CoreML gives wrong outputs or crashes for torch.diagonal

[GregoryComer]

🐛 Describe the bug

When running torch.diagonal on CoreML, it sometimes gives incorrect outputs and sometimes crashes the process. Looks like there's some sort of memory corruption happening?

Repro:

import torch

from executorch.backends.apple.coreml.partition import CoreMLPartitioner
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):
        return torch.diagonal(x)

model = Model()
inputs = (
    torch.arange(25).reshape((5, 5)).to(torch.float),
)

print(inputs)

eager_outputs = model(*inputs)
print(f"Eager: {eager_outputs.shape} {eager_outputs}")

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

print(f"EP: {ep.module()(*inputs)}")

lowered = to_edge_transform_and_lower(
    ep,
    partitioner=[CoreMLPartitioner()],
    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

Example outputs (note the last line showing the error between reference and CoreML on ET outputs):

WARNING:root:Op aten.diagonal.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.diagonal.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
(tensor([[ 0.,  1.,  2.,  3.,  4.],
        [ 5.,  6.,  7.,  8.,  9.],
        [10., 11., 12., 13., 14.],
        [15., 16., 17., 18., 19.],
        [20., 21., 22., 23., 24.]]),)
Eager: torch.Size([5]) tensor([ 0.,  6., 12., 18., 24.])
ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, x: "f32[5, 5]"):
             # File: /var/folders/90/5w9gk0fn4n3g7fw1bvq8r1_m0000gn/T/ipykernel_68275/3260929381.py:12 in forward, code: return torch.diagonal(x)
            diagonal: "f32[5]" = torch.ops.aten.diagonal.default(x);  x = None
            return (diagonal,)
            
Graph signature: 
    # inputs
    x: USER_INPUT
    
    # outputs
    diagonal: USER_OUTPUT
    
Range constraints: {}

EP: tensor([ 0.,  6., 12., 18., 24.])
Converting PyTorch Frontend ==> MIL Ops:   0%|                                                                                                                                                                               | 0[/1](http://localhost:8888/1) [00:00<?, ? ops[/s](http://localhost:8888/s)]
Running MIL frontend_pytorch pipeline: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5[/5](http://localhost:8888/5) [00:00<00:00, 18944.46 passes[/s](http://localhost:8888/s)]
Running MIL default pipeline: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 89[/89](http://localhost:8888/89) [00:00<00:00, 9404.27 passes[/s](http://localhost:8888/s)]
Running MIL backend_mlprogram pipeline: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 12[/12](http://localhost:8888/12) [00:00<00:00, 13245.17 passes[/s](http://localhost:8888/s)]
[program.cpp:135] InternalConsistency verification requested but not available
[ETCoreMLModelManager.mm:528] Cache Miss: Model with identifier=executorch_51e9ef29-cbbc-4d0f-923d-49194fa7535b_all was not found in the models cache.
ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, x: "f32[5, 5]"):
            # No stacktrace found for following nodes
            lowered_module_0 = self.lowered_module_0
            executorch_call_delegate = torch.ops.higher_order.executorch_call_delegate(lowered_module_0, x);  lowered_module_0 = x = None
            getitem: "f32[5]" = executorch_call_delegate[0];  executorch_call_delegate = None
            return (getitem,)
            
Graph signature: 
    # inputs
    x: USER_INPUT
    
    # outputs
    getitem: USER_OUTPUT
    
Range constraints: {}

tensor([0., 0., 0., 0., 0.]) <-- ET outputs
tensor([  0.,  -6., -12., -18., -24.]) <-- error

Versions

coremltools version 8.3
executorch commit 67b6009 (Jun 14)

cc @kimishpatel @YifanShenSZ @cymbalrush @metascroy

[metascroy]

Re-filed in coremltools: apple/coremltools#2565

[metascroy]

CoreML temp fix here: #13023