[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>

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -307,10 +307,12 @@ class CompressedTensorsW4A4MoeMethod(CompressedTensorsMoEMethod):
         layer.w13_weight = torch.nn.Parameter(
             layer.w13_weight_packed.data, requires_grad=False
         )
+        delattr(layer, "w13_weight_packed")
 
         layer.w2_weight = torch.nn.Parameter(
             layer.w2_weight_packed.data, requires_grad=False
         )
+        delattr(layer, "w2_weight_packed")
 
         # reorder GEMM1 weights and block scales for FlashInfer CUTLASS kernel.
         if self.allow_flashinfer: