harn-vm 0.10.16

Async bytecode virtual machine for the Harn programming language
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

# Generation defaults are inherited when a call does not set the option.
# Exact route patterns may override publisher defaults selected through
# `logical:<logical_model>`.

# OpenAI's GPT-OSS reference configuration recommends neutral sampling and
# high reasoning effort for maximum coding/reasoning quality. Every hosted
# route below shares `logical_model = "openai-gpt-oss-120b"`; provider-specific
# wire translation remains owned by the capability registry.
# https://huggingface.co/openai/gpt-oss-120b/discussions/21
[model_defaults."logical:openai-gpt-oss-120b"]
temperature = 1.0
top_p = 1.0
reasoning_effort = "high"

# Baseten's managed GPT-OSS route does not expose nucleus sampling. Keep the
# remaining logical defaults while omitting the unsupported field at the route
# boundary. https://docs.baseten.co/inference/model-apis/overview
[model_defaults."baseten/openai/gpt-oss-120b"]
_unset = ["top_p"]

# Groq compound systems publish an 8192-token completion ceiling.
[model_defaults."groq/compound*"]
max_tokens = 8192

# Moonshot's direct Kimi K2.5 route accepts only temperature 1.
[model_defaults."moonshot/kimi-k2.5"]
temperature = 1.0