fix: apply the five measured findings (#24-#28) - #32
Merged
Conversation
Closes #24, #26, #28; addresses #25 and #27 as usage-prompt caveats. #24 chrome_trace now tags torch.compile's own envelope spans [dispatch]. Verified against a fresh compiled-training capture: every one of them (Torch-Compiled Region, CompiledFunction/Backward, TorchDynamo Cache Lookup, AOTDispatcher wrappers, Inductor's ## Call CompiledFxGraph ##) arrives with cat 'cpu_op' — indistinguishable by category from a real aten op — so the existing cat rule could never see them and the match has to be by name. The autograd engine also wraps the compiled backward in an envelope of its own, so the compile-family token is matched anywhere in the name; a prefix-only rule left that 7.4ms span looking like work. Eager autograd envelopes stay untagged on purpose and are documented instead: tagging those is a separate decision about eager traces, not this finding. #26 coverage now accumulates with math.fsum. CPython 3.12 changed sum() to compensated summation, which made this the only value in the surface that differed between supported interpreters. fsum is exactly rounded, so a digest is now the same bytes everywhere — meaning a diff between two digests always says the report changed, never the runtime. Verified by hashing the same digest under 3.11/3.12/3.14: identical. #28 summarize_report reports distinct_names_total, plus distinct_names_omitted when top-N hides a name entirely. Top-N ranks units, and one name can own many units, so a variant that lost a 0.15% near-tie vanished from a summary that otherwise looked complete. #25 and #27 are usage-prompt caveats because the arithmetic was never wrong — the surface just let a false premise pass unmarked. Comparing across a fusion boundary pairs one kernel against part of its multi- kernel replacement (measured: +101% where the honest group comparison was -12.3%), and ninja's start_ms/end_ms are measured from each build's own origin, so cross-run deltas subtract unrelated clocks. New fixture is a real compiled-training trace, shrunk by dropping whole events rather than editing any. 213 tests. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The new test guarded against vacuity with `fsum(values) != sum(values)` — an assertion that is itself interpreter-dependent, which is the exact class of assumption issue #26 exists to remove. It held on 3.10/3.11 and failed on every 3.12+ leg of the matrix, because 3.12's compensated sum() returns the exact answer for these values. The guard is now an explicit left-to-right accumulation, which loses the small terms on every version. Measured while fixing: 3.12+ compensation is exact for every sequence tried, so no unit test can distinguish sum() from fsum() there — the docstring now says so rather than implying a coverage this test does not have. The cross-interpreter property itself is verified by hashing one digest under 3.11/3.12/3.14. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This was referenced Jul 19, 2026
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.
The four demanding-workload studies and the interpreter matrix each found something the tool got away with because nothing contradicted it. This applies all five.
Three are code (
[dispatch]tagging for torch.compile envelopes,math.fsumfor interpreter-stable digests, distinct-name counts insummarize_report); two are usage-prompt caveats, because in those cases the numbers were correct and only the premise was unmarked.Verified beyond the suite: the tagging rule was written from a fresh real capture (all eight envelope families arrive as
cpu_op, so a cat rule was never going to work), and the fsum fix was confirmed by hashing the same digest under 3.11/3.12/3.14 — now identical.213 tests, no version bump, no release.
🤖 Generated with Claude Code