[Kernel] Set sgl_per_token_group_quant_8bit_v2 as default choice (#22467)

2026-04-11 01:59:57 -07:00
parent 4f45472f34
commit d14d368191
5 changed files with 29 additions and 19 deletions
@@ -109,10 +109,9 @@ def sgl_per_token_group_quant_8bit(
    masked_m: Optional[torch.Tensor] = None,
    enable_v2: Optional[bool] = None,
 ) -> None:
+    _V2_KERNEL_SUPPORTED_GROUP_SIZES = [16, 32, 64, 128]
    if enable_v2 is None:
-        from sglang.srt.utils import get_bool_env_var
-
-        enable_v2 = get_bool_env_var("SGLANG_PER_TOKEN_GROUP_QUANT_8BIT_V2")
+        enable_v2 = group_size in _V2_KERNEL_SUPPORTED_GROUP_SIZES

    if enable_v2:
        return torch.ops.sgl_kernel.sgl_per_token_group_quant_8bit_v2.default(