Cross-load packed weight cache reuse for XNNPACK (#19988)

[doggeral]

Summary:

Add cross-load reuse + multi-PTE safety to the file-backed packed weight cache (D106673663). The first PTE in a session calls save_packed_index() to append a trailer; subsequent process launches mmap the file and pre-populate name_to_packed_data_metadata_ so look_up() hits for every saved weight and xnn_create_runtime skips packing entirely.

Cache file format

[packed data regions]                          (written by reserve_space)
[index entries]                                (written by save_packed_index)
  each: name_len(4B) | name(N) | file_offset(8B) | data_size(8B)
[footer: 20 bytes]
  index_start(8B) | entry_count(4B) | magic "XPWC"(4B) | version(4B)

Lifecycle invariants

• cache_loaded_ gate: load_packed_cache() runs at most once per process per path. Subsequent PTE inits for the same path reopen the write fd without re-reading the trailer.
• from_load flag: persistent entries (loaded from trailer or promoted on save) skip delete_packed_data cleanup. This keeps the mmap region and metadata alive across PTE unload/reload, so the next init hits the cache instead of repacking. Without this, every PTE destroy/recreate cycle appended a fresh copy to the file (~450 MB per cycle).
• No-op save short-circuit: save_packed_index returns early when no new reserve_space happened since the last save, avoiding the mtime churn that previously made the cache file look modified on every model load.

Multi-PTE behavior

• Multiple PTEs (or methods that don't share weights) in the same model load share one cache file. Each PTE's reserve_space extends the file; finalize_for_runtime msyncs only newly added regions; save_packed_index writes one trailer covering all PTEs at the end of the load.
• Sibling PTEs that opt out of the mmap path (caller passes empty packed_cache_path) early-return from initialize_for_runtime and fall through to heap allocation, without touching the singleton's PLLM state.
• Cross-model coexistence relies on caller-side discipline: only models that opt in set a non-empty cache path. Setting different non-empty paths concurrently is not supported by this singleton design.

Caller change

XNNPACKBackend::init always calls set_packed_cache_path (with empty string for non-opted-in PTEs). This keeps the singleton path in sync with the current PTE instead of inheriting a sibling's path.

Test Plan

buck2 test fbcode//executorch/backends/xnnpack/test:test_xnn_weights_cache  # 5 pass
buck2 build fbsource//xplat/executorch/backends/xnnpack:xnnpack_backendApple
buck2 build fbsource//xplat/executorch/backends/xnnpack:xnnpack_backend
buck2 build fbcode//executorch/backends/xnnpack:xnnpack_backend

On device (iOS Stella build, PLLM + Llama3 runner):

• Cold start: load (1184 entries) from cache, reserve_mmap=0 for cached weights
• Cache file size stable at ~593 MB across PLLM unload/reload cycles
• app_peak ~700 MB (vs ~2.5 GB pre-fix)
• compressed ~100 MB (vs ~1.7 GB pre-fix)

Reviewed By: GregoryComer

Differential Revision: D106717093

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19988

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI jobs will have longer queue times due to CI migration

❌ 5 New Failures, 1 Pending

As of commit f6f3ee8 with merge base 185bd09 ():

NEW FAILURES - The following jobs have failed:

• Cadence Build & Test / hifi-build / hifi4 (gh)
  Input required and not supplied: aws-region
• Cadence Build & Test / vision-build / vision (gh)
  Input required and not supplied: aws-region
• Lint / lintrunner (gh)
  >>> Lint for backends/xnnpack/test/runtime/test_xnn_weights_cache.cpp:
• pull / test-qnn-models-linux (dl3) / linux-job (gh)
  RuntimeError: Command docker exec -t f19df4b77893adfee648094605af790a648ffd4a37502847c695465fea5c3dca /exec failed with exit code 92
• pull / unittest-editable / windows / windows-job (gh)
  examples/models/test/test_export.py::ExportTest::test_efficient_sam_export_to_executorch

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

[meta-codesync]

@doggeral has exported this pull request. If you are a Meta employee, you can view the originating Diff in D106717093.

[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.

[digantdesai]

Review automatically exported from Phabricator review in Meta.

comments have been addressed.

[digantdesai]

Review automatically exported from Phabricator review in Meta.

addressed comments already