CoreML randint occasionally generates out of distribution values

[GregoryComer]

🐛 Describe the bug

CoreML seems to occasionally generate out of distribution values (1 above the max) for torch.randint. It's non-deterministic, but I've included a repro script below. Generating a large enough set of values (1 million, for example) will repro it regularly. Maybe some floating point rounding issue?

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.randint(0, 100, (1000, 874), dtype=torch.int64) + x

model = Model()
inputs = (
    torch.zeros(1000, 874, dtype=torch.int64),
)

eager_outputs = model(*inputs)

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

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]

flat = et_outputs.flatten()
argmax = torch.argmax(flat)
print(f"Argmax: {argmax}")
print(f"Val: {flat[argmax]}")
print(f"{et_outputs[1]}") # In my specific case, the bad value was here. See outputs below.

Output:

Argmax: 874
Val: 100
tensor([100,  39,   6,  20,  63,  69,  10,  22,  55,  39,  86,  26,  29,  40,
         40,  63,  24,  17,  31,  26,  33,  22,  36,  16,  47,  30,  86,  59,
         19,   0,  52,  64,  23,  59,  29,  49,  32,  35,  29,  89,   3,  64,
         91,  93,  86,  95,  69,  19,   7,  85,  60,  13,  57,  25,   9,  14,
         94,  86,  77,  57,   8,  52,  80,  80,  20,  42,  68,  94,   1,  87,
         ...

Versions

coremltools version 8.3
executorch commit 67b6009 (Jun 14)

cc @kimishpatel @YifanShenSZ @cymbalrush @metascroy