--- a/lib/Dialect/TritonGPU/Transforms/Prefetch.cpp
+++ b/lib/Dialect/TritonGPU/Transforms/Prefetch.cpp
@@ -433,6 +433,15 @@ LogicalResult Prefetcher::initialize() {
     while (op) {
       if (!op->getResult(0).hasOneUse())
         break;
+      // Similar to issues faced in HoistLayoutConversion pattern in
+      // OptimizeDotOperands.cpp, we can't propagate through type casts from
+      // predicates as they aren't supported in Triton when encoded with dot_op
+      // layout.
+      if (isa<arith::UIToFPOp>(op)) {
+        Type srcType = getElementTypeOrSelf(op->getOperand(0));
+        if (srcType.isInteger(1))
+          break;
+      }
       if (auto load = dyn_cast<triton::gpu::LocalLoadOp>(op)) {
         rets.push_back(load.getSrc());
         // Only handle the direct dot-operand load chain for now.