Backport of https://github.com/triton-lang/triton/pull/11245.

--- a/lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp
+++ b/lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp
@@ -877,11 +877,6 @@ public:
     bool isFp4MMA = isAFP4 && isBFP4;
     bool requiresFp4Padding =
         nvidia_gpu::TargetFeatures(computeCapability).requiresFp4Padding();
-    // mxf4/mxf4nvf4 do not support MN-major operands, so FP4 x FP4 uses
-    // mxf8f6f4 if either operand is not K-packed. The mxf8f6f4 shared-memory
-    // packing format requires padding for every FP4 operand.
-    bool isFp4MMAUsingMxf8f6f4 =
-        isFp4MMA && (!dotOp.getLhsKPack() || !dotOp.getRhsKPack());
 
     // On Blackwell, if we use mixed-precision MMA we need to pad the fp4
     // operand
@@ -893,11 +888,33 @@ public:
     // MMA_K = 64 is also not supported when BLOCK_M = 64.
     auto blockK = dotOp.getA().getType().getShape().back() * (isAFP4 ? 2 : 1);
     auto blockM = dotOp.getA().getType().getShape()[0];
+    bool hasMNMajorFp4Operand =
+        isFp4MMA && (!dotOp.getLhsKPack() || !dotOp.getRhsKPack());
+    auto aScaleType = dotOp.getAScale().getType();
+    auto bScaleType = dotOp.getBScale().getType();
+    auto aScaleElemType = aScaleType.getElementType();
+    auto bScaleElemType = bScaleType.getElementType();
+    auto isBlock16Scale = [blockK](RankedTensorType scaleType) {
+      return scaleType.getShape().back() * 16 == blockK;
+    };
+    bool hasUE4M3Scale = isa<Float8E4M3FNType>(aScaleElemType) ||
+                         isa<Float8E4M3FNType>(bScaleElemType);
+    bool hasUE5M3Scale =
+        (aScaleElemType.isInteger(8) && isBlock16Scale(aScaleType)) ||
+        (bScaleElemType.isInteger(8) && isBlock16Scale(bScaleType));
+    // mxf4nvf4 supports UE4M3/UE5M3 scales but not MN-major operands, while
+    // the mxf8f6f4 fallback for MN-major FP4 only supports E8M0 scales.
+    if (hasMNMajorFp4Operand && (hasUE4M3Scale || hasUE5M3Scale))
+      return failure();
+    // mxf4 does not support MN-major operands, so mxfp4 x mxfp4 falls back to
+    // mxf8f6f4 if either operand is not K-packed. The mxf8f6f4 shared-memory
+    // packing format requires padding for every fp4 operand, even if the
+    // operand is K packed.
     bool isMMAv5Fp4PaddedLhs =
-        isFp4MMAUsingMxf8f6f4 || // fp4  x fp4 using mxf8f6f4
+        hasMNMajorFp4Operand ||
         (IsAMixedPrecFp4 && (requiresFp4Padding || blockM == 64 ||
                              blockK == 32 || !dotOp.getLhsKPack()));
     bool isMMAv5Fp4PaddedRhs =
-        isFp4MMAUsingMxf8f6f4 || // fp4  x fp4 using mxf8f6f4
+        hasMNMajorFp4Operand ||
         (IsBMixedPrecFp4 &&
          (requiresFp4Padding || blockK == 32 || !dotOp.getRhsKPack()));