chore: update cogames mettagrid to 0.2.0.53 #4515
Conversation
yatharth
temporarily deployed
to
cogames-pypi
GitHub Actions
Inactive
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| class JobRequestUpdate(SQLModel): | ||
| status: JobStatus = Field(default=JobStatus.pending) |
There was a problem hiding this comment.
Because JobRequestUpdate.status defaults to pending, FastAPI will populate request.status even when the client does not send a status field. The update endpoint in job_routes.py treats any non-None status as a requested transition and enforces VALID_TRANSITIONS, so a request that only wants to set result or error will be forced into a pending transition and can return 409 for jobs that are already running/completed. This makes partial updates fail unless clients always include a legal status. Consider changing status to JobStatus | None = None (or using Field(default=None) with exclude_unset) so updates without a status don’t trigger transition validation.
Useful? React with 👍 / 👎.
| self._out: nn.Module = nn.Sequential(*layers) if layers else nn.Identity() | ||
| if self._storage_dtype is not None and self._storage_dtype is not torch.float32: | ||
| self._out = self._out.to(dtype=self._storage_dtype) |
There was a problem hiding this comment.
The new cast of self._out to self._storage_dtype means the output head’s weights are always in the storage dtype (e.g., bf16). If a caller sets compute_dtype to float32 (supported by the config) while dtype is bf16, the stack produces float32 activations and then self._out(y) either errors on mixed dtypes or silently downcasts, defeating the higher-precision compute path. This regression only appears when compute_dtype != dtype, but that is a documented option here. Consider keeping _out in compute dtype or casting y to the storage dtype before the head when compute_dtype differs.
Useful? React with 👍 / 👎.
yatharth
force-pushed
the
yatharth/metta-publish-cogames-version-pinning
branch
2 times, most recently
from
60f83dd to
13fa617
Compare
This PR was automatically created by
metta/setup/tools/publish.py.Asana Task