Skip to content

Open-source AI teammates in Slack, Teams, Discord, Telegram and Mattermost. Self-host it, with any agent harness and any model.

Offer Strands Decider in local compose on an NVIDIA GPU

agent-runner353a175d66ea6d8
diff --git a/deploy/local/docker-compose.yml b/deploy/local/docker-compose.yml
index 98f693d5a..fab947943 100644
--- a/deploy/local/docker-compose.yml
+++ b/deploy/local/docker-compose.yml
@@ -135,12 +135,20 @@ services:
- "${SWITCH_BIND_ADDR:-127.0.0.1}:8025:8025"
# ── Decision service (auto-route) ───────────────────────────────────────────
- # `just decision-up`, or `--profile decision` (CPU, any machine) or
- # `--profile decision-gpu` (Linux with an NVIDIA GPU and the NVIDIA Container
- # Toolkit). Either starts the decision service on :8230 and a llama.cpp
- # server for its model, with the weights downloaded once into the
- # `decision-models` volume. switch-core runs on the host, so point it at the
- # service in .env (see .env.example, "Decision routing"):
+ # `just decision-up [cpu|gpu|strands]`, or one of these profiles. Each starts
+ # the decision service on :8230 and a model server beside it:
+ # decision Qwen3.5-4B on llama.cpp, on CPU (any machine). The
+ # local default.
+ # decision-gpu the same on an NVIDIA GPU.
+ # decision-strands-gpu Strands Decider 4B, the production default (the
+ # Helm chart's): more accurate, but it needs an NVIDIA
+ # GPU with 24 GB, about 24 GiB of free RAM to load,
+ # and the first start downloads about 17 GB.
+ # The GPU profiles need Linux, an NVIDIA GPU and the NVIDIA Container
+ # Toolkit; there is no GPU passthrough on macOS. Weights download once into
+ # a volume (`decision-models`, or `decision-hf` for Strands). switch-core
+ # runs on the host, so point it at the service in .env (see .env.example,
+ # "Decision routing"):
# DECISION_ENDPOINT_URL=http://localhost:8230
# DECISION_TIMEOUT_SECONDS=30
# On CPU each decision takes several seconds (far over the one-second target
@@ -238,7 +246,35 @@ services:
decision-model-download:
condition: service_completed_successfully
- decision:
+ # Strands Decider 4B's model server (decision-strands-gpu). The checkpoint
+ # and its Gemma 4 base download from Hugging Face into `decision-hf` on the
+ # first start, then load; the port opens only once the model is loaded, so
+ # allow several minutes (much longer on the first start). HF_TOKEN is
+ # optional: for a gated checkpoint or higher Hub rate limits.
+ decision-model-strands:
+ build:
+ context: decision/model-servers/strands
+ image: switch-decision-strands:local
+ profiles: ["decision-strands-gpu"]
+ environment:
+ HF_TOKEN: ${HF_TOKEN:-}
+ volumes:
+ - decision-hf:/models
+ deploy:
+ resources:
+ reservations:
+ devices:
+ - driver: nvidia
+ count: 1
+ capabilities: [gpu]
+ healthcheck:
+ test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8400/health', timeout=3)"]
+ interval: 10s
+ timeout: 5s
+ retries: 3
+ start_period: 30m
+
+ decision: &decision
build:
context: decision
dockerfile: Dockerfile
@@ -265,7 +301,21 @@ services:
condition: service_healthy
required: false
+ # The decision service in front of Strands Decider: the same image, speaking
+ # SystemOne to its model server.
+ decision-strands:
+ <<: *decision
+ profiles: ["decision-strands-gpu"]
+ environment:
+ DECISION_BACKEND: systemone
+ DECISION_MODEL: strands-decider-E4B
+ SYSTEMONE_URL: http://decision-model-strands:8400/v1/systemone
+ depends_on:
+ decision-model-strands:
+ condition: service_healthy
+
volumes:
pgdata:
mmdata:
decision-models:
+ decision-hf:
diff --git a/justfile b/justfile
index 7f50e80bd..4048ab13d 100644
--- a/justfile
+++ b/justfile
@@ -56,18 +56,26 @@ down:
reset:
docker compose -f deploy/local/docker-compose.yml --project-directory . down -v
-# Start the decision service auto-route asks (with its model server), on CPU or,
-# with `just decision-up gpu`, an NVIDIA GPU on Linux. Weights download once
-# into a volume. docs/old/LOCAL_DEVELOPMENT.md, "Auto-route locally".
+# Start the decision service auto-route asks (with its model server): Qwen on
+# CPU by default, `gpu` for Qwen on an NVIDIA GPU, or `strands` for Strands
+# Decider 4B (the production default; an NVIDIA GPU with 24 GB). GPU modes are
+# Linux only. Weights download once into a volume.
+# docs/old/LOCAL_DEVELOPMENT.md, "Auto-route locally".
decision-up mode="cpu":
#!/usr/bin/env bash
set -euo pipefail
+ service=decision
case "{{mode}}" in
cpu) profile=decision ;;
gpu) profile=decision-gpu ;;
- *) echo "usage: just decision-up [cpu|gpu]" >&2; exit 2 ;;
+ strands) profile=decision-strands-gpu service=decision-strands ;;
+ *) echo "usage: just decision-up [cpu|gpu|strands]" >&2; exit 2 ;;
esac
- docker compose -f deploy/local/docker-compose.yml --project-directory . --profile "$profile" up -d --build --wait decision
+ if [ "$profile" = decision-strands-gpu ]; then
+ echo "Strands Decider: the first start downloads about 17 GB, then loads it;"
+ echo "this waits until the model server is up (minutes)."
+ fi
+ docker compose -f deploy/local/docker-compose.yml --project-directory . --profile "$profile" up -d --build --wait "$service"
port="${DECISION_PORT:-8230}"
echo "Decision service: http://localhost:$port (curl http://localhost:$port/health)"
if [ "$profile" = decision ]; then
@@ -80,7 +88,7 @@ decision-up mode="cpu":
echo " FEATURE_FLAGS_ENABLED=follow_the_conversation,decision_routing (plus any you had)"
decision-down:
- docker compose -f deploy/local/docker-compose.yml --project-directory . --profile decision --profile decision-gpu stop decision decision-model decision-model-gpu
+ docker compose -f deploy/local/docker-compose.yml --project-directory . --profile decision --profile decision-gpu --profile decision-strands-gpu stop decision decision-model decision-model-gpu decision-strands decision-model-strands
# ── Run switch-core locally ────────────────────────────────────────────────────
# The Python project lives in core/; `--project core` selects that environment