[BUG] Metal corrupts mx.tril output when co-evaluated with a transposed cumsum graph #4419

☑️ I understand it is strictly prohibited to use AI to write issues.

Hi, during some work on GDN kernels it appears I have found some strange behavior.

Describe the bug
On Metal, joint lazy evaluation corrupts an independent mx.tril result.

Observe the output from the reproduction script:

CPU row sums: [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0]
CPU PASS
Metal row sums: [0.0, 0.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0]
AssertionError

To Reproduce

import mlx.core as mx


def run(device):
    mx.set_default_device(device)
    x = mx.array([1.0] * 72).reshape(2, 9, 4).transpose(0, 2, 1)
    x = mx.cumsum(x, axis=-1)

    causal = mx.tril(mx.ones((9, 9)))
    relative = x[..., :, None] - x[..., None, :]
    mx.eval(relative, causal)

    return causal.tolist()


def row_sums(x):
    return [sum(row) for row in x]


expected = [[float(column <= row) for column in range(9)] for row in range(9)]

# Run Metal first so the CPU control cannot affect the failing graph.
metal = run(mx.gpu)
cpu = run(mx.cpu)

print("CPU row sums:", row_sums(cpu))
assert cpu == expected
print("CPU PASS")

print("Metal row sums:", row_sums(metal))
assert metal == expected
print("Metal PASS")

Expected behavior
CPU and Metal should both return the lower-triangular matrix, with row sums:

[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0]

Desktop (please complete the following information):

• MLX: 0.32.2
• MLX Metal: 0.32.2
• Python: 3.13.12
• macOS: 26.5.2 (25F84)
• Hardware: Apple M2 Pro, 32 GB RAM, 16-core GPU
• Metal: Metal 4
• Architecture: applegpu_g14s
• Also reproduced from clean source commit: 99e45f7

Additional context
Materializing causal early with mx.eval(causal) prevents the corruption. I was able to reproduce the failure in five out of five fresh Metal processes.