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 14, 05:30 UTC
[bugfix] remove unused parameters to reduce unnecessary vram usage (#26789)
Signed-off-by: Reinforce-II <fate@eastal.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Reinforce-II committed
Oct 22, 2025 at 23:16 UTC
980de31ca0e16018ddb871c6c1df31a0fd7db223
1 file changed
+2
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
+2
@@ -307,10 +307,12 @@ class CompressedTensorsW4A4MoeMethod(CompressedTensorsMoEMethod):
307
layer.w13_weight = torch.nn.Parameter(
308
layer.w13_weight_packed.data, requires_grad=False
309
)
310
+ delattr(layer, "w13_weight_packed")
311
312
layer.w2_weight = torch.nn.Parameter(
313
layer.w2_weight_packed.data, requires_grad=False
314
)
315
+ delattr(layer, "w2_weight_packed")
316
317
# reorder GEMM1 weights and block scales for FlashInfer CUTLASS kernel.
318
if self.allow_flashinfer: