add heuristic logic for weight padding

ROCm · Nov 14, 2024 · a7e9918 · a7e9918
1 parent 04aa1a7
commit a7e9918
Showing 1 changed file with 4 additions and 2 deletions.
diff --git a/vllm/model_executor/layers/quantization/fp8.py b/vllm/model_executor/layers/quantization/fp8.py
@@ -248,8 +248,10 @@ def process_weights_after_loading(self, layer: Module) -> None:
                 )
 
             # Pad the weight
-            if envs.VLLM_FP8_PADDING:
-                weight = F.pad(weight, (0, 256), "constant", 0)[..., :-256]
+            if envs.VLLM_FP8_PADDING and weight.stride(-1) == 1 \
+                and (weight.stride(-2) * weight.element_size()) % 512 == 0:
+                num_pad = 256 // weight.element_size()
+                weight = F.pad(weight, (0, num_pad), "constant", 0)[..., :-num_pad]
                 torch.cuda.empty_cache()
 
             # Update layer with new values.