use sglang_per_token_group_quant_fp8 from sgl-kernel instead of trion kernel (#5473)

Co-authored-by: Zhang Kaihong <zhangkaihong.zkh@alibaba-inc.com>
2025-04-18 15:05:24 +08:00
parent f28d82997a
commit b2a189dd11
2 changed files with 25 additions and 6 deletions
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -141,7 +141,7 @@ def apply_w8a8_block_fp8_linear(
        gemm_a8w8_blockscale(q_input, weight, x_scale, weight_scale, output)
    else:
        if _enable_jit_deepgemm:
-            q_input, x_scale = per_token_group_quant_fp8(
+            q_input, x_scale = sglang_per_token_group_quant_fp8(
                input_2d,
                block_size[1],
                column_major_scales=True,