Support new DeepGEMM (#7172)

2025-06-14 14:00:17 +08:00
parent ba589b88fc
commit 93cec4335f
8 changed files with 59 additions and 19 deletions
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -238,6 +238,7 @@ def deepgemm_w8a8_block_fp8_linear_with_fallback(
        block_size[1],
        column_major_scales=True,
        scale_tma_aligned=True,
+        scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
    )

    if get_bool_env_var("SGLANG_W8A8_DEEPGEMM_SANITY_CHECK_UE8M0"):