A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 05:45 UTC
[Bigfix]fix lora test by pass padded size back to the layer (#37811)
Yongye Zhu committed
Mar 22, 2026 at 15:20 UTC
c058ff44d4ccf411ca67abc0ce216571a789c6db
1 file changed
+9
vllm/model_executor/layers/quantization/mxfp4.py
+9
@@ -154,6 +154,15 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
154
)
155
self.hidden_size = hidden_size = self.moe.hidden_dim
156
157
+ # Expose padded dimensions on the layer for LoRA and Marlin code
158
+ # that reads layer.hidden_size / layer.intermediate_size_per_partition.
159
+ layer.params_dtype = params_dtype
160
+ layer.num_experts = num_experts
161
+ layer.hidden_size = hidden_size
162
+ layer.intermediate_size_per_partition = (
163
+ intermediate_size_per_partition_after_pad
164
+ )
165
+
166
# Fused gate_up_proj (column parallel)
167
w13_weight = torch.nn.Parameter(
168
torch.zeros(