[perf]optimize w4afp8 kernel on deepseek-v3-0324 (#12921)

Signed-off-by: bruce.xu <bruce.x@gmicloud.ai>
2025-12-18 18:13:22 +08:00
parent d1f0063262
commit 793c96c3d2
3 changed files with 160 additions and 264 deletions
--- a/python/sglang/srt/layers/quantization/w4afp8.py
+++ b/python/sglang/srt/layers/quantization/w4afp8.py
@@ -300,7 +300,6 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):

        x = dispatch_output.hidden_states
        topk_output = dispatch_output.topk_output
-
        topk_weights, topk_ids, _ = topk_output

        output = cutlass_w4a8_moe(