Force use of torch.compile on deterministic roi_align implementation (#…

…8436)

Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Co-authored-by: Nicolas Hug <nh.nicolas.hug@gmail.com>

test/test_ops.py

@@ -14,6 +14,7 @@
 from common_utils import assert_equal, cpu_and_cuda, cpu_and_cuda_and_mps, needs_cuda, needs_mps
 from PIL import Image
 from torch import nn, Tensor
+from torch._dynamo.utils import is_compile_supported
 from torch.autograd import gradcheck
 from torch.nn.modules.utils import _pair
 from torchvision import models, ops
@@ -529,6 +530,10 @@ def test_autocast_cpu(self, aligned, deterministic, x_dtype, rois_dtype):
     def test_backward(self, seed, device, contiguous, deterministic):
         if deterministic and device == "cpu":
             pytest.skip("cpu is always deterministic, don't retest")
+        if deterministic and device == "mps":
+            pytest.skip("no deterministic implementation for mps")
+        if deterministic and not is_compile_supported(device):
+            pytest.skip("deterministic implementation only if torch.compile supported")
         super().test_backward(seed, device, contiguous, deterministic)
 
     def _make_rois(self, img_size, num_imgs, dtype, num_rois=1000):

torchvision/ops/roi_align.py

@@ -1,9 +1,11 @@
+import functools
 from typing import List, Union
 
 import torch
 import torch._dynamo
 import torch.fx
 from torch import nn, Tensor
+from torch._dynamo.utils import is_compile_supported
 from torch.jit.annotations import BroadcastingList2
 from torch.nn.modules.utils import _pair
 from torchvision.extension import _assert_has_ops, _has_ops
@@ -12,6 +14,24 @@
 from ._utils import check_roi_boxes_shape, convert_boxes_to_roi_format
 
 
+def lazy_compile(**compile_kwargs):
+    """Lazily wrap a function with torch.compile on the first call
+
+    This avoids eagerly importing dynamo.
+    """
+
+    def decorate_fn(fn):
+        @functools.wraps(fn)
+        def compile_hook(*args, **kwargs):
+            compiled_fn = torch.compile(fn, **compile_kwargs)
+            globals()[fn.__name__] = functools.wraps(fn)(compiled_fn)
+            return compiled_fn(*args, **kwargs)
+
+        return compile_hook
+
+    return decorate_fn
+
+
 # NB: all inputs are tensors
 def _bilinear_interpolate(
     input,  # [N, C, H, W]
@@ -86,15 +106,13 @@ def maybe_cast(tensor):
         return tensor
 
 
-# This is a slow but pure Python and differentiable implementation of
-# roi_align.  It potentially is a good basis for Inductor compilation
-# (but I have not benchmarked it) but today it is solely used for the
-# fact that its backwards can be implemented deterministically,
-# which is needed for the PT2 benchmark suite.
-#
+# This is a pure Python and differentiable implementation of roi_align.  When
+# run in eager mode, it uses a lot of memory, but when compiled it has
+# acceptable memory usage.  The main point of this implementation is that
+# its backwards is deterministic.
 # It is transcribed directly off of the roi_align CUDA kernel, see
 # https://dev-discuss.pytorch.org/t/a-pure-python-implementation-of-roi-align-that-looks-just-like-its-cuda-kernel/1266
-@torch._dynamo.allow_in_graph
+@lazy_compile(dynamic=True)
 def _roi_align(input, rois, spatial_scale, pooled_height, pooled_width, sampling_ratio, aligned):
     orig_dtype = input.dtype
 
@@ -232,7 +250,9 @@ def roi_align(
     if not isinstance(rois, torch.Tensor):
         rois = convert_boxes_to_roi_format(rois)
     if not torch.jit.is_scripting():
-        if not _has_ops() or (torch.are_deterministic_algorithms_enabled() and (input.is_cuda or input.is_mps)):
+        if (
+            not _has_ops() or (torch.are_deterministic_algorithms_enabled() and (input.is_cuda or input.is_mps))
+        ) and is_compile_supported(input.device.type):
             return _roi_align(input, rois, spatial_scale, output_size[0], output_size[1], sampling_ratio, aligned)
     _assert_has_ops()
     return torch.ops.torchvision.roi_align(