extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-16 07:18:25 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp (#3084)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-04-28 16:40:23 +03:00
committed by GitHub
parent 28dcdff4c5
commit f3c42399a3
36 changed files with 16940 additions and 12400 deletions
Download Patch File Download Diff File Expand all files Collapse all files

196
examples/talk-llama/llama.h
View File

@ -60,6 +60,7 @@ extern "C" {
struct llama_model;
struct llama_context;
struct llama_sampler;
struct llama_kv_cache;
typedef int32_t llama_pos;
typedef int32_t llama_token;
@ -105,6 +106,12 @@ extern "C" {
LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26,
LLAMA_VOCAB_PRE_TYPE_MINERVA = 27,
LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28,
LLAMA_VOCAB_PRE_TYPE_GPT4O = 29,
LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30,
LLAMA_VOCAB_PRE_TYPE_TRILLION = 31,
LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32,
LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33,
LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34,
};
enum llama_rope_type {
@ -213,7 +220,7 @@ extern "C" {
LLAMA_SPLIT_MODE_ROW = 2, // split layers and KV across GPUs, use tensor parallelism if supported
};
// TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
// TODO: simplify (https://github.com/ggml-org/llama.cpp/pull/9294#pullrequestreview-2286561979)
typedef struct llama_token_data {
llama_token id; // token id
float logit; // log-odds of the token
@ -275,10 +282,18 @@ extern "C" {
};
};
struct llama_model_tensor_buft_override {
const char * pattern;
ggml_backend_buffer_type_t buft;
};
struct llama_model_params {
// NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
ggml_backend_dev_t * devices;
// NULL-terminated list of buffer types to use for tensors that match a pattern
const struct llama_model_tensor_buft_override * tensor_buft_overrides;
int32_t n_gpu_layers; // number of layers to store in VRAM
enum llama_split_mode split_mode; // how to split the model across multiple GPUs
@ -307,7 +322,7 @@ extern "C" {
};
// NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
// https://github.com/ggerganov/llama.cpp/pull/7544
// https://github.com/ggml-org/llama.cpp/pull/7544
struct llama_context_params {
uint32_t n_ctx; // text context, 0 = from model
uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode
@ -320,7 +335,7 @@ extern "C" {
enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id
enum llama_attention_type attention_type; // attention type to use for embeddings
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
// ref: https://github.com/ggml-org/llama.cpp/pull/2054
float rope_freq_base; // RoPE base frequency, 0 = from model
float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
float yarn_ext_factor; // YaRN extrapolation mix factor, negative = from model
@ -353,17 +368,18 @@ extern "C" {
// model quantization parameters
typedef struct llama_model_quantize_params {
int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
enum llama_ftype ftype; // quantize to this llama_ftype
enum ggml_type output_tensor_type; // output tensor type
enum ggml_type token_embedding_type; // token embeddings tensor type
bool allow_requantize; // allow quantizing non-f32/f16 tensors
bool quantize_output_tensor; // quantize output.weight
bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
bool pure; // quantize all tensors to the default type
bool keep_split; // quantize to the same number of shards
void * imatrix; // pointer to importance matrix data
void * kv_overrides; // pointer to vector containing overrides
int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
enum llama_ftype ftype; // quantize to this llama_ftype
enum ggml_type output_tensor_type; // output tensor type
enum ggml_type token_embedding_type; // token embeddings tensor type
bool allow_requantize; // allow quantizing non-f32/f16 tensors
bool quantize_output_tensor; // quantize output.weight
bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
bool pure; // quantize all tensors to the default type
bool keep_split; // quantize to the same number of shards
void * imatrix; // pointer to importance matrix data
void * kv_overrides; // pointer to vector containing overrides
void * tensor_types; // pointer to vector containing tensor types
} llama_model_quantize_params;
typedef struct llama_logit_bias {
@ -385,7 +401,7 @@ extern "C" {
struct llama_adapter_lora;
// Helpers for getting default parameters
// TODO: update API to start accepting pointers to params structs (https://github.com/ggerganov/llama.cpp/discussions/9172)
// TODO: update API to start accepting pointers to params structs (https://github.com/ggml-org/llama.cpp/discussions/9172)
LLAMA_API struct llama_model_params llama_model_default_params(void);
LLAMA_API struct llama_context_params llama_context_default_params(void);
LLAMA_API struct llama_sampler_chain_params llama_sampler_chain_default_params(void);
@ -468,7 +484,8 @@ extern "C" {
DEPRECATED(LLAMA_API int32_t llama_n_vocab (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");
LLAMA_API const struct llama_model * llama_get_model (const struct llama_context * ctx);
LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
LLAMA_API struct llama_kv_cache * llama_get_kv_self ( struct llama_context * ctx);
LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model);
@ -477,6 +494,7 @@ extern "C" {
LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model);
// Get the model's RoPE frequency scaling factor
LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
@ -584,7 +602,7 @@ extern "C" {
// KV cache
//
// TODO: remove llama_kv_cache_view_* API
// TODO: start using struct llama_kv_cache
// Information associated with an individual cell in the KV cache view.
struct llama_kv_cache_view_cell {
@ -639,13 +657,19 @@ extern "C" {
// Returns the number of tokens in the KV cache (slow, use only for debug)
// If a KV cell has multiple sequences assigned to it, it will be counted multiple times
LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx);
DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx),
"use llama_kv_self_n_tokens instead");
// Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx);
DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx),
"use llama_kv_self_used_cells instead");
// Clear the KV cache - both cell info is erased and KV data is zeroed
LLAMA_API void llama_kv_cache_clear(
LLAMA_API void llama_kv_self_clear(
struct llama_context * ctx);
// Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
@ -653,7 +677,7 @@ extern "C" {
// seq_id < 0 : match any sequence
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API bool llama_kv_cache_seq_rm(
LLAMA_API bool llama_kv_self_seq_rm(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
@ -663,7 +687,7 @@ extern "C" {
// Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API void llama_kv_cache_seq_cp(
LLAMA_API void llama_kv_self_seq_cp(
struct llama_context * ctx,
llama_seq_id seq_id_src,
llama_seq_id seq_id_dst,
@ -671,17 +695,17 @@ extern "C" {
llama_pos p1);
// Removes all tokens that do not belong to the specified sequence
LLAMA_API void llama_kv_cache_seq_keep(
LLAMA_API void llama_kv_self_seq_keep(
struct llama_context * ctx,
llama_seq_id seq_id);
// Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
// If the KV cache is RoPEd, the KV data is updated accordingly:
// - lazily on next llama_decode()
// - explicitly with llama_kv_cache_update()
// - explicitly with llama_kv_self_update()
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API void llama_kv_cache_seq_add(
LLAMA_API void llama_kv_self_seq_add(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
@ -691,10 +715,10 @@ extern "C" {
// Integer division of the positions by factor of `d > 1`
// If the KV cache is RoPEd, the KV data is updated accordingly:
// - lazily on next llama_decode()
// - explicitly with llama_kv_cache_update()
// - explicitly with llama_kv_self_update()
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API void llama_kv_cache_seq_div(
LLAMA_API void llama_kv_self_seq_div(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
@ -702,24 +726,76 @@ extern "C" {
int d);
// Returns the largest position present in the KV cache for the specified sequence
LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
LLAMA_API llama_pos llama_kv_self_seq_pos_max(
struct llama_context * ctx,
llama_seq_id seq_id);
// TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache
// how to avoid this?
llama_seq_id seq_id);
// Defragment the KV cache
// This will be applied:
// - lazily on next llama_decode()
// - explicitly with llama_kv_cache_update()
LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
// Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
// - explicitly with llama_kv_self_update()
LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
// Check if the context supports KV cache shifting
LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
// Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
DEPRECATED(LLAMA_API void llama_kv_cache_clear(
struct llama_context * ctx),
"use llama_kv_self_clear instead");
DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
llama_pos p1),
"use llama_kv_self_seq_rm instead");
DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp(
struct llama_context * ctx,
llama_seq_id seq_id_src,
llama_seq_id seq_id_dst,
llama_pos p0,
llama_pos p1),
"use llama_kv_self_seq_cp instead");
DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep(
struct llama_context * ctx,
llama_seq_id seq_id),
"use llama_kv_self_seq_keep instead");
DEPRECATED(LLAMA_API void llama_kv_cache_seq_add(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
llama_pos p1,
llama_pos delta),
"use llama_kv_self_seq_add instead");
DEPRECATED(LLAMA_API void llama_kv_cache_seq_div(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
llama_pos p1,
int d),
"use llama_kv_self_seq_div instead");
DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
struct llama_context * ctx,
llama_seq_id seq_id),
"use llama_kv_self_seq_pos_max instead");
DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx),
"use llama_kv_self_defrag instead");
DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx),
"use llama_kv_self_can_shift instead");
DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx),
"use llama_kv_self_update instead");
//
// State / sessions
@ -883,6 +959,10 @@ extern "C" {
// If set to true, the model will only attend to the past tokens
LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
// Set whether the model is in warmup mode or not
// If true, all model tensors are activated during llama_decode() to load and cache their weights.
LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup);
// Set abort callback
LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
@ -1040,7 +1120,7 @@ extern "C" {
/// Apply chat template. Inspired by hf apply_chat_template() on python.
/// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
/// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
/// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
/// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
/// @param chat Pointer to a list of multiple llama_chat_message
/// @param n_msg Number of llama_chat_message in this chat
@ -1114,11 +1194,12 @@ extern "C" {
};
struct llama_sampler {
struct llama_sampler_i * iface;
llama_sampler_context_t ctx;
const struct llama_sampler_i * iface;
llama_sampler_context_t ctx;
};
// mirror of llama_sampler_i:
LLAMA_API struct llama_sampler * llama_sampler_init (const struct llama_sampler_i * iface, llama_sampler_context_t ctx);
LLAMA_API const char * llama_sampler_name (const struct llama_sampler * smpl);
LLAMA_API void llama_sampler_accept( struct llama_sampler * smpl, llama_token token);
LLAMA_API void llama_sampler_apply ( struct llama_sampler * smpl, llama_token_data_array * cur_p);
@ -1148,7 +1229,7 @@ extern "C" {
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
/// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void),
"will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
"will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k);
@ -1156,7 +1237,7 @@ extern "C" {
/// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API struct llama_sampler * llama_sampler_init_top_p (float p, size_t min_keep);
/// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
/// @details Minimum P sampling as described in https://github.com/ggml-org/llama.cpp/pull/3841
LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep);
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
@ -1171,6 +1252,9 @@ extern "C" {
/// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
LLAMA_API struct llama_sampler * llama_sampler_init_xtc (float p, float t, size_t min_keep, uint32_t seed);
/// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641
LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float n);
/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@ -1194,22 +1278,38 @@ extern "C" {
float tau,
float eta);
/// @details Intializes a GBNF grammar, see grammars/README.md for details.
/// @param vocab The vocabulary that this grammar will be used with.
/// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
/// @param grammar_root The name of the start symbol for the grammar.
LLAMA_API struct llama_sampler * llama_sampler_init_grammar(
const struct llama_vocab * vocab,
const char * grammar_str,
const char * grammar_root);
/// @details Lazy grammar sampler, introduced in https://github.com/ggerganov/llama.cpp/pull/9639
/// @param trigger_words A list of words that will trigger the grammar sampler. This may be updated to a loose regex syntax (w/ ^) in a near future.
/// @param trigger_tokens A list of tokens that will trigger the grammar sampler.
LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
const struct llama_vocab * vocab,
const char * grammar_str,
const char * grammar_root,
const char ** trigger_words,
size_t num_trigger_words,
const llama_token * trigger_tokens,
size_t num_trigger_tokens);
size_t num_trigger_tokens),
"use llama_sampler_init_grammar_lazy_patterns instead");
/// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639
/// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group.
/// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included.
LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
const struct llama_vocab * vocab,
const char * grammar_str,
const char * grammar_root,
const char ** trigger_patterns,
size_t num_trigger_patterns,
const llama_token * trigger_tokens,
size_t num_trigger_tokens);
/// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
LLAMA_API struct llama_sampler * llama_sampler_init_penalties(