starred/ollama-ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-04-18 16:23:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Update vendored llama.cpp to b7847" (#14061)

Browse Source
This commit is contained in:
Jeffrey Morgan
2026-02-03 18:39:36 -08:00
committed by GitHub
parent a6355329bf
commit b1fccabb34
240 changed files with 5050 additions and 21247 deletions
Download Patch File Download Diff File Expand all files Collapse all files

229
llama/llama.cpp/common/sampling.cpp vendored
View File

@@ -104,9 +104,10 @@ struct ring_buffer {
struct common_sampler {
common_params_sampling params;
struct llama_sampler * grmr;
struct llama_sampler * chain;
bool grammar;
ring_buffer<llama_token> prev;
std::vector<llama_token_data> cur;
@@ -120,34 +121,17 @@ struct common_sampler {
}
void set_logits(struct llama_context * ctx, int idx) {
const float * sampled_probs = llama_get_sampled_probs_ith (ctx, idx);
const float * sampled_logits = llama_get_sampled_logits_ith (ctx, idx);
const llama_token * sampled_ids = llama_get_sampled_candidates_ith(ctx, idx);
const auto * logits = llama_get_logits_ith(ctx, idx);
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_vocab = llama_vocab_n_tokens(vocab);
if (sampled_probs) {
const uint32_t sampled_probs_count = llama_get_sampled_probs_count_ith(ctx, idx);
cur.resize(sampled_probs_count);
for (uint32_t i = 0; i < sampled_probs_count; ++i) {
cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], sampled_probs[i]};
}
} else if (sampled_logits) {
const uint32_t sampled_logits_count = llama_get_sampled_logits_count_ith(ctx, idx);
cur.resize(sampled_logits_count);
for (uint32_t i = 0; i < sampled_logits_count; i++) {
cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], 0.0f};
}
} else {
const auto * logits = llama_get_logits_ith(ctx, idx);
GGML_ASSERT(logits != nullptr);
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
}
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
}
cur_p = { cur.data(), cur.size(), -1, false };
@@ -167,59 +151,54 @@ std::string common_params_sampling::print() const {
"\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
"\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
"\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, top_n_sigma = %.3f, temp = %.3f\n"
"\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f, adaptive_target = %.3f, adaptive_decay = %.3f",
"\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, top_n_sigma, temp,
mirostat, mirostat_eta, mirostat_tau, adaptive_target, adaptive_decay);
mirostat, mirostat_eta, mirostat_tau);
return std::string(result);
}
struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = params.no_perf;
llama_sampler * grmr = nullptr;
llama_sampler * chain = llama_sampler_chain_init(lparams);
bool grammar = false;
std::vector<llama_sampler *> samplers;
if (params.grammar.compare(0, 11, "%llguidance") == 0) {
#ifdef LLAMA_USE_LLGUIDANCE
grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
samplers.push_back(llama_sampler_init_llg(vocab, "lark", params.grammar.c_str()));
grammar = true;
#else
GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
#endif // LLAMA_USE_LLGUIDANCE
} else {
std::vector<std::string> trigger_patterns;
std::vector<std::string> patterns_anywhere;
std::vector<llama_token> trigger_tokens;
for (const auto & trigger : params.grammar_triggers) {
switch (trigger.type) {
case COMMON_GRAMMAR_TRIGGER_TYPE_WORD:
{
const auto & word = trigger.value;
trigger_patterns.push_back(regex_escape(word));
patterns_anywhere.push_back(regex_escape(word));
break;
}
case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
{
trigger_patterns.push_back(trigger.value);
patterns_anywhere.push_back(trigger.value);
break;
}
case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
{
const auto & pattern = trigger.value;
std::string anchored = "^$";
if (!pattern.empty()) {
anchored = (pattern.front() != '^' ? "^" : "")
+ pattern
+ (pattern.back() != '$' ? "$" : "");
}
trigger_patterns.push_back(anchored);
trigger_patterns.push_back(trigger.value);
break;
}
case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -233,6 +212,10 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
}
}
if (!patterns_anywhere.empty()) {
trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
}
std::vector<const char *> trigger_patterns_c;
trigger_patterns_c.reserve(trigger_patterns.size());
for (const auto & regex : trigger_patterns) {
@@ -241,12 +224,15 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
if (!params.grammar.empty()) {
if (params.grammar_lazy) {
grmr = llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
trigger_patterns_c.data(), trigger_patterns_c.size(),
trigger_tokens.data(), trigger_tokens.size());
samplers.push_back(
llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
trigger_patterns_c.data(), trigger_patterns_c.size(),
trigger_tokens.data(), trigger_tokens.size()));
} else {
grmr = llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
samplers.push_back(llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"));
}
grammar = true;
}
}
@@ -255,9 +241,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
}
if (params.mirostat == 0) {
bool use_adaptive_p = false; // see below
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY:
@@ -267,54 +250,43 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
for (const auto & str : params.dry_sequence_breakers) {
c_breakers.push_back(str.c_str());
}
samplers.push_back(llama_sampler_init_dry(vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
samplers.push_back(llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_K:
samplers.push_back(llama_sampler_init_top_k(params.top_k));
samplers.push_back(llama_sampler_init_top_k (params.top_k));
break;
case COMMON_SAMPLER_TYPE_TOP_P:
samplers.push_back(llama_sampler_init_top_p(params.top_p, params.min_keep));
samplers.push_back(llama_sampler_init_top_p (params.top_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
samplers.push_back(llama_sampler_init_top_n_sigma(params.top_n_sigma));
break;
case COMMON_SAMPLER_TYPE_MIN_P:
samplers.push_back(llama_sampler_init_min_p(params.min_p, params.min_keep));
samplers.push_back(llama_sampler_init_min_p (params.min_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_XTC:
samplers.push_back(llama_sampler_init_xtc(params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
samplers.push_back(llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
break;
case COMMON_SAMPLER_TYPE_TYPICAL_P:
samplers.push_back(llama_sampler_init_typical(params.typ_p, params.min_keep));
samplers.push_back(llama_sampler_init_typical (params.typ_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TEMPERATURE:
samplers.push_back(llama_sampler_init_temp_ext(params.temp, params.dynatemp_range, params.dynatemp_exponent));
samplers.push_back(llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
samplers.push_back(llama_sampler_init_infill(vocab));
samplers.push_back(llama_sampler_init_infill (vocab));
break;
case COMMON_SAMPLER_TYPE_PENALTIES:
samplers.push_back(llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
break;
case COMMON_SAMPLER_TYPE_ADAPTIVE_P:
// the `adaptive-p` sampler is like `dist` and `mirostat` in that it selects
// a single token, so we will add `dist` at the end of the chain by default,
// unless the user specifically included `adaptive-p`. we set this flag here
// so we know to add the sampler at the very end.
use_adaptive_p = true;
samplers.push_back(llama_sampler_init_penalties (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
}
}
if (use_adaptive_p) {
// only if user explicitly included adaptive-p sampler
samplers.push_back(llama_sampler_init_adaptive_p(params.adaptive_target, params.adaptive_decay, params.seed));
} else {
// default: sample from distribution
samplers.push_back(llama_sampler_init_dist(params.seed));
}
samplers.push_back(llama_sampler_init_dist(params.seed));
} else if (params.mirostat == 1) {
samplers.push_back(llama_sampler_init_temp(params.temp));
samplers.push_back(llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
@@ -329,16 +301,10 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
llama_sampler_chain_add(chain, smpl);
}
if (grmr && params.backend_sampling) {
LOG_WRN("%s: backend sampling is not compatible with grammar, disabling\n", __func__);
params.backend_sampling = false;
}
auto * result = new common_sampler {
/* .params = */ params,
/* .grmr = */ grmr,
/* .chain = */ chain,
/* .grammar = */ grammar,
/* .prev = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
/* .cur = */ {},
/* .cur_p = */ {},
@@ -348,45 +314,47 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
}
void common_sampler_free(struct common_sampler * gsmpl) {
if (!gsmpl) {
return;
if (gsmpl) {
llama_sampler_free(gsmpl->chain);
delete gsmpl;
}
llama_sampler_free(gsmpl->grmr);
llama_sampler_free(gsmpl->chain);
delete gsmpl;
}
void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
if (!gsmpl) {
return;
}
const auto tm = gsmpl->tm();
if (gsmpl->grmr && accept_grammar) {
llama_sampler_accept(gsmpl->grmr, token);
}
if (gsmpl->grammar) {
const int n_smpl = llama_sampler_chain_n(gsmpl->chain);
llama_sampler_accept(gsmpl->chain, token);
for (int i = 0; i < n_smpl; i++) {
auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
// the grammar sampler is always the first one
if (i == 0) {
if (accept_grammar) {
llama_sampler_accept(smpl, token);
}
} else {
llama_sampler_accept(smpl, token);
}
}
} else {
llama_sampler_accept(gsmpl->chain, token);
}
gsmpl->prev.push_back(token);
}
void common_sampler_reset(struct common_sampler * gsmpl) {
if (!gsmpl) {
return;
}
gsmpl->reset();
}
struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
return new common_sampler {
/* .params = */ gsmpl->params,
/* .grmr = */ llama_sampler_clone(gsmpl->grmr),
/* .chain = */ llama_sampler_clone(gsmpl->chain),
/* .grammar = */ gsmpl->grammar,
/* .prev = */ gsmpl->prev,
/* .cur = */ gsmpl->cur,
/* .cur_p = */ gsmpl->cur_p,
@@ -439,14 +407,10 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
}
struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl) {
if (!gsmpl) {
return nullptr;
}
return gsmpl->chain;
}
llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx) {
llama_synchronize(ctx);
// start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
@@ -454,61 +418,11 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
llama_token id = LLAMA_TOKEN_NULL;
auto & grmr = gsmpl->grmr;
auto & chain = gsmpl->chain;
auto & cur_p = gsmpl->cur_p; // initialized by set_logits
// Check if a backend sampler has already sampled a token in which case we
// return that token id directly.
{
id = llama_get_sampled_token_ith(ctx, idx);
if (id != LLAMA_TOKEN_NULL) {
LOG_DBG("%s: Backend sampler selected token: '%d'. Will not run any CPU samplers\n", __func__, id);
GGML_ASSERT(!gsmpl->grmr && "using grammar in combination with backend sampling is not supported");
// TODO: simplify
gsmpl->cur.resize(1);
gsmpl->cur[0] = { id, 0.0f, 1.0f };
cur_p = { gsmpl->cur.data(), gsmpl->cur.size(), 0, true };
return id;
}
}
gsmpl->set_logits(ctx, idx);
if (grammar_first) {
llama_sampler_apply(grmr, &cur_p);
}
llama_sampler_apply(chain, &cur_p);
id = cur_p.data[cur_p.selected].id;
if (grammar_first) {
return id;
}
// check if it the sampled token fits the grammar (grammar-based rejection sampling)
{
llama_token_data single_token_data = { id, 1.0f, 0.0f };
llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
llama_sampler_apply(grmr, &single_token_data_array);
const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
if (is_valid) {
return id;
}
}
// resampling:
// if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
gsmpl->set_logits(ctx, idx);
llama_sampler_apply(grmr, &cur_p);
llama_sampler_apply(chain, &cur_p);
GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
@@ -518,7 +432,7 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
return id;
}
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft) {
GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
std::vector<llama_token> result;
@@ -526,7 +440,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
size_t i = 0;
for (; i < draft.size(); i++) {
const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
common_sampler_accept(gsmpl, id, true);
@@ -538,7 +452,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
}
if (i == draft.size()) {
const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
common_sampler_accept(gsmpl, id, true);
@@ -548,13 +462,13 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
return result;
}
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft) {
std::vector<int> idxs(draft.size() + 1);
for (size_t i = 0; i < idxs.size(); ++i) {
idxs[i] = i;
}
return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft);
}
uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
@@ -639,7 +553,6 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
case COMMON_SAMPLER_TYPE_XTC: return 'x';
case COMMON_SAMPLER_TYPE_INFILL: return 'i';
case COMMON_SAMPLER_TYPE_PENALTIES: return 'e';
case COMMON_SAMPLER_TYPE_ADAPTIVE_P: return 'a';
default : return '?';
}
}
@@ -656,7 +569,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
case COMMON_SAMPLER_TYPE_XTC: return "xtc";
case COMMON_SAMPLER_TYPE_INFILL: return "infill";
case COMMON_SAMPLER_TYPE_PENALTIES: return "penalties";
case COMMON_SAMPLER_TYPE_ADAPTIVE_P: return "adaptive_p";
default : return "";
}
}
@@ -673,7 +585,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
{ "xtc", COMMON_SAMPLER_TYPE_XTC },
{ "infill", COMMON_SAMPLER_TYPE_INFILL },
{ "penalties", COMMON_SAMPLER_TYPE_PENALTIES },
{ "adaptive_p", COMMON_SAMPLER_TYPE_ADAPTIVE_P },
};
// since samplers names are written multiple ways
@@ -689,7 +600,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
{ "typ", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "min-p", COMMON_SAMPLER_TYPE_MIN_P },
{ "temp", COMMON_SAMPLER_TYPE_TEMPERATURE },
{ "adaptive-p", COMMON_SAMPLER_TYPE_ADAPTIVE_P },
};
std::vector<common_sampler_type> samplers;
@@ -726,7 +636,6 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC), COMMON_SAMPLER_TYPE_XTC },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL), COMMON_SAMPLER_TYPE_INFILL },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_PENALTIES), COMMON_SAMPLER_TYPE_PENALTIES },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_ADAPTIVE_P), COMMON_SAMPLER_TYPE_ADAPTIVE_P },
};
std::vector<common_sampler_type> samplers;