add autocast for conv1/2/3d

Author: t-vi

thunder/core/transforms.py

@@ -3856,6 +3856,76 @@ def autocast_linear_rule(a, w, bias, dtype):
     return _linear_autocast_impl(a, w, bias, dtype)
 
 
+def _convolution_autocast_impl(a, w, bias, *other_args, dtype):
+    if bias is None:
+        # Don't pass `bias` to maybe_downcast_to.
+        downcast_args = maybe_downcast_to(dtype, (a, w)) + (bias,)
+    else:
+        downcast_args = maybe_downcast_to(dtype, (a, w, bias))
+
+    return prims.convolution(*downcast_args, *other_args)
+
+
+@register_autocast_rule("torch.nn.functional.conv1d")
+def autocast_ltorch_conv1d_rule(
+    a: TensorProxy,
+    /,
+    weight: TensorProxy,
+    bias: TensorProxy | None = None,
+    stride: int | Sequence[int] = 1,
+    padding: int | Sequence[int] | str = 0,
+    dilation: int = 1,
+    groups: int = 1,
+    *,
+    dtype,
+) -> TensorProxy:
+    from thunder.torch import _conv_helper
+
+    return _conv_helper(
+        1, a, weight, bias, stride, padding, dilation, groups, conv_function=_convolution_autocast_impl, dtype=dtype
+    )
+
+
+@register_autocast_rule("torch.nn.functional.conv2d")
+def autocast_ltorch_conv2d_rule(
+    a: TensorProxy,
+    /,
+    weight: TensorProxy,
+    bias: TensorProxy | None = None,
+    stride: int | Sequence[int] = 1,
+    padding: int | Sequence[int] | str = 0,
+    dilation: int = 1,
+    groups: int = 1,
+    *,
+    dtype,
+) -> TensorProxy:
+    from thunder.torch import _conv_helper
+
+    return _conv_helper(
+        2, a, weight, bias, stride, padding, dilation, groups, conv_function=_convolution_autocast_impl, dtype=dtype
+    )
+
+
+@register_autocast_rule("torch.nn.functional.conv3d")
+def autocast_ltorch_conv3d_rule(
+    a: TensorProxy,
+    /,
+    weight: TensorProxy,
+    bias: TensorProxy | None = None,
+    stride: int | Sequence[int] = 1,
+    padding: int | Sequence[int] | str = 0,
+    dilation: int = 1,
+    groups: int = 1,
+    *,
+    dtype,
+) -> TensorProxy:
+    from thunder.torch import _conv_helper
+
+    return _conv_helper(
+        3, a, weight, bias, stride, padding, dilation, groups, conv_function=_convolution_autocast_impl, dtype=dtype
+    )
+
+
 @register_autocast_rule("torch.nn.functional.scaled_dot_product_attention")
 def autocast_scaled_dot_product_attention(
     query,

thunder/tests/test_autocast.py

@@ -169,3 +169,29 @@ def foo(x, w):
         jit_out = jfoo(x, w)
 
     torch.testing.assert_close(eager_out, jit_out)
+
+
+@pytest.mark.parametrize("dim", [1, 2, 3])
+def test_autocast_convolution(dim):
+    conv_fn = getattr(torch.nn.functional, f"conv{dim}d")
+
+    def foo(x, w, b=None):
+        return conv_fn(x, w, b)
+
+    x = torch.randn(1, 2, *(dim * (8,)))
+    w = torch.randn(3, 2, *(dim * (4,)))
+    b = torch.randn(3)
+
+    jfoo = thunder.jit(foo)
+
+    with torch.autocast("cpu", torch.bfloat16):
+        eager_out = foo(x, w, b)
+        jit_out = jfoo(x, w, b)
+
+    torch.testing.assert_close(eager_out, jit_out)
+
+    with torch.autocast("cpu", torch.bfloat16):
+        eager_out = foo(x, w)
+        jit_out = jfoo(x, w)
+
+    torch.testing.assert_close(eager_out, jit_out)

thunder/torch/__init__.py

@@ -3816,6 +3816,9 @@ def _conv_helper(
     padding: int | Sequence[int] | str = 0,
     dilation: int | Sequence[int] = 1,
     groups: int = 1,
+    *,
+    conv_function=clang.convolution,
+    **extra_kwargs,
 ) -> TensorProxy:
     # a, weight rank check
     utils.check(dim + 1 <= a.ndim <= dim + 2, lambda: f"{a.ndim=} should be either {dim + 1} or {dim + 2}")
@@ -3875,8 +3878,17 @@ def pad_lo_hi_dilation_seq():
     padding, a = process_padding_str(int_to_seq(padding), stride, dilation, a)
     # }
 
-    res = clang.convolution(
-        a, weight, bias, stride, padding, dilation, False, (0,) * dim, groups  # transposed  # output_padding
+    res = conv_function(
+        a,
+        weight,
+        bias,
+        stride,
+        padding,
+        dilation,
+        False,  # transposed
+        (0,) * dim,  # output_padding
+        groups,
+        **extra_kwargs,
     )
     return res