fix: validate generation temperature

[ryanL1100]

Summary

• reject negative and non-finite generation temperatures before sampling
• preserve temperature=0 greedy decoding behavior

Tests

• PYTHONPYCACHEPREFIX=/tmp/deepseek-v3-pycache python3 -m py_compile inference/generate.py
• git diff --check

[qingkong66]

@ryanL1100

Thank you for this — it's a focused follow-up to #1586.

Validating generation temperature before sampling is the kind of edge-case handling that prevents silent failures. Rejecting negative and non-finite values early, before they reach the sampling path, saves time that would otherwise be spent debugging opaque errors.

The fact that temperature=0 greedy decoding behavior is preserved is also worth noting — this is about rejecting invalid inputs, not changing behavior for valid ones.

It's good to see this pattern of incremental edge-case validation continuing.

I'll include this in the summary under inference stability, alongside #1586.

— qingkong66