Use the correct wrapper for fp4_quantize (#27956)

Co-authored-by: Brayden Zhong <brayden@radixark.ai>
2026-06-11 20:23:09 -07:00
parent 40894be3c3
commit 8bfcc0c39c
1 changed files with 1 additions and 2 deletions
@@ -286,8 +286,7 @@ class DeepseekV2MLP(nn.Module):
            and self.swiglu_limit is None
            and not isinstance(x, tuple)
        ):
-            from flashinfer import fp4_quantize
-
+            from sglang.srt.layers.quantization.fp4_utils import fp4_quantize
            from sglang.srt.layers.quantization.nvfp4_gemm_swiglu_nvfp4_quant import (
                nvfp4_gemm_swiglu_nvfp4_quant,
            )