) {
 
-    // check if split_prefix ends with postfix
-    int size_prefix = str_split_path.size() - str_postfix.size();
-    if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) {
-        snprintf(split_prefix, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path);
-        return size_prefix;
+    const std::string str_split_path(split_path);
+
+    char postfix[32];
+    snprintf(postfix, sizeof(postfix), "-%05d-of-%05d.gguf", split_no + 1, split_count);
+
+    const std::string str_postfix(postfix);
+    if (str_split_path.size() <= str_postfix.size()) {
+        return 0;
+    }
+
+    const size_t size_prefix = str_split_path.size() - str_postfix.size();
+
+    if (str_split_path.compare(size_prefix, std::string::npos, str_postfix) == 0) {
+        const size_t copy_len = std::min(size_prefix + 1, maxlen);
+        snprintf(split_prefix, copy_len, "%s", split_path);
+
+        return (int32_t) size_prefix;
     }
 
     return 0;

src/models/deepseek2.cpp

@@ -2,14 +2,11 @@
 
 llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-    bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
-
-    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+    const bool is_mla = hparams.is_mla();
 
     // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
-    const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
-    const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+    const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
 
     const int64_t n_embd_head_qk_rope = hparams.n_rot;
     const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
@@ -43,7 +40,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
-    auto * inp_attn = build_attn_inp_kv();
+    auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;
+    auto * inp_attn_k  =  is_mla ? build_attn_inp_k()  : nullptr;
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -57,6 +55,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
         // self_attention
         {
             ggml_tensor * q = NULL;
+
+            const bool is_lite = model.layers[il].wq;
+
             if (!is_lite) {
                 q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
                 cb(q, "q", il);
@@ -124,14 +125,14 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
 
                 // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
                 // note: rope must go first for in-place context shifting in build_rope_shift()
-                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
                 cb(Qcur, "Qcur", il);
 
                 kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
                 cb(kv_cmpr, "kv_cmpr_reshape", il);
 
                 // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
-                ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
                 cb(Kcur, "Kcur", il);
 
                 // {kv_lora_rank, 1, n_tokens}
@@ -145,7 +146,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 }
 
                 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
-                cur = build_attn(inp_attn,
+                cur = build_attn(inp_attn_k,
                         model.layers[il].wo, NULL,
                         Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
             } else {
@@ -169,11 +170,10 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 Vcur = ggml_cont(ctx0, Vcur);
                 cb(Vcur, "Vcur_cont", il);
 
-                // note: rope must go first for in-place context shifting in build_rope_shift()
-                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);
                 cb(Qcur, "Qcur", il);
 
-                ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
                 cb(Kcur, "Kcur", il);
 
                 if (inp_attn_scale) {
@@ -183,7 +183,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 }
 
                 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
-                cur = build_attn(inp_attn,
+                cur = build_attn(inp_attn_kv,
                             model.layers[il].wo, NULL,
                             Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }

src/models/gemma3n-iswa.cpp

@@ -245,12 +245,12 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_altup, n_layer, n_tokens]
 ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
-    auto inp = std::make_unique<llm_graph_input_embd>();
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
     ggml_tensor * inp_per_layer;
     if (ubatch.token) {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
         ggml_set_input(inp->tokens);
-        res->t_tokens = inp->tokens;
+        res->t_inp_tokens = inp->tokens;
         inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
         inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
         inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));

src/models/minicpm3.cpp

@@ -9,6 +9,7 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
 
     const uint32_t n_embd_head_qk_rope = hparams.n_rot;
     const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
     ggml_tensor * cur;

src/models/plm.cpp

@@ -5,6 +5,7 @@ llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params &
 
     const uint32_t n_embd_head_qk_rope = hparams.n_rot;
     const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
     ggml_tensor * cur;

src/models/qwen3vl-moe.cpp

@@ -2,7 +2,8 @@
 
 llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = hparams.n_embd;
+
+    const int64_t n_embd      = hparams.n_embd;
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16,17 +17,6 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
-    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
-
-    if (ubatch.embd) {
-        // Image input: split main embd and deepstack embds
-        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
-        for (size_t i = 0; i < n_deepstack_layers; i++) {
-            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
-        }
-        inpL = inpL_main;
-    }
-
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -120,8 +110,9 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
         cur = build_cvec(cur, il);
         cb(cur, "l_out", il);
 
-        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
-            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+        if (il < (int) n_deepstack_layers) {
+            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
+            cur = ggml_add(ctx0, cur, ds);
             cb(cur, "deepstack_out", il);
         }
 

src/models/qwen3vl.cpp

@@ -2,7 +2,8 @@
 
 llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = hparams.n_embd;
+
+    const int64_t n_embd      = hparams.n_embd;
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16,17 +17,6 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
-    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
-
-    if (ubatch.embd) {
-        // Image input: split main embd and deepstack embds
-        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
-        for (size_t i = 0; i < n_deepstack_layers; i++) {
-            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
-        }
-        inpL = inpL_main;
-    }
-
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -113,8 +103,9 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
         cur = build_cvec(cur, il);
         cb(cur, "l_out", il);
 
-        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
-            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+        if (il < (int) n_deepstack_layers) {
+            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
+            cur = ggml_add(ctx0, cur, ds);
             cb(cur, "deepstack_out", il);
         }
 

tests/test-backend-ops.cpp

@@ -6122,7 +6122,19 @@ struct test_flash_attn_ext : public test_case {
         ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,         nr23[1], true); // the K tensor is usually a view of the K cache
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1], true); // the V tensor is usually a view of the V cache
+        ggml_tensor * v = nullptr;
+        if (hsk_padded == 576 && hsv_padded == 512) {
+            // TODO: this branch should become a separate test case parameter instead of hardcoding this for these head shapes
+
+            // in this branch, the V cache is sub-view of the K cache. this is used by some MLA-based models
+            // for more info:
+            //   - https://github.com/ggml-org/llama.cpp/pull/13435
+            //   - https://github.com/ggml-org/llama.cpp/pull/18953#issuecomment-3774948392
+            //   - https://github.com/ggml-org/llama.cpp/pull/18986
+            v = ggml_view_4d(ctx, k, hsv_padded, kv, nh, nr23[1], k->nb[1], k->nb[2], k->nb[3], 0);
+        } else {
+            v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1], true); // the V tensor is usually a view of the V cache
+        }
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
@@ -8204,8 +8216,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                             for (int nh : { 4, }) {
                                 for (int nr3 : { 1, 3, }) {
                                     if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes
-                                    for (int nr2 : { 1, 4, 16 }) {
-                                        if (nr2 == 16 && hsk != 128) continue;
+                                    for (int nr2 : { 1, 4, 12 }) {
+                                        if (nr2 == 12 && hsk != 128) continue;
                                         //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) {
                                         for (int kv : { 113, 512, 1024, }) {
                                             if (nr2 != 1 && kv != 512) continue;

tests/test-chat.cpp

@@ -462,9 +462,9 @@ static void test_parser_with_streaming(const common_chat_msg & expected, const s
     for (size_t i = 1; i <= raw_message.size(); ++i) {
         auto curr_msg = parse_msg(std::string(utf8_truncate_safe_view(std::string_view(raw_message).substr(0, i))));
         if (curr_msg == simple_assist_msg("")) continue;
-        LOG_INF("Streaming msg: %s\n", common_chat_msgs_to_json_oaicompat<json>({curr_msg}).dump().c_str());
+        LOG_INF("Streaming msg: %s\n", common_chat_msgs_to_json_oaicompat({curr_msg}).dump().c_str());
         for (auto diff: common_chat_msg_diff::compute_diffs(last_msg, curr_msg)) {
-            LOG_INF("Streaming diff: %s\n", common_chat_msg_diff_to_json_oaicompat<json>(diff).dump().c_str());
+            LOG_INF("Streaming diff: %s\n", common_chat_msg_diff_to_json_oaicompat(diff).dump().c_str());
             if (!diff.reasoning_content_delta.empty()) {
                 merged.reasoning_content += diff.reasoning_content_delta;
             }
@@ -480,7 +480,7 @@ static void test_parser_with_streaming(const common_chat_msg & expected, const s
                     merged.tool_calls.back().arguments += diff.tool_call_delta.arguments;
                 }
             }
-            LOG_INF("Streaming merged: %s\n", common_chat_msgs_to_json_oaicompat<json>({merged}).dump().c_str());
+            LOG_INF("Streaming merged: %s\n", common_chat_msgs_to_json_oaicompat({merged}).dump().c_str());
         }
         assert_msg_equals(curr_msg, merged, true);
         last_msg = curr_msg;
@@ -622,7 +622,7 @@ static void test_msgs_oaicompat_json_conversion() {
         message_assist_call_code_interpreter,
     };
     for (const auto & msg : msgs) {
-        auto oai_json = common_chat_msgs_to_json_oaicompat<json>({msg});
+        auto oai_json = common_chat_msgs_to_json_oaicompat({msg});
         auto msgs2 = common_chat_msgs_parse_oaicompat(oai_json);
         assert_equals((size_t) 1, msgs2.size());
         auto msg2 = msgs2[0];
@@ -646,7 +646,7 @@ static void test_msgs_oaicompat_json_conversion() {
             "  }\n"
             "]"
         ),
-        common_chat_msgs_to_json_oaicompat<json>({message_user_parts}).dump(2));
+        common_chat_msgs_to_json_oaicompat({message_user_parts}).dump(2));
 
     assert_equals(
         std::string(
@@ -666,7 +666,7 @@ static void test_msgs_oaicompat_json_conversion() {
             "  }\n"
             "]"
         ),
-        common_chat_msgs_to_json_oaicompat<json>({message_assist_call_python}).dump(2));
+        common_chat_msgs_to_json_oaicompat({message_assist_call_python}).dump(2));
 
     auto res = common_chat_msgs_parse_oaicompat(json::parse("[{\"role\": \"assistant\", \"tool_calls\": []}]"));
     assert_equals<size_t>(1, res.size());
@@ -693,7 +693,7 @@ static void test_tools_oaicompat_json_conversion() {
     };
 
     for (const auto & tool : tools) {
-        auto oai_json = common_chat_tools_to_json_oaicompat<json>({tool});
+        auto oai_json = common_chat_tools_to_json_oaicompat({tool});
         auto tools2 = common_chat_tools_parse_oaicompat(oai_json);
         assert_equals((size_t) 1, tools2.size());
         auto tool2 = tools2[0];
@@ -726,7 +726,7 @@ static void test_tools_oaicompat_json_conversion() {
             "  }\n"
             "]"
         ),
-        common_chat_tools_to_json_oaicompat<json>({special_function_tool}).dump(2));
+        common_chat_tools_to_json_oaicompat({special_function_tool}).dump(2));
 
     {
         auto tools_no_params = common_chat_tools_parse_oaicompat(json::parse(

tests/test-gguf.cpp

@@ -1,9 +1,11 @@
 #include "ggml.h"
 #include "ggml-backend.h"
 #include "../ggml/src/ggml-impl.h"
+#include "gguf.h"
 
 #include <algorithm>
 #include <array>
+#include <cmath>
 #include <cstdint>
 #include <cstdio>
 #include <random>
@@ -34,6 +36,7 @@ enum handcrafted_file_type {
     HANDCRAFTED_TENSORS_BAD_N_DIMS         =  20 + offset_has_tensors,
     HANDCRAFTED_TENSORS_BAD_SHAPE          =  30 + offset_has_tensors,
     HANDCRAFTED_TENSORS_NE_TOO_BIG         =  40 + offset_has_tensors,
+    HANDCRAFTED_TENSORS_NBYTES_TOO_BIG     =  45 + offset_has_tensors,
     HANDCRAFTED_TENSORS_BAD_TYPE           =  50 + offset_has_tensors,
     HANDCRAFTED_TENSORS_BAD_OFFSET         =  60 + offset_has_tensors,
     HANDCRAFTED_TENSORS_DUPLICATE_NAME     =  70 + offset_has_tensors,
@@ -69,6 +72,7 @@ static std::string handcrafted_file_type_name(const enum handcrafted_file_type h
         case HANDCRAFTED_TENSORS_BAD_N_DIMS:         return "TENSORS_BAD_N_DIMS";
         case HANDCRAFTED_TENSORS_BAD_SHAPE:          return "TENSORS_BAD_SHAPE";
         case HANDCRAFTED_TENSORS_NE_TOO_BIG:         return "TENSORS_NE_TOO_BIG";
+        case HANDCRAFTED_TENSORS_NBYTES_TOO_BIG:     return "TENSORS_NBYTES_TOO_BIG";
         case HANDCRAFTED_TENSORS_BAD_TYPE:           return "TENSORS_BAD_TYPE";
         case HANDCRAFTED_TENSORS_BAD_OFFSET:         return "TENSORS_BAD_OFFSET";
         case HANDCRAFTED_TENSORS_DUPLICATE_NAME:     return "TENSORS_DUPLICATE_NAME";
@@ -326,7 +330,7 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
 
     uint64_t offset = 0;
     for (int i = 0; i < int(tensor_configs.size()); ++i) {
-        const ggml_type                          type  = tensor_configs[i].first;
+        const ggml_type                          type  = hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG ? GGML_TYPE_I64 : tensor_configs[i].first;
         const std::array<int64_t, GGML_MAX_DIMS> shape = tensor_configs[i].second;
 
         std::string name = "my_tensor";
@@ -343,7 +347,7 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         }
         helper_write(file, name.data(), name.length());
 
-        uint32_t n_dims = hft == HANDCRAFTED_TENSORS_NE_TOO_BIG ? 2 : 1;
+        uint32_t n_dims = (hft == HANDCRAFTED_TENSORS_NE_TOO_BIG || hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG) ? 2 : 1;
         for (int i = GGML_MAX_DIMS-1; i >= 1; --i) {
             if (shape[i] != 1) {
                 n_dims = i + 1;
@@ -358,13 +362,19 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         }
 
         if (hft == HANDCRAFTED_TENSORS_BAD_SHAPE) {
+            const int64_t bad_dim = -1;
             for (uint32_t j = 0; j < n_dims; ++j) {
-                const int64_t bad_dim = -1;
                 helper_write(file, bad_dim);
             }
         } else if (hft == HANDCRAFTED_TENSORS_NE_TOO_BIG){
+            const int64_t big_dim = 4*int64_t(INT32_MAX);
+            for (uint32_t j = 0; j < n_dims; ++j) {
+                helper_write(file, big_dim);
+            }
+        } else if (hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG){
+            const size_t  big_ne  = SIZE_MAX/ggml_type_size(type);
+            const int64_t big_dim = GGML_PAD(int64_t(1.01f*std::pow(big_ne, 1.0f/n_dims)) + 1, ggml_blck_size(type));
             for (uint32_t j = 0; j < n_dims; ++j) {
-                const int64_t big_dim = 4*int64_t(INT32_MAX);
                 helper_write(file, big_dim);
             }
         } else {
@@ -682,6 +692,7 @@ static std::pair<int, int> test_handcrafted_file(const unsigned int seed) {
         HANDCRAFTED_TENSORS_BAD_N_DIMS,
         HANDCRAFTED_TENSORS_BAD_SHAPE,
         HANDCRAFTED_TENSORS_NE_TOO_BIG,
+        HANDCRAFTED_TENSORS_NBYTES_TOO_BIG,
         HANDCRAFTED_TENSORS_BAD_TYPE,
         HANDCRAFTED_TENSORS_BAD_OFFSET,
         HANDCRAFTED_TENSORS_DUPLICATE_NAME,

tools/cli/README.md

@@ -45,10 +45,10 @@
 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
 | `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
@@ -109,30 +109,30 @@
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.8) |
+| `--temp N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) |
-| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
-| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
-| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
 | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
+| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
@@ -173,12 +173,12 @@
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
 | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
 | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.75)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |

tools/cli/cli.cpp

@@ -84,6 +84,9 @@ struct cli_context {
             // chat template settings
             task.params.chat_parser_params = common_chat_parser_params(chat_params);
             task.params.chat_parser_params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+            if (!chat_params.parser.empty()) {
+                task.params.chat_parser_params.parser.load(chat_params.parser);
+            }
 
             rd.post_task({std::move(task)});
         }

tools/completion/README.md

@@ -128,10 +128,10 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
 | `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
@@ -192,28 +192,30 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.8) |
+| `--temp N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
-| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
-| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
 | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
+| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
@@ -251,8 +253,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
 
 <!-- HELP_END -->

tools/completion/completion.cpp

@@ -342,44 +342,51 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    // debug message about similarity of saved session, if applicable
-    size_t n_matching_session_tokens = 0;
-    if (!session_tokens.empty()) {
-        for (llama_token id : session_tokens) {
-            if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
-                break;
+    bool session_do_save = false;
+
+    {
+        size_t n_match = 0;
+
+        if (!session_tokens.empty()) {
+            for (llama_token id : session_tokens) {
+                if (n_match >= embd_inp.size() || id != embd_inp[n_match]) {
+                    break;
+                }
+                n_match++;
+            }
+            if (params.prompt.empty() && n_match == embd_inp.size()) {
+                LOG_INF("%s: using full prompt from session file\n", __func__);
+            } else if (n_match >= embd_inp.size()) {
+                LOG_INF("%s: session file has exact match for prompt!\n", __func__);
+            } else if (n_match < (embd_inp.size() / 2)) {
+                LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+                        __func__, n_match, embd_inp.size());
+            } else {
+                LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
+                        __func__, n_match, embd_inp.size());
+            }
+
+            if (session_tokens.size() == n_match) {
+                // [TAG_CONTEXT_STATE_LOGITS]
+                // in this case, we are going to reuse the logits from the session
+                // if we ever decide to remove the logits from the session, we need to handle this somehow
+                // ref: https://github.com/ggml-org/llama.cpp/pull/18862#issuecomment-3756330941
+            }
+
+            // remove any "future" tokens that we might have inherited from the previous session
+            if (session_tokens.size() > n_match) {
+                if (!llama_memory_seq_rm(mem, -1, n_match, -1)) {
+                    LOG_WRN("%s: unable to resuse common prefix (for example, when the memory is recurrent)\n", __func__);
+                    llama_memory_clear(mem, true);
+                    session_tokens.clear();
+                    n_match = 0;
+                } else {
+                    session_tokens.resize(n_match);
+                }
             }
-            n_matching_session_tokens++;
-        }
-        if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
-            LOG_INF("%s: using full prompt from session file\n", __func__);
-        } else if (n_matching_session_tokens >= embd_inp.size()) {
-            LOG_INF("%s: session file has exact match for prompt!\n", __func__);
-        } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
-            LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
-                    __func__, n_matching_session_tokens, embd_inp.size());
-        } else {
-            LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
-                    __func__, n_matching_session_tokens, embd_inp.size());
         }
 
-        // remove any "future" tokens that we might have inherited from the previous session
-        if (!llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1)) {
-            LOG_INF("%s: unable to resuse common prefix\n", __func__);
-            n_matching_session_tokens = 0;
-            llama_memory_seq_rm(mem, -1, -1, -1);
-        }
-    }
-
-    LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
-         embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
-
-    // if we will use the cache for the full prompt without reaching the end of the cache, force
-    // reevaluation of the last token to recalculate the cached logits
-    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
-        LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
-
-        session_tokens.resize(embd_inp.size() - 1);
+        session_do_save = !path_session.empty() && n_match < embd_inp.size() && !params.prompt_cache_ro;
     }
 
     // number of tokens to keep when resetting context
@@ -521,10 +528,9 @@ int main(int argc, char ** argv) {
         is_interacting = params.interactive_first;
     }
 
-    bool is_antiprompt        = false;
-    bool input_echo           = true;
-    bool display              = true;
-    bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();
+    bool is_antiprompt = false;
+    bool input_echo    = true;
+    bool display       = true;
 
     int n_past             = 0;
     int n_remain           = params.n_predict;
@@ -700,8 +706,8 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // optionally save the session on first sample (for faster prompt loading next time)
-            if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
-                need_to_save_session = false;
+            if (session_do_save) {
+                session_do_save = false;
                 llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
 
                 LOG_DBG("saved session to %s\n", path_session.c_str());

tools/fit-params/fit-params.cpp

@@ -36,7 +36,7 @@ int main(int argc, char ** argv) {
 
     LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__);
     common_log_flush(common_log_main());
-    printf("-c %" PRIu32 " -ngl %" PRIu32, cparams.n_ctx, mparams.n_gpu_layers);
+    printf("-c %" PRIu32 " -ngl %" PRIi32, cparams.n_ctx, mparams.n_gpu_layers);
 
     size_t nd = llama_max_devices();
     while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) {

tools/mtmd/mtmd.h

@@ -224,7 +224,7 @@ MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
 
 // get output embeddings from the last encode pass
 // the reading size (in bytes) is equal to:
-// llama_model_n_embd(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
+// llama_model_n_embd_inp(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 
 // Set callback for all future logging events.

tools/server/README.md

@@ -63,10 +63,10 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
 | `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
@@ -126,30 +126,30 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.8) |
+| `--temp N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) |
-| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
-| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
-| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
 | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
+| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
@@ -199,7 +199,8 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
-| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
+| `--cache-prompt, --no-cache-prompt` | whether to enable prompt caching (default: enabled)<br/>(env: LLAMA_ARG_CACHE_PROMPT) |
+| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting, requires prompt caching to be enabled (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
 | `--slots, --no-slots` | expose slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
@@ -212,8 +213,8 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--prefill-assistant, --no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/><br/>(env: LLAMA_ARG_PREFILL_ASSISTANT) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
@@ -222,7 +223,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |
 | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
 | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.75)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
@@ -781,6 +782,7 @@ By default, it is read-only. To make POST request to change global properties, y
   "total_slots": 1,
   "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
   "chat_template": "...",
+  "chat_template_caps": {},
   "modalities": {
     "vision": false
   },
@@ -793,6 +795,7 @@ By default, it is read-only. To make POST request to change global properties, y
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `model_path` - the path to model file (same with `-m` argument)
 - `chat_template` - the model's original Jinja2 prompt template
+- `chat_template_caps` - capabilities of the chat template (see `common/jinja/caps.h` for more info)
 - `modalities` - the list of supported modalities
 - `is_sleeping` - sleeping status, see [Sleeping on idle](#sleeping-on-idle)
 
@@ -1267,6 +1270,12 @@ This provides information on the performance of the server. It also allows calcu
 
 The total number of tokens in context is equal to `prompt_n + cache_n + predicted_n`
 
+*Reasoning support*
+
+The server supports parsing and returning reasoning via the `reasoning_content` field, similar to Deepseek API.
+
+Reasoning input (preserve reasoning in history) is also supported by some specific templates. For more details, please refer to [PR#18994](https://github.com/ggml-org/llama.cpp/pull/18994).
+
 ### POST `/v1/responses`: OpenAI-compatible Responses API
 
 *Options:*

tools/server/server-context.cpp

@@ -48,11 +48,8 @@ enum server_state {
 struct server_slot {
     int id;
 
-    llama_batch batch_spec = {};
-
     // TODO: change to unique_ptrs for consistency:
     llama_context * ctx = nullptr;
-    llama_context * ctx_dft = nullptr;
 
     // multimodal
     mtmd_context * mctx = nullptr;
@@ -259,7 +256,7 @@ struct server_slot {
     }
 
     bool can_speculate() const {
-        return ctx_dft;
+        return !!spec;
     }
 
     void add_token(const completion_token_output & token) {
@@ -397,6 +394,7 @@ struct server_slot {
                     draft_ratio, n_draft_accepted, n_draft_total
             );
         }
+        common_speculative_print_stats(spec);
     }
 
     json to_json(bool only_metrics = false) const {
@@ -553,18 +551,13 @@ private:
 
     // note: keep these alive - they determine the lifetime of the model, context, etc.
     common_init_result_ptr llama_init;
-    common_init_result_ptr llama_init_dft;
 
     llama_context * ctx = nullptr;
 
-    bool vocab_dft_compatible = true;
-
-    llama_model * model_dft = nullptr;
-
-    llama_context_params cparams_dft;
-
     llama_batch batch {};
 
+    llama_model_ptr model_dft;
+
     bool add_bos_token  = true;
 
     int32_t n_ctx; // total context for all clients / slots
@@ -597,13 +590,8 @@ private:
 
         // Clear any sampling context
         for (server_slot & slot : slots) {
-            llama_free(slot.ctx_dft);
-            slot.ctx_dft = nullptr;
-
             common_speculative_free(slot.spec);
             slot.spec = nullptr;
-
-            llama_batch_free(slot.batch_spec);
         }
 
         llama_batch_free(batch);
@@ -648,44 +636,26 @@ private:
 
         add_bos_token = llama_vocab_get_add_bos(vocab);
 
-        if (params_base.has_speculative()) {
+        if (params_base.speculative.has_dft()) {
             SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
 
+            const auto & params_spec = params_base.speculative;
+
             auto params_dft = params_base;
 
-            params_dft.devices      = params_base.speculative.devices;
-            params_dft.model        = params_base.speculative.model;
-            params_dft.n_ctx        = params_base.speculative.n_ctx == 0 ? llama_n_ctx_seq(ctx) : params_base.speculative.n_ctx;
-            params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
-            params_dft.n_parallel   = 1;
-            params_dft.cache_type_k = params_base.speculative.cache_type_k;
-            params_dft.cache_type_v = params_base.speculative.cache_type_v;
+            params_dft.devices      = params_spec.devices;
+            params_dft.model        = params_spec.model;
+            params_dft.n_gpu_layers = params_spec.n_gpu_layers;
 
-            params_dft.cpuparams.n_threads = params_base.speculative.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params_base.speculative.cpuparams_batch.n_threads;
-            params_dft.tensor_buft_overrides = params_base.speculative.tensor_buft_overrides;
+            params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides;
 
-            llama_init_dft = common_init_from_params(params_dft);
-
-            model_dft = llama_init_dft->model();
+            auto mparams_dft = common_model_params_to_llama(params_dft);
 
+            model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
             if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str());
+                SRV_ERR("failed to load draft model, '%s'\n", params_spec.model.path.c_str());
                 return false;
             }
-
-            vocab_dft_compatible = common_speculative_are_compatible(ctx, llama_init_dft->context());
-            if (!vocab_dft_compatible) {
-                SRV_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
-            }
-
-            const int n_ctx_dft = llama_n_ctx(llama_init_dft->context());
-
-            cparams_dft = common_context_params_to_llama(params_dft);
-            cparams_dft.n_batch = n_ctx_dft;
-
-            // the context is not needed - we will create one for each slot
-            llama_init_dft->free_context();
         }
 
         std::string & mmproj_path = params_base.mmproj.path;
@@ -695,6 +665,7 @@ private:
             }
 
             mtmd_context_params mparams = mtmd_context_params_default();
+
             mparams.use_gpu          = params_base.mmproj_use_gpu;
             mparams.print_timings    = false;
             mparams.n_threads        = params_base.cpuparams.n_threads;
@@ -702,6 +673,7 @@ private:
             mparams.warmup           = params_base.warmup;
             mparams.image_min_tokens = params_base.image_min_tokens;
             mparams.image_max_tokens = params_base.image_max_tokens;
+
             mctx = mtmd_init_from_file(mmproj_path.c_str(), model, mparams);
             if (mctx == nullptr) {
                 SRV_ERR("failed to load multimodal model, '%s'\n", mmproj_path.c_str());
@@ -718,11 +690,6 @@ private:
                 params_base.n_cache_reuse = 0;
                 SRV_WRN("%s\n", "cache_reuse is not supported by multimodal, it will be disabled");
             }
-
-            if (params_base.has_speculative()) {
-                SRV_ERR("%s\n", "err: speculative decode is not supported by multimodal");
-                return false;
-            }
         }
 
         if (!llama_memory_can_shift(llama_get_memory(ctx))) {
@@ -757,29 +724,39 @@ private:
         for (int i = 0; i < params_base.n_parallel; i++) {
             server_slot slot;
 
-            slot.id = i;
-            slot.ctx = ctx;
+            slot.id    = i;
+            slot.ctx   = ctx;
             slot.n_ctx = n_ctx_slot;
-            slot.mctx = mctx;
+
+            slot.mctx                   = mctx;
             slot.prompt.tokens.has_mtmd = mctx != nullptr;
 
-            if (model_dft) {
-                slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
+            // try speculative decoding
+            {
+                const auto & params_spec = params_base.speculative;
 
-                // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK]
-                slot.ctx_dft = llama_init_from_model(model_dft, cparams_dft);
-                if (slot.ctx_dft == nullptr) {
-                    SRV_ERR("%s", "failed to create draft context\n");
-                    return false;
-                }
+                auto params_dft = params_base;
 
-                slot.spec = common_speculative_init(slot.ctx, slot.ctx_dft);
-                if (slot.spec == nullptr) {
-                    SRV_ERR("%s", "failed to create speculator\n");
-                    return false;
-                }
-                for (auto & pair : params_base.speculative.replacements) {
-                    common_speculative_add_replacement_tgt_dft(slot.spec, pair.first.c_str(), pair.second.c_str());
+                params_dft.n_parallel   = 1;
+                params_dft.n_ctx        = params_spec.n_ctx == 0 ? llama_n_ctx_seq(ctx) : params_spec.n_ctx;
+                params_dft.n_batch      = llama_n_ctx_seq(ctx);
+                params_dft.cache_type_k = params_spec.cache_type_k;
+                params_dft.cache_type_v = params_spec.cache_type_v;
+
+                params_dft.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
+                params_dft.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
+
+                auto cparams_dft = common_context_params_to_llama(params_dft);
+
+                slot.spec = common_speculative_init(params_base.speculative, slot.ctx, cparams_dft, model_dft.get());
+                if (slot.spec) {
+                    if (mctx) {
+                        SRV_ERR("%s\n", "speculative decoding is not supported with multimodal");
+                        return false;
+                    }
+                    SRV_WRN("%s", "speculative decoding context initialized\n");
+                } else {
+                    SRV_WRN("%s", "speculative decoding context not initialized\n");
                 }
             }
 
@@ -1059,7 +1036,7 @@ private:
         return res;
     }
 
-    std::vector<common_adapter_lora_info> construct_lora_list(const std::map<int, float> & config) {
+    std::vector<common_adapter_lora_info> construct_lora_list(const std::map<int, float> & config) const {
         std::vector<common_adapter_lora_info> output = params_base.lora_adapters; // copy
         for (size_t i = 0; i < output.size(); ++i) {
             auto it = config.find(i);
@@ -1162,7 +1139,7 @@ private:
             backend_sampling &= task.params.sampling.backend_sampling;
 
             // TODO: speculative decoding requires multiple samples per batch - not supported yet
-            backend_sampling &= !(slot.ctx_dft && task.params.speculative.n_max > 0);
+            backend_sampling &= !(slot.spec && task.params.speculative.n_max > 0);
 
             // TODO: getting post/pre sampling logits is not yet supported with backend sampling
             backend_sampling &= !need_logits;
@@ -1179,14 +1156,6 @@ private:
             slot.smpl.reset();
         }
 
-        // initialize draft batch
-        // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK]
-        if (slot.ctx_dft) {
-            llama_batch_free(slot.batch_spec);
-
-            slot.batch_spec = llama_batch_init(task.params.speculative.n_max + 1, 0, 1);
-        }
-
         slot.task = std::make_unique<const server_task>(std::move(task));
 
         slot.state = slot.task->is_child()
@@ -2066,13 +2035,19 @@ private:
                     GGML_ABORT("not supported by multimodal");
                 }
 
-                struct common_speculative_params params_spec;
-                params_spec.n_draft = n_draft_max;
-                params_spec.n_reuse = llama_n_ctx(slot.ctx_dft) - slot.task->params.speculative.n_max;
-                params_spec.p_min   = slot.task->params.speculative.p_min;
+                struct common_speculative_params params_spec = {
+                    /*.params_spec.n_draft =*/ n_draft_max,
+                    /*.params_spec.p_min   =*/ slot.task->params.speculative.p_min,
+                };
+
                 const llama_tokens & cached_text_tokens = slot.prompt.tokens.get_text_tokens();
                 llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, cached_text_tokens, slot.sampled);
 
+                if (draft.size() > 0) {
+                    std::string tmp = common_detokenize(slot.ctx, draft);
+                    //LOG_WRN("XXXXXX: draft: '%s'\n", tmp.c_str());
+                }
+
                 // add the sampled token to the batch
                 slot.i_batch_dft.push_back(batch.n_tokens);
                 common_batch_add(batch, slot.sampled, slot.prompt.tokens.pos_next(), { slot.id }, true);
@@ -2813,6 +2788,9 @@ private:
                 // update how many tokens out of those tested were accepted
                 slot.n_draft_accepted += ids.size() - 1;
 
+                // inform the speculative decoding about the number of accepted tokens
+                common_speculative_accept(slot.spec, ids.size() - 1);
+
                 // rollback to the state before sampling the draft tokens
                 slot.prompt.tokens.keep_first(slot.prompt.n_tokens() - n_draft);
 
@@ -2903,6 +2881,7 @@ server_context_meta server_context::get_meta() const {
         /* pooling_type           */ llama_pooling_type(impl->ctx),
 
         /* chat_params            */ impl->chat_params,
+        /* chat_template_caps     */ common_chat_templates_get_caps(impl->chat_params.tmpls.get()),
 
         /* bos_token_str          */ bos_token_str,
         /* eos_token_str          */ eos_token_str,
@@ -3410,6 +3389,7 @@ void server_routes::init_routes() {
             { "webui",                       params.webui },
             { "webui_settings",              meta->json_webui_settings },
             { "chat_template",               tmpl_default },
+            { "chat_template_caps",          meta->chat_template_caps },
             { "bos_token",                   meta->bos_token_str },
             { "eos_token",                   meta->eos_token_str },
             { "build_info",                  meta->build_info },

tools/server/server-context.h

@@ -22,6 +22,7 @@ struct server_context_meta {
 
     // chat params
     server_chat_params & chat_params;
+    std::map<std::string, bool> chat_template_caps;
 
     // tokens
     std::string bos_token_str;

tools/server/server-http.cpp

@@ -28,14 +28,20 @@ server_http_context::server_http_context()
 server_http_context::~server_http_context() = default;
 
 static void log_server_request(const httplib::Request & req, const httplib::Response & res) {
-    // skip GH copilot requests when using default port
-    if (req.path == "/v1/health") {
+    // skip logging requests that are regularly sent, to avoid log spam
+    if (req.path == "/health"
+        || req.path == "/v1/health"
+        || req.path == "/models"
+        || req.path == "/v1/models"
+        || req.path == "/props"
+        || req.path == "/metrics"
+    ) {
         return;
     }
 
     // reminder: this function is not covered by httplib's exception handler; if someone does more complicated stuff, think about wrapping it in try-catch
 
-    SRV_INF("request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
+    SRV_INF("done request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
 
     SRV_DBG("request:  %s\n", req.body.c_str());
     SRV_DBG("response: %s\n", res.body.c_str());