llama : model loader

ggml-ci

src/CMakeLists.txt

@@ -20,6 +20,7 @@ add_library(llama
             llama-kv-cache.cpp
             llama-mmap.cpp
             llama-model.cpp
+            llama-model-loader.cpp
             llama-sampling.cpp
             llama-vocab.cpp
             unicode.h

src/llama-hparams.h

@@ -74,7 +74,8 @@ struct llama_hparams {
     float    rope_freq_scale_train;
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul;
-    int       rope_sections[4]; // TODO: actually this should be std::array (I was wrong)
+
+    std::array<int, 4> rope_sections;
 
     // for State Space Models
     uint32_t ssm_d_conv  = 0;

src/llama-impl.cpp

@@ -2,9 +2,11 @@
 
 #include "llama.h"
 
+#include <cinttypes>
 #include <climits>
 #include <cstdarg>
 #include <vector>
+#include <sstream>
 
 struct llama_logger_state {
     ggml_log_callback log_callback = llama_log_callback_default;
@@ -89,3 +91,75 @@ std::string format(const char * fmt, ...) {
     va_end(ap);
     return std::string(buf.data(), size);
 }
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
+    for (size_t i = 1; i < ne.size(); i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
+    }
+    return buf;
+}
+
+std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+    }
+    return buf;
+}
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}

src/llama-impl.h

@@ -3,6 +3,7 @@
 #include "ggml.h" // for ggml_log_level
 
 #include <string>
+#include <vector>
 
 #ifdef __GNUC__
 #ifdef __MINGW32__
@@ -33,6 +34,12 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 // helpers
 //
 
+template <typename T>
+struct no_init {
+    T value;
+    no_init() { /* do nothing */ }
+};
+
 struct time_meas {
     time_meas(int64_t & t_acc, bool disable = false);
     ~time_meas();
@@ -47,3 +54,8 @@ void replace_all(std::string & s, const std::string & search, const std::string
 // TODO: rename to llama_format ?
 LLAMA_ATTRIBUTE_FORMAT(1, 2)
 std::string format(const char * fmt, ...);
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
+std::string llama_format_tensor_shape(const struct ggml_tensor * t);
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);

src/llama-model-loader.h

@@ -0,0 +1,157 @@
+#pragma once
+
+#include "llama.h"
+
+#include "llama-impl.h"
+#include "llama-arch.h"
+#include "llama-mmap.h"
+
+#include "ggml-cpp.h"
+
+#include <cstddef>
+#include <map>
+#include <unordered_map>
+
+using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
+
+enum llama_fver {
+    GGUF_FILE_VERSION_V1 = 1,
+    GGUF_FILE_VERSION_V2 = 2,
+    GGUF_FILE_VERSION_V3 = 3,
+};
+
+const char * llama_file_version_name(llama_fver version);
+
+struct llama_model_loader {
+    // Holds information on a model weight
+    struct llama_tensor_weight {
+        uint16_t  idx; // source file index
+        size_t   offs; // tensor data offset in the original file
+
+        ggml_tensor * tensor;
+
+        llama_tensor_weight(const llama_file * file, uint16_t idx, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) {
+            const int tensor_idx = gguf_find_tensor(gguf_ctx,  ggml_get_name(tensor));
+            if (tensor_idx < 0) {
+                throw std::runtime_error(format("tensor '%s' not found in the model", ggml_get_name(tensor)));
+            }
+
+            offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx);
+            if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size()) {
+                throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", ggml_get_name(tensor)));
+            }
+        }
+    };
+
+    // custom comparator to sort weights more nicely by layer
+    struct weight_name_comparer {
+        bool operator()(const std::string & a, const std::string & b) const {
+            int a_layer = -1;
+            int b_layer = -1;
+            sscanf(a.c_str(), "blk.%d.", &a_layer);
+            sscanf(b.c_str(), "blk.%d.", &b_layer);
+            if (a_layer != b_layer) {
+                return a_layer < b_layer;
+            }
+            return a < b;
+        }
+    };
+
+    static const int TENSOR_NOT_REQUIRED = 1;
+    static const int TENSOR_DUPLICATED   = 2;
+
+    int n_kv      = 0;
+    int n_tensors = 0;
+    int n_created = 0;
+
+    uint64_t n_elements = 0;
+    size_t   n_bytes    = 0;
+
+    bool use_mmap = false;
+    bool check_tensors;
+
+    llama_files files;
+    llama_ftype ftype;
+    llama_fver  fver;
+
+    llama_mmaps mappings;
+
+    std::map<std::string, struct llama_tensor_weight, weight_name_comparer> weights_map;
+    std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;
+
+    gguf_context_ptr meta;
+    std::vector<ggml_context_ptr> contexts;
+
+    std::string arch_name;
+    LLM_KV      llm_kv    = LLM_KV(LLM_ARCH_UNKNOWN);
+
+    size_t size_done = 0;
+    size_t size_data = 0;
+    std::vector<std::pair<size_t, size_t>> mmaps_used;
+
+    llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p);
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(const std::string & key, T & result, bool required = true);
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T>
+    bool get_arr(const std::string & key, std::vector<T> & result, bool required = true);
+
+    template<typename T, size_t N_MAX>
+    bool get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required = true);
+
+    template<typename T>
+    bool get_arr(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T>
+    bool get_key(const std::string & key, T & result, bool required = true);
+
+    template<typename T>
+    bool get_key(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T, size_t N_MAX>
+    bool get_key_or_arr(const std::string & key, std::array<T, N_MAX> & result, uint32_t n, bool required = true);
+
+    template<typename T>
+    bool get_key_or_arr(enum llm_kv kid, T & result, uint32_t n, bool required = true);
+
+    std::string get_arch_name() const;
+
+    enum llm_arch get_arch() const;
+
+    const llama_tensor_weight * get_weight(const char * name) const;
+
+    const llama_tensor_weight & require_weight(const char * name) const;
+
+    struct ggml_tensor * get_tensor_meta(const char * name) const;
+
+    struct ggml_tensor * require_tensor_meta(const std::string & name) const;
+
+    const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const;
+
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags = 0);
+
+    struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required = true);
+
+    void done_getting_tensors() const;
+
+    void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr);
+
+    void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const;
+
+    // for backwards compatibility, does not support ggml-backend
+    void load_data_for(struct ggml_tensor * cur) const;
+
+    // Returns false if cancelled by progress_callback
+    bool load_all_data(
+            struct ggml_context * ctx,
+            llama_buf_map & bufs,
+            llama_mlocks * lmlocks,
+            llama_progress_callback progress_callback,
+            void * progress_callback_user_data);
+};