Fix mixed-dtype autocast regressions

Author: JPPhoto

invokeai/backend/model_manager/load/model_cache/torch_module_autocast/custom_modules/custom_linear.py

@@ -58,34 +58,40 @@ def autocast_linear_forward_sidecar_patches(
 
     # Finally, apply any remaining patches.
     if len(unprocessed_patches_and_weights) > 0:
+        weight, bias = orig_module._cast_weight_bias_for_input(input)
         # Prepare the original parameters for the patch aggregation.
-        orig_params = {"weight": orig_module.weight, "bias": orig_module.bias}
+        orig_params = {"weight": weight, "bias": bias}
         # Filter out None values.
         orig_params = {k: v for k, v in orig_params.items() if v is not None}
 
         aggregated_param_residuals = orig_module._aggregate_patch_parameters(
             unprocessed_patches_and_weights, orig_params=orig_params, device=input.device
         )
-        output += torch.nn.functional.linear(
-            input, aggregated_param_residuals["weight"], aggregated_param_residuals.get("bias", None)
-        )
+        residual_weight = orig_module._cast_tensor_for_input(aggregated_param_residuals["weight"], input)
+        residual_bias = orig_module._cast_tensor_for_input(aggregated_param_residuals.get("bias", None), input)
+        assert residual_weight is not None
+        output += torch.nn.functional.linear(input, residual_weight, residual_bias)
 
     return output
 
 
 class CustomLinear(torch.nn.Linear, CustomModuleMixin):
-    def _cast_weight_bias_for_input(self, input: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor | None]:
-        weight = cast_to_device(self.weight, input.device)
-        bias = cast_to_device(self.bias, input.device)
+    def _cast_tensor_for_input(self, tensor: torch.Tensor | None, input: torch.Tensor) -> torch.Tensor | None:
+        tensor = cast_to_device(tensor, input.device)
         if (
-            input.is_floating_point()
-            and weight.is_floating_point()
-            and not isinstance(weight, GGMLTensor)
-            and weight.dtype != input.dtype
+            tensor is not None
+            and input.is_floating_point()
+            and tensor.is_floating_point()
+            and not isinstance(tensor, GGMLTensor)
+            and tensor.dtype != input.dtype
         ):
-            weight = weight.to(dtype=input.dtype)
-            if bias is not None and not isinstance(bias, GGMLTensor):
-                bias = bias.to(dtype=input.dtype)
+            tensor = tensor.to(dtype=input.dtype)
+        return tensor
+
+    def _cast_weight_bias_for_input(self, input: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor | None]:
+        weight = self._cast_tensor_for_input(self.weight, input)
+        bias = self._cast_tensor_for_input(self.bias, input)
+        assert weight is not None
         return weight, bias
 
     def _autocast_forward_with_patches(self, input: torch.Tensor) -> torch.Tensor:
@@ -100,10 +106,14 @@ def forward(self, input: torch.Tensor) -> torch.Tensor:
             return self._autocast_forward_with_patches(input)
         elif self._device_autocasting_enabled:
             return self._autocast_forward(input)
-        elif (
-            input.is_floating_point()
-            and self.weight.is_floating_point()
-            and self.weight.dtype != input.dtype
+        elif input.is_floating_point() and (
+            (self.weight.is_floating_point() and self.weight.dtype != input.dtype)
+            or (
+                self.bias is not None
+                and self.bias.is_floating_point()
+                and not isinstance(self.bias, GGMLTensor)
+                and self.bias.dtype != input.dtype
+            )
         ):
             weight, bias = self._cast_weight_bias_for_input(input)
             return torch.nn.functional.linear(input, weight, bias)

invokeai/backend/model_manager/load/model_cache/torch_module_autocast/custom_modules/custom_module_mixin.py

@@ -49,7 +49,9 @@ def _aggregate_patch_parameters(
         # parameters. But, of course, any sub-layers that need to access the actual values of the parameters will fail.
         for param_name in orig_params.keys():
             param = orig_params[param_name]
-            if type(param) is torch.nn.Parameter and type(param.data) is torch.Tensor:
+            if isinstance(param, torch.nn.Parameter) and type(param.data) is torch.Tensor:
+                pass
+            elif type(param) is torch.Tensor:
                 pass
             elif type(param) is GGMLTensor:
                 # Move to device and dequantize here. Doing it in the patch layer can result in redundant casts /

tests/backend/model_manager/load/model_cache/torch_module_autocast/custom_modules/test_all_custom_modules.py

@@ -1,4 +1,5 @@
 import copy
+from collections.abc import Callable
 
 import gguf
 import pytest
@@ -124,6 +125,67 @@ def unwrap_single_custom_layer(layer: torch.nn.Module):
     return unwrap_custom_layer(layer, orig_layer_type)
 
 
+class ZeroParamPatch(BaseLayerPatch):
+    """A minimal parameter patch that exercises the aggregated sidecar patch path."""
+
+    def get_parameters(self, orig_parameters: dict[str, torch.Tensor], weight: float) -> dict[str, torch.Tensor]:
+        return {name: torch.zeros_like(param) for name, param in orig_parameters.items()}
+
+    def to(self, device: torch.device | None = None, dtype: torch.dtype | None = None):
+        return self
+
+    def calc_size(self) -> int:
+        return 0
+
+
+def _cpu_dtype_supported(
+    layer_factory: Callable[[], torch.nn.Module],
+    input_factory: Callable[[torch.dtype], torch.Tensor],
+    dtype: torch.dtype,
+) -> bool:
+    try:
+        layer = layer_factory().to(dtype=dtype)
+        input_tensor = input_factory(dtype)
+        with torch.no_grad():
+            _ = layer(input_tensor)
+        return True
+    except (RuntimeError, TypeError, NotImplementedError):
+        return False
+
+
+def _cpu_dtype_param(
+    dtype: torch.dtype,
+    layer_factory: Callable[[], torch.nn.Module],
+    input_factory: Callable[[torch.dtype], torch.Tensor],
+):
+    supported = _cpu_dtype_supported(layer_factory, input_factory, dtype)
+    return pytest.param(
+        dtype,
+        id=str(dtype).removeprefix("torch."),
+        marks=pytest.mark.skipif(not supported, reason=f"CPU {dtype} is not supported for this op"),
+    )
+
+
+LINEAR_CPU_MIXED_DTYPE_PARAMS = [
+    _cpu_dtype_param(torch.bfloat16, lambda: torch.nn.Linear(8, 16), lambda dtype: torch.randn(2, 8, dtype=dtype)),
+    _cpu_dtype_param(torch.float16, lambda: torch.nn.Linear(8, 16), lambda dtype: torch.randn(2, 8, dtype=dtype)),
+]
+
+
+CONV2D_CPU_MIXED_DTYPE_PARAMS = [
+    _cpu_dtype_param(
+        torch.bfloat16,
+        lambda: torch.nn.Conv2d(8, 16, 3),
+        lambda dtype: torch.randn(2, 8, 5, 5, dtype=dtype),
+    ),
+    _cpu_dtype_param(
+        torch.float16,
+        lambda: torch.nn.Conv2d(8, 16, 3),
+        lambda dtype: torch.randn(2, 8, 5, 5, dtype=dtype),
+    ),
+]
+
+
 def test_isinstance(layer_under_test: LayerUnderTest):
     """Test that isinstance() and type() behave as expected after wrapping a layer in a custom layer."""
     orig_layer, _, _ = layer_under_test
@@ -550,3 +612,67 @@ def test_quantized_linear_sidecar_patches_with_autocast_from_cpu_to_device(
 
     # Assert that the outputs with and without autocasting are the same.
     assert torch.allclose(expected_output, autocast_output, atol=1e-6)
+
+
+@pytest.mark.parametrize("dtype", LINEAR_CPU_MIXED_DTYPE_PARAMS)
+@torch.no_grad()
+def test_linear_mixed_dtype_inference_without_patches(dtype: torch.dtype):
+    layer = wrap_single_custom_layer(torch.nn.Linear(8, 16))
+    input = torch.randn(2, 8, dtype=dtype)
+
+    output = layer(input)
+
+    assert output.dtype == input.dtype
+    assert output.shape == (2, 16)
+
+
+@pytest.mark.parametrize("dtype", LINEAR_CPU_MIXED_DTYPE_PARAMS)
+@torch.no_grad()
+def test_linear_mixed_dtype_inference_without_patches_bias_only_mismatch(dtype: torch.dtype):
+    layer = torch.nn.Linear(8, 16).to(dtype=dtype)
+    layer.bias = torch.nn.Parameter(layer.bias.detach().to(torch.float32))
+    layer = wrap_single_custom_layer(layer)
+    input = torch.randn(2, 8, dtype=dtype)
+
+    output = layer(input)
+
+    assert output.dtype == input.dtype
+    assert output.shape == (2, 16)
+
+
+@pytest.mark.parametrize("dtype", CONV2D_CPU_MIXED_DTYPE_PARAMS)
+@torch.no_grad()
+def test_conv2d_mixed_dtype_inference_without_patches(dtype: torch.dtype):
+    layer = wrap_single_custom_layer(torch.nn.Conv2d(8, 16, 3))
+    input = torch.randn(2, 8, 5, 5, dtype=dtype)
+
+    output = layer(input)
+
+    assert output.dtype == input.dtype
+    assert output.shape == (2, 16, 3, 3)
+
+
+@pytest.mark.parametrize("dtype", LINEAR_CPU_MIXED_DTYPE_PARAMS)
+@torch.no_grad()
+def test_linear_mixed_dtype_sidecar_parameter_patch(dtype: torch.dtype):
+    layer = wrap_single_custom_layer(torch.nn.Linear(8, 16))
+    layer.add_patch(ZeroParamPatch(), 1.0)
+    input = torch.randn(2, 8, dtype=dtype)
+
+    output = layer(input)
+
+    assert output.dtype == input.dtype
+    assert output.shape == (2, 16)
+
+
+@pytest.mark.parametrize("dtype", CONV2D_CPU_MIXED_DTYPE_PARAMS)
+@torch.no_grad()
+def test_conv2d_mixed_dtype_sidecar_parameter_patch(dtype: torch.dtype):
+    layer = wrap_single_custom_layer(torch.nn.Conv2d(8, 16, 3))
+    layer.add_patch(ZeroParamPatch(), 1.0)
+    input = torch.randn(2, 8, 5, 5, dtype=dtype)
+
+    output = layer(input)
+
+    assert output.dtype == input.dtype
+    assert output.shape == (2, 16, 3, 3)