starred/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-04-26 02:36:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

feat: llama.cpp bump (17f7f4) for SSM performance improvements (#13408)

Browse Source 
* feat: Bump llama.cpp to the latest master (17f7f4b)

This brings in significant improvements to prefill performance for all
models using the SSM_CONV and SSM_SCAN ops (granite4, jamba, falcon-h,
nemotron-h, Qwen3 Next) on Apple Metal.

See https://github.com/ggml-org/llama.cpp/pull/17876

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patches 1-4

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Update patches 5-12

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patches 13-18

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patch 20

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patches 21-31

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Sync vendored code

The two files I'm not sure about here are the swap from gemma3-iswa.cpp to
gemma3.cpp (I chose to include this because I think it's required), and the
inclusion of `ggml-zendnn.h` which I chose to omit.

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

---------

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart
2025-12-10 13:59:27 -07:00
committed by GitHub
parent c34fc64688
commit b95693056c
115 changed files with 5176 additions and 2585 deletions
Download Patch File Download Diff File Expand all files Collapse all files

265
llama/llama.cpp/src/llama-grammar.cpp vendored
View File

@@ -181,6 +181,52 @@ static std::pair<uint32_t, const char *> parse_char(const char * src) {
throw std::runtime_error("unexpected end of input");
}
static std::pair<uint32_t, const char *> parse_token(const llama_vocab * vocab, const char * src) {
const char * pos = src;
if (*pos != '<') {
throw std::runtime_error(std::string("expecting '<' at ") + pos);
}
pos++;
// Parse <[id]>
if (*pos == '[') {
pos++;
const char * int_end = parse_int(pos);
uint32_t token_id = std::stoul(std::string(pos, int_end - pos));
pos = int_end;
if (*pos != ']') {
throw std::runtime_error(std::string("expecting ']' at ") + pos);
}
pos++;
if (*pos != '>') {
throw std::runtime_error(std::string("expecting '>' at ") + pos);
}
pos++;
return std::make_pair(token_id, pos);
}
if (vocab == nullptr) {
throw std::runtime_error(std::string("no vocab to parse token at ") + src);
}
// Parse <token> and tokenize to obtain the token id
while (*pos != 0 && *pos != '>') {
pos++;
}
if (*pos != '>') {
throw std::runtime_error(std::string("expecting '>' at ") + pos);
}
pos++;
llama_token tokens[2];
int32_t n_tokens = vocab->tokenize(src, static_cast<int32_t>(pos - src), tokens, 2, false, true);
if (n_tokens != 1) {
// must tokenize to exactly 1 token
throw std::runtime_error("invalid token '" + std::string(src, pos - src) + "'");
}
return std::make_pair(tokens[0], pos);
}
static void print_grammar_char(FILE * file, uint32_t c) {
if (0x20 <= c && c <= 0x7f) {
fprintf(file, "%c", static_cast<char>(c));
@@ -212,6 +258,8 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
case LLAMA_GRETYPE_CHAR_ALT: fprintf(file, "CHAR_ALT"); break;
case LLAMA_GRETYPE_CHAR_ANY: fprintf(file, "CHAR_ANY"); break;
case LLAMA_GRETYPE_TOKEN: fprintf(file, "TOKEN"); break;
case LLAMA_GRETYPE_TOKEN_NOT: fprintf(file, "TOKEN_NOT"); break;
}
switch (elem.type) {
case LLAMA_GRETYPE_END:
@@ -228,6 +276,17 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
print_grammar_char(file, elem.value);
fprintf(file, "\") ");
break;
case LLAMA_GRETYPE_TOKEN:
fprintf(file, "<[");
fprintf(file, "%u", elem.value);
fprintf(file, "]> ");
break;
case LLAMA_GRETYPE_TOKEN_NOT:
fprintf(file, "!");
fprintf(file, "<[");
fprintf(file, "%u", elem.value);
fprintf(file, "]> ");
break;
}
}
fprintf(file, "\n");
@@ -284,6 +343,17 @@ static void print_rule(
case LLAMA_GRETYPE_CHAR_ANY:
fprintf(file, ".");
break;
case LLAMA_GRETYPE_TOKEN:
fprintf(file, "<[");
fprintf(file, "%u", elem.value);
fprintf(file, "]> ");
break;
case LLAMA_GRETYPE_TOKEN_NOT:
fprintf(file, "!");
fprintf(file, "<[");
fprintf(file, "%u", elem.value);
fprintf(file, "]> ");
break;
}
if (is_char_element(elem)) {
switch (rule[i + 1].type) {
@@ -444,6 +514,17 @@ const char * llama_grammar_parser::parse_sequence(
}
}
pos = parse_space(pos + 1, is_nested);
} else if (*pos == '<' || *pos == '!') { // token
auto type = LLAMA_GRETYPE_TOKEN;
if (*pos == '!') { // token inverse
type = LLAMA_GRETYPE_TOKEN_NOT;
pos++;
}
auto token_pair = parse_token(vocab, pos);
const char * token_end = token_pair.second;
last_sym_start = rule.size();
rule.push_back({type, token_pair.first});
pos = parse_space(token_end, is_nested);
} else if (is_word_char(*pos)) { // rule reference
const char * name_end = parse_name(pos);
uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
@@ -691,6 +772,21 @@ static bool llama_grammar_match_partial_char(
return !is_positive_char;
}
// returns true iff token matches the rule at pos (regular or inverse)
// asserts that pos is pointing to a token element
static bool llama_grammar_match_token(
const llama_grammar_element * pos,
const llama_token token) {
GGML_ASSERT(pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT);
if (pos->type == LLAMA_GRETYPE_TOKEN) {
return pos->value == static_cast<uint32_t>(token);
}
if (pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
return pos->value != static_cast<uint32_t>(token);
}
return false;
}
// transforms a grammar pushdown stack into N possible stacks, all ending
// at a character range (terminal element)
static void llama_grammar_advance_stack(
@@ -738,6 +834,8 @@ static void llama_grammar_advance_stack(
case LLAMA_GRETYPE_CHAR:
case LLAMA_GRETYPE_CHAR_NOT:
case LLAMA_GRETYPE_CHAR_ANY:
case LLAMA_GRETYPE_TOKEN:
case LLAMA_GRETYPE_TOKEN_NOT:
if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
// only add the stack if it's not a duplicate of one we already have
new_stacks.emplace_back(stack);
@@ -831,26 +929,38 @@ llama_grammar_stacks & llama_grammar_get_stacks(struct llama_grammar * grammar)
return grammar->stacks;
}
static void llama_grammar_accept_chr(
struct llama_grammar & grammar,
const llama_grammar_stack & stack,
uint32_t chr,
llama_grammar_stacks & new_stacks) {
if (stack.empty()) {
return;
}
const llama_grammar_element * pos = stack.back();
// ignore if this turns into a token
if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
return;
}
auto match = llama_grammar_match_char(pos, chr);
if (match.first) {
llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
if (!llama_grammar_is_end_of_sequence(match.second)) {
new_stack.push_back(match.second);
}
llama_grammar_advance_stack(grammar.rules, new_stack, new_stacks);
}
}
void llama_grammar_accept(struct llama_grammar * grammar, uint32_t chr) {
llama_grammar_stacks stacks_new;
stacks_new.reserve(grammar->stacks.size());
for (const auto & stack : grammar->stacks) {
if (stack.empty()) {
continue;
}
auto match = llama_grammar_match_char(stack.back(), chr);
if (match.first) {
const llama_grammar_element * pos = match.second;
// update top of stack to next element, if any
llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
if (!llama_grammar_is_end_of_sequence(pos)) {
new_stack.push_back(pos);
}
llama_grammar_advance_stack(grammar->rules, new_stack, stacks_new);
}
llama_grammar_accept_chr(*grammar, stack, chr, stacks_new);
}
grammar->stacks = std::move(stacks_new);
@@ -875,6 +985,22 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
const llama_grammar_element * stack_pos = stack.back();
// if the top of the stack is a token rule, then we only need to check the token id
if (stack_pos->type == LLAMA_GRETYPE_TOKEN || stack_pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
for (const auto & tok : candidates) {
if (*tok.code_points == 0) {
// reached the end of a token consumed by char rules, reject iff it ended
// in a partial response
if (tok.partial_utf8.n_remain != 0) {
rejects.push_back(tok);
}
} else if (!llama_grammar_match_token(stack_pos, tok.id)) {
rejects.push_back(tok);
}
}
return rejects;
}
llama_grammar_candidates next_candidates;
next_candidates.reserve(candidates.size());
@@ -887,7 +1013,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
rejects.push_back(tok);
}
} else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) {
next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8 });
next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8, tok.id });
} else {
rejects.push_back(tok);
}
@@ -905,7 +1031,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
for (const auto & tok : next_rejects) {
rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 });
rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8, tok.id });
}
return rejects;
@@ -974,12 +1100,13 @@ struct llama_grammar * llama_grammar_init_impl(
ollama_vocab,
std::move(vec_rules),
std::move(stacks),
/* .partial_utf8 = */ {},
/* .lazy =*/ false,
/* .awaiting_trigger = */ false,
/* .trigger_buffer = */ "",
/* .trigger_tokens = */ {},
/* .trigger_patterns = */ {},
/* .partial_utf8 = */ {},
/* .lazy = */ false,
/* .awaiting_trigger = */ false,
/* .trigger_buffer = */ "",
/* .trigger_buffer_positions = */ {},
/* .trigger_tokens = */ {},
/* .trigger_patterns = */ {},
};
}
@@ -993,7 +1120,7 @@ struct llama_grammar * llama_grammar_init_impl(
size_t num_trigger_patterns,
const llama_token * trigger_tokens,
size_t num_trigger_tokens) {
llama_grammar_parser parser;
llama_grammar_parser parser(vocab);
// if there is a grammar, parse it
// rules will be empty (default) if there are parse errors
@@ -1081,10 +1208,11 @@ struct llama_grammar * llama_grammar_init_impl(
ollama_vocab,
std::move(vec_rules),
std::move(stacks),
/* .partial_utf8 = */ {},
/* .lazy = */ lazy,
/* .awaiting_trigger = */ lazy,
/* .trigger_buffer = */ "",
/* .partial_utf8 = */ {},
/* .lazy = */ lazy,
/* .awaiting_trigger = */ lazy,
/* .trigger_buffer = */ "",
/* .trigger_buffer_positions = */ {},
std::move(vec_trigger_tokens),
std::move(vec_trigger_patterns),
};
@@ -1108,6 +1236,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
grammar.lazy,
grammar.awaiting_trigger,
grammar.trigger_buffer,
grammar.trigger_buffer_positions,
grammar.trigger_tokens,
grammar.trigger_patterns,
};
@@ -1164,7 +1293,7 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
cur_p->data[i].logit = -INFINITY;
} else {
candidates_decoded.push_back(decode_utf8(piece, grammar.partial_utf8));
candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second, id });
}
}
@@ -1184,10 +1313,12 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
grammar.awaiting_trigger = false;
grammar.trigger_buffer.clear();
llama_grammar_accept_str(grammar, piece);
llama_grammar_accept_token(grammar, token, piece);
LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
return;
} else {
auto position = std::make_pair(grammar.trigger_buffer.size(), grammar.trigger_buffer.size() + piece.size());
grammar.trigger_buffer_positions.push_back(std::make_pair(token, position));
grammar.trigger_buffer += piece;
std::smatch match;
@@ -1205,10 +1336,23 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
if (start == std::string::npos) {
start = match.position(0);
}
// replay tokens that overlap with [start, end)
for (const auto & [tok, tok_pos] : grammar.trigger_buffer_positions) {
auto [tok_start, tok_end] = tok_pos;
if (tok_end <= start) {
continue;
}
size_t piece_start = (tok_start < start) ? start : tok_start; // allow for partial token pieces
size_t piece_len = tok_end - piece_start;
auto tok_piece = grammar.trigger_buffer.substr(piece_start, piece_len);
llama_grammar_accept_token(grammar, tok, tok_piece);
}
auto constrained_str = grammar.trigger_buffer.substr(start);
// std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
grammar.trigger_buffer.clear();
llama_grammar_accept_str(grammar, constrained_str);
grammar.trigger_buffer_positions.clear();
LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
return;
}
@@ -1228,7 +1372,7 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
GGML_ABORT("grammar error: end of grammar token received but grammar stack is not empty");
}
llama_grammar_accept_str(grammar, piece);
llama_grammar_accept_token(grammar, token, piece);
}
void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
@@ -1246,6 +1390,61 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
}
}
void llama_grammar_accept_token(struct llama_grammar & grammar, llama_token token, const std::string & piece) {
// Note terminating 0 in decoded string
const auto decoded = decode_utf8(piece, grammar.partial_utf8);
const auto & code_points = decoded.first;
llama_grammar_stacks stacks_new;
stacks_new.reserve(grammar.stacks.size());
for (const auto & stack : grammar.stacks) {
if (stack.empty()) {
continue;
}
const llama_grammar_element * pos = stack.back();
if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
if (llama_grammar_match_token(pos, token)) {
llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
if (!llama_grammar_is_end_of_sequence(pos + 1)) {
new_stack.push_back(pos + 1);
}
llama_grammar_advance_stack(grammar.rules, new_stack, stacks_new);
}
} else {
llama_grammar_stacks current_stacks = {stack};
for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
llama_grammar_stacks next_stacks;
for (const auto & cur_stack : current_stacks) {
llama_grammar_accept_chr(grammar, cur_stack, *it, next_stacks);
}
current_stacks = std::move(next_stacks);
if (current_stacks.empty()) {
break;
}
}
for (auto & surviving_stack : current_stacks) {
if (std::find(stacks_new.begin(), stacks_new.end(), surviving_stack) == stacks_new.end()) {
stacks_new.emplace_back(surviving_stack);
}
}
}
}
grammar.stacks = std::move(stacks_new);
grammar.partial_utf8 = decoded.second;
if (grammar.stacks.empty()) {
throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece + " (" + std::to_string(token) + ")");
}
}
const std::string & ollama_vocab::token_to_piece(const uint32_t token) const {
try {