fix(skippy): treat zero-temperature sampling as greedy - #1143
Conversation
|
This pull request is currently a draft. Reviews will not take place until the PR is marked as ready for review. |
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (2)
🚧 Files skipped from review as they are similar to previous changes (2)
📝 WalkthroughWalkthroughLinear proposal admission now evaluates the complete ChangesSampling-aware linear proposal admission
Estimated code review effort: 3 (Moderate) | ~20 minutes Possibly related PRs
Suggested labels: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
27a32e4 to
e5171e9
Compare
e5171e9 to
4d289b0
Compare
michaelneale
left a comment
There was a problem hiding this comment.
Low risk: local speculative-verification admission only, invisible across versions. Zero-temp treated as greedy is a correct, well-scoped change. LGTM.
4d289b0 to
42b7fdf
Compare
Problem
OpenAI clients commonly use
temperature: 0when they need reproducible output. Skippy still classified those requests as stochastic whenever sampling was enabled, so it rejected verified linear proposals even when no sampling option could change the winning token.That meant a deterministic request behaved differently depending on whether the client omitted sampling settings or explicitly requested zero temperature.
Fix
Admission now depends on whether the request can actually change greedy token selection. Zero temperature with neutral penalties and no logit bias is treated as deterministic. Non-zero temperature, grammar-constrained sampling, penalties, and logit bias remain excluded.
This keeps the safety boundary narrow: proposals are admitted only when verification is equivalent to greedy decoding, without special-casing a particular model or client.
Validation
Skippy Server tests pass (353/353), including deterministic and stochastic sampling cases.
Summary by CodeRabbit