Add NVIDIA provider, and improve declarative provider UX#8798
Conversation
Signed-off-by: jh-block <jhugo@block.xyz>
jh-block
force-pushed
the
jhugo/nvidia-provider
branch
from
f04b989 to
b39ff8c
Compare
jh-block
changed the title
![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.
Reviewed commit: f04b989c88
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".
| self.max_tokens = canonical | ||
| .limit | ||
| .output | ||
| .filter(|&output| output < canonical.limit.context) | ||
| .map(|output| output as i32); |
There was a problem hiding this comment.
This filter now drops every canonical output limit that is not strictly less than context, so models with output > context in the bundled registry lose their explicit limit and fall back to max_output_tokens() = 4096. In practice that silently shrinks allowed output for affected models (for example entries in canonical_models.json with larger output caps), even though this change was intended to handle the output == context sentinel case. Restricting the skip to equality (or clamping) avoids this unintended 4k cap.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Will fix the canonical data and then we can look at removing this workaround anyway.
* main: (29 commits) chore(deps): bump winreg from 0.55.0 to 0.56.0 (#8829) Fix grammar issue (#8669) colorize context window indicator (#8851) Refresh canonical model metadata from models.dev (#8838) fix(ci): prevent flaky smoke test timeouts from failing the build (#8837) updates: release 0.19.0 of the tui/sdk/etc (#8806) add a goose2 signed release flow (#8728) Port provider tests to typescript (#8237) refactor: make ACP server smaller (#8787) Add NVIDIA provider, and improve declarative provider UX (#8798) fix: removed failed provider test for deprecated providers (#8801) fix: only call cleanup when the pr is from same repo (#8799) chore: check stale for draft pr (#8803) fix: use _meta instead of meta in newSession request (#8796) fix: add missing underscore prefix in updateWorkingDir method name (#8743) feat: migrate session metadata storage from frontend overlay to backend (#8769) Add more info to BUILDING_LINUX (#8789) feat(acp): Align to new request patterns of ACP Streamable HTTP/WS transport (#8605) Dedupe and organize skills/sources (#8731) docs: add skills slash command (#8783) ...
2 participants
Fixes #8505
Summary
Testing
cargo fmt --check, targetedcargo test -p goose ...provider/model tests, and desktop TypeScript typecheck.