Disable autotuning in test - with --xla_gpu_experimental_enable_fusion_autotuner set to true, this will start to fail because the error returned is actually "No executables to profile!" because all the individual backends fail to compile (and correctly so). Remove autotuning to continue the tests previous goals of testing the native emitter's return error.

vwbaker · tensorflower-gardener · commit d09aefd4616d · 2025-10-06T22:09:27.000-07:00
PiperOrigin-RevId: 816033671
diff --git a/third_party/xla/xla/service/gpu/tests/gpu_kernel_tiling_test.cc b/third_party/xla/xla/service/gpu/tests/gpu_kernel_tiling_test.cc
@@ -399,6 +399,12 @@ TEST_F(GpuKernelTilingTest, ReductionInputTooLarge) {
   }
   )";
   auto hlo_module = ParseAndReturnVerifiedModule(kHloString).value();
+  // Disable autotuning because this is checking for an error returned by the
+  // Native Emitter. With autotuning enabled, the error is that autotuning
+  // itself fails to find a config because all the backends return failure.
+  hlo_module->mutable_config()
+      .mutable_debug_options()
+      .set_xla_gpu_autotune_level(0);
   absl::Status status = CompileToExecutable(std::move(hlo_module)).status();
 
   if (xla::PlatformUtil::CanonicalPlatformName("gpu").value() == "rocm") {
