ace-step + stable-audio-open: deploy music + SFX generation to irv-ml1

Two new audio-generation stacks alongside the TTS slate:

ace-step :8210 — Apache 2.0 music generation foundation model
(hybrid diffusion + LLM). Lyric-aware multi-minute songs. ~10-12 GB
VRAM during inference, A6000-pinned. Custom Dockerfile patches
upstream's torch/cu126 resolution bug (--extra-index-url cu126 was
falling back to pypi-default cu13 wheels, mismatching torchvision).

stable-audio-open :8211 — Stability AI 1.21B latent-diffusion SFX +
ambience. Up to 47s clips at 44.1 kHz. ~6 GB VRAM in fp16,
A6000-pinned. Custom FastAPI shim around diffusers' StableAudioPipeline
(no upstream HTTP server). Dockerfile pins torchsde explicitly —
diffusers doesn't pull it as a hard dep but
CosineDPMSolverMultistepScheduler needs it.

stacks/ace-step/compose.yaml

@@ -0,0 +1,70 @@
+# ACE-Step 1.5 — open-source music generation foundation model
+# (April 2026). Hybrid diffusion + LLM architecture, Apache 2.0.
+# ~50-80 s for a 4-minute song on A6000; under 4 GB VRAM at idle,
+# ~10-12 GB during inference. Beats YuE / DiffRhythm on the
+# speed/coherence trade.
+#
+# We launch upstream's REST API (`infer-api.py`) instead of the
+# default `gui.py` (Gradio). The REST surface is what we'll point
+# clients + automation at; Gradio is dev-time eye candy.
+#
+# Image is built locally from upstream's repo via docker buildx git
+# context, same pattern as fish-s2.
+#
+# All tunables live in .env — edit that, not this file.
+
+services:
+  ace-step:
+    image: local/ace-step:${ACE_STEP_TAG}
+    build:
+      # Build from local Dockerfile (not upstream's git context) — we
+      # ship a patched Dockerfile that fixes upstream's torch/cu126
+      # resolution bug. Playbook uploads Dockerfile alongside this
+      # compose.yaml.
+      context: .
+      dockerfile: Dockerfile
+      args:
+        ACE_STEP_REF: ${ACE_STEP_SHA}
+    container_name: ace-step
+    restart: unless-stopped
+    runtime: nvidia
+    ports:
+      # Container default for infer-api.py is 8000 (hardcoded
+      # uvicorn.run(host=0.0.0.0, port=8000) — no flags). Map host
+      # ACE_STEP_PORT to it.
+      - "${ACE_STEP_BIND:-0.0.0.0}:${ACE_STEP_PORT}:8000"
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=${ACE_STEP_GPU_DEVICES:-1}
+      # ACE_OUTPUT_DIR is read by acestep at generation time — keep
+      # in sync with the bind mount below.
+      - ACE_OUTPUT_DIR=/app/outputs
+      # HF_HOME points the HuggingFace cache at the bind mount so the
+      # ~5-10 GB checkpoint download survives container recreates.
+      - HF_HOME=/app/hf_cache
+    volumes:
+      - ${ACE_STEP_CHECKPOINTS_DIR}:/app/checkpoints
+      - ${ACE_STEP_OUTPUTS_DIR}:/app/outputs
+      - ${ACE_STEP_LOGS_DIR}:/app/logs
+      - ${ACE_STEP_CACHE_DIR}:/app/hf_cache
+    # Override upstream's default `python3 acestep/gui.py` with the
+    # REST API entry point. infer-api.py self-binds 0.0.0.0:8000 and
+    # exposes POST /generate + GET /health.
+    command: ["python3", "infer-api.py"]
+    healthcheck:
+      # /health is the cheapest signal infer-api.py exposes — returns
+      # 200 as soon as the FastAPI app is up. The pipeline lazy-loads
+      # on first /generate, so /health says "process alive" not
+      # "model warm". Good enough for a liveness signal; first
+      # /generate has the ~30-60 s warmup baked in.
+      test: ["CMD-SHELL", "python3 -c \"import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5).status==200 else 1)\""]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      # First boot pulls ACE-Step checkpoint (~5-10 GB) into HF cache.
+      start_period: 600s
+    labels:
+      - homepage.group=AI Systems
+      - homepage.name=ACE-Step
+      - homepage.icon=mdi-music-note-eighth
+      - homepage.description=Open-source music generation — 4-min song in ~60s, lyrics + style prompts (irv-ml1)
+      - homepage.href=http://10.100.79.3:${ACE_STEP_PORT}