voxtral: mount patched stage YAML to dodge hardcoded 0.8 GPU util; fish-s2: --half + streaming wins

Voxtral final fix (8th iteration):
* The bundled voxtral_tts.yaml hardcodes gpu_memory_utilization: 0.8
  on the language_model stage — overrides the CLI flag. Mounted a
  patched copy (0.4) at /etc/voxtral/voxtral_tts.yaml and pointed
  --stage-configs-path there.
* With Kyutai stopped to free 5 GB on the 3090, both stages fit
  (target 9.4 + 2.4 GB ≈ 11.8 GB; 17 GB free post-kyutai-stop).
* Voxtral now healthy on GPU 0 — bench: 1.9-2.7 s TTFB, real WAV.

Fish s2-pro optimization (per-request sweep, no model swap):
* `streaming: true` in request body drops TTFB from 7.7 s → 0.026 s
  (300×). Total time goes up ~1 s (chunked HTTP overhead) but
  perceived latency = TTFB. Use stream:true for any interactive use.
* `latency: "balanced"` actually slower than default — bad name; skip.
* `use_memory_cache: "on"` no measurable benefit.
* `chunk_length: 100` (default 200) no TTFB benefit non-streaming.
* Server-side `--half` (fp16 inference) added via compose `command`
  override — passes through start_server.sh's $@ unchanged into
  api_server.py. Should reduce total time too. Validation pending
  the post-restart bench.

Kyutai stopped to free GPU 0 budget — the bench numbers earlier
(3.4 s avg) were unimpressive vs Voxtral's 2.3 s in the same
multilingual slot. Kept the stack files for future re-deploy if
needed; just the running container is gone.

stacks/fish-s2/compose.yaml

@@ -80,6 +80,11 @@ services:
       - ${FISH_S2_REFERENCE_DIR}:/app/references
       # Persistent HF cache so model re-pull only happens on first deploy.
       - ${FISH_S2_CACHE_DIR}:/app/hf_cache
+    # Pass --half to start_server.sh → enables fp16 inference on the
+    # LLM half. Speeds up the autoregressive loop (the dominant cost
+    # in TTFB). Also enables the streaming path's faster total time.
+    # build_compile_args() echoes unknown args back to the exec line.
+    command: ["--half"]
     healthcheck:
       # Fish ships /v1/health on the API server. python urllib instead
       # of curl because the upstream image is python-based and may not

stacks/voxtral/compose.yaml

@@ -37,6 +37,12 @@ services:
     volumes:
       - ${VOXTRAL_CACHE_DIR}:/root/.cache/huggingface
       - ${VOXTRAL_VOICES_DIR}:/voices:ro
+      # Patched stage config — bundled YAML hardcodes
+      # gpu_memory_utilization: 0.8 on the language_model stage which
+      # OOMs anywhere we have other models resident on the same GPU.
+      # Mount our own copy at a custom path; the --stage-configs-path
+      # flag below points at it.
+      - /opt/docker/conf/voxtral/voxtral_tts.yaml:/etc/voxtral/voxtral_tts.yaml:ro
     # vllm/vllm-omni image has no default ENTRYPOINT or CMD — the
     # container init expected --model=... as argv[0]. Set entrypoint
     # to `vllm serve` (the standard CLI) and pass model as positional
@@ -54,7 +60,7 @@ services:
     entrypoint: ["vllm-omni", "serve"]
     command:
       - "${VOXTRAL_MODEL:-mistralai/Voxtral-4B-TTS-2603}"
-      - "--stage-configs-path=vllm_omni/model_executor/stage_configs/voxtral_tts.yaml"
+      - "--stage-configs-path=/etc/voxtral/voxtral_tts.yaml"
       - "--host=0.0.0.0"
       - "--port=8000"
       - "--gpu-memory-utilization=${VOXTRAL_GPU_UTIL:-0.45}"