-
Studio: load Kimi-K3 with thinking on, and fix its sampling defaults (#7619)
frostbyte_neo 发布于
2026-07-29 15:35:02 +00:00 Kimi-K3 loaded with reasoning disabled. Its template branches
reasoning_effort on 'none' as the disable sentinel, so the literal scan
swept 'none' into the effort ladder as the weakest entry. The chat store
ships "medium", which the ladder does not offer, so the clamp fell back
to levels[0] == 'none', and _request_reasoning_kwargs turns that into
enable_thinking=false. The Think menu hides 'none', so the level could
not be picked by hand either -- it only ever arrived from the fallback.Drop the sentinel from the enable_thinking_effort ladder. Kimi-K3 now
reports low/high/max, the same untouched clamp lands on low, and
thinking is on. Disabling is unaffected: _request_reasoning_kwargs
compares reasoning_effort == "none" directly rather than testing list
membership, so a raw caller still disables.Checked against 33 templates old vs new (29 from the Hub, plus Studio's
four fixtures): only Kimi-K3 changes. GLM-5.2 and DeepSeek-V4 keep
['high', 'max'], and the max-only and high-only fixtures are identical.Also move the Kimi-K3 sampling defaults from the model_defaults YAML
added in #7616 into inference_defaults.json. The YAML is reached by
exact alias or a one/two-component path suffix, so only the bare repo id
matched; a cache snapshot path, a .gguf path and repo:variant all fell
through to the kimi family at temperature 0.6 / min_p 0.01. Family
patterns match all of them. It also restores the training form reset:
load_model_defaults replaces default.yaml rather than merging with it,
so an inference-only YAML left the previous model's hyperparameters in
place. Diffed 394 model ids; exactly the two broken Kimi-K3 path forms
change.下载附件