Move lingvo patches into the manifest #409
Comments
6 tasks
yhtang
added a commit
that referenced
this issue
Dec 18, 2023
This introduces "the manifest" (`manifest.yaml`) that describes the complete state of the jax stack to allow for reproducible nightly builds and reproducible presubmit CI. The manifest (and patches) are staged in a "trial branch" each night, and if the build succeeds, we can merge the trial branch into main (TODO: GH Issue tracking automating merge). The presubmit CI runs on the PR's git-ref and so unless the author has committed a custom patch/manifest, they will always be running from the "last working state". # Description of `manifest.yaml` The manifest allows specifying/pinning libraries that serve different purposes: * git repos: libraries that need their source checked out at a particular SHA * pip VCS constraints: [VCS dependency specs](https://pip.pypa.io/en/stable/topics/vcs-support/) that need to be pinned like `fiddle @ git+https://github.com/google/fiddle` * pip constraints: Python packages we need to constrain or pin to address one-off fixes, e.g., addressing a CVE Here are example entries of each of these in the manifest.yaml # Git repos ```yaml jax: url: https://github.com/google/jax.git tracking_ref: main latest_verified_commit: b032a0271e3e2ea8d0df64d2f3f1a1e450a38dc9 # 2023-11-15 mode: git-clone ``` * url: This is the url to lookup the latest git ref * tracking_ref: This is the git-ref to lookup the latest commit SHA. Usually it’s main, but in some cases like lingvo, it’s master * latest_verified_commit: SHA * mode: * git-clone: Whether to clone the repo or not ## Git repos with patches ```yaml flax: url: https://github.com/google/flax.git mirror_url: https://github.com/nvjax-svc-0/flax.git extra_dir: null tracking_ref: main latest_verified_commit: a572f6af2fef565c0f9ba2fc12b781e9e3385140 mode: git-clone patches: pull/3340/head: file://patches/flax/PR-3340.patch # Add Sharding Annotations to Flax Modules ``` * mirror_url: a git url where we can pull patches that start with `^mirror/` * extra_dir: an optional local dir to pull patches not found in upstream or the mirror. Useful for private repos * patches: a dictionary where keys are git-refs and values are URIs for the local patch. The reason for the value is the git-ref can be rebased or be updated and the SHA will change; this makes patches a moving target when using git refs. So each `file://` URI will be a patch committed into Jax-Toolbox's VC to ensure reproducibility # VCS Constraint ```yaml clu: url: https://github.com/google/CommonLoopUtils.git tracking_ref: main latest_verified_commit: 89c2face3474a7482358068d7a00d9bb6e4b31fe mode: pip-constraint ``` These aren't cloned (in fact `get-source.sh` will error if you try to clone). These are used in `pip-finalize.sh` to pin VCS dependencies like ```sh clu @ git+https://github.com/google/CommonLoopUtils#egg=clu # To clu @ git+https://github.com/google/CommonLoopUtils@89c2face3474a7482358068d7a00d9bb6e4b31fe ``` # Pip constraint ```yaml pydantic: mode: pip-constraint version: X.Y.Z ``` * mode: pip-constraint * Adds a pip constraint during pip-compile * version: If ${package}.version is present, then we can treat this as a pip constraint. This allows us to hotfix any python package in case we discover a bug or CVE. # Changes to the CI * `Nightly JAX Build` bumps the manifest.yaml and patches and commits them to a trial branch. If the tests pass, this trial branch should be merged to `main` * introduces `bump.sh` that bumps the world state given the manifest * No more `REPO_*` and `REF_*` build args, since they are all specified in the `manifest.yaml` * `get-source.sh` and `create-distribution.sh` now take the manifest * Rename pip-tools "manifests" to "requirements-*.in` to avoid confusion and use the same terminology as pip-tools # Not addressed in this PR * [ ] Automating the merging of the trial branch at the end of a successful nightly run * [ ] Running presubmit on latest world state (currently only will run on last-known working state) * [ ] Move lingvo patches to the manifest (#409 ) * [ ] Move xla arm neon patch to the manifest * [ ] Add a cleanup.sh script to remove things like `~/.gitconfig` * [ ] Have not implemented `mode: pip-constraint` since there was no dep that needed a constraint --------- Co-authored-by: Yu-Hang Tang <Tang.Maxin@gmail.com>
|
No longer relevant. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
After #405 gets merged, we should move the lingvo patches into the manifest and use
create-distribution.shto build our flavor oflingvoneeded for ARM64.The text was updated successfully, but these errors were encountered: