[silencelamb]

Motivation

Requests that omit temperature currently fall back to 0.8, which sharpens the model distribution, and clients using top_k=-1 to disable top-k filtering are rejected. Use a neutral temperature fallback and accept the disabling convention used by vLLM and SGLang.

This complements #5079 and can be reviewed and merged independently.

Modification

• Change the temperature fallback to 1.0 in GenerationConfig, the PyTorch sampling parameters, and the CLI argument helper.
• Accept top_k=-1 alongside 0 in Chat Completions, Completions, Responses, Anthropic Messages, and /generate. Normalize -1 to 0 in the shared GenerationConfig before either inference backend receives it; continue rejecting values below -1.
• Cover request/model precedence, model config loading, API validation and adapter paths, Python configuration, and zero-temperature greedy conversion with regression tests.
• Document the sampling behavior and migration in English and Chinese.

BC-breaking

Sampled outputs may change when neither the request nor the effective model/server defaults supplies a temperature. Set temperature=0.8 explicitly to retain the previous fallback behavior. Explicit request/model temperatures retain their precedence; greedy decoding is unaffected. Accepting top_k=-1 is an additive compatibility change.

Validation

• Focused configuration, message, Chat/Completion, Responses, Anthropic Messages, and /generate tests: 211 passed, 4 deselected with Transformers 5.16.0.
• The same suite with Transformers 4.57.6: 211 passed, 4 deselected. The four deselected cases require unrelated tokenizer downloads.
• All applicable pre-commit hooks passed on the changed files; git diff --check passed.
• Rendered both new documentation sections with Sphinx/MyST, treating warnings as errors.
• Tests use CPU and fake serving engines; no GPU end-to-end model inference was run.

[RunningLeon]

LGTM