Transforms: stop fork splitting from stranding a permute

[rascani]

Splitting a permute over a fork is meant to be free: each branch takes its own
copy on the way down, and the copies merge again on the way up. That holds when
every branch behaves the same way -- branches that rejoin have their copies
fused at the meeting node, branches that stay separate have theirs hoisted back
to the shared producer.

It does not hold for a fork where some branches rejoin and others do not. The
rejoining side ends with a copy below the meeting node, the diverging side with
one at the source, and neither mechanism can bring those together: upward
propagation has no fork split of its own, so it cannot hoist a copy above a
rejoin. Five ops are enough to see it, with no views and no quantization:

p = x.permute(0, 2, 3, 1)
return torch.abs(p) + torch.neg(p), torch.sigmoid(p)

before: 1   after Down: 2   after Up: 2

Decline the split in exactly that shape. Uniform forks are unaffected, which is
what keeps the existing behaviour: all-rejoin still splits and fuses at the
meeting node, all-diverge still splits and merges at the producer. A sweep of 75
fan-out shapes now regresses none, and the wins on real graphs are unchanged --
MultiheadAttention 10 to 9, TransformerEncoderLayer 14 to 11.

Also stop is_swappable raising on a non-keepdim reduction. The shared pass
answers the question it was asked and declines; Arm overrides it to keep raising,
since its pipeline normalizes reductions earlier and a violation there means the
ordering is wrong rather than that the graph is unusual.

No behaviour change for Arm: the pass suite produces a failure set identical to
merge base, name for name.

Authored with Claude Code.

[rascani]

Stack from ghstack (oldest at bottom):

• Transforms: let a backend declare what counts as a permute #22125
• -> Transforms: stop fork splitting from stranding a permute #22124
• Transforms: share the permute/view propagation driver #22123
• Transforms: share the horizontal fusion passes #22122
• Transforms: share the view/permute canonicalizer #22121
• Arm backend: give the propagation pass backend extension points #22120
• Transforms: share the view/permute algebra with all backends #22119

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22124

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 8 New Failures, 1 Unrelated Failure

As of commit be06ff4 with merge base e831b3e ():

NEW FAILURES - The following jobs have failed:

• pull / test-arm-backend-no-driver (test_pytest_ops_tosa) / linux-job (gh)
  RuntimeError: Command docker exec -t ac8a28e9e7c5efc4ff01d0171989152f54985b0d83b5c4bf85568d7f0e588b85 /exec failed with exit code 1
• pull / unittest-buck / linux / linux-job (gh)
  RuntimeError: Command docker exec -t dcc42e94ece59cfe5463d5e9581d7c1c5300af98d985a3552f1d7f821a9a7dbb /exec failed with exit code 3
• pull / unittest-buck / macos / macos-job (gh)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 3
• trunk / test-arm-backend-ethos-u (test_model_smollm2_135M_ethos_u85) / linux-job (gh)
  RuntimeError: Command docker exec -t 6c0f3e153331946d1edddf95dc059a15df4b5bd19320ee2f206ab2bd965ccbbe /exec failed with exit code 1
• trunk / test-arm-backend-ethos-u (test_pytest_ops_ethos_u55) / linux-job (gh)
  RuntimeError: Command docker exec -t 4ad2b90535175f69c2a246abc6b1d2461090e2b193b492fc8166bda0b18187dd /exec failed with exit code 1
• trunk / test-arm-backend-ethos-u (test_pytest_ops_ethos_u85) / linux-job (gh)
  RuntimeError: Command docker exec -t 22f1e7f78e9d62731ffe74b8f23821e09d4d3ec5efe14160215ac55826f4f24e /exec failed with exit code 1
• trunk / test-arm-backend-vkml (test_pytest_models_vkml) / linux-job (gh)
  RuntimeError: Command docker exec -t 612e6a7e3d9f6e563d991ba4435bd442fc78c5f5aa1f416fe1808188cb442ec6 /exec failed with exit code 1
• trunk / test-arm-backend-vkml (test_pytest_ops_vkml) / linux-job (gh)
  RuntimeError: Command docker exec -t cce0f233401b2aca30db8ba8ed6e1ebc8b85ccb35b27766f1b8f6c408e20e826 /exec failed with exit code 1

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / test-arm-backend-no-driver (test_pytest_models_tosa) / linux-job (gh) (matched linux rule in flaky-rules.json)
  The runner has received a shutdown signal. This can happen when the runner service is stopped, or a manually started runner is canceled.

This comment was automatically generated by Dr. CI and updates every 15 minutes.