From c56d19b9779cc5e4b613e9fee5cb1f13fe5f1a0b Mon Sep 17 00:00:00 2001
From: MikkoParkkola <78788115+MikkoParkkola@users.noreply.github.com>
Date: Tue, 27 Jan 2026 11:43:17 +0100
Subject: [PATCH] fix(quantization): add sgl_kernel fallback for FP4 quantize
 on Blackwell GPUs (#17816)

---
 python/sglang/srt/layers/quantization/modelopt_quant.py | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/layers/quantization/modelopt_quant.py b/python/sglang/srt/layers/quantization/modelopt_quant.py
index f7fc3d026..dc9501133 100755
--- 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