Merge branch 'habana_main' into michalkuligowski-qwen-sync_after_weig…

…htload
HabanaAI · Jan 10, 2025 · 20a4450 · 20a4450
2 parents 467b333 + 73aaf71
commit 20a4450
Showing 1 changed file with 6 additions and 13 deletions.
diff --git a/vllm/model_executor/layers/quantization/utils/w8a8_utils.py b/vllm/model_executor/layers/quantization/utils/w8a8_utils.py
@@ -149,19 +149,12 @@ def apply_fp8_linear(
 
         if per_tensor_weights and per_tensor_activations:
             # Fused GEMM_DQ
-            if current_platform.is_hpu():
-                #hpu does not support torch._scaled_mm (SW-197036)
-                output = torch.ops.hpu.fp8_gemm_v2(qinput, False, weight,
-                                                   False, None, input.dtype,
-                                                   x_scale, weight_scale, None,
-                                                   False)
-            else:
-                output = torch._scaled_mm(qinput,
-                                          weight,
-                                          out_dtype=input.dtype,
-                                          scale_a=x_scale,
-                                          scale_b=weight_scale,
-                                          bias=bias)
+            output = torch._scaled_mm(qinput,
+                                      weight,
+                                      out_dtype=input.dtype,
+                                      scale_a=x_scale,
+                                      scale_b=weight_scale,
+                                      bias=bias)
 
             # A fix for discrepancy in scaled_mm which returns tuple
             # for torch < 2.5 and a single value in torch >= 2.5