From ef134d407d6965c3e0ac3c35d3323a9a8ca20033 Mon Sep 17 00:00:00 2001 From: b8zhong Date: Sat, 31 Jan 2026 10:01:29 -0500 Subject: [PATCH] [Fix] Revert back to using CUTLASS `mm_fp4` backend (#17369) --- python/sglang/srt/layers/quantization/modelopt_quant.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/quantization/modelopt_quant.py b/python/sglang/srt/layers/quantization/modelopt_quant.py index dc9501133..7a0dfed08 100755 --- a/python/sglang/srt/layers/quantization/modelopt_quant.py +++ b/python/sglang/srt/layers/quantization/modelopt_quant.py @@ -134,7 +134,11 @@ def fp4_gemm( fp4_backend = get_fp4_gemm_runner_backend() if enable_flashinfer_fp4_gemm: # Use the remapping logic to convert SGLang backend names to FlashInfer API names - backend = fp4_backend.get_flashinfer_backend() + backend = ( + fp4_backend.get_flashinfer_backend() + if not fp4_backend.is_auto() + else "cutlass" + ) return flashinfer_fp4_gemm( input, weight, input_sf, weight_sf, alpha, out_dtype, backend=backend )