fix(quantization): add sgl_kernel fallback for FP4 quantize on Blackwell GPUs (#17816)

2026-01-27 11:43:17 +01:00
parent dba264ac73
commit c56d19b977
1 changed files with 5 additions and 2 deletions
--- a/python/sglang/srt/layers/quantization/modelopt_quant.py
+++ b/python/sglang/srt/layers/quantization/modelopt_quant.py
@@ -66,12 +66,15 @@ if TYPE_CHECKING:
        StandardDispatchOutput,
    )

+fp4_quantize = None
 try:
    if is_sm120_supported():
-        from flashinfer import fp4_quantize
+        try:
+            from flashinfer import fp4_quantize
+        except ImportError:
+            from sgl_kernel import scaled_fp4_quant as fp4_quantize
    else:
        from sgl_kernel import scaled_fp4_quant as fp4_quantize
-
 except ImportError:
    fp4_quantize = None