21 lines
1.2 KiB
Plaintext
21 lines
1.2 KiB
Plaintext
# prain stack config, read by the `prain` front door (bash).
|
|
# Install: cp scripts/prain.conf.example ~/.config/prain/prain.conf
|
|
# Commands run in PRAIN_ROOT with its scripts/env.sh sourced, so `prain` in them is the
|
|
# venv's CLI. Service names must be shell identifiers.
|
|
|
|
PRAIN_ROOT=/home/fabian/coding/misc/prain
|
|
PRAIN_SERVICES="vllm web" # `prain up` starts in this order, `prain down` stops in reverse
|
|
|
|
# cmd_NAME: what service prain.NAME runs. ready_NAME: URL that answers once it can serve
|
|
# (`prain up` waits for it, PRAIN_READY_TIMEOUT seconds, default 900). requires_NAME:
|
|
# services that must already be up (their ready URL answers); if not, NAME refuses to start
|
|
# and says so. They are never started for it.
|
|
cmd_vllm='prain serve build/agent-int4.toml models/gemma-4-26B-A4B-it-Uncensored-MAX --
|
|
--served-model-name gemma4 --language-model-only --max-model-len 16384
|
|
--max-num-batched-tokens 1024 --enable-prefix-caching --port 8000
|
|
--attention-backend TRITON_ATTN --enable-auto-tool-choice --tool-call-parser gemma4'
|
|
ready_vllm=http://127.0.0.1:8000/health
|
|
|
|
cmd_web='prain web --host 0.0.0.0 --port 8080 --url http://127.0.0.1:8000 --config build/agent-int4.toml'
|
|
requires_web=vllm
|