Skip to content

GitLab

Unverified Commit 544b6739 authored by Daniel Hiltgen's avatar Daniel Hiltgen Committed by GitHub
Browse files

ggml update to b6840 (#12791)

parent c4ba257c
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 585 additions and 135 deletions
Makefile.sync
View file @ 544b6739
UPSTREAM=https://github.com/ggml-org/llama.cpp.git
WORKDIR=llama/vendor
FETCH_HEAD=7049736b2dd9011bf819e298b844ebbc4b5afdc9
FETCH_HEAD=3cfa9c3f125763305b4226bc032f1954f08990dc
.PHONY: help
help:
......
llama/build-info.cpp
View file @ 544b6739
int LLAMA_BUILD_NUMBER = 0;
char const *LLAMA_COMMIT = "7049736b2dd9011bf819e298b844ebbc4b5afdc9";
char const *LLAMA_COMMIT = "3cfa9c3f125763305b4226bc032f1954f08990dc";
char const *LLAMA_COMPILER = "";
char const *LLAMA_BUILD_TARGET = "";
llama/llama.cpp/common/json-schema-to-grammar.cpp
View file @ 544b6739
......@@ -41,9 +41,9 @@ static std::string build_repetition(const std::string & item_rule, int min_items
return result;
}
static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
auto has_min = min_value != std::numeric_limits<int>::min();
auto has_max = max_value != std::numeric_limits<int>::max();
static void _build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
auto has_min = min_value != std::numeric_limits<int64_t>::min();
auto has_max = max_value != std::numeric_limits<int64_t>::max();
auto digit_range = [&](char from, char to) {
out << "[";
......@@ -159,7 +159,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
if (has_min) {
if (min_value < 0) {
out << "\"-\" (";
_build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
_build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
out << ") | [0] | [1-9] ";
more_digits(0, decimals_left - 1);
} else if (min_value == 0) {
......@@ -194,7 +194,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
}
digit_range(c, c);
out << " (";
_build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
_build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
out << ")";
if (c < '9') {
out << " | ";
......@@ -216,7 +216,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
_build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
} else {
out << "\"-\" (";
_build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
_build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
out << ")";
}
return;
......@@ -925,17 +925,17 @@ public:
int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
} else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
int min_value = std::numeric_limits<int>::min();
int max_value = std::numeric_limits<int>::max();
int64_t min_value = std::numeric_limits<int64_t>::min();
int64_t max_value = std::numeric_limits<int64_t>::max();
if (schema.contains("minimum")) {
min_value = schema["minimum"].get<int>();
min_value = schema["minimum"].get<int64_t>();
} else if (schema.contains("exclusiveMinimum")) {
min_value = schema["exclusiveMinimum"].get<int>() + 1;
min_value = schema["exclusiveMinimum"].get<int64_t>() + 1;
}
if (schema.contains("maximum")) {
max_value = schema["maximum"].get<int>();
max_value = schema["maximum"].get<int64_t>();
} else if (schema.contains("exclusiveMaximum")) {
max_value = schema["exclusiveMaximum"].get<int>() - 1;
max_value = schema["exclusiveMaximum"].get<int64_t>() - 1;
}
std::stringstream out;
out << "(";
......
llama/llama.cpp/src/llama-arch.cpp
View file @ 544b6739
......@@ -5,6 +5,7 @@
#include <map>
static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_CLIP, "clip" }, // dummy, only used by llama-quantize
{ LLM_ARCH_LLAMA, "llama" },
{ LLM_ARCH_LLAMA4, "llama4" },
{ LLM_ARCH_DECI, "deci" },
......@@ -85,6 +86,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
{ LLM_ARCH_PLM, "plm" },
{ LLM_ARCH_BAILINGMOE, "bailingmoe" },
{ LLM_ARCH_BAILINGMOE2, "bailingmoe2" },
{ LLM_ARCH_DOTS1, "dots1" },
{ LLM_ARCH_ARCEE, "arcee" },
{ LLM_ARCH_ERNIE4_5, "ernie4_5" },
......@@ -135,6 +137,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EXPERT_COUNT, "%s.expert_count" },
{ LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" },
{ LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" },
{ LLM_KV_EXPERT_GROUP_COUNT, "%s.expert_group_count" },
{ LLM_KV_EXPERT_GROUP_USED_COUNT, "%s.expert_group_used_count" },
{ LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
{ LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
{ LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" },
......@@ -277,6 +281,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
};
static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_NAMES = {
{
LLM_ARCH_CLIP,
{},
},
{
LLM_ARCH_LLAMA,
{
......@@ -1961,6 +1969,38 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
},
},
{
LLM_ARCH_BAILINGMOE2,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
{ LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
{ LLM_TENSOR_NEXTN_EH_PROJ, "blk.%d.nextn.eh_proj" },
{ LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
{ LLM_TENSOR_NEXTN_ENORM, "blk.%d.nextn.enorm" },
{ LLM_TENSOR_NEXTN_HNORM, "blk.%d.nextn.hnorm" },
{ LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
{ LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
},
},
{
LLM_ARCH_DOTS1,
{
......
llama/llama.cpp/src/llama-arch.h
View file @ 544b6739
......@@ -9,6 +9,7 @@
//
enum llm_arch {
LLM_ARCH_CLIP,
LLM_ARCH_LLAMA,
LLM_ARCH_LLAMA4,
LLM_ARCH_DECI,
......@@ -89,6 +90,7 @@ enum llm_arch {
LLM_ARCH_WAVTOKENIZER_DEC,
LLM_ARCH_PLM,
LLM_ARCH_BAILINGMOE,
LLM_ARCH_BAILINGMOE2,
LLM_ARCH_DOTS1,
LLM_ARCH_ARCEE,
LLM_ARCH_ERNIE4_5,
......@@ -139,6 +141,8 @@ enum llm_kv {
LLM_KV_EXPERT_COUNT,
LLM_KV_EXPERT_USED_COUNT,
LLM_KV_EXPERT_SHARED_COUNT,
LLM_KV_EXPERT_GROUP_COUNT,
LLM_KV_EXPERT_GROUP_USED_COUNT,
LLM_KV_EXPERT_WEIGHTS_SCALE,
LLM_KV_EXPERT_WEIGHTS_NORM,
LLM_KV_EXPERT_GATING_FUNC,
......
llama/llama.cpp/src/llama-batch.h
View file @ 544b6739
......@@ -123,7 +123,7 @@ private:
uint32_t n_seq_max;
uint32_t n_outputs;
std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
std::array<llama_seq_id, 1> seq_id_0 = {{ 0 }}; // default sequence id
std::vector<llama_pos> pos;
std::vector<int32_t> n_seq_id;
......
llama/llama.cpp/src/llama-chat.cpp
View file @ 544b6739
......@@ -63,6 +63,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "megrez", LLM_CHAT_TEMPLATE_MEGREZ },
{ "yandex", LLM_CHAT_TEMPLATE_YANDEX },
{ "bailing", LLM_CHAT_TEMPLATE_BAILING },
{ "bailing-think", LLM_CHAT_TEMPLATE_BAILING_THINK },
{ "bailing2", LLM_CHAT_TEMPLATE_BAILING2 },
{ "llama4", LLM_CHAT_TEMPLATE_LLAMA4 },
{ "smolvlm", LLM_CHAT_TEMPLATE_SMOLVLM },
{ "hunyuan-moe", LLM_CHAT_TEMPLATE_HUNYUAN_MOE },
......@@ -191,6 +193,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_YANDEX;
} else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
return LLM_CHAT_TEMPLATE_BAILING;
} else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("\"HUMAN\"") && tmpl_contains("<think>")) {
return LLM_CHAT_TEMPLATE_BAILING_THINK;
} else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("<role>HUMAN</role>") && tmpl_contains("<|role_end|>")) {
return LLM_CHAT_TEMPLATE_BAILING2;
} else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
return LLM_CHAT_TEMPLATE_LLAMA4;
} else if (tmpl_contains("<|endofuserprompt|>")) {
......@@ -644,8 +650,8 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << " Ассистент:[SEP]";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
// Bailing (Ling) template
} else if (tmpl == LLM_CHAT_TEMPLATE_BAILING || tmpl == LLM_CHAT_TEMPLATE_BAILING_THINK) {
// Bailing (Ling/Ring) template
for (auto message : chat) {
std::string role(message->role);
......@@ -658,6 +664,33 @@ int32_t llm_chat_apply_template(
ss << "<role>" << role << "</role>" << message->content;
}
if (add_ass) {
ss << "<role>ASSISTANT</role>";
if (tmpl == LLM_CHAT_TEMPLATE_BAILING_THINK) {
ss << "<think>";
}
}
} else if (tmpl == LLM_CHAT_TEMPLATE_BAILING2) {
// Bailing2 (Ling 2.0) template
bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
if (!has_system) {
ss << "<role>SYSTEM</role>detailed thinking off<|role_end|>";
}
for (auto message : chat) {
std::string role(message->role);
if (role == "user") {
role = "HUMAN";
} else {
std::transform(role.begin(), role.end(), role.begin(), ::toupper);
}
ss << "<role>" << role << "</role>" << message->content << "<|role_end|>";
}
if (add_ass) {
ss << "<role>ASSISTANT</role>";
}
......
llama/llama.cpp/src/llama-chat.h
View file @ 544b6739
......@@ -42,6 +42,8 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_MEGREZ,
LLM_CHAT_TEMPLATE_YANDEX,
LLM_CHAT_TEMPLATE_BAILING,
LLM_CHAT_TEMPLATE_BAILING_THINK,
LLM_CHAT_TEMPLATE_BAILING2,
LLM_CHAT_TEMPLATE_LLAMA4,
LLM_CHAT_TEMPLATE_SMOLVLM,
LLM_CHAT_TEMPLATE_DOTS1,
......
llama/llama.cpp/src/llama-context.cpp
View file @ 544b6739
......@@ -2345,7 +2345,8 @@ llama_context * llama_init_from_model(
return nullptr;
}
if (params.pooling_type != model->hparams.pooling_type) {
if (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED &&
params.pooling_type != model->hparams.pooling_type) {
//user-specified pooling-type is different from the model default
LLAMA_LOG_WARN("%s: model default pooling_type is [%d], but [%d] was specified\n", __func__,
model->hparams.pooling_type, params.pooling_type);
......
llama/llama.cpp/src/llama-graph.cpp
View file @ 544b6739
......@@ -261,12 +261,17 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
}
}
static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
(swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
(swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
(swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
const char * swa_type_str = "unknown";
switch (swa_type) {
case LLAMA_SWA_TYPE_NONE: swa_type_str = "LLAMA_SWA_TYPE_NONE"; break;
case LLAMA_SWA_TYPE_STANDARD: swa_type_str = "LLAMA_SWA_TYPE_STANDARD"; break;
case LLAMA_SWA_TYPE_CHUNKED: swa_type_str = "LLAMA_SWA_TYPE_CHUNKED"; break;
case LLAMA_SWA_TYPE_SYMMETRIC: swa_type_str = "LLAMA_SWA_TYPE_SYMMETRIC"; break;
};
LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
......@@ -295,50 +300,67 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
const int64_t n_kv = ubatch->n_tokens;
const int64_t n_tokens = ubatch->n_tokens;
GGML_ASSERT(kq_mask);
GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
float * data = (float *) kq_mask->data;
// [TAG_NO_CACHE_ISWA]
GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");
for (int h = 0; h < 1; ++h) {
for (int i1 = 0; i1 < n_tokens; ++i1) {
const llama_seq_id s1 = ubatch->seq_id[i1][0];
const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) {
for (int h = 0; h < 1; ++h) {
for (int i1 = 0; i1 < n_tokens; ++i1) {
const llama_seq_id s1 = ubatch->seq_id[i1][0];
const llama_pos p1 = ubatch->pos[i1];
for (int i0 = 0; i0 < n_tokens; ++i0) {
float f = -INFINITY;
const uint64_t idst = h*(n_kv*n_tokens) + i1*n_kv;
for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
for (int i0 = 0; i0 < n_tokens; ++i0) {
const llama_seq_id s0 = ubatch->seq_id[i0][0];
const llama_pos p0 = ubatch->pos[i0];
// mask different sequences
if (s0 != s1) {
continue; // skip different sequences
continue;
}
if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
continue; // skip future tokens for causal attention
// mask future tokens
if (cparams.causal_attn && p0 > p1) {
continue;
}
// TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
//if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
// continue; // skip masked tokens for SWA
//}
// TODO: reimplement this like in llama_kv_cache_unified
if (hparams.use_alibi) {
f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
} else {
f = 0.0f;
// apply SWA if any
if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
continue;
}
data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
}
data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
}
}
};
{
GGML_ASSERT(self_kq_mask);
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
float * data = (float *) self_kq_mask->data;
std::fill(data, data + ggml_nelements(self_kq_mask), -INFINITY);
fill_mask(data, 0, LLAMA_SWA_TYPE_NONE);
if (debug) {
print_mask(data, n_tokens, n_kv, 0, LLAMA_SWA_TYPE_NONE);
}
}
if (debug) {
print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
GGML_ASSERT(self_kq_mask_swa);
GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
float * data = (float *) self_kq_mask_swa->data;
std::fill(data, data + ggml_nelements(self_kq_mask_swa), -INFINITY);
fill_mask(data, hparams.n_swa, hparams.swa_type);
if (debug) {
print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
}
}
}
......@@ -928,6 +950,31 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cb(selection_probs, "ffn_moe_probs_biased", il);
}
// select top n_group_used expert groups
// https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/e815299b0bcbac849fa540c768ef21845365c9eb/modeling_deepseek.py#L440-L457
if (hparams.n_expert_groups > 1 && n_tokens > 0) {
const int64_t n_exp_per_group = n_expert / hparams.n_expert_groups;
// organize experts into n_expert_groups
ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
// get top n_group_used expert groups
group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
cb(expert_groups, "ffn_moe_group_topk", il);
// mask out the other groups
selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
cb(selection_probs, "ffn_moe_probs_masked", il);
}
// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
cb(selected_experts->src[0], "ffn_moe_argsort", il);
......@@ -959,6 +1006,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
cb(weights_sum, "ffn_moe_weights_sum", il);
if (arch == LLM_ARCH_BAILINGMOE2) {
weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
cb(weights_sum, "ffn_moe_weights_sum_biased", il);
}
weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
cb(weights, "ffn_moe_weights_norm", il);
......@@ -1299,12 +1351,9 @@ ggml_tensor * llm_graph_context::build_attn_mha(
k = ggml_permute(ctx0, k, 0, 2, 1, 3);
v = ggml_permute(ctx0, v, 0, 2, 1, 3);
const auto n_kv = k->ne[1];
ggml_tensor * cur;
// TODO: replace hardcoded padding with ggml-provided padding
if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) {
if (cparams.flash_attn && kq_b == nullptr) {
GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");
if (v_trans) {
......@@ -1419,10 +1468,20 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
// note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
ggml_set_input(inp->kq_mask);
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
ggml_set_input(inp->self_kq_mask);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask;
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
ggml_set_input(inp->self_kq_mask_swa);
inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
} else {
inp->self_kq_mask_swa = nullptr;
inp->self_kq_mask_swa_cnv = nullptr;
}
return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp));
}
......@@ -1447,7 +1506,9 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_build_forward_expand(gf, k_cur);
ggml_build_forward_expand(gf, v_cur);
const auto & kq_mask = inp->get_kq_mask();
const bool is_swa = hparams.is_swa(il);
const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
// [TAG_NO_CACHE_PAD]
// TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
......
llama/llama.cpp/src/llama-graph.h
View file @ 544b6739
......@@ -257,10 +257,14 @@ public:
void set_input(const llama_ubatch * ubatch) override;
ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
ggml_tensor * kq_mask = nullptr; // F32 [n_tokens, n_batch, 1, 1]
ggml_tensor * kq_mask_cnv = nullptr; // [n_tokens, n_batch, 1, 1]
// n_tokens == n_batch
ggml_tensor * self_kq_mask = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_tokens, n_batch/n_stream, 1, n_stream]
const llama_hparams hparams;
const llama_cparams cparams;
......
llama/llama.cpp/src/llama-hparams.h
View file @ 544b6739
......@@ -74,6 +74,8 @@ struct llama_hparams {
uint32_t n_ff_chexp = 0;
uint32_t n_expert_shared = 0;
uint32_t n_norm_groups = 0;
uint32_t n_expert_groups = 0;
uint32_t n_group_used = 0;
uint32_t n_group_experts = 0;
float expert_group_scale = 0.05f;
......
llama/llama.cpp/src/llama-model.cpp
View file @ 544b6739
This diff is collapsed.
llama/llama.cpp/src/llama-model.h
View file @ 544b6739
......@@ -108,9 +108,12 @@ enum llm_type {
LLM_TYPE_17B_16E, // llama4 Scout
LLM_TYPE_17B_128E, // llama4 Maverick
LLM_TYPE_A13B,
LLM_TYPE_7B_A1B,
LLM_TYPE_8B_A1B, // lfm2moe
LLM_TYPE_16B_A1B,
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_100B_A6B,
LLM_TYPE_106B_A12B, // GLM-4.5-Air
LLM_TYPE_235B_A22B,
LLM_TYPE_300B_A47B, // Ernie MoE big
......
llama/llama.cpp/src/llama-quant.cpp
View file @ 544b6739
......@@ -701,6 +701,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
});
}
bool is_clip_model = false;
for (const auto * it : tensors) {
const struct ggml_tensor * tensor = it->tensor;
......@@ -714,12 +715,14 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
} else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
qs.has_output = true;
}
is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
}
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
// sanity checks for models that have attention layers
if (qs.n_attention_wv != 0)
if (qs.n_attention_wv != 0 && !is_clip_model)
{
const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
// attention layers have a non-zero number of kv heads
......@@ -881,6 +884,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
// do not quantize relative position bias (T5)
quantize &= name.find("attn_rel_b.weight") == std::string::npos;
// do not quantize specific multimodal tensors
quantize &= name.find(".position_embd.") == std::string::npos;
ggml_type new_type;
void * new_data;
size_t new_size;
......
llama/llama.cpp/src/llama-vocab.cpp
View file @ 544b6739
......@@ -1957,6 +1957,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
clean_spaces = false;
} else if (
tokenizer_pre == "bailingmoe" ||
tokenizer_pre == "bailingmoe2" ||
tokenizer_pre == "llada-moe") {
pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
clean_spaces = false;
......
llama/llama.cpp/src/llama.cpp
View file @ 544b6739
......@@ -124,6 +124,9 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
} catch(const std::exception & e) {
throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
}
if (model.arch == LLM_ARCH_CLIP) {
throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead");
}
try {
model.load_vocab(ml);
} catch(const std::exception & e) {
......@@ -314,6 +317,7 @@ struct llama_model * llama_model_load_from_splits(
LLAMA_LOG_ERROR("%s: list of splits is empty\n", __func__);
return nullptr;
}
splits.reserve(n_paths);
for (size_t i = 0; i < n_paths; ++i) {
splits.push_back(paths[i]);
}
......
llama/llama.cpp/tools/mtmd/clip-impl.h
View file @ 544b6739
......@@ -30,6 +30,7 @@
#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
// vision-specific
#define KEY_VISION_PROJ_TYPE "clip.vision.projector_type" // for models with mixed modalities
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_PREPROC_IMAGE_SIZE "clip.vision.preproc_image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
......@@ -48,6 +49,7 @@
#define KEY_MINICPMV_QUERY_NUM "clip.minicpmv_query_num"
// audio-specific
#define KEY_AUDIO_PROJ_TYPE "clip.audio.projector_type" // for models with mixed modalities
#define KEY_A_NUM_MEL_BINS "clip.audio.num_mel_bins"
#define KEY_A_PROJ_STACK_FACTOR "clip.audio.projector.stack_factor"
......
llama/llama.cpp/tools/mtmd/clip.cpp
View file @ 544b6739
......@@ -2234,15 +2234,27 @@ struct clip_model_loader {
// projector type
std::string proj_type;
{
// default key
get_string(KEY_PROJ_TYPE, proj_type, false);
if (!proj_type.empty()) {
model.proj_type = clip_projector_type_from_string(proj_type);
// for models with mixed modalities
if (proj_type.empty()) {
if (modality == CLIP_MODALITY_VISION) {
get_string(KEY_VISION_PROJ_TYPE, proj_type, false);
} else if (modality == CLIP_MODALITY_AUDIO) {
get_string(KEY_AUDIO_PROJ_TYPE, proj_type, false);
} else {
GGML_ABORT("unknown modality");
}
}
model.proj_type = clip_projector_type_from_string(proj_type);
if (model.proj_type == PROJECTOR_TYPE_UNKNOWN) {
throw std::runtime_error(string_format("%s: unknown projector type: %s\n", __func__, proj_type.c_str()));
}
// correct arch for multimodal models
// correct arch for multimodal models (legacy method)
if (model.proj_type == PROJECTOR_TYPE_QWEN25O) {
model.proj_type = modality == CLIP_MODALITY_VISION
? PROJECTOR_TYPE_QWEN25VL
......
llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch
View file @ 544b6739
......@@ -23,7 +23,7 @@ problem.
8 files changed, 21 insertions(+), 2 deletions(-)
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index ff9135fe..8ba86f82 100644
index ff9135fe2..8ba86f824 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -113,7 +113,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
......@@ -64,18 +64,18 @@ index ff9135fe..8ba86f82 100644
/* .init_tensor = */ NULL, // no initialization required
/* .memset_tensor = */ ggml_backend_cpu_buffer_memset_tensor,
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index ad1adba6..7d44f74f 100755
index 8bd5449f1..01e2df61a 100644
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -843,6 +843,7 @@ static void ggml_backend_cann_buffer_free_buffer(
ggml_backend_cann_buffer_context* ctx =
(ggml_backend_cann_buffer_context*)buffer->context;
@@ -820,6 +820,7 @@ static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
static void ggml_backend_cann_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cann_buffer_context * ctx = (ggml_backend_cann_buffer_context *) buffer->context;
delete ctx;
+ delete buffer;
}
/**
@@ -1630,6 +1631,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
@@ -1560,6 +1561,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
*/
static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
ACL_CHECK(aclrtFreeHost(buffer->context));
......@@ -84,10 +84,10 @@ index ad1adba6..7d44f74f 100755
/**
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 856e9de2..c0b1e4c1 100644
index bc396b521..aefc6935e 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -567,6 +567,7 @@ struct ggml_backend_cuda_buffer_context {
@@ -576,6 +576,7 @@ struct ggml_backend_cuda_buffer_context {
static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
delete ctx;
......@@ -95,7 +95,7 @@ index 856e9de2..c0b1e4c1 100644
}
static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
@@ -822,6 +823,7 @@ struct ggml_backend_cuda_split_buffer_context {
@@ -831,6 +832,7 @@ struct ggml_backend_cuda_split_buffer_context {
static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
delete ctx;
......@@ -103,7 +103,7 @@ index 856e9de2..c0b1e4c1 100644
}
static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -1103,6 +1105,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
@@ -1112,6 +1114,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
CUDA_CHECK(cudaFreeHost(buffer->context));
......@@ -112,7 +112,7 @@ index 856e9de2..c0b1e4c1 100644
static void * ggml_cuda_host_malloc(size_t size) {
diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
index 7afc881f..bf096227 100644
index 7afc881fa..bf0962274 100644
--- a/ggml/src/ggml-metal/ggml-metal.cpp
+++ b/ggml/src/ggml-metal/ggml-metal.cpp
@@ -25,6 +25,7 @@ static void ggml_backend_metal_buffer_shared_free_buffer(ggml_backend_buffer_t b
......@@ -132,10 +132,10 @@ index 7afc881f..bf096227 100644
static void * ggml_backend_metal_buffer_private_get_base(ggml_backend_buffer_t buffer) {
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 79d21487..38c75018 100644
index db33a4ab6..c42ee26e1 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -3212,6 +3212,7 @@ struct ggml_backend_opencl_buffer_context {
@@ -3266,6 +3266,7 @@ struct ggml_backend_opencl_buffer_context {
static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
delete ctx;
......@@ -144,7 +144,7 @@ index 79d21487..38c75018 100644
static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index aad48d62..a46c0f52 100644
index a38df5a97..fd07e4a21 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -528,6 +528,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
......@@ -156,10 +156,10 @@ index aad48d62..a46c0f52 100644
static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 45b8c216..4ec9a592 100644
index b695ba051..37e853120 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -334,6 +334,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
@@ -352,6 +352,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
ggml_sycl_set_device(ctx->device);
delete ctx;
......@@ -167,7 +167,7 @@ index 45b8c216..4ec9a592 100644
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -795,6 +796,7 @@ struct ggml_backend_sycl_split_buffer_context {
@@ -813,6 +814,7 @@ struct ggml_backend_sycl_split_buffer_context {
static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
delete ctx;
......@@ -175,7 +175,7 @@ index 45b8c216..4ec9a592 100644
}
static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -1137,6 +1139,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
@@ -1155,6 +1157,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_sycl_host_free(buffer->context);
......@@ -184,10 +184,10 @@ index 45b8c216..4ec9a592 100644
static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 3cd89c71..ed83236f 100644
index b783f7805..216dc167c 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -11600,6 +11600,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@@ -11828,6 +11828,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
ggml_vk_destroy_buffer(ctx->dev_buffer);
delete ctx;
......@@ -195,7 +195,7 @@ index 3cd89c71..ed83236f 100644
}
static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -11743,6 +11744,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
@@ -11971,6 +11972,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
ggml_vk_host_free(vk_instance.devices[0], buffer->context);
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment