EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

model : add hunyuan dense (#14878)

Browse Source 
* support hunyuan_v1_dense

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* update hunyuan_moe to hunyuan_v1_moe

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* fix rope alpha assert and bos token

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* add blank line

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* Revert "update hunyuan_moe to hunyuan_v1_moe"

This reverts commit aa973ca21913aba77f6e81a935270ef7be222e75.

* use hunyuan_dense instead of hunyuan_v1_dense

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* fix hunyuan_moe chat template

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* remove leftover code

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* update hunyuan dense chat template

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

* fix hunyuan dense vocab and chat template

Signed-off-by: stevenkuang <stevenkuang@tencent.com>

---------

Signed-off-by: stevenkuang <stevenkuang@tencent.com>
This commit is contained in:
stevenkuang
2025-08-01 21:31:12 +08:00
committed by GitHub
parent 1c872f71fb
commit 0f5ccd6fd1
10 changed files with 351 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
src/llama-arch.cpp
View File

@@ -85,6 +85,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_ERNIE4_5, "ernie4_5" },
{ LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" },
{ LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
{ LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" },
{ LLM_ARCH_SMOLLM3, "smollm3" },
{ LLM_ARCH_LFM2, "lfm2" },
{ LLM_ARCH_DREAM, "dream" },
@@ -1897,6 +1898,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
},
},
{
LLM_ARCH_HUNYUAN_DENSE,
{
{ 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, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ 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_ARCH_SMOLLM3,
{

1
src/llama-arch.h
View File

@@ -89,6 +89,7 @@ enum llm_arch {
LLM_ARCH_ERNIE4_5,
LLM_ARCH_ERNIE4_5_MOE,
LLM_ARCH_HUNYUAN_MOE,
LLM_ARCH_HUNYUAN_DENSE,
LLM_ARCH_SMOLLM3,
LLM_ARCH_LFM2,
LLM_ARCH_DREAM,

21
src/llama-chat.cpp
View File

@@ -66,6 +66,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "llama4", LLM_CHAT_TEMPLATE_LLAMA4 },
{ "smolvlm", LLM_CHAT_TEMPLATE_SMOLVLM },
{ "hunyuan-moe", LLM_CHAT_TEMPLATE_HUNYUAN_MOE },
{ "hunyuan-dense", LLM_CHAT_TEMPLATE_HUNYUAN_DENSE },
{ "kimi-k2", LLM_CHAT_TEMPLATE_KIMI_K2 },
};
@@ -193,6 +194,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_DOTS1;
} else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) {
return LLM_CHAT_TEMPLATE_HUNYUAN_MOE;
} else if (tmpl_contains("<｜hy_place▁holder▁no▁2｜>") && tmpl_contains("<｜hy_place▁holder▁no▁3｜>")) {
return LLM_CHAT_TEMPLATE_HUNYUAN_DENSE;
} else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
return LLM_CHAT_TEMPLATE_KIMI_K2;
}
@@ -698,11 +701,27 @@ int32_t llm_chat_apply_template(
if (role == "system") {
ss << "<|startoftext|>" << message->content << "<|extra_4|>";
} else if (role == "assistant") {
ss << "<|startoftext|>" << message->content << "<|eos|>";
ss << message->content << "<|eos|>";
} else {
ss << "<|startoftext|>" << message->content << "<|extra_0|>";
}
}
} else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_DENSE) {
// tencent/Hunyuan-4B-Instruct
for (size_t i = 0; i < chat.size(); i++) {
std::string role(chat[i]->role);
if (i == 0) {
if (role == "system") {
ss << chat[i]->content << "<｜hy_place▁holder▁no▁3｜>";
}
}
if (role == "assistant") {
ss << "<｜hy_Assistant｜>" << chat[i]->content << "<｜hy_place▁holder▁no▁2｜>";
} else if (role == "user") {
ss << "<｜hy_User｜>" << chat[i]->content << "<｜hy_Assistant｜>";
}
}
} else if (tmpl == LLM_CHAT_TEMPLATE_KIMI_K2) {
// moonshotai/Kimi-K2-Instruct
for (auto message : chat) {

1
src/llama-chat.h
View File

@@ -46,6 +46,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_SMOLVLM,
LLM_CHAT_TEMPLATE_DOTS1,
LLM_CHAT_TEMPLATE_HUNYUAN_MOE,
LLM_CHAT_TEMPLATE_HUNYUAN_DENSE,
LLM_CHAT_TEMPLATE_KIMI_K2,
LLM_CHAT_TEMPLATE_UNKNOWN,
};

188
src/llama-model.cpp
View File

@@ -1760,6 +1760,18 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_HUNYUAN_DENSE:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_embd) {
case 1024: type = LLM_TYPE_0_5B; break;
case 2048: type = LLM_TYPE_1_8B; break;
case 3072: type = LLM_TYPE_4B; break;
case 4096: type = LLM_TYPE_7B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_SMOLLM3:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -5195,6 +5207,39 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
}
} break;
case LLM_ARCH_HUNYUAN_DENSE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
} break;
case LLM_ARCH_SMOLLM3:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -16930,6 +16975,144 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
}
};
struct llm_build_hunyuan_dense : public llm_graph_context {
llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv_unified();
const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
// norm
cur = build_norm(inpL,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = build_norm(Kcur,
model.layers[il].attn_k_norm, nullptr,
LLM_NORM_RMS, il);
cb(Kcur, "Kcur_norm", il);
Qcur = build_norm(Qcur,
model.layers[il].attn_q_norm, nullptr,
LLM_NORM_RMS, il);
cb(Qcur, "Qcur_norm", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
// feed-forward network (non-MoE)
ggml_tensor * cur_mlp = build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur_mlp, "ffn_out", il);
cur = ggml_add(ctx0, cur_mlp, ffn_inp);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = build_norm(cur,
model.output_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
};
struct llm_build_smollm3 : public llm_graph_context {
llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -17797,6 +17980,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_hunyuan_moe>(*this, params);
} break;
case LLM_ARCH_HUNYUAN_DENSE:
{
llm = std::make_unique<llm_build_hunyuan_dense>(*this, params);
} break;
case LLM_ARCH_SMOLLM3:
{
llm = std::make_unique<llm_build_smollm3>(*this, params);
@@ -18016,6 +18203,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_MINICPM3:
case LLM_ARCH_DOTS1:
case LLM_ARCH_HUNYUAN_MOE:
case LLM_ARCH_HUNYUAN_DENSE:
case LLM_ARCH_LFM2:
case LLM_ARCH_SMALLTHINKER:
return LLAMA_ROPE_TYPE_NEOX;

5
src/llama-vocab.cpp
View File

@@ -307,6 +307,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
};
break;
case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
case LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE:
regex_exprs = {
"\\p{N}{1,3}",
"[一-龥぀-ゟ゠-ヿ]+",
@@ -1964,6 +1965,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "hunyuan") {
pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN;
clean_spaces = false;
} else if (
tokenizer_pre == "hunyuan-dense") {
pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE;
clean_spaces = false;
} else if (
tokenizer_pre == "kimi-k2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;

1
src/llama-vocab.h
View File

@@ -46,6 +46,7 @@ enum llama_vocab_pre_type {
LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35,
LLAMA_VOCAB_PRE_TYPE_HUNYUAN = 36,
LLAMA_VOCAB_PRE_TYPE_KIMI_K2 = 37,
LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE = 38,
};
struct LLM_KV;