server : fix local model name in server (#4420)

Fix small typo.

.devops/tools.sh

@@ -13,6 +13,8 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     ./main "$@"
+elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
+    ./finetune "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -34,6 +36,8 @@ else
     echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
+    echo "  --finetune (-f): Run finetune command to create a lora finetune of the model"
+    echo "              See documentation for finetune for command-line parameters"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
     echo "  --server (-s): Run a model on the server"

.github/workflows/build.yml

@@ -143,6 +143,9 @@ jobs:
           cd build
           ctest --verbose
 
+  # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
   macOS-latest-make:
     runs-on: macos-latest
 
@@ -160,14 +163,18 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          make -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: make_test
         run: |
-          make tests -j $(sysctl -n hw.logicalcpu)
-          make test -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make test  -j $(sysctl -n hw.logicalcpu)
 
+  # TODO: build with LLAMA_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
+  #       would be great if we fix these
   macOS-latest-cmake:
     runs-on: macos-latest
 
@@ -188,7 +195,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake ..
+          cmake -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test

.gitignore

@@ -88,15 +88,17 @@ poetry.lock
 poetry.toml
 
 # Test binaries
-tests/test-grammar-parser
-tests/test-llama-grammar
-tests/test-double-float
-tests/test-grad0
-tests/test-opt
-tests/test-quantize-fns
-tests/test-quantize-perf
-tests/test-sampling
-tests/test-tokenizer-0-llama
-tests/test-tokenizer-0-falcon
-tests/test-tokenizer-1-llama
-tests/test-tokenizer-1-bpe
+/tests/test-grammar-parser
+/tests/test-llama-grammar
+/tests/test-double-float
+/tests/test-grad0
+/tests/test-opt
+/tests/test-quantize-fns
+/tests/test-quantize-perf
+/tests/test-sampling
+/tests/test-tokenizer-0-llama
+/tests/test-tokenizer-0-falcon
+/tests/test-tokenizer-1-llama
+/tests/test-tokenizer-1-bpe
+/tests/test-rope
+/tests/test-backend-ops

CMakeLists.txt

@@ -97,9 +97,9 @@ option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
-option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)
+option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -116,6 +116,11 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+# enable libstdc++ assertions for debug builds
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
+endif()
+
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
@@ -657,11 +662,11 @@ add_library(ggml OBJECT
             ggml-backend.h
             ggml-quants.c
             ggml-quants.h
-            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_CUDA}   ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_METAL}  ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}    ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}  ${GGML_HEADERS_EXTRA}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

Makefile

@@ -8,7 +8,8 @@ BUILD_TARGETS = \
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
-	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
+	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
+	tests/test-backend-ops
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -30,7 +31,7 @@ ifeq '' '$(findstring clang,$(shell $(CC) --version))'
 	CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
 else
 	CC_IS_CLANG=1
-	ifeq '' '$(findstring Apple LLVM,$(shell $(CC) --version))'
+	ifeq '' '$(findstring Apple,$(shell $(CC) --version))'
 		CC_IS_LLVM_CLANG=1
 	else
 		CC_IS_APPLE_CLANG=1
@@ -174,6 +175,10 @@ ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
 	MK_LDFLAGS  += -g
+
+	ifeq ($(UNAME_S),Linux)
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
+	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
 endif
@@ -648,7 +653,7 @@ beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS)
 finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-export-lora: examples/export-lora/export-lora.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
@@ -701,28 +706,28 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-opt: tests/test-opt.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
@@ -737,5 +742,11 @@ tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMM
 tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 tests/test-c.o: tests/test-c.c llama.h
 	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
+
+tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

Package.swift

@@ -2,33 +2,14 @@
 
 import PackageDescription
 
-#if arch(arm) || arch(arm64)
-let platforms: [SupportedPlatform]? = [
-    .macOS(.v12),
-    .iOS(.v14),
-    .watchOS(.v4),
-    .tvOS(.v14)
-]
-let exclude: [String] = []
-let resources: [Resource] = [
-    .process("ggml-metal.metal")
-]
-let additionalSources: [String] = ["ggml-metal.m"]
-let additionalSettings: [CSetting] = [
-    .unsafeFlags(["-fno-objc-arc"]),
-    .define("GGML_USE_METAL")
-]
-#else
-let platforms: [SupportedPlatform]? = nil
-let exclude: [String] = ["ggml-metal.metal"]
-let resources: [Resource] = []
-let additionalSources: [String] = []
-let additionalSettings: [CSetting] = []
-#endif
-
 let package = Package(
     name: "llama",
-    platforms: platforms,
+    platforms: [
+        .macOS(.v12),
+        .iOS(.v14),
+        .watchOS(.v4),
+        .tvOS(.v14)
+    ],
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -36,25 +17,30 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: exclude,
+            exclude: [],
             sources: [
                 "ggml.c",
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",
-            ] + additionalSources,
-            resources: resources,
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
+            ],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE")
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL"),
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14
                 // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
                 // .define("ACCELERATE_NEW_LAPACK"),
                 // .define("ACCELERATE_LAPACK_ILP64")
-            ] + additionalSettings,
+            ],
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]

README.md

@@ -10,6 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- **llama.h API change for handling KV cache offloading and data type: https://github.com/ggerganov/llama.cpp/pull/4309**
 - Using `llama.cpp` with AWS instances: https://github.com/ggerganov/llama.cpp/discussions/4225
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 - Collecting Apple Silicon performance stats: https://github.com/ggerganov/llama.cpp/discussions/4167
@@ -324,7 +325,7 @@ mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
 
 ### BLAS Build
 
-Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
+Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
 
 - #### Accelerate Framework:
 

common/CMakeLists.txt

@@ -11,7 +11,12 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
     if(NOT IS_DIRECTORY "${GIT_DIR}")
         file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
         string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
-        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+        string(FIND "${REAL_GIT_DIR}" "/" SLASH_POS)
+        if (SLASH_POS EQUAL 0)
+            set(GIT_DIR "${REAL_GIT_DIR}")
+        else()
+            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+        endif()
     endif()
 
     set(GIT_INDEX "${GIT_DIR}/index")

common/common.cpp

@@ -278,8 +278,18 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--memory-f32") {
-            params.memory_f16 = false;
+        } else if (arg == "--samplers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.samplers_sequence = parse_samplers_input(argv[i]);
+        } else if (arg == "--sampling-seq") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.samplers_sequence = argv[i];
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -498,6 +508,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.infill = true;
         } else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
             params.dump_kv_cache = true;
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            params.no_kv_offload = true;
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
+            params.cache_type_k = argv[++i];
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            params.cache_type_v = argv[++i];
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -678,6 +694,47 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 std::istreambuf_iterator<char>(),
                 std::back_inserter(sparams.grammar)
             );
+        } else if (arg == "--override-kv") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            char * sep = strchr(argv[i], '=');
+            if (sep == nullptr || sep - argv[i] >= 128) {
+                fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            struct llama_model_kv_override kvo;
+            std::strncpy(kvo.key, argv[i], sep - argv[i]);
+            kvo.key[sep - argv[i]] = 0;
+            sep++;
+            if (strncmp(sep, "int:", 4) == 0) {
+                sep += 4;
+                kvo.tag = LLAMA_KV_OVERRIDE_INT;
+                kvo.int_value = std::atol(sep);
+            } else if (strncmp(sep, "float:", 6) == 0) {
+                sep += 6;
+                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
+                kvo.float_value = std::atof(sep);
+            } else if (strncmp(sep, "bool:", 5) == 0) {
+                sep += 5;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
+                if (std::strcmp(sep, "true") == 0) {
+                    kvo.bool_value = true;
+                } else if (std::strcmp(sep, "false") == 0) {
+                    kvo.bool_value = false;
+                } else {
+                    fprintf(stderr, "error: Invalid boolean value for KV override: %s\n", argv[i]);
+                    invalid_param = true;
+                    break;
+                }
+            } else {
+                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            params.kv_overrides.push_back(kvo);
 #ifndef LOG_DISABLE_LOGS
         // Parse args for logging parameters
         } else if ( log_param_single_parse( argv[i] ) ) {
@@ -721,6 +778,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         }
     }
 
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back(llama_model_kv_override());
+        params.kv_overrides.back().key[0] = 0;
+    }
+
     return true;
 }
 
@@ -761,6 +823,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
+    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
     printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
@@ -798,8 +862,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
-    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
@@ -840,6 +902,12 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --verbose-prompt      print prompt before generation\n");
     printf("  -dkvc, --dump-kv-cache\n");
     printf("                        verbose print of the KV cache\n");
+    printf("  -nkvo, --no-kv-offload\n");
+    printf("                        disable KV offload\n");
+    printf("  -ctk TYPE, --cache-type-k TYPE\n");
+    printf("                        KV cache data type for K (default: %s)\n", params.cache_type_k.c_str());
+    printf("  -ctv TYPE, --cache-type-v TYPE\n");
+    printf("                        KV cache data type for V (default: %s)\n", params.cache_type_v.c_str());
     printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
@@ -850,6 +918,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
+    printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("\n");
 #ifndef LOG_DISABLE_LOGS
     log_print_usage();
@@ -886,6 +957,48 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
     GGML_UNREACHABLE();
 }
 
+//
+// String parsing
+//
+
+std::string parse_samplers_input(std::string input) {
+    std::string output = "";
+    // since samplers names are written multiple ways
+    // make it ready for both system names and input names
+    std::unordered_map<std::string, char> samplers_symbols {
+        {"top_k",      'k'},
+        {"top-k",      'k'},
+        {"top_p",      'p'},
+        {"top-p",      'p'},
+        {"nucleus",    'p'},
+        {"typical_p",  'y'},
+        {"typical-p",  'y'},
+        {"typical",    'y'},
+        {"min_p",      'm'},
+        {"min-p",      'm'},
+        {"tfs_z",      'f'},
+        {"tfs-z",      'f'},
+        {"tfs",        'f'},
+        {"temp",       't'},
+        {"temperature",'t'}
+    };
+    // expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
+    size_t separator = input.find(';');
+    while (separator != input.npos) {
+        std::string name = input.substr(0,separator);
+        input = input.substr(separator+1);
+        separator = input.find(';');
+
+        if (samplers_symbols.find(name) != samplers_symbols.end()) {
+            output += samplers_symbols[name];
+        }
+    }
+    if (samplers_symbols.find(input) != samplers_symbols.end()) {
+        output += samplers_symbols[input];
+    }
+    return output;
+}
+
 //
 // Model utils
 //
@@ -900,10 +1013,39 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
+    if (params.kv_overrides.empty()) {
+        mparams.kv_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
+        mparams.kv_overrides = params.kv_overrides.data();
+    }
 
     return mparams;
 }
 
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    throw std::runtime_error("Invalid cache type: " + s);
+}
+
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto cparams = llama_context_default_params();
 
@@ -913,7 +1055,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
     cparams.mul_mat_q         = params.mul_mat_q;
     cparams.seed              = params.seed;
-    cparams.f16_kv            = params.memory_f16;
     cparams.logits_all        = params.logits_all;
     cparams.embedding         = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
@@ -924,6 +1065,10 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_beta_fast    = params.yarn_beta_fast;
     cparams.yarn_beta_slow    = params.yarn_beta_slow;
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
+    cparams.offload_kqv       = !params.no_kv_offload;
+
+    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
+    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
 
     return cparams;
 }
@@ -1336,7 +1481,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
-    fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);

common/common.h

@@ -86,6 +86,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
     std::string logdir            = "";  // directory in which to save YAML log files
 
+    std::vector<llama_model_kv_override> kv_overrides;
+
     // TODO: avoid tuple, use struct
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
@@ -98,7 +100,6 @@ struct gpt_params {
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
-    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
@@ -123,10 +124,14 @@ struct gpt_params {
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
+    bool no_kv_offload     = false; // disable KV offloading
+
+    std::string cache_type_k = "f16"; // KV cache data type for the K
+    std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
     std::string mmproj = ""; // path to multimodal projector
-    std::string image = ""; // path to an image file
+    std::string image  = ""; // path to an image file
 };
 
 bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
@@ -141,6 +146,12 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 
 void process_escapes(std::string& input);
 
+//
+// String parsing
+//
+
+std::string parse_samplers_input(std::string input);
+
 //
 // Model utils
 //

common/grammar-parser.cpp

@@ -190,7 +190,7 @@ namespace grammar_parser {
                 pos = parse_space(pos + 1, is_nested);
             } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
                 if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
                 }
 
                 // apply transformation to previous symbol (last_sym_start to end) according to

common/sampling.cpp

@@ -99,6 +99,56 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
     return std::string(result);
 }
 
+std::string llama_sampling_order_print(const llama_sampling_params & params) {
+    std::string result = "CFG -> Penalties ";
+    if (params.mirostat == 0) {
+        for (auto s : params.samplers_sequence) {
+            switch (s) {
+                case 'k': result += "-> top_k "; break;
+                case 'f': result += "-> tfs_z "; break;
+                case 'y': result += "-> typical_p "; break;
+                case 'p': result += "-> top_p "; break;
+                case 'm': result += "-> min_p "; break;
+                case 't': result += "-> temp "; break;
+                default : break;
+            }
+        }
+    } else {
+        result += "-> mirostat ";
+    }
+
+    return result;
+}
+
+// no reasons to expose this function in header
+static void sampler_queue(
+                   struct llama_context * ctx_main,
+            const llama_sampling_params & params,
+                 llama_token_data_array & cur_p,
+                                 size_t & min_keep) {
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+
+    const float         temp              = params.temp;
+    const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float         top_p             = params.top_p;
+    const float         min_p             = params.min_p;
+    const float         tfs_z             = params.tfs_z;
+    const float         typical_p         = params.typical_p;
+    const std::string & samplers_sequence = params.samplers_sequence;
+
+    for (auto s : samplers_sequence) {
+        switch (s){
+            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case 't': llama_sample_temp     (ctx_main, &cur_p, temp); break;
+            default : break;
+        }
+    }
+}
+
 llama_token llama_sampling_sample(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
@@ -109,11 +159,6 @@ llama_token llama_sampling_sample(
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
     const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
-    const float   top_p           = params.top_p;
-    const float   min_p           = params.min_p;
-    const float   tfs_z           = params.tfs_z;
-    const float   typical_p       = params.typical_p;
     const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
     const float   penalty_repeat  = params.penalty_repeat;
     const float   penalty_freq    = params.penalty_freq;
@@ -188,12 +233,7 @@ llama_token llama_sampling_sample(
             // temperature sampling
             size_t min_keep = std::max(1, params.n_probs);
 
-            llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep);
-            llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep);
-            llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep);
-            llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep);
-            llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep);
-            llama_sample_temp     (ctx_main, &cur_p, temp);
+            sampler_queue(ctx_main, params, cur_p, min_keep);
 
             id = llama_sample_token(ctx_main, &cur_p);
 

common/sampling.h

@@ -10,22 +10,23 @@
 
 // sampling parameters
 typedef struct llama_sampling_params {
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   min_p             = 0.05f; // 0.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typical_p         = 1.00f; // 1.0 = disabled
-    float   temp              = 0.80f; // 1.0 = disabled
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.10f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = true;  // consider newlines as a repeatable token
+    int32_t     n_prev                = 64;       // number of previous tokens to remember
+    int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     top_k                 = 40;       // <= 0 to use vocab size
+    float       top_p                 = 0.95f;    // 1.0 = disabled
+    float       min_p                 = 0.05f;    // 0.0 = disabled
+    float       tfs_z                 = 1.00f;    // 1.0 = disabled
+    float       typical_p             = 1.00f;    // 1.0 = disabled
+    float       temp                  = 0.80f;    // 1.0 = disabled
+    int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
+    float       penalty_freq          = 0.00f;    // 0.0 = disabled
+    float       penalty_present       = 0.00f;    // 0.0 = disabled
+    int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float       mirostat_tau          = 5.00f;    // target entropy
+    float       mirostat_eta          = 0.10f;    // learning rate
+    bool        penalize_nl           = true;     // consider newlines as a repeatable token
+    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 
@@ -80,6 +81,9 @@ std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama
 // Print sampling parameters into a string
 std::string llama_sampling_print(const llama_sampling_params & params);
 
+// Print sampling order into a string
+std::string llama_sampling_order_print(const llama_sampling_params & params);
+
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call

convert-hf-to-gguf.py

@@ -10,7 +10,7 @@ import re
 import sys
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional
 
 import numpy as np
 import torch
@@ -168,6 +168,8 @@ class Model:
             return PersimmonModel
         if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
             return StableLMModel
+        if model_architecture == "QWenLMHeadModel":
+            return QwenModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -203,6 +205,8 @@ class Model:
             return gguf.MODEL_ARCH.PERSIMMON
         if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
             return gguf.MODEL_ARCH.STABLELM
+        if arch == "QWenLMHeadModel":
+            return gguf.MODEL_ARCH.QWEN
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -832,6 +836,131 @@ class StableLMModel(Model):
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
         self.gguf_writer.add_layer_norm_eps(1e-5)
 
+
+class QwenModel(Model):
+    @staticmethod
+    def token_bytes_to_string(b):
+        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
+        byte_encoder = bytes_to_unicode()
+        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
+
+    @staticmethod
+    def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: Optional[int] = None) -> list[bytes]:
+        parts = [bytes([b]) for b in token]
+        while True:
+            min_idx = None
+            min_rank = None
+            for i, pair in enumerate(zip(parts[:-1], parts[1:])):
+                rank = mergeable_ranks.get(pair[0] + pair[1])
+                if rank is not None and (min_rank is None or rank < min_rank):
+                    min_idx = i
+                    min_rank = rank
+            if min_rank is None or (max_rank is not None and min_rank >= max_rank):
+                break
+            assert min_idx is not None
+            parts = parts[:min_idx] + [parts[min_idx] + parts[min_idx + 1]] + parts[min_idx + 2:]
+        return parts
+
+    def set_vocab(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer  # type: ignore[attr-defined]
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams["vocab_size"]
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+
+        merges = []
+        vocab = {}
+        mergeable_ranks = tokenizer.mergeable_ranks
+        for token, rank in mergeable_ranks.items():
+            vocab[self.token_bytes_to_string(token)] = rank
+            if len(token) == 1:
+                continue
+            merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+            assert len(merged) == 2
+            merges.append(' '.join(map(self.token_bytes_to_string, merged)))
+
+        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in vocab.items()}
+        added_vocab = tokenizer.special_tokens
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                pad_token = f"[PAD{i}]".encode("utf-8")
+                tokens.append(bytearray(pad_token))
+                toktypes.append(gguf.TokenType.USER_DEFINED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.CONTROL)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
+        special_vocab.merges = merges
+        special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
+        special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
+        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_name("Qwen")
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+
+    def write_tensors(self):
+        block_count = self.hparams["num_hidden_layers"]
+        model_kv = dict(self.get_tensors())
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
 ###### CONVERSION LOGIC ######
 
 

examples/batched-bench/batched-bench.cpp

@@ -155,7 +155,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");

examples/batched.swift/README.md

@@ -1,4 +1,4 @@
 This is a swift clone of `examples/batched`.
 
 $ `make`
-$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`
+$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched.swift/Sources/main.swift

@@ -215,9 +215,10 @@ print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end
 llama_print_timings(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
-    let n_tokens = text.count + (add_bos ? 1 : 0)
+    let utf8Count = text.utf8.count
+    let n_tokens = utf8Count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
+    let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])
@@ -230,18 +231,15 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
     var result = [CChar](repeating: 0, count: 8)
     let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
     if nTokens < 0 {
-        if result.count >= -Int(nTokens) {
-            result.removeLast(-Int(nTokens))
-        } else {
-            result.removeAll()
-        }
+        let actualTokensCount = -Int(nTokens)
+        result = .init(repeating: 0, count: actualTokensCount)
         let check = llama_token_to_piece(
             model,
             token,
             &result,
             Int32(result.count)
         )
-        assert(check == nTokens)
+        assert(check == actualTokensCount)
     } else {
         result.removeLast(result.count - Int(nTokens))
     }
@@ -259,5 +257,4 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
         buffer = []
         return bufferString
     }
-    return nil
 }

examples/llama-bench/llama-bench.cpp

@@ -53,6 +53,13 @@ static std::vector<T> split(const std::string & str, char delim) {
     return values;
 }
 
+template<typename T, typename F>
+static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
+    std::vector<std::string> str_values;
+    std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
+    return str_values;
+}
+
 template<typename T>
 static T avg(const std::vector<T> & v) {
     if (v.empty()) {
@@ -126,7 +133,8 @@ struct cmd_params {
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
-    std::vector<bool> f32_kv;
+    std::vector<ggml_type> type_k;
+    std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
     std::vector<int> main_gpu;
@@ -142,7 +150,8 @@ static const cmd_params cmd_params_defaults = {
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
     /* n_batch       */ {512},
-    /* f32_kv        */ {false},
+    /* type_k        */ {GGML_TYPE_F16},
+    /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
     /* main_gpu      */ {0},
@@ -162,7 +171,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -ctk <t>, --cache-type-k <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv <t>, --cache-type-v <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
     printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
@@ -173,9 +183,32 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
-
 }
 
+static ggml_type ggml_type_from_name(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    return GGML_TYPE_COUNT;
+}
+
+
 static cmd_params parse_cmd_params(int argc, char ** argv) {
     cmd_params params;
     std::string arg;
@@ -224,13 +257,38 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
-        } else if (arg == "--memory-f32") {
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split<int>(argv[i], split_delim);
-            params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -321,7 +379,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
-    if (params.f32_kv.empty())       { params.f32_kv = cmd_params_defaults.f32_kv; }
+    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
+    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
@@ -336,7 +395,8 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
-    bool f32_kv;
+    ggml_type type_k;
+    ggml_type type_v;
     int n_threads;
     int n_gpu_layers;
     int main_gpu;
@@ -365,7 +425,8 @@ struct cmd_params_instance {
 
         cparams.n_ctx = n_prompt + n_gen;
         cparams.n_batch = n_batch;
-        cparams.f16_kv = !f32_kv;
+        cparams.type_k = type_k;
+        cparams.type_v = type_v;
         cparams.mul_mat_q = mul_mat_q;
 
         return cparams;
@@ -380,7 +441,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         cmd_params_instance instance = {
@@ -388,7 +450,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
             /* .n_prompt     = */ n_prompt,
             /* .n_gen        = */ n_gen,
             /* .n_batch      = */ nb,
-            /* .f32_kv       = */ fk,
+            /* .type_k       = */ tk,
+            /* .type_v       = */ tv,
             /* .n_threads    = */ nt,
             /* .n_gpu_layers = */ nl,
             /* .main_gpu     = */ mg,
@@ -410,7 +473,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -422,7 +486,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
-                /* .f32_kv       = */ fk,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
                 /* .main_gpu     = */ mg,
@@ -441,7 +506,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
-                /* .f32_kv       = */ fk,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
                 /* .main_gpu     = */ mg,
@@ -489,7 +555,8 @@ struct test {
     uint64_t model_n_params;
     int n_batch;
     int n_threads;
-    bool f32_kv;
+    ggml_type type_k;
+    ggml_type type_v;
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
@@ -508,7 +575,8 @@ struct test {
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
-        f32_kv = inst.f32_kv;
+        type_k = inst.type_k;
+        type_v = inst.type_v;
         n_gpu_layers = inst.n_gpu_layers;
         main_gpu = inst.main_gpu;
         mul_mat_q = inst.mul_mat_q;
@@ -571,7 +639,7 @@ struct test {
             "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_threads", "f16_kv",
+            "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
@@ -621,7 +689,7 @@ struct test {
             std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
+            std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
@@ -805,8 +873,11 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.push_back("n_batch");
         }
-        if (params.f32_kv.size() > 1 || params.f32_kv != cmd_params_defaults.f32_kv) {
-            fields.push_back("f16_kv");
+        if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
+            fields.push_back("type_k");
+        }
+        if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
+            fields.push_back("type_v");
         }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.push_back("main_gpu");

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -11,6 +11,8 @@ actor LlamaContext {
     private var context: OpaquePointer
     private var batch: llama_batch
     private var tokens_list: [llama_token]
+    /// This variable is used to store temporarily invalid cchars
+    private var temporary_invalid_cchars: [CChar]
 
     var n_len: Int32 = 512
     var n_cur: Int32 = 0
@@ -21,6 +23,7 @@ actor LlamaContext {
         self.context = context
         self.tokens_list = []
         self.batch = llama_batch_init(512, 0, 1)
+        self.temporary_invalid_cchars = []
     }
 
     deinit {
@@ -61,6 +64,7 @@ actor LlamaContext {
         print("attempting to complete \"\(text)\"")
 
         tokens_list = tokenize(text: text, add_bos: true)
+        temporary_invalid_cchars = []
 
         let n_ctx = llama_n_ctx(context)
         let n_kv_req = tokens_list.count + (Int(n_len) - tokens_list.count)
@@ -72,7 +76,7 @@ actor LlamaContext {
         }
 
         for id in tokens_list {
-            print(token_to_piece(token: id))
+            print(String(cString: token_to_piece(token: id) + [0]))
         }
 
         // batch = llama_batch_init(512, 0) // done in init()
@@ -115,10 +119,25 @@ actor LlamaContext {
 
         if new_token_id == llama_token_eos(context) || n_cur == n_len {
             print("\n")
-            return ""
+            let new_token_str = String(cString: temporary_invalid_cchars + [0])
+            temporary_invalid_cchars.removeAll()
+            return new_token_str
         }
 
-        let new_token_str = token_to_piece(token: new_token_id)
+        let new_token_cchars = token_to_piece(token: new_token_id)
+        temporary_invalid_cchars.append(contentsOf: new_token_cchars)
+        let new_token_str: String
+        if let string = String(validatingUTF8: temporary_invalid_cchars + [0]) {
+            temporary_invalid_cchars.removeAll()
+            new_token_str = string
+        } else if (0 ..< temporary_invalid_cchars.count).contains(where: {$0 != 0 && String(validatingUTF8: Array(temporary_invalid_cchars.suffix($0)) + [0]) != nil}) {
+            // in this case, at least the suffix of the temporary_invalid_cchars can be interpreted as UTF8 string
+            let string = String(cString: temporary_invalid_cchars + [0])
+            temporary_invalid_cchars.removeAll()
+            new_token_str = string
+        } else {
+            new_token_str = ""
+        }
         print(new_token_str)
         // tokens_list.append(new_token_id)
 
@@ -144,12 +163,14 @@ actor LlamaContext {
 
     func clear() {
         tokens_list.removeAll()
+        temporary_invalid_cchars.removeAll()
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
-        let n_tokens = text.count + (add_bos ? 1 : 0)
+        let utf8Count = text.utf8.count
+        let n_tokens = utf8Count + (add_bos ? 1 : 0)
         let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-        let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, false)
+        let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
 
         var swiftTokens: [llama_token] = []
         for i in 0..<tokenCount {
@@ -161,16 +182,27 @@ actor LlamaContext {
         return swiftTokens
     }
 
-    private func token_to_piece(token: llama_token) -> String {
+    /// - note: The result does not contain null-terminator
+    private func token_to_piece(token: llama_token) -> [CChar] {
         let result = UnsafeMutablePointer<Int8>.allocate(capacity: 8)
         result.initialize(repeating: Int8(0), count: 8)
+        defer {
+            result.deallocate()
+        }
+        let nTokens = llama_token_to_piece(model, token, result, 8)
 
-        let _ = llama_token_to_piece(model, token, result, 8)
-
-        let resultStr = String(cString: result)
-
-        result.deallocate()
-
-        return resultStr
+        if nTokens < 0 {
+            let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
+            newResult.initialize(repeating: Int8(0), count: Int(-nTokens))
+            defer {
+                newResult.deallocate()
+            }
+            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens)
+            let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
+            return Array(bufferPointer)
+        } else {
+            let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nTokens))
+            return Array(bufferPointer)
+        }
     }
 }

examples/llava/convert-image-encoder-to-gguf.py

@@ -5,7 +5,7 @@ import json
 import torch
 import numpy as np
 from gguf import *
-from transformers import CLIPModel, CLIPProcessor
+from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel
 
 TEXT = "clip.text"
 VISION = "clip.vision"
@@ -78,11 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
                 help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
 ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
 
+# with proper
 args = ap.parse_args()
 
 
@@ -96,15 +104,22 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
-
-with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
-    vocab = json.load(f)
-    tokens = [key for key in vocab]
+if args.clip_model_is_vision:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
 
 with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
     config = json.load(f)
-    v_hparams = config["vision_config"]
-    t_hparams = config["text_config"]
+    if args.clip_model_is_vision:
+        v_hparams = config
+        t_hparams = None
+    else:
+        v_hparams = config["vision_config"]
+        t_hparams = config["text_config"]
 
 # possible data types
 #   ftype == 0 -> float32
@@ -117,9 +132,12 @@ ftype = 1
 if args.use_f32:
     ftype = 0
 
-
-model = CLIPModel.from_pretrained(dir_model)
-processor = CLIPProcessor.from_pretrained(dir_model)
+if args.clip_model_is_vision:
+    model = CLIPVisionModel.from_pretrained(dir_model)
+    processor = None
+else:
+    model = CLIPModel.from_pretrained(dir_model)
+    processor = CLIPProcessor.from_pretrained(dir_model)
 
 fname_middle = None
 has_text_encoder = True
@@ -128,13 +146,13 @@ has_llava_projector = False
 if args.text_only:
     fname_middle = "text-"
     has_vision_encoder = False
-elif args.vision_only:
-    fname_middle = "vision-"
-    has_text_encoder = False
 elif args.llava_projector is not None:
     fname_middle = "mmproj-"
     has_text_encoder = False
     has_llava_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
 else:
     fname_middle = ""
 
@@ -182,8 +200,12 @@ if has_vision_encoder:
     block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
 
-    image_mean = processor.image_processor.image_mean if args.image_mean is None else args.image_mean
-    image_std = processor.image_processor.image_std if args.image_std is None else args.image_std
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
     fout.add_array("clip.vision.image_mean", image_mean)
     fout.add_array("clip.vision.image_std", image_std)
 

examples/main/main.cpp

@@ -100,6 +100,12 @@ static void sigint_handler(int signo) {
 }
 #endif
 
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    LOG_TEE("%s", text);
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
     g_params = &params;
@@ -113,6 +119,7 @@ int main(int argc, char ** argv) {
     log_set_target(log_filename_generator("main", "log"));
     LOG_TEE("Log start\n");
     log_dump_cmdline(argc, argv);
+    llama_log_set(llama_log_callback_logTee, nullptr);
 #endif // LOG_DISABLE_LOGS
 
     // TODO: Dump params ?
@@ -430,6 +437,7 @@ int main(int argc, char ** argv) {
         }
     }
     LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 

examples/quantize-stats/quantize-stats.cpp

@@ -321,7 +321,6 @@ int main(int argc, char ** argv) {
         auto cparams = llama_context_default_params();
         cparams.n_ctx      = 256;
         cparams.seed       = 1;
-        cparams.f16_kv     = false;
 
         ctx = llama_new_context_with_model(model, cparams);
 

examples/server/README.md

@@ -222,7 +222,7 @@ node index.js
 
     `content`: Set the text to process.
 
-    **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+-   **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 

examples/server/api_like_OAI.py

@@ -70,6 +70,7 @@ def make_postData(body, chat=False, stream=False):
     if(is_present(body, "mirostat_tau")): postData["mirostat_tau"] = body["mirostat_tau"]
     if(is_present(body, "mirostat_eta")): postData["mirostat_eta"] = body["mirostat_eta"]
     if(is_present(body, "seed")): postData["seed"] = body["seed"]
+    if(is_present(body, "grammar")): postData["grammar"] = body["grammar"]
     if(is_present(body, "logit_bias")): postData["logit_bias"] = [[int(token), body["logit_bias"][token]] for token in body["logit_bias"].keys()]
     if (args.stop != ""):
         postData["stop"] = [args.stop]

examples/server/server.cpp

@@ -155,15 +155,23 @@ struct task_server {
     json data;
     bool infill_mode = false;
     bool embedding_mode = false;
+    int multitask_id = -1;
 };
 
 struct task_result {
     int id;
+    int multitask_id = -1;
     bool stop;
     bool error;
     json result_json;
 };
 
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
 // TODO: can become bool if we can't find use of more states
 enum slot_state
 {
@@ -406,6 +414,9 @@ struct llama_client_slot
     double t_prompt_processing; // ms
     double t_token_generation; // ms
 
+    // multitasks
+    int multitask_id = -1;
+
     void reset() {
         num_prompt_tokens      = 0;
         generated_text         = "";
@@ -529,7 +540,8 @@ struct llama_server_context
 
     std::vector<task_server> queue_tasks;
     std::vector<task_result> queue_results;
-    std::mutex mutex_tasks;
+    std::vector<task_multi>  queue_multitasks;
+    std::mutex mutex_tasks; // also guards id_gen, and queue_multitasks
     std::mutex mutex_results;
 
     ~llama_server_context()
@@ -1112,17 +1124,40 @@ struct llama_server_context
         return slot.images.size() > 0;
     }
 
-    void send_error(int id, std::string error)
+    void send_error(task_server& task, std::string error)
     {
         std::lock_guard<std::mutex> lock(mutex_results);
         task_result res;
-        res.id = id;
+        res.id = task.id;
+        res.multitask_id = task.multitask_id;
         res.stop = false;
         res.error = true;
         res.result_json = { { "content", error } };
         queue_results.push_back(res);
     }
 
+    void add_multi_task(int id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    void update_multi_task(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+
     json get_model_props()
     {
         return get_formated_generation(slots[0]);
@@ -1167,6 +1202,7 @@ struct llama_server_context
         std::lock_guard<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
+        res.multitask_id = slot.multitask_id;
         res.error = false;
         res.stop = false;
 
@@ -1206,6 +1242,7 @@ struct llama_server_context
         std::lock_guard<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
+        res.multitask_id = slot.multitask_id;
         res.error = false;
         res.stop = true;
 
@@ -1251,6 +1288,12 @@ struct llama_server_context
             res.result_json["model"] = slot.oaicompat_model;
         }
 
+        // parent multitask, if any, needs to be updated
+        if (slot.multitask_id != -1)
+        {
+            update_multi_task(slot.multitask_id, slot.task_id, res);
+        }
+
         queue_results.push_back(res);
     }
 
@@ -1259,6 +1302,7 @@ struct llama_server_context
         std::lock_guard<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
+        res.multitask_id = slot.multitask_id;
         res.error = false;
         res.stop = true;
 
@@ -1285,9 +1329,9 @@ struct llama_server_context
         queue_results.push_back(res);
     }
 
-    int request_completion(json data, bool infill, bool embedding)
+    int request_completion(json data, bool infill, bool embedding, int multitask_id)
     {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
+        std::unique_lock<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
         task.target_id = 0;
@@ -1295,6 +1339,16 @@ struct llama_server_context
         task.infill_mode = infill;
         task.embedding_mode = embedding;
         task.type = COMPLETION_TASK;
+        task.multitask_id = multitask_id;
+
+        // when a completion task's prompt array is not a singleton, we split it into multiple requests
+        if (task.data.at("prompt").size() > 1)
+        {
+            lock.unlock(); // entering new func scope
+            return split_multiprompt_task(task);
+        }
+
+        // otherwise, it's a single-prompt task, we actually queue it
         queue_tasks.push_back(task);
         return task.id;
     }
@@ -1313,8 +1367,17 @@ struct llama_server_context
 
             for (int i = 0; i < (int) queue_results.size(); i++)
             {
+                // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+                if (queue_results[i].multitask_id == task_id)
+                {
+                    update_multi_task(task_id, queue_results[i].id, queue_results[i]);
+                    queue_results.erase(queue_results.begin() + i);
+                    continue;
+                }
+
                 if (queue_results[i].id == task_id)
                 {
+                    assert(queue_results[i].multitask_id == -1);
                     task_result res = queue_results[i];
                     queue_results.erase(queue_results.begin() + i);
                     return res;
@@ -1404,6 +1467,27 @@ struct llama_server_context
         queue_tasks.push_back(task);
     }
 
+    int split_multiprompt_task(task_server& multiprompt_task)
+    {
+        int prompt_count = multiprompt_task.data.at("prompt").size();
+        assert(prompt_count > 1);
+
+        int multitask_id = id_gen++;
+        std::vector<int> subtask_ids(prompt_count);
+        for (int i = 0; i < prompt_count; i++)
+        {
+            json subtask_data = multiprompt_task.data;
+            subtask_data["prompt"] = subtask_data["prompt"][i];
+
+            // subtasks inherit everything else (infill mode, embedding mode, etc.)
+            subtask_ids[i] = request_completion(subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id);
+        }
+
+        // queue up the multitask so we can track its subtask progression
+        add_multi_task(multitask_id, subtask_ids);
+        return multitask_id;
+    }
+
     void process_tasks()
     {
         std::lock_guard<std::mutex> lock(mutex_tasks);
@@ -1419,7 +1503,7 @@ struct llama_server_context
                     {
                         LOG_TEE("slot unavailable\n");
                         // send error result
-                        send_error(task.id, "slot unavailable");
+                        send_error(task, "slot unavailable");
                         return;
                     }
 
@@ -1433,11 +1517,12 @@ struct llama_server_context
                     slot->infill = task.infill_mode;
                     slot->embedding = task.embedding_mode;
                     slot->task_id = task.id;
+                    slot->multitask_id = task.multitask_id;
 
                     if (!launch_slot_with_data(slot, task.data))
                     {
                         // send error result
-                        send_error(task.id, "internal_error");
+                        send_error(task, "internal_error");
                         break;
                     }
                 } break;
@@ -1453,6 +1538,38 @@ struct llama_server_context
                 } break;
             }
         }
+
+        // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue
+        auto queue_iterator = queue_multitasks.begin();
+        while (queue_iterator != queue_multitasks.end())
+        {
+            if (queue_iterator->subtasks_remaining.empty())
+            {
+                // all subtasks done == multitask is done
+                task_result aggregate_result;
+                aggregate_result.id = queue_iterator->id;
+                aggregate_result.stop = true;
+                aggregate_result.error = false;
+
+                // collect json results into one json result
+                std::vector<json> result_jsons;
+                for (auto& subres : queue_iterator->results)
+                {
+                    result_jsons.push_back(subres.result_json);
+                    aggregate_result.error = aggregate_result.error && subres.error;
+                }
+                aggregate_result.result_json = json{ "results", result_jsons };
+
+                std::lock_guard<std::mutex> lock(mutex_results);
+                queue_results.push_back(aggregate_result);
+
+                queue_iterator = queue_multitasks.erase(queue_iterator);
+            }
+            else
+            {
+                ++queue_iterator;
+            }
+        }
     }
 
     bool update_slots() {
@@ -1844,6 +1961,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("    -spf FNAME, --system-prompt-file FNAME\n");
     printf("                        Set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA.\n");
+    printf("  --log-disable         disables logging to a file.\n");
     printf("\n");
 }
 
@@ -1990,10 +2108,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.yarn_beta_slow = std::stof(argv[i]);
         }
-        else if (arg == "--memory-f32" || arg == "--memory_f32")
-        {
-            params.memory_f16 = false;
-        }
         else if (arg == "--threads" || arg == "-t")
         {
             if (++i >= argc)
@@ -2198,6 +2312,11 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.mmproj = argv[i];
         }
+        else if (arg == "--log-disable")
+        {
+            log_set_target(stdout);
+            LOG_INFO("logging to file is disabled.", {});
+        }
         else
         {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -2263,7 +2382,9 @@ json oaicompat_completion_params_parse(
     llama_params["__oaicompat"] = true;
 
     // Map OpenAI parameters to llama.cpp parameters
+    llama_params["model"]             = json_value(body, "model", std::string("uknown"));
     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.8);
     llama_params["top_k"]             = json_value(body, "top_k", 40);
     llama_params["top_p"]             = json_value(body, "top_p", 0.95);
@@ -2287,9 +2408,7 @@ json oaicompat_completion_params_parse(
     }
 
     // Handle 'stop' field
-    if (body["stop"].is_null()) {
-        llama_params["stop"] = json::array({});
-    } else if (body["stop"].is_string()) {
+    if (body.contains("stop") && body["stop"].is_string()) {
         llama_params["stop"] = json::array({body["stop"].get<std::string>()});
     } else {
         llama_params["stop"] = json_value(body, "stop", json::array());
@@ -2596,7 +2715,7 @@ int main(int argc, char **argv)
     svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res)
             {
                 json data = json::parse(req.body);
-                const int task_id = llama.request_completion(data, false, false);
+                const int task_id = llama.request_completion(data, false, false, -1);
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.next_result(task_id);
@@ -2685,7 +2804,7 @@ int main(int argc, char **argv)
             {
                 json data = oaicompat_completion_params_parse(json::parse(req.body));
 
-                const int task_id = llama.request_completion(data, false, false);
+                const int task_id = llama.request_completion(data, false, false, -1);
 
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
@@ -2754,7 +2873,7 @@ int main(int argc, char **argv)
     svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res)
             {
                 json data = json::parse(req.body);
-                const int task_id = llama.request_completion(data, true, false);
+                const int task_id = llama.request_completion(data, true, false, -1);
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.next_result(task_id);
@@ -2858,7 +2977,7 @@ int main(int argc, char **argv)
                 {
                     prompt = "";
                 }
-                const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true);
+                const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true, -1);
                 task_result result = llama.next_result(task_id);
                 return res.set_content(result.result_json.dump(), "application/json");
             });

examples/simple/simple.cpp

@@ -75,7 +75,7 @@ int main(int argc, char ** argv) {
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
         LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
-        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        LOG_TEE("%s:        either reduce n_len or increase n_ctx\n", __func__);
         return 1;
     }
 

examples/speculative/speculative.cpp

@@ -203,8 +203,9 @@ int main(int argc, char ** argv) {
 
             const std::string token_str = llama_token_to_piece(ctx_tgt, id);
 
-            printf("%s", token_str.c_str());
-            fflush(stdout);
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
 
             if (id == llama_token_eos(model_tgt)) {
                 has_eos = true;
@@ -236,10 +237,18 @@ int main(int argc, char ** argv) {
                     ++n_past_tgt;
                     ++n_past_dft;
                     ++i_dft;
-
+                    if (params.use_color) {
+                        // Color token according to its origin sequence
+                        printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
+                        fflush(stdout);
+                    }
                     continue;
                 }
             }
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
 
             LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
 

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -1295,10 +1295,6 @@ int main(int argc, char ** argv) {
         opt_cb_data.last_save_iter = opt->iter;
     }
 
-    if (alloc) {
-        ggml_allocr_free(alloc);
-    }
-
     ggml_free(opt->ctx);
     free_train_state(train);
     ggml_free(model.ctx);

ggml-alloc.c

@@ -137,7 +137,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
 
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, tensor);
-    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    size_t cur_max = (char*)addr - (char*)alloc->base + size;
     if (cur_max > alloc->max_size) {
         printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
         for (int i = 0; i < 1024; i++) {
@@ -168,10 +168,6 @@ static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor *
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
 
-    if (!alloc->measure) {
-        ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
-    }
-
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
 #endif
@@ -237,7 +233,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
 }
 
 ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
+    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(data, size);
 
     ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
@@ -449,7 +445,6 @@ static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * n
 static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
     ggml_tallocr_t alloc = node_tallocr(galloc, view);
 
-    //printf("init_view: %s from src %s\n", view->name, view->view_src->name);
     GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
     if (update_backend) {
         view->backend = view->view_src->backend;
@@ -459,7 +454,7 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
 
     // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
     // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
-    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
+    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
         ggml_backend_buffer_init_tensor(alloc->buffer, view);
@@ -765,3 +760,43 @@ size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
 size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
     return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
 }
+
+// utils
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+
+    size_t nbytes = 0;
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL && t->view_src == NULL) {
+            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+    }
+
+    if (nbytes == 0) {
+        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
+
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(tallocr, t);
+            } else {
+                ggml_backend_view_init(buffer, t);
+            }
+        }
+    }
+
+    ggml_tallocr_free(tallocr);
+
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}

ggml-alloc.h

@@ -8,6 +8,7 @@ extern "C" {
 
 struct ggml_backend;
 struct ggml_backend_buffer;
+struct ggml_backend_buffer_type;
 
 //
 // Legacy API
@@ -80,6 +81,12 @@ GGML_API void   ggml_gallocr_alloc_graph_n(
                     struct ggml_hash_set hash_set,
                     ggml_tallocr_t * hash_node_talloc);
 
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-backend-impl.h

@@ -12,31 +12,50 @@ extern "C" {
     // Backend buffer
     //
 
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+        void     (*free_buffer)(ggml_backend_buffer_t buffer);
+        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        void *   (*get_base)   (ggml_backend_buffer_t buffer);
+        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
     };
 
     struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i iface;
-
-        ggml_backend_t                backend;
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
-
         size_t size;
     };
 
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
+    ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t      buft,
             struct ggml_backend_buffer_i           iface,
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+
     //
     // Backend
     //
@@ -49,20 +68,17 @@ extern "C" {
         void (*free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
-        // get buffer alignment
-        size_t (*get_alignment)(ggml_backend_t backend);
-
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        // (optional) asynchroneous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        void (*synchronize)     (ggml_backend_t backend);
 
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        // (optional) asynchroneous tensor copy
+        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        void (*synchronize)     (ggml_backend_t backend);
 
         // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
@@ -82,6 +98,15 @@ extern "C" {
         ggml_backend_context_t context;
     };
 
+
+    //
+    // Backend registry
+    //
+
+    typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
+
+    void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-backend.h

@@ -7,41 +7,44 @@
 extern "C" {
 #endif
 
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
     //
     // Backend buffer
     //
 
-    struct ggml_backend_buffer;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    // buffer type
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
+    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
 
-    // backend buffer functions
+    // buffer
     GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
     GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
 
     //
     // Backend
     //
 
-    struct ggml_backend;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
 
     GGML_API const char * ggml_backend_name(ggml_backend_t backend);
     GGML_API void         ggml_backend_free(ggml_backend_t backend);
 
-    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
 
-    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
@@ -57,6 +60,7 @@ extern "C" {
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
 
     //
     // CPU backend
@@ -68,8 +72,23 @@ extern "C" {
     GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
 
     // Create a backend buffer from an existing pointer
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
 
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+    //
+    // Backend registry
+    //
+
+    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+
+    GGML_API size_t                     ggml_backend_reg_get_count(void);
+    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
+    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
+    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
+    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
 
     //
     // Backend scheduler
@@ -131,6 +150,32 @@ extern "C" {
             ggml_backend_sched_t sched,
             struct ggml_cgraph * graph);
 
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+
 #ifdef  __cplusplus
 }
 #endif

ggml-cuda.h

@@ -49,7 +49,15 @@ GGML_API int    ggml_cuda_get_device_count(void);
 GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
+GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
 #ifdef  __cplusplus
 }

ggml-impl.h

@@ -232,7 +232,7 @@ bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml
 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
 size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
-// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // return index, asserts if table is full

ggml-metal.h

@@ -99,6 +99,12 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+// helper to check if the device supports a specific family
+// ideally, the user code should be doing these checks
+// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 #ifdef __cplusplus
 }

ggml-opencl.cpp

@@ -1,20 +1,18 @@
+#include "ggml.h"
 #include "ggml-opencl.h"
 
 #include <array>
 #include <atomic>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <limits>
 #include <sstream>
 #include <vector>
-#include <limits>
 
 #define CL_TARGET_OPENCL_VERSION 110
 #include <clblast.h>
 
-#include <stdlib.h>
-#include <stdio.h>
-#include <string.h>
-
-#include "ggml.h"
-
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif

ggml.c

@@ -233,24 +233,6 @@ inline static void * ggml_aligned_malloc(size_t size) {
 #define UNUSED GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
 
-//
-// tensor access macros
-//
-
-#define GGML_TENSOR_UNARY_OP_LOCALS \
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
-
-#define GGML_TENSOR_BINARY_OP_LOCALS \
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
-    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
-
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
 #if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions
@@ -1613,6 +1595,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GROUP_NORM",
 
     "MUL_MAT",
+    "MUL_MAT_ID",
     "OUT_PROD",
 
     "SCALE",
@@ -1640,6 +1623,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "POOL_1D",
     "POOL_2D",
     "UPSCALE",
+    "ARGSORT",
 
     "FLASH_ATTN",
     "FLASH_FF",
@@ -1666,7 +1650,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68");
+static_assert(GGML_OP_COUNT == 70, "GGML_OP_COUNT != 70");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1695,6 +1679,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "group_norm(x)",
 
     "X*Y",
+    "X[i]*Y",
     "X*Y",
 
     "x*v",
@@ -1722,6 +1707,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "pool_1d(x)",
     "pool_2d(x)",
     "upscale(x)",
+    "argsort(x)",
 
     "flash_attn(x)",
     "flash_ff(x)",
@@ -1748,10 +1734,28 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68");
+static_assert(GGML_OP_COUNT == 70, "GGML_OP_COUNT != 70");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
+
+static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
+    "ABS",
+    "SGN",
+    "NEG",
+    "STEP",
+    "TANH",
+    "ELU",
+    "RELU",
+    "GELU",
+    "GELU_QUICK",
+    "SILU",
+    "LEAKY",
+};
+
+static_assert(GGML_UNARY_OP_COUNT == 11, "GGML_UNARY_OP_COUNT != 11");
+
+
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
 
@@ -1771,6 +1775,7 @@ static void ggml_setup_op_has_task_pass(void) {
 
         p[GGML_OP_ACC                    ] = true;
         p[GGML_OP_MUL_MAT                ] = true;
+        p[GGML_OP_MUL_MAT_ID             ] = true;
         p[GGML_OP_OUT_PROD               ] = true;
         p[GGML_OP_SET                    ] = true;
         p[GGML_OP_GET_ROWS_BACK          ] = true;
@@ -2023,6 +2028,20 @@ const char * ggml_op_symbol(enum ggml_op op) {
     return GGML_OP_SYMBOL[op];
 }
 
+const char * ggml_unary_op_name(enum ggml_unary_op op) {
+    return GGML_UNARY_OP_NAME[op];
+}
+
+const char * ggml_op_desc(const struct ggml_tensor * t) {
+    if (t->op == GGML_OP_UNARY) {
+        enum ggml_unary_op uop = ggml_get_unary_op(t);
+        return ggml_unary_op_name(uop);
+    }
+    else {
+        return ggml_op_name(t->op);
+    }
+}
+
 size_t ggml_element_size(const struct ggml_tensor * tensor) {
     return ggml_type_size(tensor->type);
 }
@@ -3154,9 +3173,7 @@ static struct ggml_tensor * ggml_add_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    // TODO: support less-strict constraint
-    //       GGML_ASSERT(ggml_can_repeat(b, a));
-    GGML_ASSERT(ggml_can_repeat_rows(b, a));
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3371,9 +3388,7 @@ static struct ggml_tensor * ggml_mul_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    // TODO: support less-strict constraint
-    //       GGML_ASSERT(ggml_can_repeat(b, a));
-    GGML_ASSERT(ggml_can_repeat_rows(b, a));
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3418,7 +3433,7 @@ static struct ggml_tensor * ggml_div_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    GGML_ASSERT(ggml_are_same_shape(a, b));
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -4056,6 +4071,49 @@ struct ggml_tensor * ggml_mul_mat(
     return result;
 }
 
+// ggml_mul_mat_id
+
+struct ggml_tensor * ggml_mul_mat_id(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * as[],
+        struct ggml_tensor  * ids,
+        int                   id,
+        struct ggml_tensor  * b) {
+
+    int64_t n_as = ids->ne[0];
+
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_is_vector(ids));
+    GGML_ASSERT(n_as > 0 && n_as <= GGML_MAX_SRC - 2);
+    GGML_ASSERT(id >= 0 && id < n_as);
+
+    bool is_node = false;
+
+    if (as[0]->grad || b->grad) {
+        is_node = true;
+    }
+
+    const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(as[0]->n_dims, b->n_dims), ne);
+
+    ggml_set_op_params_i32(result, 0, id);
+
+    result->op   = GGML_OP_MUL_MAT_ID;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = ids;
+    result->src[1] = b;
+
+    for (int64_t i = 0; i < n_as; i++) {
+        struct ggml_tensor * a = as[i];
+        GGML_ASSERT(ggml_are_same_shape(as[0], a));
+        GGML_ASSERT(ggml_can_mul_mat(a, b));
+        GGML_ASSERT(!ggml_is_transposed(a));
+        result->src[i + 2] = a;
+    }
+
+    return result;
+}
+
 // ggml_out_prod
 
 struct ggml_tensor * ggml_out_prod(
@@ -4209,7 +4267,7 @@ struct ggml_tensor * ggml_set_2d_inplace(
         struct ggml_tensor *  b,
         size_t                nb1,
         size_t                offset) {
-    return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false);
+    return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, true);
 }
 
 // ggml_cpy
@@ -4826,7 +4884,17 @@ struct ggml_tensor * ggml_diag_mask_zero_inplace(
 static struct ggml_tensor * ggml_soft_max_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
+        struct ggml_tensor  * mask,
+        float                 scale,
         bool                  inplace) {
+    GGML_ASSERT(ggml_is_contiguous(a));
+    if (mask) {
+        GGML_ASSERT(ggml_is_contiguous(mask));
+        GGML_ASSERT(mask->ne[2] == 1);
+        GGML_ASSERT(mask->ne[3] == 1);
+        GGML_ASSERT(ggml_can_repeat_rows(mask, a));
+    }
+
     bool is_node = false;
 
     if (a->grad) {
@@ -4835,9 +4903,13 @@ static struct ggml_tensor * ggml_soft_max_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
+    float params[] = { scale };
+    ggml_set_op_params(result, params, sizeof(params));
+
     result->op   = GGML_OP_SOFT_MAX;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
+    result->src[1] = mask;
 
     return result;
 }
@@ -4845,13 +4917,21 @@ static struct ggml_tensor * ggml_soft_max_impl(
 struct ggml_tensor * ggml_soft_max(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, false);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, false);
 }
 
 struct ggml_tensor * ggml_soft_max_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, true);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, true);
+}
+
+struct ggml_tensor * ggml_soft_max_ext(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * mask,
+        float                 scale) {
+    return ggml_soft_max_impl(ctx, a, mask, scale, false);
 }
 
 // ggml_soft_max_back
@@ -5446,6 +5526,43 @@ struct ggml_tensor * ggml_upscale(
     return ggml_upscale_impl(ctx, a, scale_factor);
 }
 
+// ggml_argsort
+
+struct ggml_tensor * ggml_argsort(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_sort_order  order) {
+    bool is_node = false;
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, a->n_dims, a->ne);
+
+    ggml_set_op_params_i32(result, 0, (int32_t) order);
+
+    result->op   = GGML_OP_ARGSORT;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_top_k
+
+struct ggml_tensor * ggml_top_k(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   k) {
+    GGML_ASSERT(a->ne[0] >= k);
+
+    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_DESC);
+
+    result = ggml_view_4d(ctx, result,
+                k, result->ne[1], result->ne[2], result->ne[3],
+                   result->nb[1], result->nb[2], result->nb[3],
+                0);
+
+    return result;
+}
+
 // ggml_flash_attn
 
 struct ggml_tensor * ggml_flash_attn(
@@ -6805,7 +6922,7 @@ static void ggml_compute_forward_add_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_can_repeat_rows(src1, src0) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -6838,16 +6955,19 @@ static void ggml_compute_forward_add_f32(
             const int64_t i13 = i03 % ne13;
             const int64_t i12 = i02 % ne12;
             const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
             float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            vDSP_vadd(src0_ptr, 1, src1_ptr, 1, dst_ptr, 1, ne00);
+                vDSP_vadd(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_add_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
+                ggml_vec_add_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
+            }
         }
     } else {
         // src1 is not contiguous
@@ -6864,8 +6984,9 @@ static void ggml_compute_forward_add_f32(
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] + *src1_ptr;
             }
@@ -7585,7 +7706,7 @@ static void ggml_compute_forward_mul_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_can_repeat_rows(src1, src0) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -7608,7 +7729,6 @@ static void ggml_compute_forward_mul_f32(
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
-    GGML_ASSERT(ne00 == ne10);
 
     if (nb10 == sizeof(float)) {
         for (int64_t ir = ith; ir < nr; ir += nth) {
@@ -7620,20 +7740,21 @@ static void ggml_compute_forward_mul_f32(
             const int64_t i13 = i03 % ne13;
             const int64_t i12 = i02 % ne12;
             const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
             float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0 ; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            UNUSED(ggml_vec_mul_f32);
+                UNUSED(ggml_vec_mul_f32);
 
-            vDSP_vmul( src0_ptr, 1, src1_ptr, 1, dst_ptr,  1, ne00);
+                vDSP_vmul(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_mul_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
+                ggml_vec_mul_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
@@ -7651,8 +7772,9 @@ static void ggml_compute_forward_mul_f32(
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            for (int64_t i0 = 0; i0 < ne00; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne00; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr);
             }
@@ -7686,14 +7808,16 @@ static void ggml_compute_forward_div_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
-    const int nr  = ggml_nrows(src0);
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -7701,41 +7825,50 @@ static void ggml_compute_forward_div_f32(
     GGML_ASSERT(nb00 == sizeof(float));
 
     if (nb10 == sizeof(float)) {
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+            for (int64_t r = 0; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            UNUSED(ggml_vec_div_f32);
+                UNUSED(ggml_vec_div_f32);
 
-            vDSP_vdiv(
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ), 1,
-                    ne0);
+                vDSP_vdiv(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_div_f32(ne0,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+                ggml_vec_div_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-            float * dst_ptr  = (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
-            float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+
+            for (int64_t i0 = 0; i0 < ne00; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr);
             }
@@ -8181,7 +8314,7 @@ static void ggml_compute_forward_repeat_f16(
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
     const int nr0 = (int)(ne0/ne00);
@@ -8326,6 +8459,7 @@ static void ggml_compute_forward_concat_f32(
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
     const int ith = params->ith;
+    const int nth = params->nth;
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -8335,7 +8469,7 @@ static void ggml_compute_forward_concat_f32(
     GGML_ASSERT(nb10 == sizeof(float));
 
     for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2++) {
+        for (int i2 = ith; i2 < ne2; i2 += nth) {
             if (i2 < ne02) { // src0
                 for (int i1 = 0; i1 < ne1; i1++) {
                     for (int i0 = 0; i0 < ne0; i0++) {
@@ -9495,6 +9629,8 @@ static void ggml_compute_forward_mul_mat(
             char * wdata = params->wdata;
             const size_t row_size = ne10*ggml_type_size(vec_dot_type)/ggml_blck_size(vec_dot_type);
 
+            assert(params->wsize >= ne11*ne12*ne13*row_size);
+
             for (int64_t i13 = 0; i13 < ne13; ++i13) {
                 for (int64_t i12 = 0; i12 < ne12; ++i12) {
                     for (int64_t i11 = 0; i11 < ne11; ++i11) {
@@ -9596,6 +9732,26 @@ static void ggml_compute_forward_mul_mat(
     }
 }
 
+// ggml_compute_forward_mul_mat_id
+
+static void ggml_compute_forward_mul_mat_id(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * ids = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    const int id = ggml_get_op_params_i32(dst, 0);
+
+    const int a_id = ((int32_t *)ids->data)[id];
+
+    GGML_ASSERT(a_id >= 0 && a_id < ids->ne[0]);
+
+    const struct ggml_tensor * src0 = dst->src[a_id + 2];
+
+    ggml_compute_forward_mul_mat(params, src0, src1, dst);
+}
+
 // ggml_compute_forward_out_prod
 
 static void ggml_compute_forward_out_prod_f32(
@@ -10551,20 +10707,25 @@ static void ggml_compute_forward_diag_mask_zero(
 static void ggml_compute_forward_soft_max_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    assert(ggml_is_contiguous(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
+    float scale = 1.0f;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+
     // TODO: handle transposed/permuted matrices
 
     const int ith = params->ith;
     const int nth = params->nth;
 
+    const int64_t ne11 = src1 ? src1->ne[1] : 1;
+
     const int nc = src0->ne[0];
     const int nr = ggml_nrows(src0);
 
@@ -10575,29 +10736,40 @@ static void ggml_compute_forward_soft_max_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
+    float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
+
     for (int i1 = ir0; i1 < ir1; i1++) {
-        float *sp = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float *dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+        float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
+        float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+
+        // broadcast the mask across rows
+        float * mp = src1 ? (float *)((char *) src1->data + (i1%ne11)*src1->nb[1]) : NULL;
+
+        ggml_vec_cpy_f32  (nc, wp, sp);
+        ggml_vec_scale_f32(nc, wp, scale);
+        if (mp) {
+            ggml_vec_acc_f32(nc, wp, mp);
+        }
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
             //printf("p[%d] = %f\n", i, p[i]);
-            assert(!isnan(sp[i]));
+            assert(!isnan(wp[i]));
         }
 #endif
 
         float max = -INFINITY;
-        ggml_vec_max_f32(nc, &max, sp);
+        ggml_vec_max_f32(nc, &max, wp);
 
         ggml_float sum = 0.0;
 
         uint16_t scvt;
         for (int i = 0; i < nc; i++) {
-            if (sp[i] == -INFINITY) {
+            if (wp[i] == -INFINITY) {
                 dp[i] = 0.0f;
             } else {
-                // const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max);
-                ggml_fp16_t s = GGML_FP32_TO_FP16(sp[i] - max);
+                // const float val = (wp[i] == -INFINITY) ? 0.0 : exp(wp[i] - max);
+                ggml_fp16_t s = GGML_FP32_TO_FP16(wp[i] - max);
                 memcpy(&scvt, &s, sizeof(scvt));
                 const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
                 sum += (ggml_float)val;
@@ -10622,11 +10794,12 @@ static void ggml_compute_forward_soft_max_f32(
 static void ggml_compute_forward_soft_max(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_soft_max_f32(params, src0, dst);
+                ggml_compute_forward_soft_max_f32(params, src0, src1, dst);
             } break;
         default:
             {
@@ -11982,6 +12155,67 @@ static void ggml_compute_forward_upscale(
     }
 }
 
+// ggml_compute_forward_argsort
+
+static void ggml_compute_forward_argsort_f32(
+    const struct ggml_compute_params * params,
+    const struct ggml_tensor * src0,
+    struct ggml_tensor * dst) {
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) ggml_get_op_params_i32(dst, 0);
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
+
+        for (int64_t j = 0; j < ne0; j++) {
+            dst_data[j] = j;
+        }
+
+        // C doesn't have a functional sort, so we do a bubble sort instead
+        for (int64_t j = 0; j < ne0; j++) {
+            for (int64_t k = j + 1; k < ne0; k++) {
+                if ((order == GGML_SORT_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
+                    (order == GGML_SORT_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
+                    int32_t tmp = dst_data[j];
+                    dst_data[j] = dst_data[k];
+                    dst_data[k] = tmp;
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_argsort(
+    const struct ggml_compute_params * params,
+    const struct ggml_tensor * src0,
+    struct ggml_tensor * dst) {
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_argsort_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_flash_attn
 
 static void ggml_compute_forward_flash_attn_f32(
@@ -13805,6 +14039,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
             } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                ggml_compute_forward_mul_mat_id(params, tensor);
+            } break;
         case GGML_OP_OUT_PROD:
             {
                 ggml_compute_forward_out_prod(params, tensor->src[0], tensor->src[1], tensor);
@@ -13863,7 +14101,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                ggml_compute_forward_soft_max(params, tensor->src[0], tensor);
+                ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
@@ -13909,6 +14147,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_upscale(params, tensor->src[0], tensor);
             } break;
+        case GGML_OP_ARGSORT:
+            {
+                ggml_compute_forward_argsort(params, tensor->src[0], tensor);
+            } break;
         case GGML_OP_FLASH_ATTN:
             {
                 const int32_t t = ggml_get_op_params_i32(tensor, 0);
@@ -14559,6 +14801,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 zero_table);
                 }
             } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_OUT_PROD:
             {
                 GGML_ASSERT(false); // TODO: not implemented
@@ -14897,6 +15143,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_ARGSORT:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_FLASH_ATTN:
             {
                 struct ggml_tensor * flash_grad = NULL;
@@ -15257,12 +15507,8 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
     return ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, false);
 }
 
-struct ggml_cgraph * ggml_graph_view(struct ggml_context * ctx, struct ggml_cgraph * cgraph0, int i0, int i1) {
-    const size_t obj_size = sizeof(struct ggml_cgraph);
-    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size);
-    struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
-
-    *cgraph = (struct ggml_cgraph) {
+struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) {
+    struct ggml_cgraph cgraph = {
         /*.size         =*/ 0,
         /*.n_nodes      =*/ i1 - i0,
         /*.n_leafs      =*/ 0,
@@ -15497,7 +15743,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_SUB:
-        case GGML_OP_DIV:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_LOG:
@@ -15530,10 +15775,13 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     {
                         n_tasks = n_threads;
                     } break;
+                default:
+                    GGML_ASSERT(false);
             }
             break;
         case GGML_OP_SILU_BACK:
         case GGML_OP_MUL:
+        case GGML_OP_DIV:
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
@@ -15571,6 +15819,11 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 }
 #endif
             } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                // FIXME: blas
+                n_tasks = n_threads;
+            } break;
         case GGML_OP_OUT_PROD:
             {
                 n_tasks = n_threads;
@@ -15590,7 +15843,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_DIAG_MASK_ZERO:
         case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK:
@@ -15606,6 +15858,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = 1; //TODO
             } break;
+        case GGML_OP_SOFT_MAX:
+            {
+                n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
+            } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
                 n_tasks = n_threads;
@@ -15627,6 +15883,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = n_threads;
             } break;
+        case GGML_OP_ARGSORT:
+            {
+                n_tasks = n_threads;
+            } break;
         case GGML_OP_FLASH_ATTN:
             {
                 n_tasks = n_threads;
@@ -15689,6 +15949,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = 1;
             } break;
+        case GGML_OP_COUNT:
+            {
+                GGML_ASSERT(false);
+            } break;
         default:
             {
                 fprintf(stderr, "%s: op not implemented: ", __func__);
@@ -15837,18 +16101,16 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
 
     // thread scheduling for the different operations + work buffer size estimation
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        int n_tasks = 1;
-
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+
         size_t cur = 0;
 
         switch (node->op) {
             case GGML_OP_CPY:
             case GGML_OP_DUP:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                     }
@@ -15856,16 +16118,12 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
             case GGML_OP_ADD:
             case GGML_OP_ADD1:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->src[0]->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
             case GGML_OP_ACC:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->src[0]->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
                     }
@@ -15891,14 +16149,33 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         cur = ggml_type_size(vec_dot_type)*ggml_nelements(node->src[1])/ggml_blck_size(vec_dot_type);
                     }
                 } break;
+            case GGML_OP_MUL_MAT_ID:
+                {
+                    const struct ggml_tensor * a = node->src[2];
+                    const struct ggml_tensor * b = node->src[1];
+                    const enum ggml_type vec_dot_type = type_traits[a->type].vec_dot_type;
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
+                    if (ggml_compute_forward_mul_mat_use_blas(a, b, node)) {
+                        if (a->type != GGML_TYPE_F32) {
+                            // here we need memory just for single 2D matrix from src0
+                            cur = ggml_type_size(GGML_TYPE_F32)*(a->ne[0]*a->ne[1]);
+                        }
+                    } else
+#endif
+                    if (b->type != vec_dot_type) {
+                        cur = ggml_type_size(vec_dot_type)*ggml_nelements(b)/ggml_blck_size(vec_dot_type);
+                    }
+                } break;
             case GGML_OP_OUT_PROD:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->src[0]->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
+            case GGML_OP_SOFT_MAX:
+                {
+                    cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
+                } break;
             case GGML_OP_CONV_TRANSPOSE_1D:
                 {
                     GGML_ASSERT(node->src[0]->ne[3] == 1);
@@ -15924,10 +16201,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         GGML_ASSERT(false);
                     }
                 } break;
-            case GGML_OP_IM2COL:
-                {
-                    n_tasks = n_threads;
-                } break;
             case GGML_OP_CONV_TRANSPOSE_2D:
                 {
                     const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -15944,8 +16217,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_FLASH_ATTN:
                 {
-                    n_tasks = n_threads;
-
                     const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
 
                     if (node->src[1]->type == GGML_TYPE_F32) {
@@ -15958,8 +16229,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_FLASH_FF:
                 {
-                    n_tasks = n_threads;
-
                     if (node->src[1]->type == GGML_TYPE_F32) {
                         cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
                         cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
@@ -15970,8 +16239,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_FLASH_ATTN_BACK:
                 {
-                    n_tasks = n_threads;
-
                     const int64_t    D = node->src[0]->ne[0];
                     const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
                     const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
@@ -15986,8 +16253,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
 
             case GGML_OP_CROSS_ENTROPY_LOSS:
                 {
-                    n_tasks = n_threads;
-
                     cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
                 } break;
             case GGML_OP_COUNT:
@@ -17774,8 +18039,8 @@ size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int j = 0; j < QK5_0; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12));
+                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
+                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
@@ -17804,8 +18069,8 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int j = 0; j < QK5_1; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12));
+                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
+                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
@@ -17995,6 +18260,7 @@ struct gguf_kv {
 
 struct gguf_header {
     char magic[4];
+
     uint32_t version;
     uint64_t n_tensors; // GGUFv2
     uint64_t n_kv;      // GGUFv2
@@ -18084,7 +18350,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
         for (uint32_t i = 0; i < sizeof(magic); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                fprintf(stderr, "%s: invalid magic characters %s.\n", __func__, magic);
+                fprintf(stderr, "%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                 fclose(file);
                 return NULL;
             }
@@ -18099,7 +18365,6 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     {
         strncpy(ctx->header.magic, magic, 4);
 
-
         ctx->kv    = NULL;
         ctx->infos = NULL;
         ctx->data  = NULL;

ggml.h

@@ -215,7 +215,7 @@
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
 #define GGML_MAX_DIMS           4
-#define GGML_MAX_PARAMS         1024
+#define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            6
 #define GGML_MAX_NAME           64
@@ -283,6 +283,20 @@
     const type prefix##3 = (pointer)->array[3]; \
     GGML_UNUSED(prefix##3);
 
+#define GGML_TENSOR_UNARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
+#define GGML_TENSOR_BINARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -381,6 +395,7 @@ extern "C" {
         GGML_OP_GROUP_NORM,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_MUL_MAT_ID,
         GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
@@ -407,8 +422,8 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
-
         GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_ARGSORT,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
@@ -448,7 +463,9 @@ extern "C" {
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
-        GGML_UNARY_OP_LEAKY
+        GGML_UNARY_OP_LEAKY,
+
+        GGML_UNARY_OP_COUNT,
     };
 
     enum ggml_object_type {
@@ -631,6 +648,9 @@ extern "C" {
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
+    GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
+
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
     GGML_API bool    ggml_is_quantized(enum ggml_type type);
@@ -1027,6 +1047,15 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // indirect matrix multiplication
+    //  ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
+    GGML_API struct ggml_tensor * ggml_mul_mat_id(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * as[],
+            struct ggml_tensor  * ids,
+            int                   id,
+            struct ggml_tensor  * b);
+
     // A: m columns, n rows,
     // B: p columns, n rows,
     // result is m columns, p rows
@@ -1282,6 +1311,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // fused soft_max(a*scale + mask)
+    // mask is optional
+    GGML_API struct ggml_tensor * ggml_soft_max_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * mask,
+            float                 scale);
+
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1512,6 +1549,23 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   scale_factor);
 
+    // sort rows
+    enum ggml_sort_order {
+        GGML_SORT_ASC,
+        GGML_SORT_DESC,
+    };
+
+    GGML_API struct ggml_tensor * ggml_argsort(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_sort_order  order);
+
+    // top k elements per row
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1573,7 +1627,6 @@ extern "C" {
             int                   kh);
 
     // used in sam
-
     GGML_API struct ggml_tensor * ggml_add_rel_pos(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1748,7 +1801,7 @@ extern "C" {
     GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
     GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
-    GGML_API struct ggml_cgraph * ggml_graph_view        (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1);
+    GGML_API struct ggml_cgraph   ggml_graph_view        (struct ggml_cgraph * cgraph, int i0, int i1);
     GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
     GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);

gguf-py/gguf/constants.py

@@ -92,6 +92,7 @@ class MODEL_ARCH(IntEnum):
     BERT      = auto()
     BLOOM     = auto()
     STABLELM  = auto()
+    QWEN      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -132,6 +133,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BERT:           "bert",
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
+    MODEL_ARCH.QWEN:           "qwen",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -317,6 +319,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.QWEN: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.GPT2: [
         # TODO
     ],
@@ -336,6 +352,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     MODEL_ARCH.PERSIMMON: [
         MODEL_TENSOR.ROPE_FREQS,
     ],
+    MODEL_ARCH.QWEN: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #

gguf-py/gguf/tensor_mapping.py

@@ -10,7 +10,7 @@ class TensorNameMap:
         # Token embeddings
         MODEL_TENSOR.TOKEN_EMBD: (
             "gpt_neox.embed_in",                         # gptneox
-            "transformer.wte",                           # gpt2 gpt-j mpt refact
+            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
             "model.embed_tokens",                        # llama-hf
@@ -38,7 +38,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen
             "output",                    # llama-pth bloom
             "word_embeddings_for_head",  # persimmon
         ),
@@ -51,7 +51,7 @@ class TensorNameMap:
             "norm",                                    # llama-pth
             "embeddings.LayerNorm",                    # bert
             "transformer.norm_f",                      # mpt
-            "ln_f",                                    # refact bloom
+            "ln_f",                                    # refact bloom qwen
             "language_model.encoder.final_layernorm",  # persimmon
         ),
 
@@ -65,7 +65,7 @@ class TensorNameMap:
         # Attention norm
         MODEL_TENSOR.ATTN_NORM: (
             "gpt_neox.layers.{bid}.input_layernorm",                # gptneox
-            "transformer.h.{bid}.ln_1",                             # gpt2 gpt-j refact
+            "transformer.h.{bid}.ln_1",                             # gpt2 gpt-j refact qwen
             "transformer.blocks.{bid}.norm_1",                      # mpt
             "transformer.h.{bid}.input_layernorm",                  # falcon7b
             "h.{bid}.input_layernorm",                              # bloom
@@ -85,7 +85,7 @@ class TensorNameMap:
         # Attention query-key-value
         MODEL_TENSOR.ATTN_QKV: (
             "gpt_neox.layers.{bid}.attention.query_key_value",                     # gptneox
-            "transformer.h.{bid}.attn.c_attn",                                     # gpt2
+            "transformer.h.{bid}.attn.c_attn",                                     # gpt2 qwen
             "transformer.blocks.{bid}.attn.Wqkv",                                  # mpt
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
@@ -119,7 +119,7 @@ class TensorNameMap:
         # Attention output
         MODEL_TENSOR.ATTN_OUT: (
             "gpt_neox.layers.{bid}.attention.dense",                     # gptneox
-            "transformer.h.{bid}.attn.c_proj",                           # gpt2 refact
+            "transformer.h.{bid}.attn.c_proj",                           # gpt2 refact qwen
             "transformer.blocks.{bid}.attn.out_proj",                    # mpt
             "transformer.h.{bid}.self_attention.dense",                  # falcon
             "h.{bid}.self_attention.dense",                              # bloom
@@ -139,7 +139,7 @@ class TensorNameMap:
         # Feed-forward norm
         MODEL_TENSOR.FFN_NORM: (
             "gpt_neox.layers.{bid}.post_attention_layernorm",                # gptneox
-            "transformer.h.{bid}.ln_2",                                      # gpt2 refact
+            "transformer.h.{bid}.ln_2",                                      # gpt2 refact qwen
             "h.{bid}.post_attention_layernorm",                              # bloom
             "transformer.blocks.{bid}.norm_2",                               # mpt
             "model.layers.{bid}.post_attention_layernorm",                   # llama-hf
@@ -161,18 +161,20 @@ class TensorNameMap:
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
+            "transformer.h.{bid}.mlp.w1",                             # qwen
         ),
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
             "model.layers.{bid}.mlp.gate_proj",  # llama-hf refact
             "layers.{bid}.feed_forward.w1",      # llama-pth
+            "transformer.h.{bid}.mlp.w2",        # qwen
         ),
 
         # Feed-forward down
         MODEL_TENSOR.FFN_DOWN: (
             "gpt_neox.layers.{bid}.mlp.dense_4h_to_h",                # gptneox
-            "transformer.h.{bid}.mlp.c_proj",                         # gpt2 refact
+            "transformer.h.{bid}.mlp.c_proj",                         # gpt2 refact qwen
             "transformer.blocks.{bid}.ffn.down_proj",                 # mpt
             "transformer.h.{bid}.mlp.dense_4h_to_h",                  # falcon
             "h.{bid}.mlp.dense_4h_to_h",                              # bloom

llama.h

@@ -42,7 +42,7 @@
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 2
+#define LLAMA_SESSION_VERSION 3
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@@ -158,6 +158,22 @@ extern "C" {
         llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
+    enum llama_model_kv_override_type {
+        LLAMA_KV_OVERRIDE_INT,
+        LLAMA_KV_OVERRIDE_FLOAT,
+        LLAMA_KV_OVERRIDE_BOOL,
+    };
+
+    struct llama_model_kv_override {
+        char key[128];
+        enum llama_model_kv_override_type tag;
+        union {
+            int64_t int_value;
+            double float_value;
+            bool bool_value;
+        };
+    };
+
     struct llama_model_params {
         int32_t n_gpu_layers; // number of layers to store in VRAM
         int32_t main_gpu;     // the GPU that is used for scratch and small tensors
@@ -165,9 +181,13 @@ extern "C" {
 
         // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
+
         // context pointer passed to the progress callback
         void * progress_callback_user_data;
 
+        // override key-value pairs of the model meta data
+        const struct llama_model_kv_override * kv_overrides;
+
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
@@ -191,11 +211,14 @@ extern "C" {
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
 
+        enum ggml_type type_k; // data type for K cache
+        enum ggml_type type_v; // data type for V cache
+
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
-        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
-        bool embedding;  // embedding mode only
+        bool mul_mat_q;   // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
+        bool logits_all;  // the llama_eval() call computes all logits, not just the last one
+        bool embedding;   // embedding mode only
+        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
     };
 
     // model quantization parameters

prompts/chat-with-qwen.txt

@@ -0,0 +1 @@
+You are a helpful assistant.

requirements-hf-to-gguf.txt

@@ -0,0 +1,3 @@
+-r requirements.txt
+torch==2.1.1
+transformers==4.35.2

scripts/sync-ggml.sh

@@ -20,5 +20,6 @@ cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
 cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h
 cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h
 
-cp -rpv ../ggml/tests/test-opt.cpp    ./tests/test-opt.cpp
-cp -rpv ../ggml/tests/test-grad0.cpp  ./tests/test-grad0.cpp
+cp -rpv ../ggml/tests/test-opt.cpp         ./tests/test-opt.cpp
+cp -rpv ../ggml/tests/test-grad0.cpp       ./tests/test-grad0.cpp
+cp -rpv ../ggml/tests/test-backend-ops.cpp ./tests/test-backend-ops.cpp

tests/CMakeLists.txt

@@ -22,26 +22,32 @@ endfunction()
 llama_build_and_test_executable(test-quantize-fns.cpp)
 llama_build_and_test_executable(test-quantize-perf.cpp)
 llama_build_and_test_executable(test-sampling.cpp)
+
 llama_build_executable(test-tokenizer-0-llama.cpp)
 llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+
 llama_build_executable(test-tokenizer-0-falcon.cpp)
 llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+
 llama_build_executable(test-tokenizer-1-llama.cpp)
-llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
-llama_test_executable(test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+llama_test_executable (test-tokenizer-1-llama    test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test_executable (test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+
 llama_build_executable(test-tokenizer-1-bpe.cpp)
-llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
-llama_test_executable(test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-llama_test_executable(test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf)
-llama_test_executable(test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
-llama_test_executable(test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-llama_test_executable(test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
-# llama_test_executable(test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG
+llama_test_executable (test-tokenizer-1-falcon           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_test_executable (test-tokenizer-1-aquila           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
+llama_test_executable (test-tokenizer-1-mpt              test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+llama_test_executable (test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf)
+llama_test_executable (test-tokenizer-1-gpt-neox         test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
+llama_test_executable (test-tokenizer-1-refact           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+llama_test_executable (test-tokenizer-1-starcoder        test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+# llama_test_executable (test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG
+
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
-llama_build_and_test_executable(test-grad0.cpp) # SLOW
+llama_build_and_test_executable(test-grad0.cpp)
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
+llama_build_and_test_executable(test-backend-ops.cpp)
 
 llama_build_and_test_executable(test-rope.cpp)