Fix Mistral Large 3 nightly test (#25407)

Co-authored-by: b8zhong <b8zhong@users.noreply.github.com>
2026-05-15 17:28:35 -07:00
parent 5ba69f50fb
commit d523ae127f
1 changed files with 2 additions and 2 deletions
@@ -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