Merge remote-tracking branch 'origin/v0.7.2-dev' into v0.7.2_zero_overhead

vllm/model_executor/layers/fused_moe/layer.py

@@ -363,6 +363,9 @@ class FusedMoE(torch.nn.Module):
         if (self.quant_method.__class__.__name__ ==
                 "CompressedTensorsWNA16MoEMethod"):
             moe_quant_params["intermediate_size_full"] = intermediate_size
+        
+        if (self.quant_method.__class__.__name__ in ("BlockInt8MoEMethod")):
+            moe_quant_params["intermediate_size"] = self.intermediate_size_per_partition
 
         self.quant_method.create_weights(layer=self, **moe_quant_params)