fix: use deepgemm only on hopper (#5310)

2025-04-11 20:48:24 -07:00
parent 4f288113ce
commit 611720919d
1 changed files with 1 additions and 1 deletions
--- a/python/sglang/srt/layers/quantization/fp8_kernel.py
+++ b/python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -45,7 +45,7 @@ if _is_cuda:
    from sgl_kernel import sgl_per_token_group_quant_fp8, sgl_per_token_quant_fp8

    sm_version = get_device_sm()
-    if sm_version >= 90 and get_bool_env_var("SGL_ENABLE_JIT_DEEPGEMM", default="true"):
+    if sm_version == 90 and get_bool_env_var("SGL_ENABLE_JIT_DEEPGEMM", default="true"):
        _enable_jit_deepgemm = True