Cache backend FakeTensor dispatches during ExportPass replay (#21700) - #21700
Open
apullin wants to merge 2 commits into
Open
Cache backend FakeTensor dispatches during ExportPass replay (#21700)#21700apullin wants to merge 2 commits into
apullin wants to merge 2 commits into
Conversation
apullin
requested review from
JacobSzwejbka,
digantdesai and
larryliu0820
as code owners
August 9, 2026 21:49
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21700
Note: Links to docs will display an error until the docs builds have been completed. ❌ 7 New FailuresAs of commit 173c7bd with merge base fb5eedc ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@apullin has exported this pull request. If you are a Meta employee, you can view the originating Diff in D115374097. |
This PR needs a
|
meta-codesync
Bot
changed the title
apullin
added a commit
to apullin/executorch
that referenced
this pull request
Aug 9, 2026
…#21700) Summary: Split from D97528110 so the FakeTensor cache extension can be reviewed independently from cold-node fast-copy. `FakeTensorMode` normally caches only `aten`, `prim`, and `prims` operations. ExportPass replay repeatedly dispatches deterministic backend operators from `quantized_decomposed`, `tosa`, and `cortex_m`, but PyTorch stores `non-builtin` negative cache entries for them. This change temporarily extends `torch._library.utils.is_builtin` while ExportPass runs, evicts only matching `non-builtin` bypass entries from the class-global and active ShapeEnv caches, and restores the exact original predicate in `finally`. Positive entries remain available across successive passes, which is the source of the speedup. This intentionally remains a narrow process-global monkeypatch. Current `FakeTensorMode` has no per-instance cache-policy hook, and its non-symbolic cache is class-global. The reentrant lock serializes participating ExportPass contexts, but unrelated threads can observe the extended predicate while the context is active. Fully isolated behavior requires an upstream policy hook plus cache-policy identity or per-mode cache storage; keeping this in a separate diff makes that tradeoff explicit. A/B benchmark against the fast-copy-only parent on the same host, each side run twice: - CombinedControl U55 lowering: 124.343s / 126.491s with this change vs 148.465s / 149.726s in the parent. Warm speedup: 15.5%; two-run mean speedup: 15.9%. - Synthetic U55 suite: 80.915s / 80.905s vs 81.238s / 82.371s. The 1.1% mean difference is within run-to-run noise; CombinedControl is the representative workload. Differential Revision: D115374097
apullin
force-pushed
the
export-D115374097
branch
from
August 9, 2026 23:53
4e3427a to
639de3e
Compare
Summary: After ARM pass-skipping and targeted-op ownership landed separately in D106781989, this diff optimizes the remaining `ExportPass` replay cost for passes that declare `target_ops` or `targeted_ops`. Cold operator nodes are copied with `graph.node_copy` instead of being re-dispatched through FakeTensor. Old-to-new node remapping preserves dependencies and `get_attr` values. Fast-copy is disabled when `call()` is overridden, for exact convolution or linear targets, or after a hot node changes nested tensor metadata. The fast path preflights every input before mutating the new graph or module tree, so a remapping fallback cannot leave orphaned nodes, attributes, or remap entries. An explicitly empty `targeted_ops` remains authoritative instead of falling back to legacy `target_ops`. Nested ARM control-flow submodules use the established `ArmPass.should_run_pass()` contract. This diff does not duplicate pass auto-discovery or generic skipping logic owned by D106781989. Per review from `jrstevens`, the process-global FakeTensor cache extension is now isolated in child D115374097 so its monkeypatching design and incremental performance can be reviewed independently. A/B benchmark against the parent revision on the same host, each side run twice: - CombinedControl U55 lowering: 148.465s / 149.726s with fast-copy vs 165.886s / 171.181s before it. Warm speedup: 12.5%; two-run mean speedup: 11.5%. - Synthetic U55 suite: 81.238s / 82.371s vs 83.816s / 80.802s. The difference is within run-to-run noise; the large model is the representative workload. Differential Revision: D97528110
…#21700) Summary: Split from D97528110 so the FakeTensor cache extension can be reviewed independently from cold-node fast-copy. `FakeTensorMode` normally caches only `aten`, `prim`, and `prims` operations. ExportPass replay repeatedly dispatches deterministic backend operators from `quantized_decomposed`, `tosa`, and `cortex_m`, but PyTorch stores `non-builtin` negative cache entries for them. This change temporarily extends `torch._library.utils.is_builtin` while ExportPass runs, evicts only matching `non-builtin` bypass entries from the class-global and active ShapeEnv caches, and restores the exact original predicate in `finally`. Positive entries remain available across successive passes, which is the source of the speedup. This intentionally remains a narrow process-global monkeypatch. Current `FakeTensorMode` has no per-instance cache-policy hook, and its non-symbolic cache is class-global. The reentrant lock serializes participating ExportPass contexts, but unrelated threads can observe the extended predicate while the context is active. Fully isolated behavior requires an upstream policy hook plus cache-policy identity or per-mode cache storage; keeping this in a separate diff makes that tradeoff explicit. A/B benchmark against the fast-copy-only parent on the same host, each side run twice: - CombinedControl U55 lowering: 124.343s / 126.491s with this change vs 148.465s / 149.726s in the parent. Warm speedup: 15.5%; two-run mean speedup: 15.9%. - Synthetic U55 suite: 80.915s / 80.905s vs 81.238s / 82.371s. The 1.1% mean difference is within run-to-run noise; CombinedControl is the representative workload. Differential Revision: D115374097
apullin
force-pushed
the
export-D115374097
branch
from
August 10, 2026 00:28
639de3e to
173c7bd
Compare
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
1 participant
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:
Split from D97528110 so the FakeTensor cache extension can be reviewed independently from cold-node fast-copy.
FakeTensorModenormally caches onlyaten,prim, andprimsoperations. ExportPass replay repeatedly dispatches deterministic backend operators fromquantized_decomposed,tosa, andcortex_m, but PyTorch storesnon-builtinnegative cache entries for them. This change temporarily extendstorch._library.utils.is_builtinwhile ExportPass runs, evicts only matchingnon-builtinbypass entries from the class-global and active ShapeEnv caches, and restores the exact original predicate infinally. Positive entries remain available across successive passes, which is the source of the speedup.This intentionally remains a narrow process-global monkeypatch. Current
FakeTensorModehas no per-instance cache-policy hook, and its non-symbolic cache is class-global. The reentrant lock serializes participating ExportPass contexts, but unrelated threads can observe the extended predicate while the context is active. Fully isolated behavior requires an upstream policy hook plus cache-policy identity or per-mode cache storage; keeping this in a separate diff makes that tradeoff explicit.A/B benchmark against the fast-copy-only parent on the same host, each side run twice:
Differential Revision: D115374097