[ROCm] Fix fp8 quantization accuracy issue. (#10558)

2025-09-18 08:44:59 +08:00
parent 1ba137e98f
commit c32fb7a24d
1 changed files with 1 additions and 1 deletions
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -732,7 +732,7 @@ def apply_fp8_linear(
                # final solution should be: 1. add support to per-tensor activation scaling.
                # 2. solve the torch.compile error from weight_scale.numel() == 1 and x_scale.numel() > 1 (below line#308)
                if _is_hip and weight_scale.numel() == 1:
-                    qinput, x_scale = ops.scaled_fp8_quant(
+                    qinput, x_scale = scaled_fp8_quant(
                        input_2d,
                        input_scale,
                        use_per_token_if_dynamic=use_per_token_if_dynamic,