whispercpp 1.3.5 → 1.3.6

data/ext/sources/examples/talk-llama/llama-model-loader.h

@@ -4,17 +4,22 @@
 
 #include "llama-impl.h"
 #include "llama-arch.h"
+#include "llama-hparams.h"
 #include "llama-mmap.h"
 
 #include "ggml-cpp.h"
 
 #include <cstddef>
+#include <cstring>
 #include <map>
 #include <stdexcept>
 #include <unordered_map>
 
 using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
 
+// lists of buffer types used for each layer
+using buft_list_t = std::vector<std::pair<ggml_backend_dev_t, ggml_backend_buffer_type_t>>;
+
 enum llama_fver {
     GGUF_FILE_VERSION_V1 = 1,
     GGUF_FILE_VERSION_V2 = 2,
@@ -58,9 +63,10 @@ struct llama_model_loader {
         }
     };
 
-    static const int TENSOR_NOT_REQUIRED = 1 << 0;
-    static const int TENSOR_DUPLICATED   = 1 << 1;
-    static const int TENSOR_SKIP         = 1 << 2;
+    static const int TENSOR_NOT_REQUIRED    = 1 << 0;
+    static const int TENSOR_DUPLICATED      = 1 << 1;
+    static const int TENSOR_SKIP            = 1 << 2;
+    static const int TENSOR_SKIP_IF_VIRTUAL = 1 << 3;
 
     int n_kv      = 0;
     int n_tensors = 0;
@@ -84,7 +90,10 @@ struct llama_model_loader {
     std::unordered_map<std::string, llama_model_kv_override> kv_overrides;
     const llama_model_tensor_buft_override * tensor_buft_overrides;
 
-    gguf_context_ptr meta;
+    gguf_context_ptr metadata_ptr;
+    struct gguf_context * metadata; // either metadata_ptr.get() or externally set
+    llama_model_set_tensor_data_t set_tensor_data;
+    void * set_tensor_data_ud;
     std::vector<ggml_context_ptr> contexts;
 
     std::string arch_name;
@@ -94,7 +103,26 @@ struct llama_model_loader {
     size_t size_data = 0;
     std::vector<std::pair<size_t, size_t>> mmaps_used;
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
+    // track tensors that had to be moved for debugging:
+    size_t n_tensors_moved = 0;
+    std::string first_tensor_moved_name;
+    std::string first_tensor_moved_type_name;
+    ggml_backend_buffer_type_t first_moved_from_buft = nullptr;
+    ggml_backend_buffer_type_t first_moved_to_buft = nullptr;
+
     llama_model_loader(
+        struct gguf_context * metadata,
+        llama_model_set_tensor_data_t set_tensor_data,
+        void * set_tensor_data_ud,
         const std::string & fname,
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
@@ -149,7 +177,9 @@ struct llama_model_loader {
 
     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(
+        const llama_hparams & hparams, const buft_list_t * buft_list_cpu, const buft_list_t * buft_list_input, const buft_list_t * buft_list_output,
+        const buft_list_t * buft_list_layer, const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags);
 
     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);
 

data/ext/sources/examples/talk-llama/llama-model-saver.cpp

@@ -7,14 +7,19 @@
 #include "llama-model.h"
 #include "llama-vocab.h"
 
+#include <cstdint>
 #include <string>
 
-llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) {
-    gguf_ctx = gguf_init_empty();
-}
+llama_model_saver::llama_model_saver(const struct llama_model * model) :
+    gguf_ctx(gguf_init_empty()), gguf_ctx_owned(true), model(model), llm_kv(model->arch) {}
+
+llama_model_saver::llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx) :
+        gguf_ctx(gguf_ctx == nullptr ? gguf_init_empty() : gguf_ctx), gguf_ctx_owned(gguf_ctx == nullptr), model(nullptr), llm_kv(arch) {}
 
 llama_model_saver::~llama_model_saver() {
-    gguf_free(gguf_ctx);
+    if (gguf_ctx_owned) {
+        gguf_free(gguf_ctx);
+    }
 }
 
 void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
@@ -46,7 +51,8 @@ void llama_model_saver::add_kv(const enum llm_kv key, const char value) {
 
 template <typename Container>
 void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {
-    const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size();
+    GGML_ASSERT(model != nullptr || !per_layer);
+    const size_t n_values = per_layer ? size_t(model->hparams.n_layer) : value.size();
     GGML_ASSERT(n_values <= value.size());
 
     if (n_values == 0) {
@@ -83,6 +89,8 @@ void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, c
         GGML_ABORT("fatal error");
     }
 }
+// instantiate for external usage:
+template void llama_model_saver::add_kv<std::vector<uint32_t>>(const enum llm_kv, const std::vector<uint32_t> &, const bool);
 
 void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
     std::vector<const char *> tmp(value.size());
@@ -104,37 +112,39 @@ void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {
 }
 
 void llama_model_saver::add_kv_from_model() {
-    const llama_hparams & hparams = model.hparams;
-    const llama_vocab   & vocab   = model.vocab;
+    const llama_hparams & hparams = model->hparams;
+    const llama_vocab   & vocab   = model->vocab;
 
     const int32_t n_vocab = vocab.n_tokens();
     std::vector<std::string> tokens(n_vocab);
     std::vector<float>       scores(n_vocab);
     std::vector<int32_t>     token_types(n_vocab);
 
-    for (int32_t id = 0; id < n_vocab; ++id) {
-        const llama_vocab::token_data & token_data = vocab.get_token_data(id);
-
-        tokens[id] = token_data.text;
-        scores[id] = token_data.score;
-
-        switch(token_data.attr) {
-            case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;
-            case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;
-            case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;
-            case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;
-            case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
-            case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;
-            case LLAMA_TOKEN_ATTR_UNDEFINED:
-            default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;
+    if (vocab.get_type() != LLAMA_VOCAB_TYPE_NONE) {
+        for (int32_t id = 0; id < n_vocab; ++id) {
+            const llama_vocab::token_data & token_data = vocab.get_token_data(id);
+
+            tokens[id] = token_data.text;
+            scores[id] = token_data.score;
+
+            switch(token_data.attr) {
+                case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;
+                case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;
+                case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;
+                case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;
+                case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
+                case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;
+                case LLAMA_TOKEN_ATTR_UNDEFINED:
+                default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;
+            }
         }
     }
 
     // add_kv(LLM_KV_GENERAL_TYPE,                      ???);
-    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model.arch_name());
+    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model->arch_name());
     // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION,      ???);
     // add_kv(LLM_KV_GENERAL_ALIGNMENT,                 ???);
-    add_kv(LLM_KV_GENERAL_NAME,                      model.name);
+    add_kv(LLM_KV_GENERAL_NAME,                      model->name);
     // add_kv(LLM_KV_GENERAL_AUTHOR,                    ???);
     // add_kv(LLM_KV_GENERAL_VERSION,                   ???);
     // add_kv(LLM_KV_GENERAL_URL,                       ???);
@@ -146,8 +156,8 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
     add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
     add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
-    if (hparams.n_embd_out > 0) {
-        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out);
+    if (hparams.n_embd_out_impl > 0) {
+        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out_impl);
     }
     add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
     add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
@@ -176,8 +186,10 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
     add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
     add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
-    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
-    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k_full);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v_full);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA,          hparams.n_embd_head_k_swa);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,        hparams.n_embd_head_v_swa);
     add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
     add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
     add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
@@ -189,7 +201,8 @@ void llama_model_saver::add_kv_from_model() {
 
     const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
 
-    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot_full);
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT_SWA,          hparams.n_rot_swa);
     add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
     // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
     add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));
@@ -255,24 +268,25 @@ void llama_model_saver::add_kv_from_model() {
 }
 
 void llama_model_saver::add_tensors_from_model() {
-    if (std::string(model.output->name) != std::string(model.tok_embd->name)) {
-        add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output
+    if (std::string(model->output->name) != std::string(model->tok_embd->name)) {
+        add_tensor(model->tok_embd); // some models use the same tensor for tok_embd and output
     }
-    add_tensor(model.type_embd);
-    add_tensor(model.pos_embd);
-    add_tensor(model.tok_norm);
-    add_tensor(model.tok_norm_b);
-    add_tensor(model.output_norm);
-    add_tensor(model.output_norm_b);
-    add_tensor(model.output);
-    add_tensor(model.output_b);
-    add_tensor(model.output_norm_enc);
-    add_tensor(model.cls);
-    add_tensor(model.cls_b);
-    add_tensor(model.cls_out);
-    add_tensor(model.cls_out_b);
-
-    for (const struct llama_layer & layer : model.layers) {
+    add_tensor(model->type_embd);
+    add_tensor(model->pos_embd);
+    add_tensor(model->tok_norm);
+    add_tensor(model->tok_norm_b);
+    add_tensor(model->output_norm);
+    add_tensor(model->output_norm_b);
+    add_tensor(model->output);
+    add_tensor(model->output_b);
+    add_tensor(model->output_norm_enc);
+    add_tensor(model->cls);
+    add_tensor(model->cls_b);
+    add_tensor(model->cls_out);
+    add_tensor(model->cls_out_b);
+    add_tensor(model->cls_norm);
+
+    for (const struct llama_layer & layer : model->layers) {
         for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
             add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);
         }

data/ext/sources/examples/talk-llama/llama-model-saver.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include "gguf.h"
 #include "llama.h"
 #include "llama-arch.h"
 
@@ -7,10 +8,12 @@
 
 struct llama_model_saver {
     struct gguf_context * gguf_ctx = nullptr;
-    const struct llama_model & model;
+    const bool gguf_ctx_owned;
+    const struct llama_model * model;
     const struct LLM_KV llm_kv;
 
-    llama_model_saver(const struct llama_model & model);
+    llama_model_saver(const struct llama_model * model);
+    llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx);
     ~llama_model_saver();
 
     void add_kv(enum llm_kv key, uint32_t     value);