metal : switch to execution barriers + fix one of the barriers

ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92

.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

@@ -101,3 +101,4 @@ poetry.toml
 /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)
@@ -662,11 +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-rope
+	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
@@ -395,6 +396,11 @@ ifdef LLAMA_CUBLAS
 	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS         += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+
+ifdef LLAMA_DEBUG
+	NVCCFLAGS += -lineinfo
+endif
+
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
@@ -746,3 +752,6 @@ tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
 
 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)

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

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/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

@@ -77,8 +77,18 @@ class Model:
             self.gguf_writer.add_embedding_length(n_embd)
         if (n_ff := self.hparams.get("intermediate_size")) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_head")) is not None:
+        if (n_head := self.hparams.get("num_attention_heads")) is not None:
             self.gguf_writer.add_head_count(n_head)
+        if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
+            self.gguf_writer.add_head_count_kv(n_head_kv)
+
+        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        if (n_experts := self.hparams.get("num_local_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+
         self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
 
     def write_tensors(self):
@@ -170,6 +180,8 @@ class Model:
             return StableLMModel
         if model_architecture == "QWenLMHeadModel":
             return QwenModel
+        if model_architecture == "MixtralForCausalLM":
+            return MixtralModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -207,6 +219,8 @@ class Model:
             return gguf.MODEL_ARCH.STABLELM
         if arch == "QWenLMHeadModel":
             return gguf.MODEL_ARCH.QWEN
+        if arch == "MixtralForCausalLM":
+            return gguf.MODEL_ARCH.LLAMA
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -837,6 +851,11 @@ class StableLMModel(Model):
         self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
+class MixtralModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+
 class QwenModel(Model):
     @staticmethod
     def token_bytes_to_string(b):

convert.py

@@ -42,6 +42,7 @@ NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
+
 #
 # data types
 #
@@ -62,10 +63,10 @@ class UnquantizedDataType(DataType):
     pass
 
 
-DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
-DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
+DT_F16  = UnquantizedDataType('F16',  dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32',  dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32',  dtype = np.dtype(np.int16),   valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16),  valid_conversions = ['F32', 'F16', 'Q8_0'])
 
 
 @dataclass(frozen=True)
@@ -151,14 +152,16 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
 
 @dataclass
 class Params:
-    n_vocab:    int
-    n_embd:     int
-    n_layer:    int
-    n_ctx:      int
-    n_ff:       int
-    n_head:     int
-    n_head_kv:  int
-    f_norm_eps: float
+    n_vocab:        int
+    n_embd:         int
+    n_layer:        int
+    n_ctx:          int
+    n_ff:           int
+    n_head:         int
+    n_head_kv:      int
+    n_experts:      int | None = None
+    n_experts_used: int | None = None
+    f_norm_eps:     float | None = None
 
     rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
@@ -233,6 +236,13 @@ class Params:
             raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
+        n_experts      = None
+        n_experts_used = None
+
+        if "num_local_experts" in config:
+            n_experts = config["num_local_experts"]
+            n_experts_used = config["num_experts_per_tok"]
+
         return Params(
             n_vocab           = config["vocab_size"],
             n_embd            = config["hidden_size"],
@@ -241,6 +251,8 @@ class Params:
             n_ff              = config["intermediate_size"],
             n_head            = (n_head := config["num_attention_heads"]),
             n_head_kv         = config.get("num_key_value_heads", n_head),
+            n_experts         = n_experts,
+            n_experts_used    = n_experts_used,
             f_norm_eps        = config["rms_norm_eps"],
             f_rope_freq_base  = config.get("rope_theta"),
             rope_scaling_type = rope_scaling_type,
@@ -255,8 +267,15 @@ class Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
+        n_experts      = None
+        n_experts_used = None
+        f_rope_freq_base = None
+
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if config.get("rope_theta") == 1000000:
+        if config.get("moe"):
+            # Mixtral
+            n_ctx = 32768
+        elif config.get("rope_theta") == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -266,16 +285,27 @@ class Params:
             # LLaMA v1
             n_ctx = 2048
 
+        if "layers.0.feed_forward.w1.weight" in model:
+            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
+
+        if config.get("moe"):
+            n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
+            n_experts      = config["moe"]["num_experts"]
+            n_experts_used = config["moe"]["num_experts_per_tok"]
+            f_rope_freq_base = 1e6
+
         return Params(
             n_vocab          = model["tok_embeddings.weight"].shape[0],
             n_embd           = config["dim"],
             n_layer          = config["n_layers"],
             n_ctx            = n_ctx,
-            n_ff             = model["layers.0.feed_forward.w1.weight"].shape[0],
+            n_ff             = n_ff,
             n_head           = (n_head := config["n_heads"]),
             n_head_kv        = config.get("n_kv_heads", n_head),
+            n_experts        = n_experts,
+            n_experts_used   = n_experts_used,
             f_norm_eps       = config["norm_eps"],
-            f_rope_freq_base = config.get("rope_theta"),
+            f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
         )
 
     @staticmethod
@@ -832,7 +862,17 @@ class OutputFile:
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
-        self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
+
+        if params.n_experts:
+            self.gguf.add_expert_count(params.n_experts)
+
+        if params.n_experts_used:
+            self.gguf.add_expert_used_count(params.n_experts_used)
+
+        if params.f_norm_eps:
+            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
+        else:
+            raise ValueError('f_norm_eps is None')
 
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
@@ -956,7 +996,7 @@ class OutputFile:
 
 
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
-    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) +".weight"].data_type
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
     if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32

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/main/main.cpp

@@ -437,6 +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/server.cpp

@@ -2108,10 +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)
@@ -2387,6 +2383,7 @@ json oaicompat_completion_params_parse(
 
     // Map OpenAI parameters to llama.cpp parameters
     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);

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

@@ -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.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 +4071,51 @@ 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  * const as[],
+        int                   n_as,
+        struct ggml_tensor  * ids,
+        int                   id,
+        struct ggml_tensor  * b) {
+
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1);
+    GGML_ASSERT(ids->ne[1] == b->ne[1]);
+    GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
+    GGML_ASSERT(n_as > 0 && n_as <= GGML_MAX_SRC - 2);
+    GGML_ASSERT(id >= 0 && id < ids->ne[0]);
+
+    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);
+    ggml_set_op_params_i32(result, 1, n_as);
+
+    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 (int 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 +4269,7 @@ 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
@@ -4673,7 +4733,9 @@ struct ggml_tensor * ggml_get_rows(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b) {
-    GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32);
+    GGML_ASSERT(a->ne[2] == b->ne[1]);
+    GGML_ASSERT(b->ne[3] == 1);
+    GGML_ASSERT(b->type == GGML_TYPE_I32);
 
     bool is_node = false;
 
@@ -4683,7 +4745,7 @@ struct ggml_tensor * ggml_get_rows(
 
     // TODO: implement non F32 return
     //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]);
-    struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, a->ne[0], b->ne[0]);
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
 
     result->op   = GGML_OP_GET_ROWS;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5468,6 +5530,43 @@ 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(
@@ -6827,7 +6926,7 @@ 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;
@@ -6860,16 +6959,19 @@ 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
@@ -6886,8 +6988,9 @@ 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;
             }
@@ -7607,7 +7710,7 @@ 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;
@@ -7630,7 +7733,6 @@ 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) {
@@ -7642,20 +7744,21 @@ 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
@@ -7673,8 +7776,9 @@ 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);
             }
@@ -7708,14 +7812,16 @@ 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
 
@@ -7723,41 +7829,50 @@ 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);
             }
@@ -8203,7 +8318,7 @@ 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);
@@ -8348,6 +8463,7 @@ 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
 
@@ -8357,7 +8473,7 @@ 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++) {
@@ -9392,8 +9508,11 @@ static bool ggml_compute_forward_mul_mat_use_blas(
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
+    // NOTE: with GGML_OP_MUL_MAT_ID we don't want to go through the BLAS branch because it will dequantize (to_float)
+    //       all the experts for each batch element and the processing would become incredibly slow
     // TODO: find the optimal values for these
-    if (ggml_is_contiguous(src0) &&
+    if (dst->op != GGML_OP_MUL_MAT_ID &&
+        ggml_is_contiguous(src0) &&
         ggml_is_contiguous(src1) &&
       //src0->type == GGML_TYPE_F32 &&
         src1->type == GGML_TYPE_F32 &&
@@ -9407,11 +9526,16 @@ static bool ggml_compute_forward_mul_mat_use_blas(
 }
 #endif
 
+// off1 = offset in i11 and i1
+// cne1 = ne11 and ne1
+// in a normal matrix multiplication, off1 = 0 and cne1 = ne1
+// during GGML_TASK_INIT, the full src1 is converted regardless of off1 and cne1
 static void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
+              struct ggml_tensor * dst,
+              int64_t off1, int64_t cne1) {
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -9479,10 +9603,9 @@ static void ggml_compute_forward_mul_mat(
                 const int64_t i03 = i13/r3;
                 const int64_t i02 = i12/r2;
 
-                const void  * x = (char *)            src0->data + i02*nb02 + i03*nb03;
-                const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
-
-                float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                const void  * x = (char *)            src0->data +             i02*nb02 + i03*nb03;
+                const float * y = (float *) ((char *) src1->data + off1*nb11 + i12*nb12 + i13*nb13);
+                      float * d = (float *) ((char *)  dst->data + off1*nb1  + i12*nb2  + i13*nb3);
 
                 if (type != GGML_TYPE_F32) {
                             float * const wdata    = params->wdata;
@@ -9499,10 +9622,10 @@ static void ggml_compute_forward_mul_mat(
                 }
 
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne00,
-                        0.0f,    d, ne01);
+                         cne1, ne01, ne10,
+                         1.0f,    y, ne10,
+                                  x, ne00,
+                         0.0f,    d, ne01);
             }
         }
 
@@ -9517,6 +9640,9 @@ 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);
+            assert(src1->type == GGML_TYPE_F32);
+
             for (int64_t i13 = 0; i13 < ne13; ++i13) {
                 for (int64_t i12 = 0; i12 < ne12; ++i12) {
                     for (int64_t i11 = 0; i11 < ne11; ++i11) {
@@ -9538,7 +9664,7 @@ static void ggml_compute_forward_mul_mat(
     const size_t row_size = ne10*ggml_type_size(vec_dot_type)/ggml_blck_size(vec_dot_type);
 
     const int64_t nr0 = ne01;           // src0 rows
-    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
+    const int64_t nr1 = cne1*ne12*ne13; // src1 rows
 
     //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
@@ -9580,9 +9706,9 @@ static void ggml_compute_forward_mul_mat(
     for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
         for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
             for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                const int64_t i13 = (ir1/(ne12*ne11));
-                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+                const int64_t i13 = (ir1/(ne12*cne1));
+                const int64_t i12 = (ir1 - i13*ne12*cne1)/cne1;
+                const int64_t i11 = (ir1 - i13*ne12*cne1 - i12*cne1) + off1;
 
                 // broadcast src0 into src1
                 const int64_t i03 = i13/r3;
@@ -9618,6 +9744,34 @@ 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,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        // during GGML_TASK_INIT the entire src1 is converted to vec_dot_type
+        ggml_compute_forward_mul_mat(params, dst->src[2], src1, dst, 0, dst->ne[1]);
+        return;
+    }
+
+    const struct ggml_tensor * ids = src0;
+    const int id   = ggml_get_op_params_i32(dst, 0);
+    const int n_as = ggml_get_op_params_i32(dst, 1);
+
+    for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+        const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
+
+        GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+        const struct ggml_tensor * src0_row = dst->src[row_id + 2];
+        ggml_compute_forward_mul_mat(params, src0_row, src1, dst, i01, 1);
+    }
+}
+
 // ggml_compute_forward_out_prod
 
 static void ggml_compute_forward_out_prod_f32(
@@ -10191,21 +10345,30 @@ static void ggml_compute_forward_get_rows_q(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == ggml_type_size(type));
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == ggml_type_size(type));
+    assert(ggml_nrows(dst) == nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
 
-        dequantize_row_q(
-                (const void *) ((char *) src0->data + r*src0->nb[1]),
-                     (float *) ((char *)  dst->data + i*dst->nb[1]), nc);
+                dequantize_row_q(
+                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
+            }
+        }
     }
 }
 
@@ -10220,19 +10383,26 @@ static void ggml_compute_forward_get_rows_f16(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == sizeof(ggml_fp16_t));
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(ggml_fp16_t));
+    assert(ggml_nrows(dst) == nr);
 
-        for (int j = 0; j < nc; ++j) {
-            ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + r*src0->nb[1]))[j];
-            ((float *) ((char *)  dst->data + i*dst->nb[1]))[j] = GGML_FP16_TO_FP32(v);
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                ggml_fp16_to_fp32_row(
+                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
+            }
         }
     }
 }
@@ -10248,19 +10418,27 @@ static void ggml_compute_forward_get_rows_f32(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == sizeof(float));
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(float));
+    assert(ggml_nrows(dst) == nr);
 
-        ggml_vec_cpy_f32(nc,
-                (float *) ((char *)  dst->data + i*dst->nb[1]),
-                (float *) ((char *) src0->data + r*src0->nb[1]));
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                ggml_vec_cpy_f32(nc,
+                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
+                        (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
+            }
+        }
     }
 }
 
@@ -12021,6 +12199,67 @@ 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(
@@ -13842,7 +14081,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_MUL_MAT:
             {
-                ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor, 0, tensor->ne[1]);
+            } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                ggml_compute_forward_mul_mat_id(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_OUT_PROD:
             {
@@ -13948,6 +14191,10 @@ 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);
@@ -14598,6 +14845,10 @@ 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
@@ -14936,6 +15187,10 @@ 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;
@@ -15296,12 +15551,8 @@ 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,
@@ -15536,7 +15787,6 @@ 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:
@@ -15569,10 +15819,13 @@ 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:
@@ -15610,6 +15863,11 @@ 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;
@@ -15669,6 +15927,10 @@ 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;
@@ -15731,6 +15993,10 @@ 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__);
@@ -15927,6 +16193,23 @@ 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:
                 {
                     if (ggml_is_quantized(node->src[0]->type)) {
@@ -15962,9 +16245,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         GGML_ASSERT(false);
                     }
                 } break;
-            case GGML_OP_IM2COL:
-                {
-                } break;
             case GGML_OP_CONV_TRANSPOSE_2D:
                 {
                     const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -17803,8 +18083,8 @@ 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;
@@ -17833,8 +18113,8 @@ 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;
@@ -18024,6 +18304,7 @@ struct gguf_kv {
 
 struct gguf_header {
     char magic[4];
+
     uint32_t version;
     uint64_t n_tensors; // GGUFv2
     uint64_t n_kv;      // GGUFv2
@@ -18113,7 +18394,7 @@ 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;
             }
@@ -18128,7 +18409,6 @@ 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

@@ -217,7 +217,7 @@
 #define GGML_MAX_DIMS           4
 #define GGML_MAX_PARAMS         1024
 #define GGML_MAX_CONTEXTS       64
-#define GGML_MAX_SRC            6
+#define GGML_MAX_SRC            10
 #define GGML_MAX_NAME           64
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
@@ -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 +1047,16 @@ 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  * const as[],
+            int                   n_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
@@ -1234,6 +1264,7 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // supports 3D: a->ne[2] == b->ne[1]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1520,6 +1551,23 @@ 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,
@@ -1581,7 +1629,6 @@ extern "C" {
             int                   kh);
 
     // used in sam
-
     GGML_API struct ggml_tensor * ggml_add_rel_pos(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1756,7 +1803,7 @@ 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

@@ -38,6 +38,8 @@ class Keys:
         FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
         USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
         TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
+        EXPERT_COUNT          = "{arch}.expert_count"
+        EXPERT_USED_COUNT     = "{arch}.expert_used_count"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -111,10 +113,14 @@ class MODEL_TENSOR(IntEnum):
     ATTN_NORM       = auto()
     ATTN_NORM_2     = auto()
     ATTN_ROT_EMBD   = auto()
+    FFN_GATE_INP    = auto()
+    FFN_NORM        = auto()
     FFN_GATE        = auto()
     FFN_DOWN        = auto()
     FFN_UP          = auto()
-    FFN_NORM        = auto()
+    FFN_GATE_EXP    = auto()
+    FFN_DOWN_EXP    = auto()
+    FFN_UP_EXP      = auto()
     ATTN_Q_NORM     = auto()
     ATTN_K_NORM     = auto()
 
@@ -154,10 +160,14 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ATTN_ROT_EMBD:   "blk.{bid}.attn_rot_embd",
     MODEL_TENSOR.ATTN_Q_NORM:     "blk.{bid}.attn_q_norm",
     MODEL_TENSOR.ATTN_K_NORM:     "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.FFN_GATE_INP:    "blk.{bid}.ffn_gate_inp",
     MODEL_TENSOR.FFN_NORM:        "blk.{bid}.ffn_norm",
     MODEL_TENSOR.FFN_GATE:        "blk.{bid}.ffn_gate",
     MODEL_TENSOR.FFN_DOWN:        "blk.{bid}.ffn_down",
     MODEL_TENSOR.FFN_UP:          "blk.{bid}.ffn_up",
+    MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate.{xid}",
+    MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down.{xid}",
+    MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up.{xid}",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -172,10 +182,14 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
     ],
     MODEL_ARCH.GPTNEOX: [
         MODEL_TENSOR.TOKEN_EMBD,

gguf-py/gguf/gguf_writer.py

@@ -339,6 +339,12 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
+    def add_expert_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
+
+    def add_expert_used_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
+
     def add_layer_norm_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -149,6 +149,11 @@ class TensorNameMap:
             "model.layers.{bid}.ln2",                                        # yi
         ),
 
+        MODEL_TENSOR.FFN_GATE_INP: (
+            "layers.{bid}.feed_forward.gate",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate", # mixtral
+        ),
+
         # Feed-forward up
         MODEL_TENSOR.FFN_UP: (
             "gpt_neox.layers.{bid}.mlp.dense_h_to_4h",                # gptneox
@@ -164,11 +169,21 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.w1",                             # qwen
         ),
 
+        MODEL_TENSOR.FFN_UP_EXP: (
+            "layers.{bid}.feed_forward.experts.{xid}.w3",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
+        ),
+
         # 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
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact
+            "layers.{bid}.feed_forward.w1",               # llama-pth
+            "transformer.h.{bid}.mlp.w2",                 # qwen
+        ),
+
+        MODEL_TENSOR.FFN_GATE_EXP: (
+            "layers.{bid}.feed_forward.experts.{xid}.w1",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral
         ),
 
         # Feed-forward down
@@ -185,6 +200,11 @@ class TensorNameMap:
             "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
         ),
 
+        MODEL_TENSOR.FFN_DOWN_EXP: (
+            "layers.{bid}.feed_forward.experts.{xid}.w2",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
+        ),
+
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
         ),
@@ -213,11 +233,14 @@ class TensorNameMap:
             for tensor, keys in self.block_mappings_cfg.items():
                 if tensor not in MODEL_TENSORS[arch]:
                     continue
-                tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
-                self.mapping[tensor_name] = (tensor, tensor_name)
-                for key in keys:
-                    key = key.format(bid = bid)
-                    self.mapping[key] = (tensor, tensor_name)
+                # TODO: make this configurable
+                n_experts = 8
+                for xid in range(n_experts):
+                    tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
+                    self.mapping[tensor_name] = (tensor, tensor_name)
+                    for key in keys:
+                        key = key.format(bid = bid, xid = xid)
+                        self.mapping[key] = (tensor, tensor_name)
 
     def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
         result = self.mapping.get(key)

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.6.0"
+version = "0.7.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

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

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)