Repository navigation
Release intermediate export artifacts once they are consumed (#21748) - #21748
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21748
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 0e0a5e3 with merge base 8426a13 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D115607218. |
meta-codesync
Bot
changed the title
JakeStevens
added a commit
to JakeStevens/executorch
that referenced
this pull request
Sep 17, 2026
…#21748) Summary: Add an opt-in `release_intermediate_artifacts` recipe setting that drops each stage’s output after the next stage consumes it, reducing peak memory usage for large models. Released artifacts produce descriptive accessor errors, and stages remain reusable across subsequent exports. Recipe combination and positional compatibility are preserved. Authored with Codex. Differential Revision: D115607218
JakeStevens
force-pushed
the
export-D115607218
branch
from
September 17, 2026 18:43
76982b0 to
29ea3a0
Compare
This PR needs a
|
JakeStevens
added a commit
to JakeStevens/executorch
that referenced
this pull request
Sep 17, 2026
…#21748) Summary: Add an opt-in `release_intermediate_artifacts` recipe setting that drops each stage’s output after the next stage consumes it, reducing peak memory usage for large models. Released artifacts produce descriptive accessor errors, and stages remain reusable across subsequent exports. Recipe combination and positional compatibility are preserved. Authored with Codex. Differential Revision: D115607218
JakeStevens
force-pushed
the
export-D115607218
branch
from
September 17, 2026 18:53
29ea3a0 to
1bff8c2
Compare
rascani
approved these changes
Sep 17, 2026
JakeStevens
added a commit
to JakeStevens/executorch
that referenced
this pull request
Sep 18, 2026
…#21748) Summary: Add an opt-in `release_intermediate_artifacts` recipe setting that drops each stage’s output after the next stage consumes it, reducing peak memory usage for large models. Released artifacts produce descriptive accessor errors, and stages remain reusable across subsequent exports. Recipe combination and positional compatibility are preserved. Authored with Codex. Differential Revision: D115607218
JakeStevens
force-pushed
the
export-D115607218
branch
from
September 18, 2026 20:32
1bff8c2 to
d82206d
Compare
JakeStevens
added a commit
to JakeStevens/executorch
that referenced
this pull request
Sep 18, 2026
…#21748) Summary: Add an opt-in `release_intermediate_artifacts` recipe setting that drops each stage’s output after the next stage consumes it, reducing peak memory usage for large models. Released artifacts produce descriptive accessor errors, and stages remain reusable across subsequent exports. Recipe combination and positional compatibility are preserved. Authored with Codex. Reviewed By: psiddh Differential Revision: D115607218
JakeStevens
force-pushed
the
export-D115607218
branch
from
September 18, 2026 21:08
d82206d to
372839e
Compare
…#21748) Summary: Pull Request resolved: pytorch#21748 Add an opt-in `release_intermediate_artifacts` recipe setting that drops each stage’s output after the next stage consumes it, reducing peak memory usage for large models. Released artifacts produce descriptive accessor errors, and stages remain reusable across subsequent exports. Recipe combination and positional compatibility are preserved. Authored with Codex. Reviewed By: psiddh Differential Revision: D115607218
JakeStevens
force-pushed
the
export-D115607218
branch
from
September 18, 2026 21:11
372839e to
0e0a5e3
Compare
JakeStevens
pushed a commit
that referenced
this pull request
Oct 14, 2026
…, and Samsung conv1d (#22373) ### Summary Error-guard sites that construct a failure which cannot actually fire. This PR inserts the missing `raise` (or replaces an always-true `assert` with the file's existing `check_or_raise`) so the documented failure mode is the one callers see. Updated for `main` at 5e21c13: the two `ExportSession` guards were fixed upstream in the meantime by #21748, so the only code change left in `export/export.py` is the message typo. The regression tests for those two guards are kept, since nothing else covers them. | File | What happens on `main` today | Reachability | | --- | --- | --- | | `export/export.py` `ExportSession.print_delegation_info` | Both guards already `raise` (#21748). Remaining here: the message says "atleast", and no test pins either guard. | **Reachable.** `ExportSession` is in `executorch.export.__all__` and is `@experimental`. A pipeline with no lowering stage, or one that lists `TO_EDGE_TRANSFORM_AND_LOWER` / `TO_BACKEND` without having run it, reaches each guard. This PR fixes the wording and adds `export/tests/test_print_delegation_info.py` for both. | | `backends/xnnpack/operators/node_visitor.py` per-channel axis `else` | `assert f"Unsupported weight per channel quantization axis ..."`: the f-string is always truthy, so the assertion never fires. An `assert` would also disappear under `python -O`. | **Not reachable on a normal export.** The only call that passes `swap_in_out_for_weights=True` is `backends/xnnpack/operators/op_conv2d.py:102`, and that visitor already `check_or_raise`s the axis to `0` (depthwise) or `1` (transpose) before `define_tensor`. This change restores a guard that can actually raise if that `else` is ever entered; it does not fix a mis-serialized `.pte` on the current conv2d path. | | `backends/samsung/_passes/conv1d_to_conv2d.py` `update_kernel` `else` | `RuntimeError("Weight of 1d conv should be constant tensor or Parameter obj")` is built and discarded; `weight_node.meta["val"]` is still unsqueezed. | **Not reachable on a normal Parameter / lifted-constant / buffer conv1d export.** Those cases are handled by the preceding branches (the buffer branch was added on `main` by #22725). The `else` is the leftover for a weight that is none of those. This change only makes that documented `RuntimeError` actually raise if the branch is entered. | ### Test plan Tests cover only the two reachable `ExportSession` guards. There is intentionally no test for the XNNPACK `else` and none for the Samsung `else`: both are unreachable on a normal export, and a test that injects that state would exercise a situation the production callers do not produce. `export/tests` is already in `pytest.ini` `testpaths`: ``` pytest export/tests/test_print_delegation_info.py ``` Measured on 89cf698 (merge of `main` 5e21c13), CPython 3.12, torch 2.15.0.dev20261004+cpu: - `export/tests/test_print_delegation_info.py`: 2 passed. - Negative control, after `git checkout origin/main -- export/export.py`, same run: `AssertionError: 'at least one of the lowering stages' not found in 'No delegation info available, atleast one of the lowering stages should be present'`, 1 failed / 1 passed. - `pytest export/tests --ignore=export/tests/test_target_recipes.py`: 8 failed / 172 passed on this branch, and the same 8 failed / 170 passed on `main` alone. The 8 are a local-environment gap (`exir/_serialize/program.fbs` is copied in by the build, which this checkout has not run), not a regression. `test_target_recipes.py` does not collect locally (no `coremltools`). ### CI Merged `main` into the branch. The conflict in `export/export.py` was resolved in favour of `main`'s `_raise_if_released` call, keeping the typo fix. The six reds triaged earlier on 12c3858 came from the merge base and their upstream fixes (#22802, #22754) are now in this branch. Two checks are still red on 89cf698, neither of them from this change: - `pull / android / run-emulator`: the emulator boots (`sys.boot_completed` returns 1), and `reactivecircus/android-emulator-runner` then fails on its own post-boot `adb shell input keyevent 82` with `cmd: Failure calling service input: Broken pipe (32)`, exit code 224. That step runs before the job script, and the log carries no output from `scripts/run_android_emulator.sh`. The same job also failed on `main` at efc7f2e when checked on 2026-10-06, and passed on the seven `main` runs before it. - `MLTEC_elastic - GitHubPOC` (NXP Bamboo): outside this repository's workflows, and red on other PRs too, not just this one: of eight open PRs untouched since 2026-10-05, two carry this status and both are red. Here the status went `pending` at 2026-10-05T07:01:25Z and `failure` one second later, which reads as a trigger-side failure rather than anything built from this branch, but the build log is on an internal host so I cannot confirm that from here. This change touches three Python files and adds one test. cc @nil-is-all @digantdesai Co-authored-by: Axel.Cffrd.Dnty <AxelNoun@users.noreply.github.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Add an opt-in
release_intermediate_artifactsrecipe setting that drops each stage’s output after the next stage consumes it, reducing peak memory usage for large models. Released artifacts produce descriptive accessor errors, and stages remain reusable across subsequent exports. Recipe combination and positional compatibility are preserved.Authored with Codex.
Reviewed By: psiddh
Differential Revision: D115607218