Vulkan missing partition constraints for unsupported dtypes

[GregoryComer]

🐛 Describe the bug

When passing a model containing tensors with an unsupported type, the Vulkan lowering code currently errors out. It should ideally include a partitioner constraint to allow these ops to fall back to CPU.

Affected dtypes:

• float64
• int16
• uint16
• uint32
• uint64

Repro:

import torch

from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
from executorch.exir import to_edge_transform_and_lower, EdgeCompileConfig, to_edge
from executorch.extension.pybindings.portable_lib import _load_for_executorch_from_buffer

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        
    def forward(self, x, y):
        return x + y

model = Model()
inputs = (
    torch.randn(5, 5).to(torch.float64),
    torch.randn(5, 5).to(torch.float64),
)

eager_outputs = model(*inputs)

ep = torch.export.export(model.eval(), inputs)
print(ep)

lowered = to_edge_transform_and_lower(
    ep,
    partitioner=[VulkanPartitioner()],
    compile_config=EdgeCompileConfig(_check_ir_validity=False)
).to_executorch()

print(lowered.exported_program())

et_model = _load_for_executorch_from_buffer(lowered.buffer)
et_outputs = et_model([*inputs])[0]

print(et_outputs)

et_outputs - eager_outputs

Output:

File [~/src/executorch/src/executorch/backends/vulkan/serialization/vulkan_graph_builder.py:82](http://localhost:8888/~/src/executorch/src/executorch/backends/vulkan/serialization/vulkan_graph_builder.py#line=81), in VkGraphBuilder.get_vk_datatype(torch_dtype)
     80     return vk_graph_schema.VkDataType.INT32
     81 else:
---> 82     raise AssertionError(f"Invalid dtype for vulkan_preprocess ({torch_dtype})")

AssertionError: Invalid dtype for vulkan_preprocess (torch.float64)

Versions

executorch commit 67b6009 (Jun 14)

cc @SS-JIA @manuelcandales @cbilgin