Fix InsertIOQDQ KeyError for dequantize encodings (#18622)

[abhinaykukkadapu]

Summary:
q_dq_map only contained quantize ops as keys, so when a node with a dequantize encoding (e.g. a pre-quantized LLM parameter) feeds the output node, the lookup crashes with a KeyError.

Add dequantize ops as keys in q_dq_map, mapping them to the correct dequantize target for output boundary insertion (dq.default -> dq.tensor, matching the existing quantize convention).

Since dequantize targets are now keys, _create_node transfers QCOM_QUANT_ATTRS to inserted dequant nodes. To prevent the live iterator from revisiting these nodes, iterate over a snapshot via list(graph_module.graph.nodes).

Fixes #17732

Differential Revision: D98977887

Pulled By: abhinaykukkadapu

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18622

• 📄 Preview Python docs built from this PR

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

✅ You can merge normally! (3 Unrelated Failures)

As of commit 7efb4e5 with merge base 8b30cfe ():

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

• pull / test-openvino-linux / linux-job (gh) (detected as infra flaky with no log or failing log classifier)
• pull / unittest / windows / windows-job (gh) (matched win rule in flaky-rules.json)
  ##[error]The operation was canceled.

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / unittest-editable / windows / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.

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

[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.

[meta-codesync]

@abhinaykukkadapu has imported this pull request. If you are a Meta employee, you can view this in D98977887.

[abhinaykukkadapu]

Moving the discussion from other PR.

@haowhsu-quic

Hi @abhinaykukkadapu, the root cause of #17732 was having too many shards due to unsupported op like #16690. I believe the solution of #17732 has already been addressed in #17194.

Thanks for taking a look, right this task is not 1-1 mapping for the exact issue, but there is a comment on the task that refers to unavail dq in the map: #17732 (comment)

Also #17194 got reverted in #17385 this change with list() semantics should pass the tests.

[meta-codesync]

@abhinaykukkadapu has exported this pull request. If you are a Meta employee, you can view the originating Diff in D98977887.

[haowhsu-quic]

Feel like we could just add a check like n.target != exir_ops.edge.quantized_decomposed.dequantize_per_tensor.tensor for any dequantize nodes that already be added.
Then we don't need the list approach (probably will have smaller memory footprint). But I still wonder if this is really an issue? Looks like the root cause of #17782 is because they tweaked the codebase.

[abhinaykukkadapu]

The issue this diff is handling is not just with the snapshot of the nodes which is solved by the list. Here is what we want:

1. For input we want to pop the quant attrs, and insert q node
2. For output, we insert dq node but we don't want to pop the attrs

The target for example a conv node has q attrs, weight node if it feeds output will have dq attrs, we want to insert dq node for both without popping. I will put up a patch where we can reuse q_ops may be.

[abhinaykukkadapu]

@haowhsu-quic i'm referring to the comments on this task, which is irrelevant to the task they commented on but this is the root cause: #17732 (comment)

[abhinaykukkadapu]

This is to avoid pop on target node if we are inserting dq.