@fugood/llama.node 0.3.17 → 0.4.1

package/src/llama.cpp/ggml/src/gguf.cpp

@@ -299,10 +299,10 @@ bool gguf_read_emplace_helper(const struct gguf_reader & gr, std::vector<struct
                 return false;
             }
         } catch (std::length_error &) {
-            fprintf(stderr, "%s: encountered length_error while reading value for key '%s'\n", __func__, key.c_str());
+            GGML_LOG_ERROR("%s: encountered length_error while reading value for key '%s'\n", __func__, key.c_str());
             return false;
         } catch (std::bad_alloc &) {
-            fprintf(stderr, "%s: encountered bad_alloc error while reading value for key '%s'\n", __func__, key.c_str());
+            GGML_LOG_ERROR("%s: encountered bad_alloc error while reading value for key '%s'\n", __func__, key.c_str());
             return false;
         }
         kv.emplace_back(key, value);
@@ -328,14 +328,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         ok = ok && gr.read(magic, 4);
 
         if (!ok) {
-            fprintf(stderr, "%s: failed to read magic\n", __func__);
+            GGML_LOG_ERROR("%s: failed to read magic\n", __func__);
             gguf_free(ctx);
             return nullptr;
         }
 
         for (uint32_t i = 0; i < magic.size(); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                fprintf(stderr, "%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
+                GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                 gguf_free(ctx);
                 return nullptr;
             }
@@ -348,11 +348,11 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
     if (ok && gr.read(ctx->version)) {
         if (ctx->version == 1) {
-            fprintf(stderr, "%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
+            GGML_LOG_ERROR("%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
             ok = false;
         }
         if (ctx->version > GGUF_VERSION) {
-            fprintf(stderr, "%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
+            GGML_LOG_ERROR("%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
                 __func__, ctx->version, GGUF_VERSION);
             ok = false;
         }
@@ -363,7 +363,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     if (ok && gr.read(n_tensors)) {
         static_assert(sizeof(size_t) <= 8 && sizeof(gguf_tensor_info) >= 2, "int64_t insufficient for indexing");
         if (n_tensors < 0 || n_tensors > int64_t(SIZE_MAX/sizeof(gguf_tensor_info))) {
-            fprintf(stderr, "%s: number of tensors is %" PRIi64 " but must be in [0, %zu]\n",
+            GGML_LOG_ERROR("%s: number of tensors is %" PRIi64 " but must be in [0, %zu]\n",
                 __func__, n_tensors, SIZE_MAX/sizeof(gguf_tensor_info));
             ok = false;
         }
@@ -374,7 +374,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     if (ok && gr.read(n_kv)) {
         static_assert(sizeof(size_t) <= 8 && sizeof(gguf_tensor_info) >= 2, "int64_t insufficient for indexing");
         if (n_kv < 0 || n_kv > int64_t(SIZE_MAX/sizeof(gguf_kv))) {
-            fprintf(stderr, "%s: number of key value pairs is %" PRIi64 " but must be in [0, %zu]\n",
+            GGML_LOG_ERROR("%s: number of key value pairs is %" PRIi64 " but must be in [0, %zu]\n",
                     __func__, n_kv, SIZE_MAX/sizeof(gguf_kv));
             ok = false;
         }
@@ -383,7 +383,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     }
 
     if (!ok) {
-        fprintf(stderr, "%s: failed to read header\n", __func__);
+        GGML_LOG_ERROR("%s: failed to read header\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
@@ -399,15 +399,15 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             try {
                 ok = ok && gr.read(key);
             } catch (std::length_error &) {
-                fprintf(stderr, "%s: encountered length_error while reading key %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered length_error while reading key %" PRIi64 "\n", __func__, i);
                 ok = false;
             } catch (std::bad_alloc &) {
-                fprintf(stderr, "%s: encountered bad_alloc error while reading key %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered bad_alloc error while reading key %" PRIi64 "\n", __func__, i);
                 ok = false;
             }
             for (size_t j = 0; ok && j < ctx->kv.size(); ++j) {
                 if (key == ctx->kv[j].key) {
-                    fprintf(stderr, "%s: duplicate key '%s' for tensors %zu and %" PRIi64 " \n", __func__, key.c_str(), j, i);
+                    GGML_LOG_ERROR("%s: duplicate key '%s' for tensors %zu and %" PRIi64 " \n", __func__, key.c_str(), j, i);
                     ok = false;
                 }
             }
@@ -441,14 +441,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
                 case GGUF_TYPE_ARRAY:
                 default:
                     {
-                        fprintf(stderr, "%s: key '%s' has invalid GGUF type %d\n", __func__, key.c_str(), type);
+                        GGML_LOG_ERROR("%s: key '%s' has invalid GGUF type %d\n", __func__, key.c_str(), type);
                         ok = false;
                     } break;
             }
         }
 
         if (!ok) {
-            fprintf(stderr, "%s: failed to read key-value pairs\n", __func__);
+            GGML_LOG_ERROR("%s: failed to read key-value pairs\n", __func__);
             gguf_free(ctx);
             return nullptr;
         }
@@ -458,7 +458,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         ctx->alignment = alignment_idx == -1 ? GGUF_DEFAULT_ALIGNMENT : gguf_get_val_u32(ctx, alignment_idx);
 
         if (ctx->alignment == 0 || (ctx->alignment & (ctx->alignment - 1)) != 0) {
-            fprintf(stderr, "%s: alignment %zu is not a power of 2\n", __func__, ctx->alignment);
+            GGML_LOG_ERROR("%s: alignment %zu is not a power of 2\n", __func__, ctx->alignment);
             gguf_free(ctx);
             return nullptr;
         }
@@ -474,14 +474,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             try {
                 ok = ok && gr.read(name);
             } catch (std::length_error &) {
-                fprintf(stderr, "%s: encountered length_error while reading tensor name %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered length_error while reading tensor name %" PRIi64 "\n", __func__, i);
                 ok = false;
             } catch (std::bad_alloc &) {
-                fprintf(stderr, "%s: encountered bad_alloc error while reading tensor name %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered bad_alloc error while reading tensor name %" PRIi64 "\n", __func__, i);
                 ok = false;
             }
             if (name.length() >= GGML_MAX_NAME) {
-                fprintf(stderr, "%s: tensor name %" PRIi64 " is too long: %zu >= %d\n", __func__, i, name.length(), GGML_MAX_NAME);
+                GGML_LOG_ERROR("%s: tensor name %" PRIi64 " is too long: %zu >= %d\n", __func__, i, name.length(), GGML_MAX_NAME);
                 ok = false;
                 break;
             }
@@ -490,7 +490,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             // make sure there are no duplicate tensor names
             for (int64_t j = 0; ok && j < i; ++j) {
                 if (strcmp(info.t.name, ctx->info[j].t.name) == 0) {
-                    fprintf(stderr, "%s: duplicate tensor name '%s' for tensors %" PRIi64 " and %" PRIi64 "\n", __func__, info.t.name, j, i);
+                    GGML_LOG_ERROR("%s: duplicate tensor name '%s' for tensors %" PRIi64 " and %" PRIi64 "\n", __func__, info.t.name, j, i);
                     ok = false;
                     break;
                 }
@@ -505,7 +505,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             uint32_t n_dims = -1;
             ok = ok && gr.read(n_dims);
             if (n_dims > GGML_MAX_DIMS) {
-                fprintf(stderr, "%s: tensor '%s' has invalid number of dimensions: %" PRIu32 " > %" PRIu32 "\n",
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid number of dimensions: %" PRIu32 " > %" PRIu32 "\n",
                     __func__, info.t.name, n_dims, GGML_MAX_DIMS);
                 ok = false;
                 break;
@@ -518,7 +518,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
                 // check that all ne are non-negative
                 if (info.t.ne[j] < 0) {
-                    fprintf(stderr, "%s: tensor '%s' dimension %" PRIu32 " has invalid number of elements: %" PRIi64 " < 0\n",
+                    GGML_LOG_ERROR("%s: tensor '%s' dimension %" PRIu32 " has invalid number of elements: %" PRIi64 " < 0\n",
                         __func__, info.t.name, j, info.t.ne[j]);
                     ok = false;
                     break;
@@ -530,7 +530,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
                        (INT64_MAX/info.t.ne[2] <= info.t.ne[0]*info.t.ne[1]) ||
                        (INT64_MAX/info.t.ne[3] <= info.t.ne[0]*info.t.ne[1]*info.t.ne[2]))) {
 
-                fprintf(stderr, "%s: total number of elements in tensor '%s' with shape "
+                GGML_LOG_ERROR("%s: total number of elements in tensor '%s' with shape "
                     "(%" PRIi64 ", %" PRIi64 ", %" PRIi64 ", %" PRIi64 ") is >= %" PRIi64 "\n",
                     __func__, info.t.name, info.t.ne[0], info.t.ne[1], info.t.ne[2], info.t.ne[3], INT64_MAX);
                 ok = false;
@@ -547,7 +547,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             // check that tensor type is within defined range
             if (info.t.type < 0 || info.t.type >= GGML_TYPE_COUNT) {
-                fprintf(stderr, "%s: tensor '%s' has invalid ggml type %d (%s)\n",
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d (%s)\n",
                     __func__, info.t.name, info.t.type, ggml_type_name(info.t.type));
                 ok = false;
                 break;
@@ -557,7 +557,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             // check that row size is divisible by block size
             if (blck_size == 0 || info.t.ne[0] % blck_size != 0) {
-                fprintf(stderr, "%s: tensor '%s' of type %d (%s) has %" PRId64 " elements per row, "
+                GGML_LOG_ERROR("%s: tensor '%s' of type %d (%s) has %" PRId64 " elements per row, "
                     "not a multiple of block size (%" PRId64 ")\n",
                     __func__, info.t.name, (int) info.t.type, ggml_type_name(info.t.type), info.t.ne[0], blck_size);
                 ok = false;
@@ -582,7 +582,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     }
 
     if (!ok) {
-        fprintf(stderr, "%s: failed to read tensor info\n", __func__);
+        GGML_LOG_ERROR("%s: failed to read tensor info\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
@@ -590,7 +590,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
     // we require the data section to be aligned, so take into account any padding
     if (fseek(file, GGML_PAD(ftell(file), ctx->alignment), SEEK_SET) != 0) {
-        fprintf(stderr, "%s: failed to seek to beginning of data section\n", __func__);
+        GGML_LOG_ERROR("%s: failed to seek to beginning of data section\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
@@ -604,9 +604,9 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         for (size_t i = 0; i < ctx->info.size(); ++i) {
             const gguf_tensor_info & ti = ctx->info[i];
             if (ti.offset != ctx->size) {
-                fprintf(stderr, "%s: tensor '%s' has offset %" PRIu64 ", expected %zu\n",
+                GGML_LOG_ERROR("%s: tensor '%s' has offset %" PRIu64 ", expected %zu\n",
                     __func__, ti.t.name, ti.offset, ctx->size);
-                fprintf(stderr, "%s: failed to read tensor data\n", __func__);
+                GGML_LOG_ERROR("%s: failed to read tensor data\n", __func__);
                 gguf_free(ctx);
                 return nullptr;
             }
@@ -634,7 +634,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
         *params.ctx = ggml_init(pdata);
         if (*params.ctx == nullptr) {
-            fprintf(stderr, "%s: failed to initialize ggml context for storing tensors\n", __func__);
+            GGML_LOG_ERROR("%s: failed to initialize ggml context for storing tensors\n", __func__);
             gguf_free(ctx);
             return nullptr;
         }
@@ -656,7 +656,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             ok = ok && gr.read(data->data, ctx->size);
 
             if (!ok) {
-                fprintf(stderr, "%s: failed to read tensor data binary blob\n", __func__);
+                GGML_LOG_ERROR("%s: failed to read tensor data binary blob\n", __func__);
                 ggml_free(ctx_data);
                 *params.ctx = nullptr;
                 gguf_free(ctx);
@@ -689,7 +689,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         }
 
         if (!ok) {
-            fprintf(stderr, "%s: failed to create tensors\n", __func__);
+            GGML_LOG_ERROR("%s: failed to create tensors\n", __func__);
             ggml_free(ctx_data);
             *params.ctx = nullptr;
             gguf_free(ctx);
@@ -706,7 +706,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     FILE * file = ggml_fopen(fname, "rb");
 
     if (!file) {
-        fprintf(stderr, "%s: failed to open GGUF file '%s'\n", __func__, fname);
+        GGML_LOG_ERROR("%s: failed to open GGUF file '%s'\n", __func__, fname);
         return nullptr;
     }
 
@@ -1305,7 +1305,7 @@ bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, boo
     FILE * file = ggml_fopen(fname, "wb");
 
     if (!file) {
-        fprintf(stderr, "%s: failed to open file '%s' for writing GGUF data\n", __func__, fname);
+        GGML_LOG_ERROR("%s: failed to open file '%s' for writing GGUF data\n", __func__, fname);
         return false;
     }
 

package/src/llama.cpp/include/llama.h

@@ -4,6 +4,7 @@
 #include "ggml.h"
 #include "ggml-cpu.h"
 #include "ggml-backend.h"
+#include "ggml-opt.h"
 
 #include <stddef.h>
 #include <stdint.h>
@@ -112,6 +113,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
         LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
+        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
     };
 
     enum llama_rope_type {
@@ -343,7 +345,7 @@ extern "C" {
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
+        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
 
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
@@ -351,19 +353,18 @@ extern "C" {
         enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
         enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
 
-        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
-        // TODO: move at the end of the struct
-        bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
-        bool embeddings;  // if true, extract embeddings (together with logits)
-        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
-        bool no_perf;     // whether to measure performance timings
-
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
         // currently works only with CPU execution
         ggml_abort_callback abort_callback;
         void *              abort_callback_data;
+
+        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
+        bool embeddings;  // if true, extract embeddings (together with logits)
+        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
+        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
+        bool no_perf;     // whether to measure performance timings
+        bool op_offload;  // whether to offload host tensor operations to device
     };
 
     // model quantization parameters
@@ -445,6 +446,10 @@ extern "C" {
                                  size_t    n_paths,
               struct llama_model_params    params);
 
+    LLAMA_API void llama_model_save_to_file(
+            const struct llama_model * model,
+                        const char * path_model);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -924,14 +929,19 @@ extern "C" {
     // Frees a batch of tokens allocated with llama_batch_init()
     LLAMA_API void llama_batch_free(struct llama_batch batch);
 
-    // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
-    // Stores the encoder output internally for later use by the decoder cross-attention layers.
+    // Process a batch of tokens.
+    // In contrast to llama_decode() - this call does not use KV cache.
+    // For encode-decoder contexts, processes the batch using the encoder.
+    // Can store the encoder output internally for later use by the decoder's cross-attention layers.
     //   0 - success
     // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
+    // Process a batch of tokens.
+    // Requires KV cache.
+    // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
@@ -1428,6 +1438,37 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
+    //
+    // training
+    //
+
+    // function that returns whether or not a given tensor contains trainable parameters
+    typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata);
+
+    // always returns true
+    LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata);
+
+    struct llama_opt_params {
+        uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0
+
+        llama_opt_param_filter param_filter; // callback for determining which tensors contain trainable parameters
+        void * param_filter_ud;              // userdata for determining which tensors contain trainable parameters
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
+
+    LLAMA_API void llama_opt_epoch(
+            struct llama_context    * lctx,
+            ggml_opt_dataset_t        dataset,
+            ggml_opt_result_t         result_train,
+            ggml_opt_result_t         result_eval,
+            int64_t                   idata_split,
+            ggml_opt_epoch_callback   callback_train,
+            ggml_opt_epoch_callback   callback_eval);
+
 #ifdef __cplusplus
 }
 #endif

package/src/llama.cpp/requirements/requirements-all.txt

@@ -1,6 +1,6 @@
--r ../examples/llava/requirements.txt
--r ../examples/server/bench/requirements.txt
--r ../examples/server/tests/requirements.txt
+-r ../tools/mtmd/requirements.txt
+-r ../tools/server/bench/requirements.txt
+-r ../tools/server/tests/requirements.txt
 
 -r ./requirements-compare-llama-bench.txt
 -r ./requirements-pydantic.txt

package/src/llama.cpp/scripts/xxd.cmake

@@ -1,5 +1,5 @@
 # CMake equivalent of `xxd -i ${INPUT} ${OUTPUT}`
-# Usage: cmake -DINPUT=examples/server/public/index.html -DOUTPUT=examples/server/index.html.hpp -P scripts/xxd.cmake
+# Usage: cmake -DINPUT=tools/server/public/index.html -DOUTPUT=tools/server/index.html.hpp -P scripts/xxd.cmake
 
 SET(INPUT "" CACHE STRING "Input File")
 SET(OUTPUT "" CACHE STRING "Output File")

package/src/llama.cpp/src/CMakeLists.txt

@@ -23,6 +23,7 @@ add_library(llama
             llama-memory.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
+            llama-model-saver.cpp
             llama-model.cpp
             llama-quant.cpp
             llama-sampling.cpp

package/src/llama.cpp/src/llama-adapter.cpp

@@ -253,6 +253,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
     std::vector<ggml_backend_buffer_type_t> buft_extra;
     {
         auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (!cpu_dev) {
+            throw std::runtime_error(format("%s: no CPU backend found", __func__));
+        }
         auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
 
         auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
@@ -291,6 +294,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
                 LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
 
                 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                if (!cpu_dev) {
+                    throw std::runtime_error(format("%s: no CPU backend found", __func__));
+                }
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
 
                 break;

package/src/llama.cpp/src/llama-arch.cpp

@@ -1481,6 +1481,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
             { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
         },
     },
     {

package/src/llama.cpp/src/llama-batch.cpp

@@ -189,7 +189,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
     GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
@@ -203,6 +203,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
     for (size_t i = 0; i < n_tokens; ++i) {
         ids[i] = i;
     }
+
     if (simple_split) {
         seq.resize(1);
         llama_sbatch_seq & s = seq[0];
@@ -212,6 +213,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         s.length = n_tokens;
         return;
     }
+
     std::sort(ids.begin(), ids.end(),
             [&batch](size_t a, size_t b) {
                 int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
@@ -239,6 +241,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
                 return n_seq_a > n_seq_b;
             }
     );
+
     // init seq
     llama_sbatch_seq * last_seq = nullptr;
 
@@ -262,6 +265,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         seq.push_back(new_seq);
         last_seq = &seq.back();
     }
+
     // keep shared prompts first at the end, then sort by length descending.
     std::sort(seq.begin(), seq.end(),
             [](llama_sbatch_seq & a, llama_sbatch_seq & b) {

package/src/llama.cpp/src/llama-batch.h

@@ -70,7 +70,8 @@ struct llama_sbatch {
     // sequence-wise split
     llama_ubatch split_seq(size_t n_ubatch);
 
-    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
 };
 
 // temporary allocate memory for the input batch if needed

package/src/llama.cpp/src/llama-chat.cpp

@@ -35,6 +35,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "mistral-v3",        LLM_CHAT_TEMPLATE_MISTRAL_V3        },
     { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
     { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
+    { "mistral-v7-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN },
     { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
     { "phi4",              LLM_CHAT_TEMPLATE_PHI_4             },
     { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
@@ -202,19 +203,20 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|im_start|>assistant\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN) {
         // Official mistral 'v7' template
         // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7
+        //      https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503#basic-instruct-template-v7-tekken
+        const char * trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 ? " " : "";
         for (auto message : chat) {
             std::string role(message->role);
             std::string content(message->content);
             if (role == "system") {
-                ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";
+                ss << "[SYSTEM_PROMPT]" << trailing_space << content << "[/SYSTEM_PROMPT]";
             } else if (role == "user") {
-                ss << "[INST] " << content << "[/INST]";
-            }
-            else {
-                ss << " " << content << "</s>";
+                ss << "[INST]" << trailing_space << content << "[/INST]";
+            } else {
+                ss << trailing_space << content << "</s>";
             }
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1
@@ -447,8 +449,16 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         for (auto message : chat) {
             std::string role(message->role);
             ss << "<|" << role << "|>" << "\n" << message->content;

package/src/llama.cpp/src/llama-chat.h

@@ -14,6 +14,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3,
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
+    LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN,
     LLM_CHAT_TEMPLATE_PHI_3,
     LLM_CHAT_TEMPLATE_PHI_4,
     LLM_CHAT_TEMPLATE_FALCON_3,