typical_p is dead code and should be removed

[seayang-nv]

Priority Level

Medium

Task Summary

typical_p is dead code and should be removed. Here is the evidence:

• Not a user-facing parameter. typical_p does not appear in any config model (GenerateParameters), CLI option, YAML preset, or DEFAULT_SAMPLING_PARAMETERS.
• Never passed to prepare_params. The only call site that invokes prepare_params is in vllm_backend.py:513-526 — the sampling_kwargs dict it constructs contains temperature, repetition_penalty, top_p, top_k, min_p, logits_processors, max_tokens, etc., but not typical_p.
• _get_api_param_mapping maps it, but nothing triggers it. The "typical_p" key in the mapping dict line 295 would only fire if typical_p appeared in the kwargs — it never does.
• No other backend uses it. The timeseries backend and HuggingFace training backend do not reference typical_p.
• The TypicalLogitsWarper import from transformers is only used by this dead path.

raised in this comment: PR #342

Technical Details & Implementation Plan

No response

Dependencies

No response