From d523ae127f3cf5008bd54d99f582ed31d621e8fc Mon Sep 17 00:00:00 2001 From: Brayden Zhong Date: Fri, 15 May 2026 20:28:35 -0400 Subject: [PATCH] Fix Mistral Large 3 nightly test (#25407) Co-authored-by: b8zhong --- .../schemes/compressed_tensors_w4a4_nvfp4_moe.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py index 69e572498..0a9921874 100644 --- a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py +++ b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py @@ -311,10 +311,10 @@ class CompressedTensorsW4A4Nvfp4MoE(CompressedTensorsMoEScheme): router_logits = topk_output.router_logits topk_config = topk_output.topk_config - # Quantize input hidden states using fp4_quantize + # global_scale must be shape [1] (strict in cute-dsl backend). hs_fp4_bytes, hs_sf_bytes = fp4_quantize( x, - layer.w13_input_scale_quant, + layer.w13_input_scale_quant[:1], self.group_size, # sf_vec_size False, # use_ue8m0 False, # is_sf_swizzled_layout