biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bigfix]fix lora test by pass padded size back to the layer (#37811)

Browse Source
This commit is contained in:
Yongye Zhu
2026-03-22 15:20:13 -04:00
committed by GitHub
parent ce9b1d76cf
commit c058ff44d4
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/layers/quantization/mxfp4.py
View File

@@ -154,6 +154,15 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
) )
self.hidden_size = hidden_size = self.moe.hidden_dim self.hidden_size = hidden_size = self.moe.hidden_dim
# Expose padded dimensions on the layer for LoRA and Marlin code
# that reads layer.hidden_size / layer.intermediate_size_per_partition.
layer.params_dtype = params_dtype
layer.num_experts = num_experts
layer.hidden_size = hidden_size
layer.intermediate_size_per_partition = (
intermediate_size_per_partition_after_pad
)
# Fused gate_up_proj (column parallel) # Fused gate_up_proj (column parallel)
w13_weight = torch.nn.Parameter( w13_weight = torch.nn.Parameter(
torch.zeros( torch.zeros(