feat(morpheus): permanent mOrpheus TTS stack (vLLM bf16 + SNAC/FastAPI wrapper) on irv-ml1

Two-container stack serving MrDragonFox/mOrpheus (uncensored Orpheus TTS, Llama-3.2-3B
-> SNAC 24kHz). vllm-morpheus (GPU/3090) emits Orpheus audio tokens; morpheus-tts (CPU)
SNAC-decodes them to WAV and exposes POST /tts (baddy voice + zero-shot cloning). Deployed
+ tested end-to-end (28/28 valid frames, valid WAV, reachable over WG).

Hard-won config, all encoded in compose/README:
- bf16 REQUIRED: --quantization fp8 destroys audio-token generation (0 valid SNAC frames
  even at greedy). Footprint ~7.9GB.
- Image PINNED to v0.23.0: 'latest' ships Blackwell oink/aiter kernels that crash on Ampere
  import.
- 3090 (not the comfy-contended A6000); --enforce-eager to fit the shared card.
- RTF ~1.0 end-to-end (gen ~98 tok/s / RTF 0.84 + CPU decode + HTTP).

INTERNAL RESEARCH ONLY (CC-BY-NC-4.0); do not expose externally.

stacks/mOrpheus/compose.yaml

@@ -0,0 +1,91 @@
+# mOrpheus — uncensored Orpheus TTS, permanent stack on irv-ml1.
+#
+# Two services:
+#   vllm-morpheus (GPU, 3090, FP8) — serves the mOrpheus 3B LLM; emits Orpheus audio
+#     tokens. Real-time: ~165 tok/s single-stream => RTF ~0.50 (2x real-time), measured.
+#   morpheus-tts  (CPU) — SNAC-decodes those tokens to 24kHz WAV; the public /tts endpoint
+#     (baddy voice + optional zero-shot cloning). Kept off the GPU to avoid contention.
+#
+# PLACEMENT / GOTCHAS (irv-ml1, INTERNAL RESEARCH — CC-BY-NC, do NOT expose externally):
+#   - GPU = the 3090 (audio-zoo GPU, stabler); the A6000 is comfy's and spikes to ~41GB
+#     without warning (it OOM'd two launch attempts). Pin by UUID via MORPHEUS_GPU_UUID.
+#   - Image PINNED to vllm/vllm-openai:v0.23.0 — `latest` ships Blackwell-only kernels
+#     (oink/aiter) that crash on Ampere import. Do NOT bump to latest on this Ampere box.
+#   - bf16 REQUIRED, NOT FP8. FP8 (--quantization fp8) destroys audio-token generation on
+#     this model — 0 valid SNAC frames even at greedy (quant-breaks-TTS). bf16 weights are
+#     ~6.6GB; with --enforce-eager (no CUDA-graph capture) it fits the shared 3090 (~8GB
+#     free after nothing else vLLM). util 0.32 (~7.7GB of the 24GB 3090).
+#   All tunables in .env — edit that, not this file.
+
+name: morpheus
+
+services:
+  vllm-morpheus:
+    image: vllm/vllm-openai:v0.23.0
+    container_name: vllm-morpheus
+    restart: unless-stopped
+    runtime: nvidia
+    ipc: host
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=${MORPHEUS_GPU_UUID:-GPU-12c15b3c-0827-0bab-b0cd-31aaa37a8e72}
+      - CUDA_DEVICE_ORDER=PCI_BUS_ID
+    volumes:
+      - ${MORPHEUS_MODEL_DIR:-/home/lkraven/morpheus/models/mOrpheus}:/model:ro
+    command:
+      - /model
+      - --served-model-name
+      - morpheus
+      - --dtype
+      - bfloat16
+      # NO --quantization: FP8 breaks audio-token generation on this model (see header).
+      - --enforce-eager
+      - --max-model-len
+      - "${MORPHEUS_MAX_LEN:-4096}"
+      - --gpu-memory-utilization
+      - "${MORPHEUS_GPU_UTIL:-0.32}"
+      - --max-num-seqs
+      - "${MORPHEUS_MAX_SEQS:-8}"
+    healthcheck:
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 300s
+    networks: [morpheus-net]
+
+  morpheus-tts:
+    build: ./tts
+    image: morpheus-tts:local
+    container_name: morpheus-tts
+    restart: unless-stopped
+    depends_on:
+      vllm-morpheus:
+        condition: service_healthy
+    environment:
+      - VLLM_URL=http://vllm-morpheus:8000/v1/completions
+      - VLLM_MODEL=morpheus
+      - SNAC_DEVICE=cpu
+      - MORPHEUS_DEFAULT_VOICE=${MORPHEUS_DEFAULT_VOICE:-baddy}
+      - MORPHEUS_VOICES=${MORPHEUS_VOICES:-baddy}
+    volumes:
+      - ${MORPHEUS_MODEL_DIR:-/home/lkraven/morpheus/models/mOrpheus}:/model:ro
+      - ${SNAC_DIR:-/home/lkraven/morpheus/models/snac_24khz}:/snac:ro
+    ports:
+      - "${MORPHEUS_TTS_PORT:-8299}:8000"
+    healthcheck:
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 60s
+    labels:
+      - homepage.group=AI Systems
+      - homepage.name=mOrpheus TTS
+      - homepage.icon=mdi-account-voice
+      - homepage.description=Uncensored Orpheus TTS (baddy + zero-shot clone), real-time, irv-ml1
+      - homepage.href=http://10.100.79.3:${MORPHEUS_TTS_PORT:-8299}/docs
+    networks: [morpheus-net]
+
+networks:
+  morpheus-net:
+    name: morpheus-net