This patch fixes issues that broke internal benchmarks as well as individual
fusions reported.

--- a/lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp
+++ b/lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp
@@ -351,6 +351,16 @@ getWarpsPerTile(DotOpInterface dotOp, const ArrayRef<int64_t> shape,
 }
 
 static bool bwdFilter(Operation *op) {
+  // Dot operand layout assignment to Predicates are not currently supported
+  // during lowering from TritonGPU to LLVM in Triton for MMA cases. This
+  // condition limits visibility of the original bit-width so that predicate
+  // are not considered, hence, kwidth can never be = 32.
+  if (isa<arith::UIToFPOp>(op)) {
+    Type srcType = getElementTypeOrSelf(op->getOperand(0));
+    if (srcType.isInteger(1))
+      return false;
+  }
+
   return (op->hasTrait<OpTrait::Elementwise>() && isMemoryEffectFree(op)) ||
          isView(op) ||
          isa<Fp4ToFpOp, LoadOp, DescriptorLoadLikeOpInterface, BroadcastOp,