CoreML BatchNorm3d crashes process

[GregoryComer]

🐛 Describe the bug

When running BatchNorm3d ops on Core ML, it appears to crash the process pretty regularly.

Repro:

import torch

from executorch.backends.apple.coreml.partition import CoreMLPartitioner
from executorch.exir import to_edge_transform_and_lower, EdgeCompileConfig
from executorch.extension.pybindings.portable_lib import _load_for_executorch_from_buffer

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.norm = torch.nn.BatchNorm3d(3)
        
    def forward(self, x):
        return self.norm(x)

model = Model()
inputs = (
    torch.randn(1, 3, 4, 4, 4),
)

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

ep = torch.export.export(model.eval(), inputs)
lowered = to_edge_transform_and_lower(
    ep,
    partitioner=[CoreMLPartitioner()],
    compile_config=EdgeCompileConfig(_check_ir_validity=False)
).to_executorch()

print(ep)
print(lowered.exported_program())

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

et_outputs - eager_outputs

I captured the following output pre-crash (with slightly different shapes):

loc("tensor<fp16, [5, 10, 4, 4, 4]> aten__native_batch_norm_legit_no_training_default_cast_fp16 = batch_norm(beta = tensor<fp16, [10]>(BLOBFILE(path = string(\22/Users/gjcomer/Library/Caches/executorchcoreml/models/executorch_c8e2de39-f155-452a-af5b-13f794eb4932_all.mlmodelc/weights/weight.bin\22), offset = uint64(448))), epsilon = fp16(1.00135803e-05), gamma = tensor<fp16, [10]>(BLOBFILE(path = string(\22/Users/gjcomer/Library/Caches/executorchcoreml/models/executorch_c8e2de39-f155-452a-af5b-13f794eb4932_all.mlmodelc/weights/weight.bin\22), offset = uint64(320))), mean = tensor<fp16, [10]>(BLOBFILE(path = string(\22/Users/gjcomer/Library/Caches/executorchcoreml/models/executorch_c8e2de39-f155-452a-af5b-13f794eb4932_all.mlmodelc/weights/weight.bin\22), offset = uint64(64))), variance = tensor<fp16, [10]>(BLOBFILE(path = string(\22/Users/gjcomer/Library/Caches/executorchcoreml/models/executorch_c8e2de39-f155-452a-af5b-13f794eb4932_all.mlmodelc/weights/weight.bin\22), offset = uint64(192))), x = x_to_fp16)[milId = uint64(1), name = string(\22aten__native_batch_norm_legit_no_training_default_cast_fp16\22)]; - /Users/gjcomer/Library/Caches/executorchcoreml/models/executorch_c8e2de39-f155-452a-af5b-13f794eb4932_all.mlmodelc/model.mil":12:12): error: output type 'tensor<5x10x4x4x4xf16>' and mean type 'tensor<1x0x1x1x519870560xf16>' are not broadcast compatible
LLVM ERROR: Failed to infer result type(s).

Versions

coremltools version 8.3
executorch commit 67b6009 (Jun 14)

cc @kimishpatel @YifanShenSZ @cymbalrush @metascroy

[metascroy]

Refiled here: apple/coremltools#2572