[release/2.12] CI UT fixes

test/test_cuda.py

@@ -5045,12 +5045,16 @@ def test_throw_on_cudamalloc_oom(self):
             # preemptively rejected with OutOfMemoryError.
             # Both settings must go through _accelerator_setAllocatorSettings so
             # they are read from CUDAAllocatorConfig.
+            fraction = 0.005
             torch._C._accelerator_setAllocatorSettings(
-                "throw_on_cudamalloc_oom:True,per_process_memory_fraction:0.01"
+                f"throw_on_cudamalloc_oom:True,per_process_memory_fraction:{fraction}"
             )
 
+            total_mem = torch.cuda.get_device_properties(0).total_memory
+            # Allocate the allowed threshold + 1 MiB to guarantee rejection
+            alloc_bytes = int(total_mem * fraction) + 1024 * 1024
             with self.assertRaises(torch.cuda.OutOfMemoryError):
-                torch.empty(1024 * 1024 * 1024, dtype=torch.int8, device="cuda")
+                torch.empty(alloc_bytes, dtype=torch.int8, device="cuda")
 
             # Check that rejection counter was incremented
             stats = torch.cuda.memory_stats()

test/test_torch.py

@@ -828,6 +828,7 @@ def test_cpp_warnings_have_python_context(self, device):
         s = ".+Triggered internally at.+RangeFactories.+"
         # nvfuser deprecation warning filter
         warnings.filterwarnings("ignore", "torch::jit::fuser::cuda", UserWarning)
+        warnings.filterwarnings("ignore", ".*", DeprecationWarning) # ignore all deprecation warnings
 
         def cpp_warn_fn():
             out = torch.empty((5,))