stacks/qwen3-tts: target=production + user=root + correct HF model id

Three fixes from the first deploy attempt on irv-ml1:

- build.target=production. Upstream Dockerfile is multistage; the last
  stage `cpu-base` was selected by default, producing a CPU-only image
  with no flash-attn and `torch ... whl/cpu`.
- user: "0:0". Upstream image declares USER appuser but writes runtime
  state under /root (mode 0700). appuser cannot traverse /root, so
  /v1/voices 500s on PermissionError. Run as root to sidestep.
- QWEN3_TTS_MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice. The bare
  `1.7B` id we had isn't a real HF identifier; upstream publishes
  -CustomVoice / -Base variants of each size. Use -CustomVoice so
  `voice="clone:<name>"` works.

Tag bumped to v2 to keep the v1 cpu image distinguishable in the local
registry.

After: all 5 verify steps pass, GPU synthesis ~5s for 3-4s of audio,
three contrasting English `instructions` produce three distinct
hashes — emotion steering actually works (unlike CosyVoice's English
path).

stacks/qwen3-tts/.env.example

@@ -6,8 +6,10 @@
 QWEN3_TTS_SHA=10323ce778c48a75dbda93d0a4891983fb371f58
 
 # Local image tag — bump when you change build context to force a
-# fresh layer build.
+# fresh layer build. v2 = first GPU build (target=production); v1
-QWEN3_TTS_TAG=v1
+# was the accidental CPU-only image (last stage of upstream's
+# multi-stage Dockerfile).
+QWEN3_TTS_TAG=v2
 
 # ── network ──────────────────────────────────────────────────────────
 # Host port (container listens on 8880 internally).
@@ -23,10 +25,14 @@ QWEN3_TTS_BIND=0.0.0.0
 # more VRAM); `pytorch` = bare pytorch path.
 QWEN3_TTS_BACKEND=official
 
-# Model variant. 1.7B = flagship, 6–8 GB VRAM with bfloat16, best
+# Model variant. Upstream publishes four checkpoints on HF:
-# quality + control. 0.6B = lightweight, ~2–3 GB VRAM, faster, slightly
+#   Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice — flagship, voice cloning
-# less expressive.
+#   Qwen/Qwen3-TTS-12Hz-1.7B-Base        — flagship, no cloning
-QWEN3_TTS_MODEL=Qwen/Qwen3-TTS-12Hz-1.7B
+#   Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice — lightweight, voice cloning
+#   Qwen/Qwen3-TTS-12Hz-0.6B-Base        — lightweight, no cloning
+# 1.7B = ~6–8 GB VRAM bfloat16, best quality. 0.6B = ~2–3 GB.
+# Use -CustomVoice for `voice="clone:<name>"` to work.
+QWEN3_TTS_MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
 
 # Warm the model on container start so the first synthesis request
 # doesn't pay the load latency. Adds ~30 s to startup. Recommended.

stacks/qwen3-tts/compose.yaml

@@ -29,9 +29,22 @@ services:
     build:
       context: https://github.com/groxaxo/Qwen3-TTS-Openai-Fastapi.git#${QWEN3_TTS_SHA}
       dockerfile: Dockerfile
+      # Upstream Dockerfile defines five stages — without an explicit
+      # target, buildkit picks the LAST named stage (`cpu-base`) and
+      # produces a CPU-only image with no GPU torch + onnxruntime
+      # (not -gpu). Pin to `production` to get the CUDA stack with
+      # flash-attn that the model actually needs.
+      target: production
     container_name: qwen3-tts
     restart: unless-stopped
     runtime: nvidia
+    # Upstream image declares `USER appuser` (uid 1000) but writes
+    # state under /root (mode 0700) — appuser can't traverse it, so
+    # /v1/voices and any FS-touching endpoint 500s. The model loads
+    # only because warmup happens in an early root phase. Pin to root
+    # so all paths are reachable. Files written into the host bind
+    # mounts become root-owned; that's fine for restic + sudo reads.
+    user: "0:0"
     ports:
       - "${QWEN3_TTS_BIND:-0.0.0.0}:${QWEN3_TTS_PORT}:8880"
     environment: