CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/refactor-kv-cache

Browse Source
This commit is contained in:
Francis Couture-Harpin
2024-11-25 10:40:20 -05:00
parent 8006f3b3c8 cce5a90075
commit 691698e152
375 changed files with 86333 additions and 53885 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
include/llama.h
View File

@@ -2,6 +2,7 @@
#define LLAMA_H
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-backend.h"
#include <stddef.h>
@@ -205,7 +206,7 @@ extern "C" {
enum llama_split_mode {
LLAMA_SPLIT_MODE_NONE = 0, // single GPU
LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs
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)
@@ -217,6 +218,7 @@ extern "C" {
typedef struct llama_token_data_array {
// TODO: consider SoA
// NOTE: this pointer can be modified by the samplers
llama_token_data * data;
size_t size;
int64_t selected; // this is the index in the data array (i.e. not the token id)
@@ -232,8 +234,11 @@ extern "C" {
// - token : the token ids of the input (used when embd is NULL)
// - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
// - pos : the positions of the respective token in the sequence
// (if set to NULL, the token position will be tracked automatically by llama_decode)
// - seq_id : the sequence to which the respective token belongs
// (if set to NULL, the sequence ID will be assumed to be 0)
// - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
// (if set to NULL, only the logits for last token will be returned)
//
typedef struct llama_batch {
int32_t n_tokens;
@@ -244,15 +249,6 @@ extern "C" {
int32_t * n_seq_id;
llama_seq_id ** seq_id;
int8_t * logits; // TODO: rename this to "output"
// NOTE: helpers for smooth API transition - can be deprecated in the future
// for future-proof code, use the above fields instead and ignore everything below
//
// pos[i] = all_pos_0 + i*all_pos_1
//
llama_pos all_pos_0; // used if pos == NULL
llama_pos all_pos_1; // used if pos == NULL
llama_seq_id all_seq_id; // used if seq_id == NULL
} llama_batch;
enum llama_model_kv_override_type {
@@ -279,10 +275,7 @@ extern "C" {
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
// main_gpu interpretation depends on split_mode:
// LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
// LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
// LLAMA_SPLIT_MODE_LAYER: ignored
// the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
int32_t main_gpu;
// proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -683,6 +676,9 @@ extern "C" {
// Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
// Check if the context supports KV cache shifting
LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
//
// State / sessions
//
@@ -785,15 +781,15 @@ extern "C" {
// Decoding
//
// Return batch for single sequence of tokens starting at pos_0
// Return batch for single sequence of tokens
// The sequence ID will be fixed to 0
// The position of the tokens will be tracked automatically by llama_decode
//
// NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
//
LLAMA_API struct llama_batch llama_batch_get_one(
llama_token * tokens,
int32_t n_tokens,
llama_pos pos_0,
llama_seq_id seq_id);
int32_t n_tokens);
// Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
// Each token can be assigned up to n_seq_max sequence ids
@@ -813,7 +809,7 @@ extern "C" {
// 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.
// 0 - success
// < 0 - error
// < 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);
@@ -821,7 +817,7 @@ extern "C" {
// 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)
// < 0 - error
// < 0 - error. the KV cache state is restored to the state before this call
LLAMA_API int32_t llama_decode(
struct llama_context * ctx,
struct llama_batch batch);
@@ -1084,12 +1080,13 @@ extern "C" {
// available samplers:
LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void);
LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed);
LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed);
/// @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.
LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void);
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)");
/// @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);
@@ -1100,16 +1097,18 @@ extern "C" {
/// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep);
/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
LLAMA_API struct llama_sampler * llama_sampler_init_tail_free (float z, size_t min_keep);
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep);
/// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t);
/// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext (float t, float delta, float exponent);
/// @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 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.
@@ -1149,11 +1148,43 @@ extern "C" {
bool penalize_nl, // consider newlines as a repeatable token
bool ignore_eos); // ignore the end-of-sequence token
/// @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
LLAMA_API struct llama_sampler * llama_sampler_init_dry(
const struct llama_model * model,
float dry_multiplier,
float dry_base,
int32_t dry_allowed_length,
int32_t dry_penalty_last_n,
const char ** seq_breakers,
size_t num_breakers);
LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
int32_t n_vocab,
int32_t n_logit_bias,
const llama_logit_bias * logit_bias);
// this sampler is meant to be used for fill-in-the-middle infilling
// it's supposed to be used after top_k + top_p sampling
//
// 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
// 2. combine probs of tokens that have the same prefix
//
// example:
//
// - before:
// "hel": 0.5
// "hell": 0.2
// "hello": 0.1
// "dummy": 0.1
//
// - after:
// "hel": 0.8
// "dummy": 0.1
//
// 3. discard non-EOG tokens with low prob
// 4. if no tokens are left -> pick EOT
//
LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
// Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
@@ -1225,8 +1256,6 @@ 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);
LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
#ifdef __cplusplus
}
#endif