[AUTOGENERATED] [release/2.7] NAVI32 specific fixes (#2466)

Cherry-pick of #2450

---------

Co-authored-by: iupaikov-amd <[email protected]>

Author: okakarpa

test/inductor/test_flex_decoding.py

@@ -21,7 +21,10 @@
 )
 from torch.testing import FileCheck
 from torch.testing._internal import common_utils
-from torch.testing._internal.common_cuda import PLATFORM_SUPPORTS_BF16
+from torch.testing._internal.common_cuda import (
+    PLATFORM_SUPPORTS_BF16,
+    PLATFORM_SUPPORTS_FLASH_ATTENTION,
+)
 from torch.testing._internal.common_utils import skipIfRocm
 from torch.utils._triton import has_triton
 
@@ -1421,6 +1424,7 @@ def mask_mod(b, h, q, kv):
         self.assertEqual(query.grad[:, :, M:, :].sum(), 0)
 
     @supported_platform
+    @unittest.skipIf(not PLATFORM_SUPPORTS_FLASH_ATTENTION, "Some archs don't support SDPA")
     def test_windowed_no_mask_vs_sdpa(self):
         score_mod = _generate_windowed(1000)
         attention = functools.partial(flex_attention, score_mod=score_mod)

test/inductor/test_max_autotune.py

@@ -27,6 +27,7 @@
     TritonTemplateCaller,
 )
 from torch.testing._internal.common_cuda import PLATFORM_SUPPORTS_FP8
+from torch.testing._internal.common_device_type import largeTensorTest
 from torch.testing._internal.common_utils import (
     instantiate_parametrized_tests,
     IS_WINDOWS,
@@ -44,7 +45,12 @@
 from torch.fx.experimental.proxy_tensor import make_fx
 from torch.testing import FileCheck
 from torch.testing._internal.common_utils import skipIfRocm, skipIfXpu
-from torch.testing._internal.inductor_utils import GPU_TYPE, HAS_CPU, HAS_CUDA, HAS_GPU
+from torch.testing._internal.inductor_utils import (
+    GPU_TYPE,
+    HAS_CPU,
+    HAS_CUDA,
+    HAS_GPU,
+)
 
 
 torch.set_float32_matmul_precision("high")
@@ -981,6 +987,8 @@ def test_conv_backend(self):
 
         self.assertIn("NoValidChoicesError", str(context.exception))
 
+    # Some ROCm GPUs don't have enough VRAM to run all autotune configurations and padding benchmarks
+    @largeTensorTest("30 GB", device=GPU_TYPE)
     def test_non_contiguous_input_mm(self):
         """
         Make sure the triton template can work with non-contiguous inputs without crash.
@@ -1033,6 +1041,8 @@ def f(x, y):
     # TODO: fix accuracy failure of the triton template on XPU.
     # and enable this test case.
     @skipIfXpu
+    # Some ROCm GPUs don't have enough VRAM to run all autotune configurations and padding benchmarks
+    @largeTensorTest("30 GB", device=GPU_TYPE)
     def test_non_contiguous_input_mm_plus_mm(self):
         x1 = rand_strided((50257, 32768), (1, 50304), device=GPU_TYPE)
         y1 = rand_strided((32768, 768), (768, 1), device=GPU_TYPE)