[wip] Support multimethod in export_llama_lib

[lucylq]

Stack from ghstack (oldest at bottom):

• -> [wip] Support multimethod in export_llama_lib #17231
• [wip] Introduce MultimethodLoraConfig #17230
• Introduce LoraConfig #17229
• Support multimethod in runner #17228

TODO: add CI test.

Differential Revision: D92315602

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17231

• 📄 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, 4 Unrelated Failures

As of commit 1de0598 with merge base eac0673 ():

NEW FAILURES - The following jobs have failed:

• Lint / lintrunner / linux-job (gh)
  >>> Lint for examples/models/llama/export_llama_lib.py:
• pull / test-lora-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 822e81a1f1b90aca906de2ea1d0dee90bcc1e54ec283830ee50182e210731981 /exec failed with exit code 1
• trunk / test-static-llama-qnn-eval-linux (baseline, 62) / linux-job (gh)
  RuntimeError: Command docker exec -t 6756f4566a889322c2225f6f059f84227e5027922fdaa3b23c40351bc89a9a70 /exec failed with exit code 1

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

• Test CUDA Builds / test-model-cuda-e2e (openai, whisper-large-v3-turbo, quantized-int4-tile-packed) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)
• Test CUDA Builds / test-model-cuda-e2e (openai, whisper-small, quantized-int4-weight-only) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

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

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

• pull / test-samsung-models-linux / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / test-samsung-quantmodels-linux / linux-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.

[Copilot]

Pull request overview

This pull request adds multimethod export support to the Llama export library, enabling the export of multiple methods (base model and LoRA variants) into a single .pte file. This is part of a stack of PRs (#17228-#17231) that collectively add multimethod and LoRA support to the ExecuTorch runtime.

Changes:

• Added _export_llama_multimethod() function to handle exporting multiple methods to a single .pte file
• Added helper functions _get_xnnpack_partitioners() and _get_output_filename() to support multimethod export
• Added validation logic to ensure multimethod export only works with XNNPACK backend or portable ops
• Added configuration file for Qwen3 multimethod export with LoRA
• Added comprehensive test script for validating multimethod export functionality
• Updated build dependencies to include required tokenizer and weight conversion modules

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
examples/models/llama/export_llama_lib.py	Core implementation of multimethod export logic, including validation, helper functions, and the main export function
examples/models/qwen3/config/qwen3_multimethod.yaml	Configuration file demonstrating multimethod export with LoRA and base methods using environment variable interpolation
examples/models/llama/BUCK	Added convert_weights.py to export library sources for proper dependency resolution
examples/models/llama/runner/targets.bzl	Added regex_lookahead tokenizer dependency for enhanced tokenizer support
.ci/scripts/test_lora_multimethod.sh	Comprehensive test script validating both LoRA and base method execution with expected output verification

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Missing period between sentences in error message. Should be: "Multimethod export only supports XNNPACK backend or portable ops. Please disable other backends..."

Suggested change

	"Multimethod export only supports XNNPACK backend or portable ops"
	"Please disable other backends (coreml, vulkan, qnn, mps, openvino)."
	"Multimethod export only supports XNNPACK backend or portable ops."
	" Please disable other backends (coreml, vulkan, qnn, mps, openvino)."

[Copilot]

Return type annotation is too generic. The function returns either a list of partitioners or None, so the type should be Optional[List[Partitioner]] instead of Optional[List]. This would provide better type safety and match the usage pattern where partitioners are passed to to_edge_transform_and_lower.

[Copilot]

The generate_etrecord configuration option is not being passed to the to_edge_transform_and_lower call. This means that ETRecord generation will not work with multimethod export even if llm_config.debug.generate_etrecord is set to True. Add generate_etrecord=llm_config.debug.generate_etrecord, to the function call.

Suggested change

	constant_methods=first_builder.metadata,
	constant_methods=first_builder.metadata,
	generate_etrecord=llm_config.debug.generate_etrecord,

[Copilot]

This script installs the huggingface_hub package directly from PyPI at runtime without any version pinning or integrity verification, which introduces a supply-chain risk: if the package (or the index it is fetched from) is compromised, arbitrary code could execute in the CI environment with access to repository data and any configured secrets. To reduce this risk, pin huggingface_hub to a known-good version (e.g., via a requirements/constraints file) and, where possible, enable hash-based verification or use a vetted internal package mirror instead of an unpinned direct install.