Fuse routed scaling factor in topk_reduce kernel (#6220)

2025-06-08 02:06:50 +08:00
parent f5599ef124
commit 515ef4facb
10 changed files with 331 additions and 9 deletions
--- a/python/sglang/srt/layers/quantization/w8a8_fp8.py
+++ b/python/sglang/srt/layers/quantization/w8a8_fp8.py
@@ -328,4 +328,5 @@ class W8A8FP8MoEMethod:
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
            no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
        )