Repository navigation
fix(openai_model): bound refused provider connection retries - #5199
Conversation
Signed-off-by: Adam Rajfer <arajfer@nvidia.com>
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
|
Could you update the PR’s validation paragraph with the closed-judge, healthy-judge, and invalid-key results, identifying 34b1a26 as the combined #5072 + #5199 proof source? The three changed files match this PR’s head, but the surrounding failure-reporting code differs, so the saved attempt IDs and structured cause links should be attributed to that combined source. |
Signed-off-by: Adam Rajfer <arajfer@nvidia.com>
|
Updated the validation paragraph with the closed-judge, healthy-judge and invalid-key results, explicitly attributing the saved attempts and structured cause evidence to combined #5072 + #5199 proof source 34b1a26. The adapter implementation and tests match, while the surrounding failure-reporting code differs. No new inference run was performed. |
|
🌿 Preview your docs: https://nvidia-preview-pr-5199-ed4e535d1cb6.docs.buildwithfern.com/nemo/gym |
svcnvidia-nemo-ci
left a comment
There was a problem hiding this comment.
SHIP — LGTM — no concerns found.
Reviewed all four changed files and traced the adapter wiring through the shared client and transport retry branches. The connection limit preserves explicit single-attempt mode, and the regressions cover bounded refusal failures and HTTP 500 with the connection cause. Both documentation clarifications requested in the history are addressed.
Static inspection only; tests and documentation builds were not executed. The runtime topic was independently completed by the parent reviewer because its delegated terminal report was unavailable.
Note: formal approval is not available for this request, so this review is posted as a comment only.
Review generated by codex. This was a full review. Request a full review with /review reset=true.
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/40880755061 |
A provider that refused connections could leave the default openai_model client retrying indefinitely. Apply max_http_attempts to its connection retry budget as well, so the request eventually fails with its connection cause. Explicit single-attempt mode remains a single attempt.
Real loopback closed-port regressions fail on unchanged main (three failures and one passing single-attempt control) and pass with the fix: the default makes three attempts, a configured limit makes five, and the server returns HTTP 500 with ClientConnectorError. Native adjacent suites: 725 passed; all pre-commit hooks passed.
Additional fault validation used the combined #5072 + #5199 source 34b1a268b65c1356c9370664bc0e26feca637c64: the closed-judge path stopped after three connection attempts per call, the healthy-judge control completed 20/20, and the invalid-key control stopped after one judge attempt per call. The adapter implementation and tests match that source; its surrounding failure-reporting code comes from #5072. Saved attempt IDs and structured cause links therefore belong to that combined proof, rather than this PR alone.
The documentation follow-up explains the shared connection retry budget and that values of one or two lower request-timeout attempts from three to one or two; the default stays three. All repository pre-commit hooks passed. It changes no runtime code and required no new inference run.
Tested on main c155b77. This changes the provider client in openai_model; other model servers and readiness probes retain their policies.