Merge branch 'dev' into codeql_update

Author: ericspod

monai/networks/nets/swin_unetr.py

@@ -98,6 +98,7 @@ def __init__(
         hyena_omega_0: float = 10.0,
         hyena_l_cache: int = 32,
         hyena_short_conv_fft_chunks: int = 0,
+        use_flash_attention: bool = False,
     ) -> None:
         """
         Args:
@@ -149,6 +150,7 @@ def __init__(
             hyena_omega_0: SIREN frequency. Default 10.0 (stable).
             hyena_l_cache: SIREN coordinate-grid cache size per spatial dim.
             hyena_short_conv_fft_chunks: channel chunk size for the FFT short conv (0 = no chunking).
+            use_flash_attention: use flash attention (scaled dot product attention) at inference.
 
         Examples::
 
@@ -224,6 +226,7 @@ def __init__(
             hyena_omega_0=hyena_omega_0,
             hyena_l_cache=hyena_l_cache,
             hyena_short_conv_fft_chunks=hyena_short_conv_fft_chunks,
+            use_flash_attention=use_flash_attention,
         )
 
         self.encoder1 = UnetrBasicBlock(
@@ -513,6 +516,7 @@ def __init__(
         qkv_bias: bool = False,
         attn_drop: float = 0.0,
         proj_drop: float = 0.0,
+        use_flash_attention: bool = False,
     ) -> None:
         """
         Args:
@@ -522,12 +526,17 @@ def __init__(
             qkv_bias: add a learnable bias to query, key, value.
             attn_drop: attention dropout rate.
             proj_drop: dropout rate of output.
+            use_flash_attention: if True, use ``torch.nn.functional.scaled_dot_product_attention`` for the
+                windowed attention. Equivalent to the default path but faster at inference; only used when
+                autograd is disabled (e.g. under ``torch.no_grad()`` or ``torch.inference_mode()``, not
+                ``eval()`` alone) and the module is not scripted.
         """
 
         super().__init__()
         self.dim = dim
         self.window_size = window_size
         self.num_heads = num_heads
+        self.use_flash_attention = use_flash_attention
         head_dim = dim // num_heads
         self.scale = head_dim**-0.5
         mesh_args = torch.meshgrid.__kwdefaults__
@@ -584,12 +593,26 @@ def forward(self, x, mask):
         b, n, c = x.shape
         qkv = self.qkv(x).reshape(b, n, 3, self.num_heads, c // self.num_heads).permute(2, 0, 3, 1, 4)
         q, k, v = qkv[0], qkv[1], qkv[2]
-        q = q * self.scale
-        attn = q @ k.transpose(-2, -1)
         relative_position_bias = self.relative_position_bias_table[
             self.relative_position_index.clone()[:n, :n].reshape(-1)  # type: ignore[operator]
         ].reshape(n, n, -1)
         relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
+        if self.use_flash_attention and not torch.jit.is_scripting() and not torch.is_grad_enabled():
+            # additive bias combines the relative position bias and, for shifted windows, the attention mask
+            if mask is not None:
+                nw = mask.shape[0]
+                bias = relative_position_bias.view(1, 1, self.num_heads, n, n) + mask.reshape(1, nw, 1, n, n)
+                bias = bias.expand(b // nw, nw, self.num_heads, n, n).reshape(b, self.num_heads, n, n)
+            else:
+                bias = relative_position_bias.unsqueeze(0)
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=bias.to(q.dtype), dropout_p=0.0, scale=self.scale
+            )
+            x = x.transpose(1, 2).reshape(b, n, c)
+            return self.proj_drop(self.proj(x))
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
         attn = attn + relative_position_bias.unsqueeze(0)
         if mask is not None:
             nw = mask.shape[0]
@@ -628,6 +651,7 @@ def __init__(
         act_layer: str = "GELU",
         norm_layer: type[LayerNorm] = nn.LayerNorm,
         use_checkpoint: bool = False,
+        use_flash_attention: bool = False,
     ) -> None:
         """
         Args:
@@ -643,6 +667,7 @@ def __init__(
             act_layer: activation layer.
             norm_layer: normalization layer.
             use_checkpoint: use gradient checkpointing for reduced memory usage.
+            use_flash_attention: use flash attention (scaled dot product attention) at inference.
         """
 
         super().__init__()
@@ -660,6 +685,7 @@ def __init__(
             qkv_bias=qkv_bias,
             attn_drop=attn_drop,
             proj_drop=drop,
+            use_flash_attention=use_flash_attention,
         )
 
         self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
@@ -924,6 +950,7 @@ def __init__(
         hyena_omega_0: float = 10.0,
         hyena_l_cache: int = 32,
         hyena_short_conv_fft_chunks: int = 0,
+        use_flash_attention: bool = False,
     ) -> None:
         """
         Args:
@@ -946,6 +973,7 @@ def __init__(
                 hyena_use_chunked_fft, hyena_use_fft_short_conv, hyena_omega_0, hyena_l_cache,
                 hyena_short_conv_fft_chunks: forwarded to :class:`HyenaTransformerBlock`. See its
                 docstring for semantics.
+            use_flash_attention: use flash attention (scaled dot product attention) at inference.
         """
 
         super().__init__()
@@ -996,6 +1024,7 @@ def __init__(
                         drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                         norm_layer=norm_layer,
                         use_checkpoint=use_checkpoint,
+                        use_flash_attention=use_flash_attention,
                     )
                     for i in range(depth)
                 ]
@@ -1079,6 +1108,7 @@ def __init__(
         hyena_omega_0: float = 10.0,
         hyena_l_cache: int = 32,
         hyena_short_conv_fft_chunks: int = 0,
+        use_flash_attention: bool = False,
     ) -> None:
         """
         Args:
@@ -1112,6 +1142,7 @@ def __init__(
                 hyena_use_chunked_fft, hyena_use_fft_short_conv, hyena_omega_0, hyena_l_cache,
                 hyena_short_conv_fft_chunks: HyenaND configuration. See
                 :class:`monai.networks.blocks.HyenaTransformerBlock` for semantics.
+            use_flash_attention: use flash attention (scaled dot product attention) at inference.
         """
 
         super().__init__()
@@ -1193,6 +1224,7 @@ def __init__(
                 hyena_omega_0=hyena_omega_0,
                 hyena_l_cache=hyena_l_cache,
                 hyena_short_conv_fft_chunks=hyena_short_conv_fft_chunks,
+                use_flash_attention=use_flash_attention,
             )
             if i_layer == 0:
                 self.layers1.append(layer)

monai/transforms/croppad/array.py

@@ -342,6 +342,24 @@ def compute_pad_width(self, spatial_shape: Sequence[int]) -> tuple[tuple[int, in
         return spatial_pad.compute_pad_width(spatial_shape)
 
 
+def _to_int_list(data: Sequence[int] | int | NdarrayOrTensor) -> list[int]:
+    """Coerce an ROI spec (scalar, sequence, tensor or ndarray) to a list of Python ints."""
+    if isinstance(data, (str, bytes)):
+        raise TypeError("ROI specs must be integers or sequences of integers, not strings.")
+    return [int(i) for i in ensure_tuple(data)]
+
+
+def _broadcast_int_pair(
+    a: Sequence[int] | int | NdarrayOrTensor, b: Sequence[int] | int | NdarrayOrTensor
+) -> tuple[list[int], list[int]]:
+    """Coerce a pair of ROI specs to two equal-length int lists, broadcasting a scalar to match."""
+    list_a, list_b = _to_int_list(a), _to_int_list(b)
+    n = max(len(list_a), len(list_b))
+    if len(list_a) not in (1, n) or len(list_b) not in (1, n):
+        raise ValueError(f"ROI specs must have matching lengths or be scalar, got {len(list_a)} and {len(list_b)}.")
+    return (list_a * n if len(list_a) == 1 else list_a), (list_b * n if len(list_b) == 1 else list_b)
+
+
 class Crop(InvertibleTransform, LazyTransform):
     """
     Perform crop operations on the input image.
@@ -379,31 +397,22 @@ def compute_slices(
             roi_slices: list of slices for each of the spatial dimensions.
 
         """
-        roi_start_t: torch.Tensor
-
         if roi_slices:
             if not all(s.step is None or s.step == 1 for s in roi_slices):
                 raise ValueError(f"only slice steps of 1/None are currently supported, got {roi_slices}.")
             return ensure_tuple(roi_slices)
         else:
             if roi_center is not None and roi_size is not None:
-                roi_center_t = convert_to_tensor(data=roi_center, dtype=torch.int16, wrap_sequence=True, device="cpu")
-                roi_size_t = convert_to_tensor(data=roi_size, dtype=torch.int16, wrap_sequence=True, device="cpu")
-                _zeros = torch.zeros_like(roi_center_t)
-                half = torch.divide(roi_size_t, 2, rounding_mode="floor")
-                roi_start_t = torch.maximum(roi_center_t - half, _zeros)
-                roi_end_t = torch.maximum(roi_start_t + roi_size_t, roi_start_t)
+                centers, sizes = _broadcast_int_pair(roi_center, roi_size)
+                starts = [max(c - s // 2, 0) for c, s in zip(centers, sizes)]
+                ends = [st + s for st, s in zip(starts, sizes)]
             else:
                 if roi_start is None or roi_end is None:
                     raise ValueError("please specify either roi_center, roi_size or roi_start, roi_end.")
-                roi_start_t = convert_to_tensor(data=roi_start, dtype=torch.int16, wrap_sequence=True)
-                roi_start_t = torch.maximum(roi_start_t, torch.zeros_like(roi_start_t))
-                roi_end_t = convert_to_tensor(data=roi_end, dtype=torch.int16, wrap_sequence=True)
-                roi_end_t = torch.maximum(roi_end_t, roi_start_t)
-            # convert to slices (accounting for 1d)
-            if roi_start_t.numel() == 1:
-                return ensure_tuple([slice(int(roi_start_t.item()), int(roi_end_t.item()))])
-            return ensure_tuple([slice(int(s), int(e)) for s, e in zip(roi_start_t.tolist(), roi_end_t.tolist())])
+                starts, ends = _broadcast_int_pair(roi_start, roi_end)
+                starts = [max(s, 0) for s in starts]
+            # clamp each end to its own start so no slice has negative width
+            return ensure_tuple(slice(s, max(e, s)) for s, e in zip(starts, ends))
 
     def __call__(  # type: ignore[override]
         self, img: torch.Tensor, slices: tuple[slice, ...], lazy: bool | None = None

tests/networks/nets/test_swin_unetr.py

@@ -111,6 +111,19 @@ def test_invalid_input_shape(self):
         with self.assertRaises(ValueError):
             net_2d(torch.randn(1, 1, 48, 33))  # 33 is not divisible by 32
 
+    @skipUnless(has_einops, "Requires einops")
+    def test_flash_attention(self):
+        input_param = {"in_channels": 1, "out_channels": 2, "feature_size": 12, "spatial_dims": 3}
+        net_ref = SwinUNETR(use_flash_attention=False, **input_param).double()
+        net_flash = SwinUNETR(use_flash_attention=True, **input_param).double()
+        net_flash.load_state_dict(net_ref.state_dict())
+        x = torch.randn(1, 1, 64, 64, 64, dtype=torch.float64)
+        with eval_mode(net_ref, net_flash):
+            ref = net_ref.swinViT(x, net_ref.normalize)
+            out = net_flash.swinViT(x, net_flash.normalize)
+        for a, b in zip(ref, out, strict=True):
+            assert_allclose(a, b, atol=1e-6, rtol=1e-6, type_test=False)
+
     def test_patch_merging(self):
         dim = 10
         t = PatchMerging(dim)(torch.zeros((1, 21, 20, 20, dim)))

tests/transforms/test_center_spatial_crop.py

@@ -14,9 +14,12 @@
 import unittest
 
 import numpy as np
+import torch
 from parameterized import parameterized
 
+from monai.data.meta_obj import get_track_meta, set_track_meta
 from monai.transforms import CenterSpatialCrop
+from monai.transforms.croppad.array import Crop
 from tests.croppers import CropTest
 
 TEST_SHAPES = [
@@ -50,6 +53,28 @@ def test_value(self, input_param, input_arr, expected_arr):
     def test_pending_ops(self, input_param, input_shape, _, align_corners):
         self.crop_test_pending_ops(input_param, input_shape, align_corners)
 
+    def test_compute_slices_broadcast(self):
+        self.assertEqual(Crop.compute_slices(roi_center=2, roi_size=(4, 6, 8)), (slice(0, 4), slice(0, 6), slice(0, 8)))
+        self.assertEqual(Crop.compute_slices(roi_start=1, roi_end=(3, 5, 7)), (slice(1, 3), slice(1, 5), slice(1, 7)))
+        with self.assertRaises(ValueError):
+            Crop.compute_slices(roi_center=(2, 3), roi_size=(4, 5, 6))
+        with self.assertRaises(ValueError):
+            Crop.compute_slices(roi_start=(1, 2), roi_end=(3, 5, 7))
+        with self.assertRaises(TypeError):
+            Crop.compute_slices(roi_center="10", roi_size=(4, 6))
+
+    def test_torch_compile(self):
+        prev_track_meta = get_track_meta()
+        set_track_meta(False)
+        try:
+            # eager backend traces the transform without needing the Inductor C++ compiler
+            cropper = torch.compile(CenterSpatialCrop(roi_size=(1, 16, 16)), backend="eager")
+            img = torch.rand(1, 1, 32, 32, dtype=torch.float32)
+            self.assertEqual(tuple(cropper(img).shape), (1, 1, 16, 16))
+        finally:
+            set_track_meta(prev_track_meta)
+            torch._dynamo.reset()
+
 
 if __name__ == "__main__":
     unittest.main()