[ROCm] Prevent accidental enablement of efficient attention. (pytorch#134531)

pytorchbot · xinyazhang · xinyazhang · commit b4adece5e0a8 · 2024-08-28T16:29:21.000Z
[ROCm] Prevent accidental enablement of efficient attention. (pytorch#133331) Currently Efficient attention and Flash attention share the same set of GPU kernels on ROCM and have common limitations on head sizes. Fixes pytorch#132004 Pull Request resolved: pytorch#133331 Approved by: https://github.com/malfet, https://github.com/jithunnair-amd (cherry picked from commit 46ecc67) Co-authored-by: Xinya Zhang <Xinya.Zhang@amd.com>
diff --git a/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp b/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
@@ -620,7 +620,12 @@ bool can_use_mem_efficient_attention(sdp_params const& params, bool debug) {
       check_all_tensors_on_device,
       check_mem_efficient_hardware_support,
       check_tensor_shapes,
-      check_head_dim_size_mem_efficient);
+#ifdef USE_ROCM
+      check_head_dim_size_flash
+#else
+      check_head_dim_size_mem_efficient
+#endif
+  );
   for (auto& constraint : general_constraints) {
     if (!constraint(params, debug)) {
       return false;
diff --git a/test/test_transformers.py b/test/test_transformers.py
@@ -1457,6 +1457,8 @@ def test_invalid_fused_inputs_head_dim(self, device, kernel: SDPBackend):
             dtype = torch.float16
             make_tensor = partial(torch.rand, device=device, dtype=dtype)
             size = SdpaShape(2, 2, 3, 9) if kernel == SDPBackend.EFFICIENT_ATTENTION else SdpaShape(2, 2, 3, 257)
+            if TEST_WITH_ROCM:  # On ROCM, FA and EA share the backend GPU kernels
+                size = SdpaShape(2, 2, 3, 257)
             q, k, v = make_tensor(size), make_tensor(size), make_tensor(size)
             self.assertRaises(RuntimeError, lambda: torch.nn.functional.scaled_dot_product_attention(
                 q, k, v, None, 0.0, False))
@@ -1499,8 +1501,9 @@ def test_unaligned_tensors(self, device):
         make_tensor = partial(torch.rand, size, device=device, dtype=dtype)
         q, k, v = make_tensor(), make_tensor(), make_tensor()
         with sdpa_kernel(backends=[SDPBackend.EFFICIENT_ATTENTION]):
-            self.assertRaises(RuntimeError, lambda: torch.nn.functional.scaled_dot_product_attention(
-                q, k, v, None, 0.0, False))
+            ctxmgr = self.assertRaises(RuntimeError) if not TEST_WITH_ROCM else contextlib.nullcontext()
+            with ctxmgr:
+                torch.nn.functional.scaled_dot_product_attention(q, k, v, None, 0.0, False)
 
     @onlyCUDA
     @unittest.skipIf(not PLATFORM_SUPPORTS_FLASH_ATTENTION, "Does not support fused SDPA or pre-SM80 hardware")
