vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

format file

Browse source
This commit is contained in:
Michal Moskal 2025-01-26 10:11:39 -08:00
parent 44e1973af0
commit c9e9853e6c
1 changed files with 46 additions and 50 deletions
Download patch file Download diff file Expand all files Collapse all files

96
common/llguidance.cpp
View file

@ -1,24 +1,24 @@
#ifdef LLAMA_USE_LLGUIDANCE #ifdef LLAMA_USE_LLGUIDANCE
#include "common.h" # include "common.h"
#include "sampling.h" # include "sampling.h"
#include "log.h" # include "log.h"
#include "llama.h" # include "llama.h"
#include "llguidance.h" # include "llguidance.h"
struct llama_sampler_llg { struct llama_sampler_llg {
const llama_vocab * vocab; const llama_vocab * vocab;
std::string grammar_kind; std::string grammar_kind;
std::string grammar_data; std::string grammar_data;
LlgTokenizer *tokenizer; LlgTokenizer * tokenizer;
LlgConstraint *grammar; LlgConstraint * grammar;
LlgMaskResult llg_res; LlgMaskResult llg_res;
bool has_llg_res; bool has_llg_res;
}; };
static LlgConstraint *llama_sampler_llg_new(LlgTokenizer *tokenizer, static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
const char * grammar_kind, const char * grammar_data) { const char * grammar_data) {
LlgConstraintInit cinit; LlgConstraintInit cinit;
llg_constraint_init_set_defaults(&cinit, tokenizer); llg_constraint_init_set_defaults(&cinit, tokenizer);
// cinit.log_stderr_level = 2; // cinit.log_stderr_level = 2;
@ -64,7 +64,7 @@ static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array
} }
} }
} else { } else {
const uint32_t *mask = ctx->llg_res.sample_mask; const uint32_t * mask = ctx->llg_res.sample_mask;
for (size_t i = 0; i < cur_p->size; ++i) { for (size_t i = 0; i < cur_p->size; ++i) {
auto token = cur_p->data[i].id; auto token = cur_p->data[i].id;
if ((mask[token / 32] & (1 << (token % 32))) == 0) { if ((mask[token / 32] & (1 << (token % 32))) == 0) {
@ -84,7 +84,7 @@ static void llama_sampler_llg_reset(llama_sampler * smpl) {
auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str()); auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
llg_free_constraint(ctx->grammar); llg_free_constraint(ctx->grammar);
ctx->grammar = grammar_new; ctx->grammar = grammar_new;
ctx->has_llg_res = false; ctx->has_llg_res = false;
} }
@ -100,8 +100,8 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
if (ctx->grammar) { if (ctx->grammar) {
result_ctx->grammar_kind = ctx->grammar_kind; result_ctx->grammar_kind = ctx->grammar_kind;
result_ctx->grammar_data = ctx->grammar_data; result_ctx->grammar_data = ctx->grammar_data;
result_ctx->grammar = llg_clone_constraint(ctx->grammar); result_ctx->grammar = llg_clone_constraint(ctx->grammar);
result_ctx->tokenizer = llg_clone_tokenizer(ctx->tokenizer); result_ctx->tokenizer = llg_clone_tokenizer(ctx->tokenizer);
} }
} }
@ -128,45 +128,42 @@ static llama_sampler_i llama_sampler_llg_i = {
/* .free = */ llama_sampler_llg_free, /* .free = */ llama_sampler_llg_free,
}; };
static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,
static size_t llama_sampler_llg_tokenize_fn(const void *user_data, uint32_t * output_tokens, size_t output_tokens_len) {
const uint8_t *bytes, const llama_vocab * vocab = (const llama_vocab *) user_data;
size_t bytes_len, int r = 0;
uint32_t *output_tokens,
size_t output_tokens_len)
{
const llama_vocab *vocab = (const llama_vocab *)user_data;
int r = 0;
try { try {
r = llama_tokenize(vocab, (const char *) bytes, bytes_len, r = llama_tokenize(vocab, (const char *) bytes, bytes_len, (int32_t *) output_tokens, output_tokens_len, false,
(int32_t*)output_tokens, output_tokens_len, false, true); true);
} catch (const std::exception &e) { } catch (const std::exception & e) {
GGML_ABORT("llama_tokenize failed: %s\n", e.what()); GGML_ABORT("llama_tokenize failed: %s\n", e.what());
} }
if (r < 0) if (r < 0) {
return -r; return -r;
}
return r; return r;
} }
static LlgTokenizer *llama_sampler_llg_new_tokenizer(const llama_vocab * vocab) { static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab) {
// TODO store the tokenizer in the vocab somehow // TODO store the tokenizer in the vocab somehow
static const llama_vocab *vocab_cache; static const llama_vocab * vocab_cache;
static LlgTokenizer *tokenizer_cache; static LlgTokenizer * tokenizer_cache;
if (vocab_cache == vocab) { if (vocab_cache == vocab) {
return llg_clone_tokenizer(tokenizer_cache); return llg_clone_tokenizer(tokenizer_cache);
} }
auto tok_eos = llama_vocab_eot(vocab); auto tok_eos = llama_vocab_eot(vocab);
if (tok_eos == LLAMA_TOKEN_NULL) if (tok_eos == LLAMA_TOKEN_NULL) {
tok_eos = llama_vocab_eos(vocab); tok_eos = llama_vocab_eos(vocab);
}
size_t vocab_size = llama_vocab_n_tokens(vocab); size_t vocab_size = llama_vocab_n_tokens(vocab);
auto token_lens = new uint32_t[vocab_size]; auto token_lens = new uint32_t[vocab_size];
// we typically have ~7 bytes per token; let's go on the safe side here // we typically have ~7 bytes per token; let's go on the safe side here
auto token_bytes_size = vocab_size * 16 + 1024 * 1024; auto token_bytes_size = vocab_size * 16 + 1024 * 1024;
auto token_bytes = new uint8_t[token_bytes_size]; auto token_bytes = new uint8_t[token_bytes_size];
size_t offset = 0; size_t offset = 0;
for (size_t i = 0; i < vocab_size; i++) { for (size_t i = 0; i < vocab_size; i++) {
@ -176,8 +173,8 @@ static LlgTokenizer *llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
} }
llama_token token = i; llama_token token = i;
auto dp = (char *) token_bytes + offset; auto dp = (char *) token_bytes + offset;
auto size = llama_detokenize(vocab, &token, 1, dp, max_token, false, false); auto size = llama_detokenize(vocab, &token, 1, dp, max_token, false, false);
if (size < 0) { if (size < 0) {
GGML_ABORT("llama_detokenize failed\n"); GGML_ABORT("llama_detokenize failed\n");
} }
@ -187,7 +184,7 @@ static LlgTokenizer *llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
GGML_ABORT("llama_detokenize failed\n"); GGML_ABORT("llama_detokenize failed\n");
} }
if (size != 0) { if (size != 0) {
*dp = '\xff'; // special token prefix marker *dp = '\xff'; // special token prefix marker
size += 1; size += 1;
} }
} }
@ -196,10 +193,9 @@ static LlgTokenizer *llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
offset += size; offset += size;
} }
LlgTokenizerInit tinit = { LlgTokenizerInit tinit = {
/* .vocab_size = */ (uint32_t)vocab_size, /* .vocab_size = */ (uint32_t) vocab_size,
/* .tok_eos = */ (uint32_t)tok_eos, /* .tok_eos = */ (uint32_t) tok_eos,
/* .token_lens = */ token_lens, /* .token_lens = */ token_lens,
/* .token_bytes = */ token_bytes, /* .token_bytes = */ token_bytes,
/* .tokenizer_json = */ nullptr, /* .tokenizer_json = */ nullptr,
@ -209,8 +205,8 @@ static LlgTokenizer *llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
/* .tokenize_user_data = */ vocab, /* .tokenize_user_data = */ vocab,
}; };
char error_buffer[1024]; char error_buffer[1024];
LlgTokenizer *tokenizer = llg_new_tokenizer(&tinit, error_buffer, sizeof(error_buffer)); LlgTokenizer * tokenizer = llg_new_tokenizer(&tinit, error_buffer, sizeof(error_buffer));
delete[] token_bytes; delete[] token_bytes;
delete[] token_lens; delete[] token_lens;
@ -223,19 +219,19 @@ static LlgTokenizer *llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
if (tokenizer_cache) { if (tokenizer_cache) {
llg_free_tokenizer(tokenizer_cache); llg_free_tokenizer(tokenizer_cache);
} }
vocab_cache = vocab; vocab_cache = vocab;
tokenizer_cache = tokenizer; tokenizer_cache = tokenizer;
return llg_clone_tokenizer(tokenizer_cache); return llg_clone_tokenizer(tokenizer_cache);
} }
llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * grammar_kind,
const char * grammar_kind, const char * grammar_data) { const char * grammar_data) {
auto * ctx = new llama_sampler_llg; auto * ctx = new llama_sampler_llg;
if (grammar_kind != nullptr && grammar_kind[0] != '\0') { if (grammar_kind != nullptr && grammar_kind[0] != '\0') {
auto tokenizer = llama_sampler_llg_new_tokenizer(vocab); auto tokenizer = llama_sampler_llg_new_tokenizer(vocab);
*ctx = { *ctx = {
/* .vocab = */ vocab, /* .vocab = */ vocab,
/* .grammar_kind = */ grammar_kind, /* .grammar_kind = */ grammar_kind,
/* .grammar_data = */ grammar_data, /* .grammar_data = */ grammar_data,
@ -256,10 +252,10 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,
}; };
} }
return new llama_sampler { return new llama_sampler{
/* .iface = */ &llama_sampler_llg_i, /* .iface = */ &llama_sampler_llg_i,
/* .ctx = */ ctx, /* .ctx = */ ctx,
}; };
} }
#endif // LLAMA_USE_LLGUIDANCE #endif // LLAMA_USE_LLGUIDANCE