Repository navigation
Slim the slow checks: CPU-torch CI installs, a one-file GPU suite, no provenance-only sidecar fields - #79
Merged
Conversation
… provenance-only sidecar fields - CI installs with uv and the CPU torch wheel: the default Linux torch pulled ~2.5 GB of CUDA libraries that pyright and the CPU suite never touch (291 s of the test job's 571 s, 97 s of typecheck's 169 s). - tests/gpu is test_smoke.py alone: one extract → steer → generate smoke and the four throughput gates. The session, jlens, units, curved-stack, numerics, SAELens and skip-helper modules duplicated what the tiny sessions run on CPU (MPS numerics keep their `mps` params in tests/model); the extraction time budget, hook-cleanup, save/load, monitor, chat-input and real-model discover smokes go with them. `gpu_gemma` and DROWSE_TEST_SAELENS are gone. - The fitted sidecar drops share_metric and subspace_metric (always "mahalanobis"; nothing read them), and its reader ignores keys the contract has retired instead of demanding an exact key set, so a 6.1.0 fit still loads and dropping provenance never strands one. - tests/invariants/test_user_message.py (a repo scan that every exception subclasses DrowseError and every RequestError is a 400) goes; the route-level request-error tests keep the statuses clients see. - "Every bug fix lands with a regression test" becomes "when a bug could quietly come back".
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
~2.5 GB of CUDA libraries that pyright and the CPU suite never touch
(291 s of the test job's 571 s, 97 s of typecheck's 169 s).
the four throughput gates. The session, jlens, units, curved-stack, numerics,
SAELens and skip-helper modules duplicated what the tiny sessions run on CPU
(MPS numerics keep their
mpsparams in tests/model); the extraction timebudget, hook-cleanup, save/load, monitor, chat-input and real-model discover
smokes go with them.
gpu_gemmaand DROWSE_TEST_SAELENS are gone."mahalanobis"; nothing read them), and its reader ignores keys the contract
has retired instead of demanding an exact key set, so a 6.1.0 fit still
loads and dropping provenance never strands one.
subclasses DrowseError and every RequestError is a 400) goes; the route-level
request-error tests keep the statuses clients see.
quietly come back".