replica
/
llama.cpp
mirror of https://github.com/ggml-org/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

model: add Ernie 4.5 MoE support (#14658) 

* Add Ernie4.5 MoE

* Fix Flake errors.

* Properly encode/decode MoE layer step

* Correct tensor mappings (.weight)

* Pass and read n_ff_exp

* n_ff_shexp calculation and further minor changes

* Rope fixes.

* .gitignore fix

* Add unit32 cast for Linux builds

* Apply suggestions from code review

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Further fixes from code review

* Fix trailing whitespace

* Reenable missing experts error

* Code style from code review

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Fix non-MoE regression

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Piotr Wilkin (ilintar) 2025-07-17 23:15:32 +02:00 committed by GitHub
parent d6fb3f6b49
commit cb887f1bc1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 373 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

89
convert_hf_to_gguf.py
View File

@ -2861,7 +2861,8 @@ class Ernie4_5Model(TextModel):
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
num_heads = self.hparams["num_attention_heads"]
num_kv_heads = self.hparams["num_key_value_heads"]
head_dim = self.hparams["head_dim"]
if (head_dim := self.hparams.get("head_dim")) is None:
head_dim = self.hparams["hidden_size"] // num_heads
if "ernie." in name:
name = name.replace("ernie.", "model.")
@ -2894,6 +2895,92 @@ class Ernie4_5Model(TextModel):
return [(self.map_tensor_name(name), data_torch)]
@ModelBase.register("Ernie4_5_MoeForCausalLM")
class Ernie4_5MoeModel(Ernie4_5Model):
model_arch = gguf.MODEL_ARCH.ERNIE4_5_MOE
_experts: list[dict[str, Tensor]] | None = None
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self._experts = [{} for _ in range(self.block_count)]
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"])
self.gguf_writer.add_expert_used_count(self.hparams["moe_k"])
self.gguf_writer.add_interleave_moe_layer_step(self.hparams["moe_layer_interval"])
self.gguf_writer.add_leading_dense_block_count(self.hparams["moe_layer_start_index"])
self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
if (shared_expert_intermediate_size := self.hparams.get('intermediate_size')) is not None and (num_key_value_heads := self.hparams.get('num_key_value_heads')) is not None:
self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size // num_key_value_heads)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# Modify correction bias name as in DeepseekV2
if name.endswith("e_score_correction_bias"):
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
# skip Multi-Token Prediction (MTP) layers (again, same as DeepseekV2)
match = re.match(r"model.mtp_block.(\d+)", name)
if match:
return []
# skip all other MTP tensors for now
match = re.match(r"model.mtp_emb_norm.(\d+)", name)
if match:
return []
match = re.match(r"model.mtp_hidden_norm.(\d+)", name)
if match:
return []
match = re.match(r"model.mtp_linear_proj.(\d+)", name)
if match:
return []
# process the experts separately
if name.find("mlp.experts") != -1:
n_experts = self.hparams["moe_num_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["gate_proj", "up_proj", "down_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
datas.append(self._experts[bid][ename_to_retrieve])
del self._experts[bid][ename_to_retrieve]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
def prepare_tensors(self):
super().prepare_tensors()
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register(
"Qwen2VLModel",
"Qwen2VLForConditionalGeneration",

24
gguf-py/gguf/constants.py
View File

@ -364,6 +364,7 @@ class MODEL_ARCH(IntEnum):
DOTS1 = auto()
ARCEE = auto()
ERNIE4_5 = auto()
ERNIE4_5_MOE = auto()
HUNYUAN_MOE = auto()
SMOLLM3 = auto()
LFM2 = auto()
@ -680,6 +681,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.DOTS1: "dots1",
MODEL_ARCH.ARCEE: "arcee",
MODEL_ARCH.ERNIE4_5: "ernie4_5",
MODEL_ARCH.ERNIE4_5_MOE: "ernie4_5-moe",
MODEL_ARCH.FALCON_H1: "falcon-h1",
MODEL_ARCH.HUNYUAN_MOE: "hunyuan-moe",
MODEL_ARCH.SMOLLM3: "smollm3",
@ -2022,6 +2024,28 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
MODEL_ARCH.ERNIE4_5_MOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
MODEL_ARCH.PLM: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,

45
gguf-py/gguf/tensor_mapping.py
View File

@ -324,7 +324,8 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_EXP_PROBS_B: (
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
"model.layers.{bid}.mlp.moe_statics.e_score_correction", # ernie4.5-moe
),
# Feed-forward up
@ -364,13 +365,13 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.up_proj", # llama4
"encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged) ernie4.5-moe
"model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.up_proj", # llama4
"encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe
),
MODEL_TENSOR.FFN_UP_SHEXP: (
@ -403,12 +404,12 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged) ernie4.5-moe
"model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
),
MODEL_TENSOR.FFN_GATE_SHEXP: (
@ -450,14 +451,14 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
"model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.down_proj", # llama4
"encoder.layers.{bid}.mlp.experts.mlp.w2", # nomic-bert-moe
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged) ernie4.5-moe
"model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
"model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.down_proj", # llama4
"encoder.layers.{bid}.mlp.experts.mlp.w2", # nomic-bert-moe
),
MODEL_TENSOR.FFN_DOWN_SHEXP: (

26
src/llama-arch.cpp
View File

@ -82,6 +82,7 @@ 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_ERNIE4_5_MOE, "ernie4_5-moe" },
{ LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
{ LLM_ARCH_SMOLLM3, "smollm3" },
{ LLM_ARCH_LFM2, "lfm2" },
@ -1825,6 +1826,31 @@ 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_ERNIE4_5_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_K, "blk.%d.attn_k" },
{ 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_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ 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_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
},
},
{
LLM_ARCH_HUNYUAN_MOE,
{

1
src/llama-arch.h
View File

@ -86,6 +86,7 @@ enum llm_arch {
LLM_ARCH_DOTS1,
LLM_ARCH_ARCEE,
LLM_ARCH_ERNIE4_5,
LLM_ARCH_ERNIE4_5_MOE,
LLM_ARCH_HUNYUAN_MOE,
LLM_ARCH_SMOLLM3,
LLM_ARCH_LFM2,

212
src/llama-model.cpp
View File

@ -107,8 +107,10 @@ const char * llm_type_name(llm_type type) {
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_21B_A3B: return "21B.A3B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_235B_A22B: return "235B.A22B";
case LLM_TYPE_300B_A47B: return "300B.A47B";
case LLM_TYPE_E2B: return "E2B";
case LLM_TYPE_E4B: return "E4B";
default: return "?B";
@ -1649,10 +1651,20 @@ void llama_model::load_hparams(llama_model_loader & ml) {
}
} break;
case LLM_ARCH_ERNIE4_5:
case LLM_ARCH_ERNIE4_5_MOE:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
if (arch == LLM_ARCH_ERNIE4_5_MOE) {
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, false);
ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP, hparams.n_moe_layer_step);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
}
switch (hparams.n_layer) {
case 18: type = LLM_TYPE_0_3B; break;
case 28: type = LLM_TYPE_21B_A3B; break;
case 54: type = LLM_TYPE_300B_A47B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
@ -4858,6 +4870,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
} break;
case LLM_ARCH_ERNIE4_5:
case LLM_ARCH_ERNIE4_5_MOE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@ -4886,9 +4899,27 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
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);
if (arch == LLM_ARCH_ERNIE4_5_MOE && static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
int n_ff_exp = hparams.n_ff_exp;
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
// Shared expert (if present)
if (hparams.n_ff_shexp > 0) {
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_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);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, hparams.n_ff_shexp}, 0);
}
} else { // Dense layers
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_FALCON_H1:
@ -15569,6 +15600,176 @@ struct llm_build_ernie4_5 : public llm_graph_context {
}
};
struct llm_build_ernie4_5_moe : public llm_graph_context {
llm_build_ernie4_5_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();
ggml_tensor * inp_out_ids = build_inp_out_ids();
GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
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
{
// 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, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
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);
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 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);
// feed-forward network
bool is_moe_layer = static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
if (!is_moe_layer) {
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
cur = 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, "ffn_out", il);
} else {
// MoE branch
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
ggml_tensor * moe_out = 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,
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(moe_out, "ffn_moe_out", il);
// Shared expert (if present)
if (hparams.n_ff_shexp > 0) {
ggml_tensor * ffn_shexp = 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(ffn_shexp, "ffn_shexp", il);
cur = ggml_add(ctx0, moe_out, ffn_shexp);
} else {
cur = moe_out;
}
cb(cur, "ffn_out", il);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);
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_falcon_h1 : public llm_graph_context_mamba {
llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context_mamba(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@ -17034,6 +17235,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_ernie4_5>(*this, params, gf);
} break;
case LLM_ARCH_ERNIE4_5_MOE:
{
llm = std::make_unique<llm_build_ernie4_5_moe>(*this, params, gf);
} break;
case LLM_ARCH_HUNYUAN_MOE:
{
llm = std::make_unique<llm_build_hunyuan_moe>(*this, params, gf);
@ -17206,6 +17411,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_SMOLLM3:
case LLM_ARCH_ARCEE:
case LLM_ARCH_ERNIE4_5:
case LLM_ARCH_ERNIE4_5_MOE:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2

2
src/llama-model.h
View File

@ -99,8 +99,10 @@ enum llm_type {
LLM_TYPE_17B_16E, // llama4 Scout
LLM_TYPE_17B_128E, // llama4 Maverick
LLM_TYPE_A13B,
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_235B_A22B,
LLM_TYPE_300B_A47B, // Ernie MoE big
LLM_TYPE_E2B,
LLM_TYPE_E4B,
};