Skip certain unit tests on NAVI (#1950)

This PR is to skip certain unit tests on NAVI only.
Fixes SWDEV-509011 - test_sac_ilp.py::TestSACILP::test_sac_ilp_case1
Fixes SWDEV-509311 -
test_max_autotune.py::TestMaxAutotune::test_non_contiguous_input_addmm
Fixes SWDEV-510738
test_fsdp_sharded_grad_scaler.py::TestShardedGradScalerParityWithDDP::test_sharded_grad_scaler_found_inf

Author: BLOrange-AMD

test/distributed/_tools/test_sac_ilp.py

@@ -19,7 +19,7 @@
     sac_milp,
 )
 from torch.testing._internal.common_cuda import TEST_CUDA
-from torch.testing._internal.common_utils import run_tests, skipIfTorchDynamo, TestCase
+from torch.testing._internal.common_utils import NAVI_ARCH, run_tests, skipIfTorchDynamo, TestCase, skipIfRocmArch
 from torch.testing._internal.distributed._tensor.common_dtensor import (
     ModelArgs,
     Transformer,
@@ -131,6 +131,7 @@ def _collect_module_info_with_fake_tensor_mode(self) -> ModuleInfo:
 
     @skipIfTorchDynamo("https://github.com/pytorch/pytorch/issues/115653")
     @unittest.skipIf(not TEST_CUDA, "CUDA not available")
+    @skipIfRocmArch(NAVI_ARCH)
     def test_sac_ilp_case1(self):
         """
         This is a case where the memory budget is either binding or too tight,

test/distributed/fsdp/test_fsdp_sharded_grad_scaler.py

@@ -36,6 +36,8 @@
     run_tests,
     TEST_WITH_DEV_DBG_ASAN,
     TestCase,
+    NAVI_ARCH,
+    skipIfRocmArch,
 )
 
 
@@ -236,6 +238,7 @@ def _build_model_and_optim(
         return model, optim, ref_model, ref_optim
 
     @skip_if_lt_x_gpu(2)
+    @skipIfRocmArch(NAVI_ARCH)
     def test_sharded_grad_scaler_found_inf(self):
         self.run_subtests(
             {

test/inductor/test_max_autotune.py

@@ -27,6 +27,8 @@
     instantiate_parametrized_tests,
     parametrize,
     TEST_WITH_ROCM,
+    NAVI_ARCH,
+    skipIfRocmArch,
 )
 from torch.utils._triton import has_triton_tma_device
 
@@ -881,6 +883,7 @@ def f(x, y):
         act = f(x, y)
         torch.testing.assert_close(act, ref, atol=2e-2, rtol=1e-2)
 
+    @skipIfRocmArch(NAVI_ARCH)
     def test_non_contiguous_input_addmm(self):
         b = torch.randn((768), dtype=torch.bfloat16, device="cuda")
         x = rand_strided(

torch/testing/_internal/common_utils.py

@@ -108,6 +108,7 @@
 
 
 MI300_ARCH = ("gfx940", "gfx941", "gfx942")
+NAVI_ARCH = ("gfx1030", "gfx1100", "gfx1101", "gfx1200", "gfx1201")
 
 
 def freeze_rng_state(*args, **kwargs):