Enable GPU: LLM + Whisper on the RTX 5050, pick qwen3:8b

GPU acceleration is now on by default and verified end-to-end on the
Blackwell RTX 5050 (sm_120):

- Ollama offloads 100% to GPU (log: library=CUDA compute=12.0,
  BLACKWELL_NATIVE_FP4=1). compose passes GPU via CDI
  (devices: nvidia.com/gpu=all) to both ollama and javis.
- Whisper STT on GPU: faster-whisper>=1.1.0 + nvidia-cublas/cudnn cu12,
  LD_LIBRARY_PATH baked into the image. Verified float16 transcribe on
  sm_120; bridge auto-falls back to CPU when no GPU is present.
- Model: default chat model -> qwen3:8b (best 8GB-VRAM tool-calling,
  ~5GB Q4). Embed stays nomic-embed-text.
- README documents the host one-time setup (nvidia-container-toolkit +
  `nvidia-ctk cdi generate`) and GPU on/off.

Verified: image builds; GPU visible in both containers via compose;
ollama ps = 100% GPU; faster-whisper cuda OK + CPU fallback OK;
bridge /health 200.

.env.example

@@ -20,9 +20,10 @@ BRIDGE_HOST=127.0.0.1
 BRIDGE_PORT=8765
 JARVIS_BRAIN_ENABLED=1
 JARVIS_TTS_ENABLED=1
-# faster-whisper device/compute. On this RTX 5050 box: cuda / float16.
-WHISPER_DEVICE=auto
-WHISPER_COMPUTE_TYPE=auto
+# faster-whisper device/compute. GPU by default (RTX 5050 / sm_120, verified).
+# Falls back to CPU automatically if no GPU is passed to the container.
+WHISPER_DEVICE=cuda
+WHISPER_COMPUTE_TYPE=float16
 # Optional explicit Piper voice model (.onnx). If empty, the jarvis default is used.
 TTS_PIPER_MODEL_PATH=
 
@@ -32,7 +33,8 @@ TTS_PIPER_MODEL_PATH=
 # ---------------------------------------------------------------------------
 # In docker-compose this is overridden to http://ollama:11434 automatically.
 OLLAMA_BASE_URL=http://127.0.0.1:11434
-OLLAMA_CHAT_MODEL=llama3.1:8b
+# qwen3:8b — best 8GB-VRAM pick: strongest tool-calling, ~5GB Q4, fits the RTX 5050.
+OLLAMA_CHAT_MODEL=qwen3:8b
 OLLAMA_EMBED_MODEL=nomic-embed-text
 WHISPER_MODEL=small
 

Dockerfile

@@ -8,7 +8,10 @@ FROM ubuntu:24.04
 ENV DEBIAN_FRONTEND=noninteractive \
     LANG=C.UTF-8 \
     DISPLAY=:1 \
-    PATH=/opt/venv/bin:/root/.bun/bin:/usr/local/bin:/usr/bin:/bin
+    PATH=/opt/venv/bin:/root/.bun/bin:/usr/local/bin:/usr/bin:/bin \
+    NVIDIA_VISIBLE_DEVICES=all \
+    NVIDIA_DRIVER_CAPABILITIES=compute,utility \
+    LD_LIBRARY_PATH=/opt/venv/lib/python3.12/site-packages/nvidia/cublas/lib:/opt/venv/lib/python3.12/site-packages/nvidia/cudnn/lib
 
 # --- System packages: desktop, VNC, Chrome deps, ffmpeg, python, ocr ---
 RUN apt-get update && apt-get install -y --no-install-recommends \

README.md

@@ -75,7 +75,24 @@ docker compose up -d          # 봇이 시작되고 /자비스 명령 등록
 
 디스코드에서 `/자비스 join` 으로 호출하세요. (`OLLAMA_CHAT_MODEL` 등 모델을 바꾸려면 `.env`에서 지정 후 `docker compose up -d`.)
 
-- GPU(RTX 5050) 가속: 호스트에 nvidia-container-toolkit 설치 후 `docker-compose.yml`의 GPU 블록 주석 해제, `.env`에서 `WHISPER_DEVICE=cuda` / `WHISPER_COMPUTE_TYPE=float16`.
+### GPU 가속 (기본 ON)
+
+LLM(Ollama)과 Whisper STT가 **기본적으로 GPU(RTX 5050, Blackwell sm_120)** 에서 돕니다. 검증 완료: Ollama 100% GPU 오프로드, faster-whisper float16 GPU 동작.
+
+호스트 사전 준비(1회):
+
+```bash
+# nvidia-container-toolkit 설치 후 CDI 스펙 생성 (Docker 29 CDI 방식, 데몬 재시작 불필요)
+sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
+docker run --rm --device nvidia.com/gpu=all ubuntu nvidia-smi -L   # GPU 보이면 OK
+```
+
+`docker-compose.yml`은 두 컨테이너에 `devices: ["nvidia.com/gpu=all"]`(CDI)로 GPU를 넣습니다.
+
+- 모델: 기본 `qwen3:8b` — 8GB VRAM에서 도구호출(tool calling)이 가장 안정적이고 ~5GB(Q4)로 잘 맞습니다. 더 가볍게/무겁게 쓰려면 `.env`의 `OLLAMA_CHAT_MODEL` 변경.
+- Whisper는 `WHISPER_DEVICE=cuda`/`float16` 기본. **GPU가 없으면 자동으로 CPU로 폴백**하므로 안전합니다.
+- GPU가 아예 없는 호스트라면 `docker-compose.yml`의 두 `devices:` 블록을 지우고 `.env`에 `WHISPER_DEVICE=cpu`를 두면 됩니다.
+
 - 데이터(메모리 DB), Whisper 캐시, Piper 음성은 named volume에 영속됩니다.
 - 셀프봇 영상 송출 의존성은 이미지에 기본 포함하지 않습니다. 쓰려면 컨테이너에서 `cd /app/bot && bun add discord.js-selfbot-v13 @dank074/discord-video-stream` 후 재시작(또는 Dockerfile에 추가).
 

bridge/requirements-bridge.txt

@@ -5,12 +5,19 @@
 
 # --- Brain runtime (imported when the reply engine loads) ---
 python-dotenv==1.0.1
-faster-whisper==1.0.3
+# >=1.1.0 pulls a ctranslate2 with Blackwell (sm_120) CUDA kernels.
+faster-whisper>=1.1.0
 mcp==1.13.1
 numpy<2.0.0
 rapidfuzz==3.6.1
 requests==2.32.3
 
+# --- CUDA libraries for GPU-accelerated Whisper (RTX 5050 / sm_120) ---
+# ctranslate2 dlopens these at transcribe time; LD_LIBRARY_PATH is set in the
+# Dockerfile to point at them. Verified working on Blackwell sm_120.
+nvidia-cublas-cu12
+nvidia-cudnn-cu12
+
 # --- Bridge HTTP service ---
 flask>=3.0.0
 

bridge/server.py

@@ -90,7 +90,16 @@ def _ensure_brain():
             )
             device = os.environ.get("WHISPER_DEVICE", "auto")
             compute = os.environ.get("WHISPER_COMPUTE_TYPE", "auto")
-            whisper = WhisperModel(cfg.whisper_model, device=device, compute_type=compute)
+            try:
+                whisper = WhisperModel(cfg.whisper_model, device=device, compute_type=compute)
+            except Exception as ge:
+                # GPU not available / unsupported -> fall back to CPU so the
+                # bridge still works without a GPU passed to the container.
+                if device != "cpu":
+                    print(f"[bridge] whisper device='{device}' failed ({ge}); falling back to CPU", flush=True)
+                    whisper = WhisperModel(cfg.whisper_model, device="cpu", compute_type="int8")
+                else:
+                    raise
 
             _cfg, _db, _dialogue_memory, _whisper = cfg, db, dialogue_memory, whisper
             print(f"[bridge] brain ready (chat={cfg.ollama_chat_model}, whisper={cfg.whisper_model})", flush=True)

docker-compose.yml

@@ -19,14 +19,10 @@ services:
     restart: unless-stopped
     volumes:
       - ollama_models:/root/.ollama
-    # --- GPU (optional): needs nvidia-container-toolkit on the host ---
-    # deploy:
-    #   resources:
-    #     reservations:
-    #       devices:
-    #         - driver: nvidia
-    #           count: all
-    #           capabilities: [gpu]
+    # GPU: needs nvidia-container-toolkit on the host (CDI). Verified on the
+    # RTX 5050 (Blackwell sm_120) — Ollama offloads 100% to GPU.
+    devices:
+      - "nvidia.com/gpu=all"
 
   # Auto-pull the models the brain needs, then exit. Idempotent (re-runnable).
   ollama-init:
@@ -36,7 +32,7 @@ services:
     restart: "no"
     environment:
       OLLAMA_HOST: http://ollama:11434
-      CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-llama3.1:8b}
+      CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-qwen3:8b}
       EMBED_MODEL: ${OLLAMA_EMBED_MODEL:-nomic-embed-text}
     entrypoint: ["/bin/sh", "-c"]
     command:
@@ -58,12 +54,18 @@ services:
     environment:
       # Point the brain at the ollama service and the bot at the in-container bridge.
       OLLAMA_BASE_URL: http://ollama:11434
-      OLLAMA_CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-llama3.1:8b}
+      OLLAMA_CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-qwen3:8b}
       OLLAMA_EMBED_MODEL: ${OLLAMA_EMBED_MODEL:-nomic-embed-text}
       WHISPER_MODEL: ${WHISPER_MODEL:-small}
+      WHISPER_DEVICE: ${WHISPER_DEVICE:-cuda}
+      WHISPER_COMPUTE_TYPE: ${WHISPER_COMPUTE_TYPE:-float16}
       BRIDGE_URL: http://127.0.0.1:8765
     depends_on:
       - ollama
+    # GPU: accelerates Whisper STT (and anything else CUDA) in this container.
+    # Verified: faster-whisper float16 works on the RTX 5050 (sm_120).
+    devices:
+      - "nvidia.com/gpu=all"
     shm_size: "1gb"          # Chrome needs a larger /dev/shm
     ports:
       # Host ports are overridable. If the HOST already runs VNC on 5901
@@ -75,7 +77,6 @@ services:
       - javis_data:/data                         # jarvis db + memory
       - whisper_cache:/root/.cache/huggingface   # cached Whisper models
       - piper_voices:/opt/piper-voices           # TTS voices
-    # --- GPU (optional): mirror the ollama GPU block above to accelerate Whisper ---
 
 volumes:
   ollama_models:

docker/entrypoint.sh

@@ -8,11 +8,11 @@ set -euo pipefail
 : "${VNC_PASSWORD:=javis123}"
 : "${VNC_RESOLUTION:=1920x1080}"
 : "${OLLAMA_BASE_URL:=http://ollama:11434}"
-: "${OLLAMA_CHAT_MODEL:=llama3.1:8b}"
+: "${OLLAMA_CHAT_MODEL:=qwen3:8b}"
 : "${OLLAMA_EMBED_MODEL:=nomic-embed-text}"
 : "${WHISPER_MODEL:=small}"
-: "${WHISPER_DEVICE:=cpu}"
-: "${WHISPER_COMPUTE_TYPE:=int8}"
+: "${WHISPER_DEVICE:=cuda}"
+: "${WHISPER_COMPUTE_TYPE:=float16}"
 : "${JARVIS_DB_PATH:=/data/jarvis.db}"
 : "${BRIDGE_HOST:=0.0.0.0}"
 : "${BRIDGE_PORT:=8765}"