Run the output shape matcher for delegates

[pssrawat]

Summary:
The Inspector's shape/dtype output matcher was gated behind num_outputs == 1.
Dropping that one term is the whole functional change:

  if (
-     num_outputs == 1
-     and len(runtime_intermediate_output) > 1
+     len(runtime_intermediate_output) > 1
      and isinstance(aot_intermediate_output, torch.Tensor)
  ):

The matcher was written for multi-output aten ops (native_layer_norm.out,
native_dropout.out), where the runtime logs several tensors but AOT captured
only the primary one, and the gate confined it to exactly that. Delegates
returning more than one tensor have the same problem for a different reason and
were left on positional pairing. However, positional pairing is unsound here.

Take a fused FFN block -- five nodes in one partition, the last two escaping it:

handle  node          shape
10      linear_gate   [1, 4, 1408]
11      linear_up     [1, 4, 1408]
12      silu          [1, 4, 1408]
13      mul           [1, 4, 1408]   escapes
14      linear_down   [1, 4,  512]   escapes

The delegate carries the whole tuple (10, 11, 12, 13, 14) as its debug handle
and returns two tensors, so num_outputs is 2 and the walk is:

i=0   negative_index = -1
      AOT       handle[-1]  = 14, linear_down   [1, 4,  512]
      runtime   outputs[-1] =     mul           [1, 4, 1408]   mispaired
i=1   negative_index = -2
      AOT       handle[-2]  = 13, mul           [1, 4, 1408]
      runtime   outputs[-2] =     linear_down   [1, 4,  512]   mispaired

Both AOT picks are the right nodes. Only the runtime side is crossed, because
the two sides index different lists:

• AOT side: _combine_aot_overlapped_intermediate_outputs indexes the
  delegate's debug-handle tuple --
  last_int = runtime_debug_handle[negative_index].
• runtime side: _process_single_runtime_output indexes the delegate's
  output list -- negative_index = -1 * (output_index + 1).

The handle tuple is ordered by the backend's node serialization; the output list
by the partition's output spec. Nothing ties the two orders together, and nothing
in the debug handles records which fused node produced which output. With a
single output they trivially agree, which is why the gate hid this. With several
they usually do not, and the result is either

• a shape mismatch, raising ValueError: Error computing SNR difference between tensors: The size of tensor a (N) must match the size of tensor b (M) at non-singleton dimension K. NumericalComparatorBase.compare builds
  its rows in a plain loop, so one bad pair aborts the entire DataFrame rather
  than the offending row.
• a coincidental shape match, and a silently wrong SNR.

Letting the matcher run replaces the runtime half of each pair by content:
exactly one runtime output is [1, 4, 512], so the AOT linear_down row takes
that one instead of outputs[-1], and mul gets the other. Shape settles this
example; where two outputs share a shape the matcher narrows by dtype, which is
what separates the int8 and float32 tensors a quantized partition returns.

Differential Revision: D117735521

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22236

• 📄 Preview Python docs built from this PR

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

❌ 3 New Failures

As of commit 39ab85a with merge base 41db862 ():

NEW FAILURES - The following jobs have failed:

• Build Presets / windows (windows) / build (gh)
• Cadence Build & Test / hifi-build / hifi4 (gh)
  ##[error]Refusing to check out fork pull request code from a 'pull_request_target' workflow. This workflow runs with the base repository's GITHUB_TOKEN, secrets, default-branch cache scope, and runner access. Fetching and executing a fork's code in that trusted context commonly leads to "pwn request" vulnerabilities. To opt in, review the risks at https://gh.io/securely-using-pull_request_target and set 'allow-unsafe-pr-checkout: true' on the actions/checkout step.
• Cadence Build & Test / vision-build / vision (gh)
  ##[error]Refusing to check out fork pull request code from a 'pull_request_target' workflow. This workflow runs with the base repository's GITHUB_TOKEN, secrets, default-branch cache scope, and runner access. Fetching and executing a fork's code in that trusted context commonly leads to "pwn request" vulnerabilities. To opt in, review the risks at https://gh.io/securely-using-pull_request_target and set 'allow-unsafe-pr-checkout: true' on the actions/checkout step.

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

[meta-codesync]

@pssrawat has exported this pull request. If you are a Meta employee, you can view the originating Diff in D117735521.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.