[AUTOGENERATED] [release/2.7] NAVI32 specific fixes

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

@@ -32,6 +32,7 @@
     IS_WINDOWS,
     parametrize,
     TEST_WITH_ROCM,
+    skipIfRocmNotEnoughMemory,
 )
 from torch.utils._triton import has_triton_tma_device
 
@@ -981,6 +982,8 @@ 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
+    @skipIfRocmNotEnoughMemory(30)
     def test_non_contiguous_input_mm(self):
         """
         Make sure the triton template can work with non-contiguous inputs without crash.
@@ -1033,6 +1036,8 @@ 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
+    @skipIfRocmNotEnoughMemory(30)
     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)

torch/testing/_internal/common_utils.py

@@ -1902,6 +1902,31 @@ def wrap_fn(self, *args, **kwargs):
         return wrap_fn
     return dec_fn
 
+# Checks if current ROCm device has enough VRAM against the required amount in GB
+def skipIfRocmNotEnoughMemory(required_amount):
+    def dec_fn(fn):
+        @wraps(fn)
+        def wrap_fn(self, *args, **kwargs):
+            if TEST_WITH_ROCM:
+                device = torch.cuda.current_device()
+                props = torch.cuda.get_device_properties(device)
+
+                total = props.total_memory / (1024 ** 3)  # in GB
+                # This will probably return 0 because it only counts tensors
+                # and doesn't take into account any small supporting allocations
+                allocated = torch.cuda.memory_allocated(device) / (1024 ** 3)
+                free_global = total - allocated
+
+                result = free_global > required_amount
+
+                if not result:
+                    reason = f"skipIfRocm: Not enough free VRAM on current ROCm device. " \
+                        f"Available: {free_global:.2f} GB | Required: {required_amount:.2f} GB."
+                    raise unittest.SkipTest(reason)
+            return fn(self, *args, **kwargs)
+        return wrap_fn
+    return dec_fn
+
 def runOnRocm(fn):
     @wraps(fn)
     def wrapper(*args, **kwargs):