---
# Matrix-deploy variant overrides for this role. Each list entry is
# one variant deep-merged on top of `meta/services.yml`. The development
# deploy wrapper iterates the global round count (max variants across
# all primary apps) and runs cleanup between rounds for any app whose
# variant changed. Apps with fewer variants stay on variant 0 in the
# extra rounds.
# See docs/contributing/artefact/files/role/variants.md for the file format.
#
# Variant 0 runs the round that every other role's variant 0 runs with, so it
# preloads nothing: the gateway then serves its deterministic mock and no row
# pays for pulling or running a model. Variant 1 preloads for real and is the
# round that proves this role actually serves.
#
# Variant 1 serves three models whose declared windows differ, because that is
# what makes the gateway's router observable: a prompt larger than the smallest
# window can only be answered by a larger one, so the model that answers says
# which route was chosen. A declared window below what the tag supports is
# deliberate; it bounds the KV cache and it is the number the router filters on.
#
# Order is load-bearing. The CLI probe sizes its prompt off the smallest
# declared window, so whichever model holds that window is the one the request
# must exclude, and the model answering instead is what shows the filter ran.
- services:
docs:
enabled: true
shared: true
container_backup:
enabled: true
ollama:
mem_reservation: 1g
mem_limit: 2g
preload_models: []
- services:
docs:
enabled: false
shared: false
container_backup:
enabled: true
ollama:
mem_reservation: 3g
mem_limit: 6g
preload_models:
- alias: qwen2.5:0.5b
context: 4096
- alias: smollm2:135m
context: 8192
- alias: llama3.2:1b
context: 65536
