Removing the padding again after it had been overwritten by upstream merge (ROCm#485)

Author: gshtras

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -146,7 +146,8 @@ def __init__(self,
         # as it breaks with dynamic shapes.
         if pad_output is None:
             config = get_current_vllm_config().compilation_config
-            pad_output = config.level < CompilationLevel.PIECEWISE
+            pad_output = (not current_platform.is_rocm()
+                          and config.level < CompilationLevel.PIECEWISE)
         self.output_padding = 17 if pad_output else None
 
     def apply(