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

model : add hunyuan moe (#14425)

Browse Source 
* model : add hunyuan moe

* tokenizer ok

* fix tensor name

* cgraph init

* chat template

* wip

* almost working

* skip embed, fix bos

* cleanup

* yarn scaling

* cleanup

* correct rope type

* failed token fix

* ntk alpha freq_base

* tokenization working

* cleanup and pr changes

* vocab_size sanity check

* ntk alpha generic

* Update convert_hf_to_gguf.py

* Apply suggestions from code review

* fix regression

* fix style

---------

Co-authored-by: kooshi <1934337+kooshi@users.noreply.github.com>
This commit is contained in:
Xuan-Son Nguyen
2025-07-08 10:24:06 +02:00
committed by GitHub
parent 53903ae6fa
commit 8f22dc0a53
12 changed files with 449 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
src/llama-arch.cpp
View File

@@ -78,6 +78,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_DOTS1, "dots1" },
{ LLM_ARCH_ARCEE, "arcee" },
{ LLM_ARCH_ERNIE4_5, "ernie4_5" },
{ LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@@ -1694,6 +1695,29 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_HUNYUAN_MOE,
{
{ 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_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ 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_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_ARCH_UNKNOWN,
{

1
src/llama-arch.h
View File

@@ -82,6 +82,7 @@ enum llm_arch {
LLM_ARCH_DOTS1,
LLM_ARCH_ARCEE,
LLM_ARCH_ERNIE4_5,
LLM_ARCH_HUNYUAN_MOE,
LLM_ARCH_UNKNOWN,
};

18
src/llama-chat.cpp
View File

@@ -64,6 +64,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "bailing", LLM_CHAT_TEMPLATE_BAILING },
{ "llama4", LLM_CHAT_TEMPLATE_LLAMA4 },
{ "smolvlm", LLM_CHAT_TEMPLATE_SMOLVLM },
{ "hunyuan-moe", LLM_CHAT_TEMPLATE_HUNYUAN_MOE },
};
llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -185,6 +186,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_LLAMA4;
} else if (tmpl_contains("<|endofuserprompt|>")) {
return LLM_CHAT_TEMPLATE_DOTS1;
} else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) {
return LLM_CHAT_TEMPLATE_HUNYUAN_MOE;
}
return LLM_CHAT_TEMPLATE_UNKNOWN;
}
@@ -665,6 +668,21 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << "<|response|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_MOE) {
// tencent/Hunyuan-A13B-Instruct
for (auto message : chat) {
std::string role(message->role);
if (role == "system") {
ss << "<|startoftext|>" << message->content << "<|extra_4|>";
} else if (role == "assistant") {
ss << "<|startoftext|>" << message->content << "<|eos|>";
} else {
ss << "<|startoftext|>" << message->content << "<|extra_0|>";
}
}
if (add_ass) {
ss << "<|startoftext|>";
}
} else {
// template not supported
return -1;

1
src/llama-chat.h
View File

@@ -44,6 +44,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_LLAMA4,
LLM_CHAT_TEMPLATE_SMOLVLM,
LLM_CHAT_TEMPLATE_DOTS1,
LLM_CHAT_TEMPLATE_HUNYUAN_MOE,
LLM_CHAT_TEMPLATE_UNKNOWN,
};

216
src/llama-model.cpp
View File

@@ -102,6 +102,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_57B_A14B: return "57B.A14B";
case LLM_TYPE_17B_16E: return "17Bx16E (Scout)";
case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)";
case LLM_TYPE_A13B: return "A13B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_235B_A22B: return "235B.A22B";
case LLM_TYPE_E2B: return "E2B";
@@ -1549,6 +1550,17 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_HUNYUAN_MOE:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp);
switch (hparams.n_layer) {
case 32: type = LLM_TYPE_A13B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
default: throw std::runtime_error("unsupported model architecture");
}
@@ -4475,6 +4487,43 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
} break;
case LLM_ARCH_HUNYUAN_MOE:
{
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_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@@ -14645,6 +14694,168 @@ struct llm_build_arcee : public llm_graph_context {
}
};
struct llm_build_hunyuan_moe : public llm_graph_context {
llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : 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, gf,
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_shexp, NULL, NULL,
model.layers[il].ffn_gate_shexp, NULL, NULL,
model.layers[il].ffn_down_shexp, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur_mlp, "ffn_mlp", il);
// MoE branch
ggml_tensor * cur_moe = build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU,
true, // norm_topk_prob
false,
0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(cur_moe, "ffn_moe_out", il);
ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);
cb(ffn_out, "ffn_out", il);
cur = ggml_add(ctx0, ffn_out, 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);
}
};
llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
llama_memory_i * res;
@@ -15025,6 +15236,10 @@ llm_graph_result_ptr llama_model::build_graph(
{
llm = std::make_unique<llm_build_ernie4_5>(*this, params, gf);
} break;
case LLM_ARCH_HUNYUAN_MOE:
{
llm = std::make_unique<llm_build_hunyuan_moe>(*this, params, gf);
} break;
default:
GGML_ABORT("fatal error");
}
@@ -15213,6 +15428,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_EXAONE:
case LLM_ARCH_MINICPM3:
case LLM_ARCH_DOTS1:
case LLM_ARCH_HUNYUAN_MOE:
return LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_QWEN2VL:

1
src/llama-model.h
View File

@@ -94,6 +94,7 @@ enum llm_type {
LLM_TYPE_57B_A14B,
LLM_TYPE_17B_16E, // llama4 Scout
LLM_TYPE_17B_128E, // llama4 Maverick
LLM_TYPE_A13B,
LLM_TYPE_30B_A3B,
LLM_TYPE_235B_A22B,
LLM_TYPE_E2B,

5
src/llama-vocab.cpp
View File

@@ -351,6 +351,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
break;
case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
case LLAMA_VOCAB_PRE_TYPE_QWEN2:
case LLAMA_VOCAB_PRE_TYPE_HUNYUAN:
regex_exprs = {
// original regex from tokenizer.json
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
@@ -1656,6 +1657,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "seed-coder") {
pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
clean_spaces = false;
} else if (
tokenizer_pre == "hunyuan") {
pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN;
clean_spaces = false;
} else {
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
}