EngineX-MetaX/enginex-c_series-vllm
9
1
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Add support for Qwen3 MoE+GPTQ

Browse Source
This commit is contained in:
sunjichen
2025-11-15 20:14:45 +08:00
parent b296c44ae0
commit 8152e24cb2
35 changed files with 6468 additions and 574 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/qwen3_moe.py
View File

@@ -122,7 +122,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
self.gate = ReplicatedLinear(config.hidden_size,
config.num_experts,
bias=False,
quant_config=None,
quant_config=quant_config,
prefix=f"{prefix}.gate")
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -294,7 +294,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
positions: torch.Tensor,
hidden_states: torch.Tensor,
residual: Optional[torch.Tensor],
) -> torch.Tensor:
) -> tuple[torch.Tensor, torch.Tensor]:
# Self Attention
if residual is None:
residual = hidden_states
@@ -532,4 +532,4 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP):
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(self)
return loader.load_weights(weights)
return loader.load_weights(weights)