Repository navigation
[executorch][cuda] Write a weight storage the methods share only once - #23950
Merged
Merged
Conversation
Pull Request resolved: #23738 A multi-method CUDA export compiles every method against one `CudaWeightCollector`, so a storage the methods share (all of the model's weights, for a batching export) is stored under one `cuda_fqn_weight:<device>:<FQN>` key. But every method's `materialize` still copied each storage from the GPU to the host in 8 MB chunks, SHA-256'd it on one core, and wrote it to a temp file; only then did the `NamedDataStore` drop the duplicate. For an 18 GB model that is ~90 s per method (measured on 2 GiB: copy 1.9 s, SHA-256 5.9 s, write 1.3 s), all of it thrown away after the first method. The collector now remembers a fingerprint of each CUDA storage it has written, by storage key. A later method's storage under the same key is fingerprinted on the GPU first; if it matches, its write is skipped and the method's manifest points at the stored entry. Otherwise it is written and added exactly as before, so a changed storage still fails with the store's "Duplicate key ... with different data". The fingerprint is the storage's size, two independent random linear hashes of its 8-byte words (int64, chunk order kept significant) and its trailing bytes: ~0.8 s for 18 GB on an A100, and a single flipped bit changes it. Library-local constants (`_tensor_constant*`, keyed per shared library), off-graph KV entries and CPU storages keep the existing path. Full Muse Glimmer 30B batching export (`export_solo_batching.py`, 10 methods; A100; empty Triton and Inductor caches), stacked on D124496888: | | before | this diff | |---|---|---| | total | 135.1 min | 125.7 min | | main thread in `cuda_weight_collector` | ~21.4 min | ~7.7 min | | per method after the first (AOTI), min | 13.2 / 13.3 / 9.7 / 12.3 / 14.5 / 19.0 / 10.6 / 2.6 | 13.6 / 12.1 / 8.3 / 11.3 / 13.2 / 17.0 / 8.9 / 1.0 | (forward_2 .. forward_64, forward_others, sample. forward_1 writes everything once, as before.) Most of what is left in later methods is the collector's `trim_host_memory` (`gc.collect` + `malloc_trim`). The weight file is byte-identical with and without this diff (`aoti_cuda_blob.ptd`, 17,856,903,936 bytes, same SHA-256), and so is every method's weight manifest; only the methods' compiled libraries differ, as between any two cold exports. The export generates the same tokens (greedy, a 512-token prompt and 4 x 2k prompts) at the same speed (62.5 vs 62.7 tok/s; 110.0 vs 111.3). ghstack-source-id: 451021040 @exported-using-ghexport Differential Revision: [D124713681](https://our.internmc.facebook.com/intern/diff/D124713681/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23950
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Gasoonjia
added this pull request to stack #23954
October 18, 2026 07:47
meta-cla
Bot
added
the
CLA Signed
Gasoonjia
approved these changes
Oct 18, 2026
This PR needs a
|
This branch was successfully deployed
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
2 participants
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.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #23738 by @Gasoonjia
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/Gasoonjia/54/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/Gasoonjia/54/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/Gasoonjia/53/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/Gasoonjia/54/orig
Differential Revision: D124713681
@diff-train-skip-merge