[ROCm][CI] adjust tf32 tolerance for test_compile_kernel_advanced (pytorch#163783)

Fixes #ISSUE_NUMBER

Pull Request resolved: pytorch#163783
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <[email protected]>

Author: jeffdaily

test/test_cuda.py

@@ -6923,7 +6923,7 @@ def test_compile_kernel_large_shared_memory(self):
         with self.assertRaises(RuntimeError):
             kernel.set_shared_memory_config(excessive_shared_mem)
 
-    @tf32_on_and_off(0.005)
+    @tf32_on_and_off(0.05 if TEST_WITH_ROCM else 0.005)
     @unittest.skipIf(not TEST_CUDA, "No CUDA")
     def test_compile_kernel_advanced(self):
         # Test matrix multiplication