xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-07 22:09:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix][Qwen3-Next] add prefixes to shared_expert in qwen3-next and mlp in qwen2moe to successfully load ignored params in quantized models (#24960)

Browse Source 
Signed-off-by: toncao <cpatonn@gmail.com>
Co-authored-by: toncao <cpatonn@gmail.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
toncao 2025-09-18 11:08:50 +07:00 committed by GitHub
parent b98219670f
commit 027d37df38
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 26 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
vllm/model_executor/models/qwen2_moe.py
View File

@ -72,17 +72,20 @@ class Qwen2MoeMLP(nn.Module):
hidden_act: str, hidden_act: str,
quant_config: Optional[QuantizationConfig] = None, quant_config: Optional[QuantizationConfig] = None,
reduce_results: bool = True, reduce_results: bool = True,
prefix: str = "",
) -> None: ) -> None:
super().__init__() super().__init__()
self.gate_up_proj = MergedColumnParallelLinear( self.gate_up_proj = MergedColumnParallelLinear(
hidden_size, [intermediate_size] * 2, hidden_size, [intermediate_size] * 2,
bias=False, bias=False,
quant_config=quant_config) quant_config=quant_config,
prefix=f"{prefix}.gate_up_proj")
self.down_proj = RowParallelLinear(intermediate_size, self.down_proj = RowParallelLinear(intermediate_size,
hidden_size, hidden_size,
bias=False, bias=False,
quant_config=quant_config, quant_config=quant_config,
reduce_results=reduce_results) reduce_results=reduce_results,
prefix=f"{prefix}.down_proj")
if hidden_act != "silu": if hidden_act != "silu":
raise ValueError(f"Unsupported activation: {hidden_act}. " raise ValueError(f"Unsupported activation: {hidden_act}. "
"Only silu is supported for now.") "Only silu is supported for now.")
@ -123,7 +126,8 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
self.gate = ReplicatedLinear(config.hidden_size, self.gate = ReplicatedLinear(config.hidden_size,
config.num_experts, config.num_experts,
bias=False, bias=False,
quant_config=None) quant_config=None,
prefix=f"{prefix}.gate")
if config.shared_expert_intermediate_size > 0: if config.shared_expert_intermediate_size > 0:
self.shared_expert = Qwen2MoeMLP( self.shared_expert = Qwen2MoeMLP(
hidden_size=config.hidden_size, hidden_size=config.hidden_size,
@ -132,6 +136,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
quant_config=quant_config, quant_config=quant_config,
reduce_results=self.experts.must_reduce_shared_expert_outputs( reduce_results=self.experts.must_reduce_shared_expert_outputs(
), ),
prefix=f"{prefix}.shared_expert",
) )
else: else:
self.shared_expert = None self.shared_expert = None
@ -203,21 +208,19 @@ class Qwen2MoeAttention(nn.Module):
self.max_position_embeddings = max_position_embeddings self.max_position_embeddings = max_position_embeddings
self.dual_chunk_attention_config = dual_chunk_attention_config self.dual_chunk_attention_config = dual_chunk_attention_config
self.qkv_proj = QKVParallelLinear( self.qkv_proj = QKVParallelLinear(hidden_size,
hidden_size, self.head_dim,
self.head_dim, self.total_num_heads,
self.total_num_heads, self.total_num_kv_heads,
self.total_num_kv_heads, bias=True,
bias=True, quant_config=quant_config,
quant_config=quant_config, prefix=f"{prefix}.qkv_proj")
)
self.o_proj = RowParallelLinear( self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
self.total_num_heads * self.head_dim, hidden_size,
hidden_size, bias=False,
bias=False, quant_config=quant_config,
quant_config=quant_config, prefix=f"{prefix}.o_proj")
)
self.rotary_emb = get_rope( self.rotary_emb = get_rope(
self.head_dim, self.head_dim,
@ -296,12 +299,11 @@ class Qwen2MoeDecoderLayer(nn.Module):
quant_config=quant_config, quant_config=quant_config,
prefix=f"{prefix}.mlp") prefix=f"{prefix}.mlp")
else: else:
self.mlp = Qwen2MoeMLP( self.mlp = Qwen2MoeMLP(hidden_size=config.hidden_size,
hidden_size=config.hidden_size, intermediate_size=config.intermediate_size,
intermediate_size=config.intermediate_size, hidden_act=config.hidden_act,
hidden_act=config.hidden_act, quant_config=quant_config,
quant_config=quant_config, prefix=f"{prefix}.mlp")
)
self.input_layernorm = RMSNorm(config.hidden_size, self.input_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps) eps=config.rms_norm_eps)
self.post_attention_layernorm = RMSNorm(config.hidden_size, self.post_attention_layernorm = RMSNorm(config.hidden_size,

1
vllm/model_executor/models/qwen3_next.py
View File

@ -138,6 +138,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
quant_config=quant_config, quant_config=quant_config,
reduce_results=self.experts.must_reduce_shared_expert_outputs( reduce_results=self.experts.must_reduce_shared_expert_outputs(
), ),
prefix=f"{prefix}.shared_expert",
) )
else: else:
self.shared_expert = None self.shared_expert = None