feat: add top_k, min_p, repetition_penalty, presence_penalty

[chethanuk]

📋 Summary

ChatCompletionInferenceParams only takes temperature, top_p and max_tokens, so today top_k, min_p, repetition_penalty and presence_penalty have to be written into extra_body by hand. This adds them as optional, range-checked fields sent as top-level request fields. vLLM and SGLang accept all four; a strict OpenAI endpoint may reject the first three.

params = dd.ChatCompletionInferenceParams(
    temperature=0.7,
    top_k=20,
    min_p=0.05,
    repetition_penalty=1.05,
)

🔗 Related Issue

Closes #354

🔄 Changes

• top_k, min_p and repetition_penalty are fields on the engine's ChatCompletionRequest and in _COMPLETION_REQUEST_FIELDS, like presence_penalty. A per-call extra_body can't drop them, and a same-named extra_body key still wins.
• The Anthropic adapter also drops min_p and repetition_penalty. top_k is a Messages API param and is still sent.
• The fingerprint skips the new keys while they're None, so existing config_hash values don't change and older runs still resume.
• frequency_penalty is left out since Add other common params to ChatCompletionInferenceParams #354 doesn't ask for it. The engine request already has the field, so it's a small follow-up.

🧪 Testing

• make test passes
• Unit tests added/updated
• E2E tests added/updated (if applicable)

Ran a 2-record preview against a local mock server with one LLM text column per provider type, all four params set and extra_body={"chat_template_kwargs": {...}}. The /v1/chat/completions body had all four fields and chat_template_kwargs at the top level. The /v1/messages body had top_k and no min_p or repetition_penalty, and both columns generated. ModelFacade.generate(..., extra_body={"reasoning_effort": "high"}) still sends all four fields.

✅ Checklist

• Follows commit message conventions
• Commits are signed off (DCO)
• Architecture docs updated (if applicable)

[github-actions]

Linked Issue Check

Issue #354 has not been triaged yet. A maintainer needs to review
the issue and add the triaged label for this check to pass.

You can continue working on the PR in the meantime. The check will
re-run automatically once the issue is triaged.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge, with the new sampling parameters consistently represented across configuration, fingerprinting, request construction, provider adaptation, tests, and documentation.

Summary

This PR adds typed sampling controls to chat-completion configuration and carries them through configuration identity and provider request construction.

• Adds validated presence_penalty, top_k, min_p, and repetition_penalty fields.
• Preserves existing fingerprints when the optional fields are unset.
• Routes the fields through canonical OpenAI-compatible requests while filtering unsupported Anthropic fields.
• Adds synchronous, asynchronous, validation, fingerprint, and documentation coverage.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[ChatCompletionInferenceParams] --> B[generate_kwargs]
    B --> C[ModelFacade request consolidation]
    C --> D[ChatCompletionRequest]
    D --> E{Provider adapter}
    E -->|OpenAI-compatible| F[Forward all configured sampling fields]
    E -->|Anthropic| G[Forward top_k and exclude unsupported fields]
    A --> H[Fingerprint normalization]
    H --> I[Drop new fields only when unset]

Loading

_{Reviews (1) · Last reviewed commit: "fix: send top_k, min_p, repetition_penal..."}