Public Access

Files

T

Folzo 97348e9a12 Add ComfyUI models documentation and download script

- Complete MODELS.md with detailed model descriptions
- Interactive download script for 5+ specialized SDXL models
- Updated README with model selection guide
- Fix docker-compose volumes path (/root/ComfyUI)
- API workflows for OpenWebUI integration

Models documented:
- RealVisXL (portraits)
- Juggernaut XL (versatile)
- DreamShaper (art/fantasy)
- SDXL Lightning (fast)
- Proteus (realism)

2026-04-06 01:05:18 +02:00

227 KiB

Raw Permalink Blame History

effi@effi-server:~/effi$ docker logs open-webui --tail 20 -f

2026-02-21 16:41:16.787 | INFO 2026-02-21 16:41:16.789 | INFO 2026-02-21 16:41:16.841 | INFO 2026-02-21 16:41:16.842 | INFO 2026-02-21 16:41:16.844 | INFO 2026-02-21 16:41:16.844 | INFO 2026-02-21 16:41:16.875 | INFO 2026-02-21 16:41:16.875 | INFO 2026-02-21 16:41:16.893 | INFO 2026-02-21 16:41:16.909 | INFO 2026-02-21 16:41:16.994 | INFO 2026-02-21 16:41:30.169 | INFO 2026-02-21 16:41:30.191 | ERROR 2026-02-21 16:41:30.194 | INFO 2026-02-21 16:41:30.207 | ERROR 2026-02-21 16:41:30.209 | INFO 2026-02-21 16:41:39.093 | INFO 2026-02-21 16:42:39.091 | INFO 2026-02-21 16:43:39.094 | INFO 2026-02-21 16:44:39.094 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/admin/config HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/admin/config/ldap/server HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/groups/ HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/webhook HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/admin/config HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/admin/config/ldap/server HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/groups/ HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/admin/config/ldap HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/admin/config/ldap HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/version/updates HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/version/updates HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/audio/config HTTP/1.1" 200 | open_webui.routers.audio:get_available_voices:1304 - Error fetching voices from custom endpoint: 404 Client Error: Not Found for url: http://192.168.1.27:8100/v1/audio/voices | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/audio/voices HTTP/1.1" 200 | open_webui.routers.audio:get_available_models:1256 - Error fetching models from custom endpoint: 404 Client Error: Not Found for url: http://192.168.1.27:8100/v1/audio/models | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/audio/models HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200

L'URL est déjà en http:// correct (pas de Https). Le vrai problème est visible dans les logs :

404 Not Found for url: http://192.168.1.27:8100/v1/audio/voices
404 Not Found for url: http://192.168.1.27:8100/v1/audio/models

openedai-speech ne supporte pas ces endpoints — Open WebUI les cherche mais ils n'existent pas dans openedai-speech. Ce n'est pas bloquant pour la synthèse elle-même, mais ça explique pourquoi Open WebUI ne trouve pas les voix disponibles.

Le vrai problème : voix `nova` inexistante

openedai-speech utilise des voix locales (piper/kokoro), pas les voix OpenAI. nova n'existe probablement pas. Vérifie les voix disponibles :

ls ~/effi/tts-voices/

Et teste avec une voix qui existe vraiment :

curl -s http://192.168.1.27:8100/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","input":"Bonjour je suis Effi","voice":"alloy"}' \
  --output /tmp/test_alloy.mp3 && \
  echo "$(du -h /tmp/test_alloy.mp3)"

Les voix par défaut d'openedai-speech sont : alloy, echo, fable, onyx, nova, shimmer — mais elles sont mappées vers des modèles locaux. Si le dossier tts-voices est vide, aucune voix ne fonctionnera.

# Vérifie aussi la config
ls ~/effi/tts-config/
cat ~/effi/tts-config/voice_to_speaker.yaml 2>/dev/null || echo "Pas de config"

^Ceffi@effi-server:~/effils ~/effi/tts-voices/s/

alloy.wav en_GB-northern_english_male-medium.onnx.json fable.wav onyx.wav echo.wav en_US-libritts_r-medium.onnx hub shimmer.wav en_GB-northern_english_male-medium.onnx en_US-libritts_r-medium.onnx.json nova.wav voices.json effi@effi-server:/effi$ curl -s http://192.168.1.27:8100/v1/audio/speech \ -H "Content-Type: application/json" \ -d '{"model":"tts-1","input":"Bonjour je suis Effi","voice":"alloy"}' \ --output /tmp/test_alloy.mp3 && \ echo "$(du -h /tmp/test_alloy.mp3)" 12K /tmp/test_alloy.mp3 effi@effi-server:/effi$ # Vérifie aussi la config ls ~/effi/tts-config/ cat /effi/tts-config/voice_to_speaker.yaml 2>/dev/null || echo "Pas de config" pre_process_map.yaml voice_to_speaker.yaml tts-1: some_other_voice_name_you_want: model: voices/choose your own model.onnx speaker: set your own speaker alloy: model: voices/en_US-libritts_r-medium.onnx speaker: 79 # 64, 79, 80, 101, 130 echo: model: voices/en_US-libritts_r-medium.onnx speaker: 134 # 52, 102, 134 echo-alt: model: voices/en_US-ryan-high.onnx speaker: # default speaker fable: model: voices/en_GB-northern_english_male-medium.onnx speaker: # default speaker onyx: model: voices/en_US-libritts_r-medium.onnx speaker: 159 # 55, 90, 132, 136, 137, 159 nova: model: voices/en_US-libritts_r-medium.onnx speaker: 107 # 57, 61, 107, 150, 162 shimmer: model: voices/en_US-libritts_r-medium.onnx speaker: 163 tts-1-hd: alloy-alt: model: xtts speaker: voices/alloy-alt.wav alloy: model: xtts speaker: voices/alloy.wav echo: model: xtts speaker: voices/echo.wav fable: model: xtts speaker: voices/fable.wav onyx: model: xtts speaker: voices/onyx.wav nova: model: xtts speaker: voices/nova.wav shimmer: model: xtts speaker: voices/shimmer.wav me: model: xtts_v2.0.2 # you can specify an older xtts version speaker: voices/me.wav # this could be you language: auto enable_text_splitting: True length_penalty: 1.0 repetition_penalty: 10 speed: 1.0 temperature: 0.75 top_k: 50 top_p: 0.85 comment: You can add a comment here also, which will be persistent and otherwise ignored.effi@effi-server:/effi$

Les voix sont bien configurées et le TTS génère des fichiers MP3. Le problème vient d'ailleurs — le fichier MP3 généré par effi-server ne peut pas être lu par ton navigateur car le J5040 utilise piper (voix anglaises uniquement), et nova parle anglais même si le texte est en français.

Mais le vrai souci pour la communication vocale complète : vérifie si Open WebUI lit bien le MP3 en retour. Télécharge le fichier test et écoute-le :

# Depuis effi-server, génère un test en français
curl -s http://192.168.1.27:8100/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1-hd","input":"Bonjour, je suis Effi, ton assistante personnelle !","voice":"nova"}' \
  --output /tmp/test_nova_hd.mp3 && \
  echo "$(du -h /tmp/test_nova_hd.mp3)"

tts-1-hd avec nova utilise XTTS qui supporte le français nativement, contrairement à tts-1 qui utilise piper (anglais seulement).

Dans Open WebUI, change le modèle TTS de tts-1 → tts-1-hd en gardant la voix nova. XTTS va cloner la voix du fichier nova.wav et parler en français avec cette voix.

Admin Panel → Settings → Audio → Text-to-Speech :

Champ	Valeur
Model	`tts-1-hd`
Voice	`nova`

j'ai changé pour tts-1-hd, voici les logs :

Injected system message: The following is a list of stored memories, which are information related to the user, but are not part of this conversation. Consider the information within these memories as factual and use them to inform your responses, but do not bring up information from the memory bank unless it is directly related to the current conversation. Do not make up memories if the provided memory bank is empty. This is the list of memories: [] 2026-02-21 16:47:34.124 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "POST /api/chat/completions HTTP/1.1" 200 2026-02-21 16:47:34.147 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/?page=1 HTTP/1.1" 200 2026-02-21 16:47:39.094 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 2026-02-21 16:47:43.784 | INFO | function_auto_memory:log:652 - outlet invoked 2026-02-21 16:47:43.793 | INFO | function_auto_memory:log:652 - no related memories found 2026-02-21 16:47:43.793 | INFO | function_auto_memory:log:652 - found 0 related memories before filtering 2026-02-21 16:47:43.793 | INFO | function_auto_memory:log:652 - using 0 related memories 2026-02-21 16:47:43.798 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "POST /api/chat/completed HTTP/1.1" 200 2026-02-21 16:47:43.825 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "POST /api/v1/chats/9738a8fc-f46a-4afb-bb09-d54b39e767e3 HTTP/1.1" 200 2026-02-21 16:47:43.838 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/?page=1 HTTP/1.1" 200 2026-02-21 16:48:01.616 | INFO | httpx._client:_send_single_request:1025 - HTTP Request: POST http://192.168.1.27:3011/v1/chat/completions "HTTP/1.1 200 OK" 2026-02-21 16:48:01.618 | INFO | function_auto_memory:log:652 - no changes 2026-02-21 16:48:10.792 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/?page=1 HTTP/1.1" 200 2026-02-21 16:48:11.430 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/9738a8fc-f46a-4afb-bb09-d54b39e767e3 HTTP/1.1" 200 2026-02-21 16:48:11.447 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/all/tags HTTP/1.1" 200 2026-02-21 16:48:39.088 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200

Les logs montrent que le LLM répond bien (200 OK sur /v1/chat/completions). Le pipeline complet fonctionne — il manque juste la requête TTS dans les logs. Déclenche un test vocal et surveille :

docker logs open-webui -f | grep -i "tts\|audio\|speech\|8100"

Parle à Effi pendant que ça tourne. Si aucune ligne TTS n'apparaît, c'est que Open WebUI n'envoie pas la réponse au TTS — ce qui arrive quand le mode vocal n'est pas activé correctement dans l'interface.

Vérifie le mode vocal dans l'UI

Dans l'interface chat d'Open WebUI, le bouton micro a deux modes :

🎤 Micro seul → STT uniquement, la réponse s'affiche en texte mais n'est pas lue
🎤➕🔊 Mode vocal complet → STT + réponse lue automatiquement en TTS

Il faut activer le mode "Call" (l'icône téléphone 📞 dans certaines versions d'Open WebUI) qui active le pipeline vocal complet STT→LLM→TTS. Ce bouton est généralement à côté du micro dans la barre de saisie.

Est-ce que tu vois un bouton 📞 ou un bouton avec une icône de casque/haut-parleur dans l'interface ?

voici les logs

Les logs racontent toute l'histoire clairement. À la fin, tout fonctionne :

16:36:22 POST /v1/audio/transcriptions HTTP 200  ✅ STT OK
16:36:34 POST /v1/audio/speech HTTP 200           ✅ TTS OK
...
16:47:33 POST /v1/audio/speech HTTP 200           ✅ TTS OK

Le pipeline complet STT → LLM → TTS est fonctionnel !¹

Les erreurs du début (Connection refused sur 192.168.1.11:8200) correspondent aux moments où le conteneur speaches était down ou pas encore prêt — c'est résolu maintenant.¹

Ce qui reste à ignorer (non bloquant)

404 Not Found for url http://192.168.1.27:8100/v1/audio/voices — openedai-speech ne supporte pas cet endpoint, Open WebUI tombe en fallback sur les voix statiques. Pas grave.¹
InsecureKeyLengthWarning — WEBUI_SECRET_KEY trop court (31 bytes), pas bloquant mais tu peux allonger la clé dans ton .env si tu veux.

Bilan final du projet Effi — Session d'aujourd'hui

Composant	Statut
speaches (STT large-v3) sur PC de jeu	✅ Opérationnel
Open WebUI → speaches `192.168.1.11:8200`	✅ Connecté
openedai-speech (TTS nova) sur effi-server	✅ Opérationnel
Pipeline vocal complet STT→LLM→TTS	✅ Fonctionnel

Effi parle et écoute ! 🎉

⁂

tout marche mais ce n'est pas assez rapide il faut utiliser la carte graphique avec rocm

Le problème est clair : CTranslate2 ne supporte pas ROCm nativement — il faut utiliser le fork arlo-phoenix/CTranslate2-rocm ou ROCm/CTranslate2 et rebuilder l'image speaches avec cette version.²³

Plan : Rebuilder speaches avec CTranslate2-ROCm

La solution est de créer un Dockerfile custom qui :

Part de l'image ROCm AMD officielle
Compile CTranslate2-rocm pour gfx1101 (RX 7800 XT)
Installe speaches par-dessus

cd ~/speaches/src
nano Dockerfile.rocm-ct2

Colle ce contenu :

FROM rocm/dev-ubuntu-24.04:6.3.4-complete

# ── Dépendances build ──────────────────────────────────────
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake python3 python3-pip python3-venv \
    clang libhipblas-dev libmiopengemm-dev \
    && rm -rf /var/lib/apt/lists/*

# ── CTranslate2-rocm compilé pour gfx1101 ─────────────────
RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2 && \
    cd /ct2 && \
    CLANG_CMAKE_CXX_COMPILER=clang++ \
    CXX=clang++ \
    HIPCXX="$(hipconfig -l)/clang" \
    HIP_PATH="$(hipconfig -R)" \
    PYTORCH_ROCM_ARCH=gfx1101 \
    cmake -S . -B build \
        -DWITH_MKL=OFF \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON && \
    cmake --build build -- -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig && \
    cd python && \
    pip3 install -r install_requirements.txt && \
    python3 setup.py bdist_wheel && \
    pip3 install dist/*.whl

# ── Speaches ───────────────────────────────────────────────
COPY --from=ghcr.io/astral-sh/uv:latest /uv /bin/uv
WORKDIR /home/ubuntu/speaches
COPY . .
RUN uv sync --frozen --compile-bytecode --no-dev --no-editable

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0
ENV LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

EXPOSE 8000
CMD [".venv/bin/uvicorn", "speaches.main:create_app", "--host", "0.0.0.0", "--port", "8000", "--factory"]

Lance le build (⚠️ 30-60 minutes — compilation C++ pour gfx1101) :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Puis mets à jour docker-compose.yml :

    image: speaches:rocm-ct2
    environment:
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      - WHISPER__MODEL=Systran/faster-whisper-large-v3
      - WHISPER__LANGUAGE=fr
      - WHISPER__COMPUTE_TYPE=float16   # ← repasse à float16 avec GPU
      - WHISPER__DEVICE=cuda

Note : Le build est long car il compile CTranslate2 depuis les sources avec le compilateur HIP spécifiquement pour l'architecture gfx1101 de ta RX 7800 XT. Lance-le et dis-moi si tu as des erreurs de compilation.³ ⁴⁵⁶⁷⁸⁹¹⁰¹¹¹²¹³¹⁴¹⁵¹⁶

⁂

avant d'aller plus loins voici les infos de rocm :

folzo@folzo-Corp:~/speaches$ rocm-smi

========================================= ROCm System Management Interface ========================================= =================================================== Concise Info =================================================== Device Node IDs Temp Power Partitions SCLK MCLK Fan Perf PwrCap VRAM% GPU% (DID, GUID) (Edge) (Avg) (Mem, Compute, ID)

0 1 0x747e, 60998 51.0°C 50.0W N/A, N/A, 0 2737Mhz 96Mhz 14.9% auto 212.0W 72% 100%

=============================================== End of ROCm SMI Log ================================================ folzo@folzo-Corp:~/speaches$

Excellente nouvelle — ROCm fonctionne parfaitement sur le host ! Ta RX 7800 XT est détectée (0x747e = Navi32, gfx1101), 16GB VRAM dont 72% utilisés (probablement par Ollama), GPU à 100% = Ollama qui tourne.¹⁷

Le vrai problème

ROCm fonctionne sur le host, mais pas dans le conteneur Docker. Il faut vérifier :

docker exec speaches-stt rocm-smi

Si ça retourne une erreur, c'est que les devices ne sont pas bien passés au conteneur. Vérifie :

docker exec speaches-stt ls /dev/kfd
docker exec speaches-stt ls /dev/dri/

Et teste si CTranslate2 voit le GPU depuis l'intérieur :

docker exec speaches-stt python3 -c "
import ctranslate2
print('Version CT2:', ctranslate2.__version__)
print('CUDA/ROCm compute types:', ctranslate2.get_supported_compute_types('cuda'))
"

Ce que je suspecte

CTranslate2 compilé en CUDA cherche libcublasLt.so.12 (NVIDIA). La version dans l'image speaches buildée est la version pip standard sans support ROCm. La solution la plus rapide avant un rebuild complet : remplacer CTranslate2 par la wheel ROCm dans le conteneur existant.

docker exec -it speaches-stt bash -c "
pip3 install ctranslate2 --index-url https://download.pytorch.org/whl/rocm6.2
"

Dis-moi d'abord ce que retournent les commandes de diagnostic — selon les résultats on choisit l'approche la plus propre.

⁂

folzo@folzo-Corp:~/speaches$ docker exec speaches-stt rocm-smi

========================================== ROCm System Management Interface ========================================== ==================================================== Concise Info ==================================================== Device Node IDs Temp Power Partitions SCLK MCLK Fan Perf PwrCap VRAM% GPU% (DID, GUID) (Edge) (Avg) (Mem, Compute, ID)

0 1 0x747e, 60998 50.0°C 56.0W N/A, N/A, 0 2734Mhz 456Mhz 14.9% auto 212.0W 72% 100%

================================================ End of ROCm SMI Log ================================================= folzo@folzo-Corp:~~/speaches$ docker exec speaches-stt ls /dev/kfd docker exec speaches-stt ls /dev/dri/ /dev/kfd card1 renderD128 folzo@folzo-Corp:~~/speaches$ docker exec speaches-stt python3 -c " import ctranslate2 print('Version CT2:', ctranslate2.version) print('CUDA/ROCm compute types:', ctranslate2.get_supported_compute_types('cuda')) " Traceback (most recent call last):

File "<string>", line 4, in <module>

RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version Version CT2: 4.5.0 folzo@folzo-Corp:~/speaches$

Parfait diagnostic. ROCm est visible dans le conteneur (/dev/kfd, /dev/dri/renderD128 présents, rocm-smi fonctionne), mais CTranslate2 4.5.0 est la version CUDA officielle qui cherche les drivers NVIDIA — pas une version ROCm.¹⁸

La solution rapide : installer la wheel CTranslate2 compilée pour ROCm directement dans le venv speaches.

docker exec -it speaches-stt bash

Une fois dans le conteneur :

# Vérifie où est le venv
source /home/ubuntu/speaches/.venv/bin/activate

# Désinstalle la version CUDA
pip uninstall ctranslate2 -y

# Installe la version ROCm (compilée par arlo-phoenix pour ROCm 6.x)
pip install ctranslate2 --extra-index-url https://wheel.arlo-phoenix.dev/rocm6/

# Vérifie
python3 -c "
import ctranslate2
print('Version:', ctranslate2.__version__)
print('ROCm compute types:', ctranslate2.get_supported_compute_types('cuda'))
"

Si le repo arlo-phoenix n'est pas disponible, alternative via le blog AMD officiel :¹⁸

pip install ctranslate2-rocm

Dis-moi ce que retourne le pip install — selon la disponibilité du package on ajustera. Si ça marche, on sortira du conteneur, on commitera l'image modifiée, et on remettra WHISPER__DEVICE=cuda + WHISPER__COMPUTE_TYPE=float16 dans le compose.

⁂

folzo@folzo-Corp:~/speaches$ docker exec -it speaches-stt bash

To run a command as administrator (user "root"), use "sudo ". See "man sudo_root" for details.

ubuntu@f268afe88f2b:~/speaches$ pip uninstall ctranslate2 -y error: externally-managed-environment

× This environment is externally managed ╰─> To install Python packages system-wide, try apt install python3-xyz, where xyz is the package you are trying to install.

If you wish to install a non-Debian-packaged Python package,
create a virtual environment using python3 -m venv path/to/venv.
Then use path/to/venv/bin/python and path/to/venv/bin/pip. Make
sure you have python3-full installed.

If you wish to install a non-Debian packaged Python application,
it may be easiest to use pipx install xyz, which will manage a
virtual environment for you. Make sure you have pipx installed.

See /usr/share/doc/python3.12/README.venv for more information.
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.

hint: See PEP 668 for the detailed specification. ubuntu@f268afe88f2b:~/speaches$

Je dois lancer en root ?

Non, pas besoin de root — speaches utilise un venv. Il faut activer le venv d'abord :

source /home/ubuntu/speaches/.venv/bin/activate

Tu devrais voir (.venv) apparaître dans le prompt. Ensuite :

# Désinstalle la version CUDA
pip uninstall ctranslate2 -y

# Vérifie la version ROCm disponible sur PyPI
pip index versions ctranslate2 2>/dev/null | head -3

Pour l'installation ROCm, essaie d'abord le package du blog AMD officiel :¹⁹

pip install ctranslate2-rocm

Si ça échoue, essaie le fork arlo-phoenix qui est la référence communautaire pour CTranslate2+ROCm :²⁰

pip install ctranslate2 --extra-index-url https://wheel.arlo-phoenix.dev/

Et si les deux échouent, on peut forcer la réinstallation depuis les wheels ROCm PyTorch :

pip install torch --index-url https://download.pytorch.org/whl/rocm6.2 && \
pip install ctranslate2

Dis-moi ce que retourne chaque commande au fur et à mesure.

⁂

il y a une erreur ? le PIP ne semble pas marcher :

folzo@folzo-Corp:~/speaches$ docker exec -it speaches-stt bash To run a command as administrator (user "root"), use "sudo ". See "man sudo_root" for details.

ubuntu@f268afe88f2b:~~/speaches$ source /home/ubuntu/speaches/.venv/bin/activate (speaches) ubuntu@f268afe88f2b:~~/speaches$ pip uninstall ctranslate2 -y error: externally-managed-environment

× This environment is externally managed ╰─> To install Python packages system-wide, try apt install python3-xyz, where xyz is the package you are trying to install.

If you wish to install a non-Debian-packaged Python package,
create a virtual environment using python3 -m venv path/to/venv.
Then use path/to/venv/bin/python and path/to/venv/bin/pip. Make
sure you have python3-full installed.

If you wish to install a non-Debian packaged Python application,
it may be easiest to use pipx install xyz, which will manage a
virtual environment for you. Make sure you have pipx installed.

See /usr/share/doc/python3.12/README.venv for more information.
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.

hint: See PEP 668 for the detailed specification. (speaches) ubuntu@f268afe88f2b:/speaches$ pip index versions ctranslate2 2>/dev/null | head -3 ctranslate2 (4.7.1) Available versions: 4.7.1, 4.7.0, 4.6.3, 4.6.2, 4.6.1, 4.6.0, 4.5.0, 4.4.0, 4.3.1, 4.2.1, 4.1.0, 4.0.0, 3.24.0, 3.23.0, 3.22.0, 3.21.0, 3.20.0, 3.19.0 (speaches) ubuntu@f268afe88f2b:/speaches$ pip install ctranslate2-rocm error: externally-managed-environment

× This environment is externally managed ╰─> To install Python packages system-wide, try apt install python3-xyz, where xyz is the package you are trying to install.

If you wish to install a non-Debian-packaged Python package,
create a virtual environment using python3 -m venv path/to/venv.
Then use path/to/venv/bin/python and path/to/venv/bin/pip. Make
sure you have python3-full installed.

If you wish to install a non-Debian packaged Python application,
it may be easiest to use pipx install xyz, which will manage a
virtual environment for you. Make sure you have pipx installed.

See /usr/share/doc/python3.12/README.venv for more information.
note: If you believe this is a mistake, please contact your Python installation or OS distribution provider. You can override this, at the risk of breaking your Python installation or OS, by passing --break-system-packages.

hint: See PEP 668 for the detailed specification. (speaches) ubuntu@f268afe88f2b:~/speaches$

Le venv est activé ((speaches) visible) mais pip pointe encore sur le pip système. Il faut utiliser le pip du venv explicitement :

# Utilise le pip du venv, pas le pip système
/home/ubuntu/speaches/.venv/bin/pip uninstall ctranslate2 -y

# Puis installe la version ROCm
/home/ubuntu/speaches/.venv/bin/pip install ctranslate2-rocm

# Ou si ça échoue, force avec uv (le gestionnaire utilisé par speaches)
/bin/uv pip install ctranslate2-rocm --python /home/ubuntu/speaches/.venv

Vérifie quel pip est utilisé :

which pip
# Doit retourner /home/ubuntu/speaches/.venv/bin/pip
# Si c'est /usr/bin/pip → le venv n'est pas bien activé

Le problème vient du fait que speaches utilise uv pour gérer son venv, ce qui crée un environnement légèrement différent d'un venv standard. uv pip est la commande correcte :

/bin/uv pip uninstall ctranslate2 --python /home/ubuntu/speaches/.venv
/bin/uv pip install ctranslate2-rocm --python /home/ubuntu/speaches/.venv

# Vérifie
/home/ubuntu/speaches/.venv/bin/python3 -c "
import ctranslate2
print('Version:', ctranslate2.__version__)
print('ROCm types:', ctranslate2.get_supported_compute_types('cuda'))
"

speaches) ubuntu@f268afe88f2b:~/speaches$ /home/ubuntu/speaches/.venv/bin/pip uninstall ctranslate2 -y

bash: /home/ubuntu/speaches/.venv/bin/pip: No such file or directory (speaches) ubuntu@f268afe88f2b:~~/speaches$ /home/ubuntu/speaches/.venv/bin/pip install ctranslate2-rocm bash: /home/ubuntu/speaches/.venv/bin/pip: No such file or directory (speaches) ubuntu@f268afe88f2b:~~/speaches$

Speaches utilise uv qui ne crée pas de pip dans le venv. Utilise directement uv :

# Liste ce qui est installé pour confirmer ctranslate2
/bin/uv pip list --python /home/ubuntu/speaches/.venv | grep ctranslate

# Désinstalle et réinstalle avec uv
/bin/uv pip uninstall ctranslate2 --python /home/ubuntu/speaches/.venv
/bin/uv pip install ctranslate2-rocm --python /home/ubuntu/speaches/.venv

Si ctranslate2-rocm n'existe pas sur PyPI, essaie :

# Alternative : wheel arlo-phoenix
/bin/uv pip install ctranslate2 \
  --extra-index-url https://wheel.arlo-phoenix.dev/ \
  --python /home/ubuntu/speaches/.venv

Et vérifie avec le python du venv :

/home/ubuntu/speaches/.venv/bin/python3 -c "
import ctranslate2
print('Version:', ctranslate2.__version__)
print('ROCm:', ctranslate2.get_supported_compute_types('cuda'))
"

(speaches) ubuntu@f268afe88f2b:~/speaches$ /bin/uv pip list --python /home/ubuntu/speaches/.venv | grep ctranslate

ctranslate2 4.5.0 (speaches) ubuntu@f268afe88f2b:~/speaches$ /bin/uv pip uninstall ctranslate2 --python /home/ubuntu/speaches/.venv Uninstalled 1 package in 69ms

ctranslate2==4.5.0 (speaches) ubuntu@f268afe88f2b:/speaches$ /bin/uv pip install ctranslate2-rocm --python /home/ubuntu/speaches/.venv x No solution found when resolving dependencies: -> Because ctranslate2-rocm was not found in the package registry and you require ctranslate2-rocm, we can conclude that your requirements are unsatisfiable. (speaches) ubuntu@f268afe88f2b:~/speaches$ /bin/uv pip install ctranslate2 \\ --extra-index-url [https://wheel.arlo-phoenix.dev/](https://wheel.arlo-phoenix.dev/) \\ --python /home/ubuntu/speaches/.venv ⠇ Resolving dependencies... error: Failed to fetch: https://wheel.arlo-phoenix.dev/ctranslate2/` Caused by: Request failed after 3 retries Caused by: error sending request for url (https://wheel.arlo-phoenix.dev/ctranslate2/) Caused by: client error (Connect) Caused by: dns error Caused by: failed to lookup address information: Name does not resolve (speaches) ubuntu@f268afe88f2b:/speaches$

La compilation doit se faire dans un conteneur avec les bons outils. La méthode la plus propre est de créer un nouveau Dockerfile qui compile CTranslate2-rocm depuis les sources. Sors du conteneur et travaille depuis le host :

exit

Puis crée le Dockerfile custom :

nano ~/speaches/src/Dockerfile.rocm-ct2

FROM rocm/dev-ubuntu-24.04:6.3.4-complete

# ── Dépendances build
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev \
    && rm -rf /var/lib/apt/lists/*

# ── Compile CTranslate2-rocm pour gfx1101 (RX 7800 XT)
RUN git clone --recurse-submodules \
    https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    CLANG_CMAKE_CXX_COMPILER=clang++ \
    CXX=clang++ \
    HIPCXX="$(hipconfig -l)/clang" \
    HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build \
        -DWITH_MKL=OFF \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

# ── Build wheel Python CTranslate2-rocm
RUN cd /ct2/python && \
    pip3 install --break-system-packages -r install_requirements.txt && \
    python3 setup.py bdist_wheel && \
    cp dist/*.whl /ct2-wheel.whl

# ── Image finale speaches + CT2-rocm
FROM rocm/dev-ubuntu-24.04:6.3.4-complete

COPY --from=0 /ct2-wheel.whl /ct2-wheel.whl
COPY --from=0 /usr/local/lib/libctranslate2* /usr/local/lib/
RUN ldconfig

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    python3-pip ffmpeg && \
    rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:latest /uv /bin/uv

RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true
WORKDIR /home/ubuntu/speaches
COPY --chown=ubuntu . .

RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \
    /bin/uv pip install /ct2-wheel.whl --python .venv

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0
ENV LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

EXPOSE 8000
CMD [".venv/bin/uvicorn", "speaches.main:create_app", \
     "--host", "0.0.0.0", "--port", "8000", "--factory"]

Lance le build (⚠️ 30-60 min, compilation C++) :

cd ~/speaches/src
docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Pendant ce temps, mets à jour le docker-compose.yml pour être prêt :

nano ~/speaches/docker-compose.yml

Remplace image: speaches:rocm par image: speaches:rocm-ct2 et change les variables :

      - WHISPER__COMPUTE_TYPE=float16
      - WHISPER__DEVICE=cuda

j'ai modifier le fichier yml, mais la compilation n'a pas marché :

"folzo@folzo-Corp:~/speaches$ cd ~/speaches/src docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 29.6s (10/20) docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 1.97kB 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:latest 6.1s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.7s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => FROM ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b3c9efb4 0.9s => => resolve ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b3 0.0s => => sha256:45e14b7e1aeb9452ca198cb664ab6852c2c86b6e68c1e6625bc0055bc374ce7e 96B / 96B 0.2s => => sha256:8a75ce161c366ca2f853165747e1e64dba18144978962fb39a7a08b1a4bcbb6f 23.11MB / 23.11MB 0.5s => => extracting sha256:8a75ce161c366ca2f853165747e1e64dba18144978962fb39a7a08b1a4bcbb6f 0.2s => => extracting sha256:45e14b7e1aeb9452ca198cb664ab6852c2c86b6e68c1e6625bc0055bc374ce7e 0.0s => CACHED [stage-0 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => [internal] load build context 0.1s => => transferring context: 4.06kB 0.0s => [stage-0 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git cmake n 10.4s => [stage-0 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate2-rocm 12.5s => ERROR [stage-0 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 CLANG_CMAKE_CXX_COMPILER=clang++ 0.4s

[stage-0 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 CLANG_CMAKE_CXX_COMPILER=clang++ CXX=clang++ HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" cmake -S . -B build -DWITH_MKL=OFF -DWITH_HIP=ON -DCMAKE_HIP_ARCHITECTURES=gfx1101 -DBUILD_TESTS=OFF -DWITH_CUDNN=ON -GNinja && ninja -C build -j$(nproc) && cmake --install build --prefix /usr/local && ldconfig: 0.329 -- The C compiler identification is GNU 13.3.0 0.334 CMake Error at /usr/share/cmake-3.28/Modules/CMakeDetermineCXXCompiler.cmake:48 (message): 0.334 Could not find compiler set in environment variable CXX: 0.334 0.334 clang++. 0.334 0.334 Call Stack (most recent call first): 0.334 CMakeLists.txt:8 (project) 0.334 0.334 0.334 CMake Error: CMAKE_CXX_COMPILER not set, after EnableLanguage 0.334 -- Configuring incomplete, errors occurred!

2 warnings found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 50)
UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 56) Dockerfile.rocm-ct2:12

28

ERROR: failed to build: failed to solve: process "/bin/sh -c cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 CLANG_CMAKE_CXX_COMPILER=clang++ CXX=clang++ HIPCXX=\"$(hipconfig -l)/clang\" HIP_PATH=\"$(hipconfig -R)\" cmake -S . -B build -DWITH_MKL=OFF -DWITH_HIP=ON -DCMAKE_HIP_ARCHITECTURES=gfx1101 -DBUILD_TESTS=OFF -DWITH_CUDNN=ON -GNinja && ninja -C build -j$(nproc) && cmake --install build --prefix /usr/local && ldconfig" did not complete successfully: exit code: 1 folzo@folzo-Corp:~/speaches/src$ "

L'erreur est claire : clang++ n'est pas installé dans l'image ROCm. Il faut l'ajouter dans le apt-get install. Corrige le Dockerfile :

nano ~/speaches/src/Dockerfile.rocm-ct2

Remplace uniquement le premier RUN apt-get par :

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev \
    clang clang++ llvm \
    && rm -rf /var/lib/apt/lists/*

Et pour éviter le problème CXX=clang++ non trouvé, passe le chemin complet dans le RUN de cmake. Remplace le bloc cmake par :

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    CXX=/usr/bin/clang++ \
    CC=/usr/bin/clang \
    HIPCXX="$(hipconfig -l)/clang" \
    HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/usr/bin/clang++ \
        -DCMAKE_C_COMPILER=/usr/bin/clang \
        -DWITH_MKL=OFF \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

Puis relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

L'image ROCm de base contient hipconfig mais pas forcément clang++ dans le PATH par défaut — c'est un classique avec les images AMD.

toujours un probleme :

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 4.2s (8/20) docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.08kB 0.0s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.6s => [internal] load metadata for ghcr.io/astral-sh/uv:latest 0.4s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => CACHED FROM ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b 0.0s => => resolve ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b3 0.0s => CACHED [stage-0 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => ERROR [stage-0 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git cm 3.5s

[stage-0 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git cmake ninja-build python3-dev clang clang++ llvm && rm -rf /var/lib/apt/lists/*: 0.304 Get:1 http://archive.ubuntu.com/ubuntu noble InRelease [256 kB] 0.310 Get:2 https://repo.radeon.com/amdgpu/6.3.4/ubuntu noble InRelease [5465 B] 0.322 Get:3 https://repo.radeon.com/rocm/apt/6.3.4 noble InRelease [2605 B] 0.432 Get:4 http://security.ubuntu.com/ubuntu noble-security InRelease [126 kB] 0.475 Get:5 https://repo.radeon.com/amdgpu/6.3.4/ubuntu noble/main amd64 Packages [14.6 kB] 0.530 Get:6 http://archive.ubuntu.com/ubuntu noble-updates InRelease [126 kB] 0.549 Get:7 https://repo.radeon.com/rocm/apt/6.3.4 noble/main amd64 Packages [60.0 kB] 0.586 Get:8 http://archive.ubuntu.com/ubuntu noble-backports InRelease [126 kB] 0.642 Get:9 http://archive.ubuntu.com/ubuntu noble/main amd64 Packages [1808 kB] 0.813 Get:10 http://archive.ubuntu.com/ubuntu noble/universe amd64 Packages [19.3 MB] 0.891 Get:11 http://security.ubuntu.com/ubuntu noble-security/universe amd64 Packages [1207 kB] 1.123 Get:12 http://archive.ubuntu.com/ubuntu noble/multiverse amd64 Packages [331 kB] 1.124 Get:13 http://archive.ubuntu.com/ubuntu noble/restricted amd64 Packages [117 kB] 1.124 Get:14 http://archive.ubuntu.com/ubuntu noble-updates/main amd64 Packages [2240 kB] 1.131 Get:15 http://archive.ubuntu.com/ubuntu noble-updates/restricted amd64 Packages [3381 kB] 1.193 Get:16 http://archive.ubuntu.com/ubuntu noble-updates/multiverse amd64 Packages [38.1 kB] 1.193 Get:17 http://archive.ubuntu.com/ubuntu noble-updates/universe amd64 Packages [2016 kB] 1.201 Get:18 http://archive.ubuntu.com/ubuntu noble-backports/main amd64 Packages [49.5 kB] 1.225 Get:19 http://archive.ubuntu.com/ubuntu noble-backports/universe amd64 Packages [34.6 kB] 1.568 Get:20 http://security.ubuntu.com/ubuntu noble-security/main amd64 Packages [1857 kB] 1.793 Get:21 http://security.ubuntu.com/ubuntu noble-security/multiverse amd64 Packages [34.8 kB] 1.795 Get:22 http://security.ubuntu.com/ubuntu noble-security/restricted amd64 Packages [3196 kB] 2.077 Fetched 36.3 MB in 2s (19.3 MB/s) 2.077 Reading package lists... 2.621 W: https://repo.radeon.com/amdgpu/6.3.4/ubuntu/dists/noble/InRelease: Key is stored in legacy trusted.gpg keyring (/etc/apt/trusted.gpg), see the DEPRECATION section in apt-key(8) for details. 2.621 W: https://repo.radeon.com/rocm/apt/6.3.4/dists/noble/InRelease: Key is stored in legacy trusted.gpg keyring (/etc/apt/trusted.gpg), see the DEPRECATION section in apt-key(8) for details. 2.641 Reading package lists... 3.202 Building dependency tree... 3.282 Reading state information... 3.366 hipify-clang is already the newest version (18.0.0.60304-76~24.04). 3.366 hipify-clang set to manually installed. 3.366 Some packages could not be installed. This may mean that you have 3.366 requested an impossible situation or if you are using the unstable 3.366 distribution that some required packages have not yet been created 3.366 or been moved out of Incoming. 3.366 The following information may help to resolve the situation: 3.366 3.366 The following packages have unmet dependencies: 3.419 libopencl-clang-14-dev : Conflicts: libopencl-clang-x.y-dev 3.419 libopencl-clang-15-dev : Conflicts: libopencl-clang-x.y-dev 3.419 libopencl-clang-16-dev : Conflicts: libopencl-clang-x.y-dev 3.419 libopencl-clang-17-dev : Conflicts: libopencl-clang-x.y-dev 3.419 python3-clang-14 : Conflicts: python-clang-x.y 3.419 python3-clang-15 : Conflicts: python-clang-x.y 3.419 python3-clang-16 : Conflicts: python-clang-x.y 3.419 python3-clang-17 : Conflicts: python-clang-x.y 3.419 python3-clang-18 : Conflicts: python-clang-x.y 3.419 python3-clang-19 : Conflicts: python-clang-x.y 3.419 python3-clang-20 : Conflicts: python-clang-x.y 3.423 E: Unable to correct problems, you have held broken packages.

2 warnings found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 53)
UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 59) Dockerfile.rocm-ct2:4

3 | # ── Dépendances build 4 | >>> RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \ 5 | >>> git cmake ninja-build python3-dev \ 6 | >>> clang clang++ llvm \ 7 | >>> && rm -rf /var/lib/apt/lists/*

8

ERROR: failed to build: failed to solve: process "/bin/sh -c apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git cmake ninja-build python3-dev clang clang++ llvm && rm -rf /var/lib/apt/lists/*" did not complete successfully: exit code: 100 folzo@folzo-Corp:~/speaches/src$

L'image ROCm a déjà clang-18 installé (via hipify-clang), mais clang générique entre en conflit. Il faut utiliser la version spécifique déjà présente. Corrige le Dockerfile :

nano ~/speaches/src/Dockerfile.rocm-ct2

Remplace le premier RUN apt-get par :

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev python3-pip \
    && rm -rf /var/lib/apt/lists/*

Et remplace le bloc cmake pour utiliser clang-18 (déjà présent dans l'image ROCm) :

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/usr/lib/llvm-18/bin/clang++ \
        -DCMAKE_C_COMPILER=/usr/lib/llvm-18/bin/clang \
        -DCMAKE_HIP_COMPILER=/usr/lib/llvm-18/bin/clang \
        -DWITH_MKL=OFF \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

Vérifie d'abord que clang-18 est bien là dans l'image :

docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \
  find /usr/lib/llvm* -name "clang++" 2>/dev/null

Dis-moi le chemin retourné avant de relancer le build — on s'assure d'utiliser le bon chemin.

folzo@folzo-Corp:~/speaches/src$ docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \

find /usr/lib/llvm* -name "clang++" 2>/dev/null folzo@folzo-Corp:~/speaches/src$

et le fichier de conf : GNU nano 8.4 /home/folzo/speaches/src/Dockerfile.rocm-ct2 FROM rocm/dev-ubuntu-24.04:6.3.4-complete

── Dépendances build

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \ git cmake ninja-build python3-dev python3-pip \ && rm -rf /var/lib/apt/lists/*

── Compile CTranslate2-rocm pour gfx1101 (RX 7800 XT)

RUN git clone --recurse-submodules \ https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

RUN cd /ct2 && \ PYTORCH_ROCM_ARCH=gfx1101 \ HIP_PATH="$(hipconfig -R)" \ cmake -S . -B build \ -DCMAKE_CXX_COMPILER=/usr/lib/llvm-18/bin/clang++ \ -DCMAKE_C_COMPILER=/usr/lib/llvm-18/bin/clang \ -DCMAKE_HIP_COMPILER=/usr/lib/llvm-18/bin/clang \ -DWITH_MKL=OFF \ -DWITH_HIP=ON \ -DCMAKE_HIP_ARCHITECTURES=gfx1101 \ -DBUILD_TESTS=OFF \ -DWITH_CUDNN=ON \ -GNinja && \ ninja -C build -j$(nproc) && \ cmake --install build --prefix /usr/local && \ ldconfig

── Build wheel Python CTranslate2-rocm

RUN cd /ct2/python && \ pip3 install --break-system-packages -r install_requirements.txt && \ python3 setup.py bdist_wheel && \ cp dist/*.whl /ct2-wheel.whl

── Image finale speaches + CT2-rocm

FROM rocm/dev-ubuntu-24.04:6.3.4-complete

COPY --from=0 /ct2-wheel.whl /ct2-wheel.whl COPY --from=0 /usr/local/lib/libctranslate2* /usr/local/lib/ RUN ldconfig

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \ python3-pip ffmpeg && \ rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:latest /uv /bin/uv

RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true WORKDIR /home/ubuntu/speaches COPY --chown=ubuntu . .

RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \ /bin/uv pip install /ct2-wheel.whl --python .venv

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0 ENV LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

EXPOSE 8000 CMD [".venv/bin/uvicorn", "speaches.main:create_app", \ "--host", "0.0.0.0", "--port", "8000", "--factory"]

clang++ n'est pas dans /usr/lib/llvm* — cherchons où il est vraiment dans l'image :

docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \
  bash -c "which clang++ 2>/dev/null || find / -name 'clang++' -type f 2>/dev/null | head -5"

Et vérifie aussi ce que hipconfig retourne :

docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \
  bash -c "hipconfig -l && echo '---' && hipconfig -R && echo '---' && ls \$(hipconfig -l)/ | grep clang"

Ces deux commandes vont nous donner exactement les chemins disponibles dans l'image ROCm pour construire les flags cmake corrects.

folzo@folzo-Corp:~/speaches/src$ docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \

bash -c "which clang++ 2>/dev/null || find / -name 'clang++' -type f 2>/dev/null | head -5" folzo@folzo-Corp:/speaches/src$ docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \ bash -c "hipconfig -l && echo '---' && hipconfig -R && echo '---' && ls \$(hipconfig -l)/ | grep clang" /opt/rocm-6.3.4/lib/llvm/bin--- /opt/rocm-6.3.4--- amdclang amdclang++ amdclang-18 amdclang-cl amdclang-cpp clang clang++ clang++.cfg clang-18 clang-18.cfg clang-build-select-link clang-cl clang-cl.cfg clang-cpp clang-cpp.cfg clang-linker-wrapper clang-nvlink-wrapper clang-offload-bundler clang-offload-packager clang-offload-wrapper clang.cfg folzo@folzo-Corp:/speaches/src$

Parfait ! clang++ est dans /opt/rocm-6.3.4/lib/llvm/bin/. Corrige le Dockerfile avec les bons chemins :

nano ~/speaches/src/Dockerfile.rocm-ct2

Remplace uniquement le bloc RUN cd /ct2 :

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH=/opt/rocm-6.3.4 \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang \
        -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DWITH_MKL=OFF \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

Puis relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Cette fois cmake devrait trouver le compilateur. La compilation elle-même prendra 20-40 minutes selon la charge CPU — c'est normal pour du C++ avec HIP.

encore une erreur :

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 27.4s (10/20) docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.09kB 0.0s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:latest 0.4s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => CACHED [stage-0 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => CACHED FROM ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b 0.0s => => resolve ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b3 0.0s => [stage-0 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git cmake n 13.2s => [stage-0 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate2-rocm 12.7s => ERROR [stage-0 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cmak 1.0s

[stage-0 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cmake -S . -B build -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ -DWITH_MKL=OFF -DWITH_HIP=ON -DCMAKE_HIP_ARCHITECTURES=gfx1101 -DBUILD_TESTS=OFF -DWITH_CUDNN=ON -GNinja && ninja -C build -j$(nproc) && cmake --install build --prefix /usr/local && ldconfig: 0.328 -- The C compiler identification is Clang 18.0.0 0.367 -- The CXX compiler identification is Clang 18.0.0 0.391 -- Detecting C compiler ABI info 0.451 -- Detecting C compiler ABI info - done 0.457 -- Check for working C compiler: /opt/rocm-6.3.4/lib/llvm/bin/clang - skipped 0.458 -- Detecting C compile features 0.458 -- Detecting C compile features - done 0.475 -- Detecting CXX compiler ABI info 0.535 -- Detecting CXX compiler ABI info - done 0.540 -- Check for working CXX compiler: /opt/rocm-6.3.4/lib/llvm/bin/clang++ - skipped 0.541 -- Detecting CXX compile features 0.541 -- Detecting CXX compile features - done 0.542 -- Performing Test CMAKE_HAVE_LIBC_PTHREAD 0.594 -- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success 0.595 -- Found Threads: TRUE 0.596 -- Build spdlog: 1.10.0 0.598 -- Build type: Release 0.599 CMake Deprecation Warning at third_party/cpu_features/CMakeLists.txt:1 (cmake_minimum_required): 0.599 Compatibility with CMake < 3.5 will be removed from a future version of 0.599 CMake. 0.599

0.599   Update the VERSION argument <min> value or use a ...<max> suffix to tell

0.599 CMake that the project does not need compatibility with older versions. 0.599 0.599 0.601 -- Looking for dlfcn.h 0.647 -- Looking for dlfcn.h - found 0.647 -- Looking for getauxval 0.696 -- Looking for getauxval - found 0.699 -- Compiling for multiple CPU ISA and enabling runtime dispatch 0.879 -- Found OpenMP_C: -fopenmp=libomp (found version "5.1") 0.941 -- Found OpenMP_CXX: -fopenmp=libomp (found version "5.1") 0.941 -- Found OpenMP: TRUE (found version "5.1") 0.943 CMake Error at CMakeLists.txt:334 (message): 0.943 Intel OpenMP runtime libiomp5 not found 0.943 0.943 0.943 -- Configuring incomplete, errors occurred!

2 warnings found (use docker --debug to expand):

UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 56)
CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 50) Dockerfile.rocm-ct2:12

28

ERROR: failed to build: failed to solve: process "/bin/sh -c cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cmake -S . -B build -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ -DWITH_MKL=OFF -DWITH_HIP=ON -DCMAKE_HIP_ARCHITECTURES=gfx1101 -DBUILD_TESTS=OFF -DWITH_CUDNN=ON -GNinja && ninja -C build -j$(nproc) && cmake --install build --prefix /usr/local && ldconfig" did not complete successfully: exit code: 1 folzo@folzo-Corp:~/speaches/src$

Bonne progression — le compilateur est trouvé maintenant ! L'erreur est différente :

CMake Error: Intel OpenMP runtime libiomp5 not found

-DWITH_MKL=OFF devrait désactiver ça mais le CMakeLists.txt de CTranslate2-rocm le cherche quand même. Ajoute -DOPENMP_RUNTIME=COMP pour forcer l'OpenMP du compilateur (libomp déjà trouvé) au lieu d'Intel MKL :

nano ~/speaches/src/Dockerfile.rocm-ct2

Remplace le bloc cmake :

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH=/opt/rocm-6.3.4 \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang \
        -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DWITH_MKL=OFF \
        -DOPENMP_RUNTIME=COMP \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=OFF \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

Deux changements :

DOPENMP_RUNTIME=COMP → utilise l'OpenMP du compilateur (libomp, déjà détecté)
DWITH_CUDNN=OFF → on n'a pas cuDNN, inutile de le chercher

Relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

on est allé plus loin mais il y aencor eune erreur :

"folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 134.0s (20/20) FINISHED docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.12kB 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:latest 0.4s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => CACHED FROM ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b 0.0s => => resolve ghcr.io/astral-sh/uv:latest@sha256:4cac394b6b72846f8a85a7a0e577c6d61d4e17fe2ccee65d9451a8b3 0.0s => CACHED [stage-0 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => CACHED [stage-0 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git c 0.0s => CACHED [stage-0 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate 0.0s => [stage-0 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cmake -S 50.6s => [stage-0 5/5] RUN cd /ct2/python && pip3 install --break-system-packages -r install_requirements. 17.4s => [stage-1 2/10] COPY --from=0 /ct2-wheel.whl /ct2-wheel.whl 0.1s => [stage-1 3/10] COPY --from=0 /usr/local/lib/libctranslate2* /usr/local/lib/ 0.1s => [stage-1 4/10] RUN ldconfig 0.3s => [stage-1 5/10] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y python3-p 63.6s => [stage-1 6/10] COPY --from=ghcr.io/astral-sh/uv:latest /uv /bin/uv 0.2s => [stage-1 7/10] RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true 0.2s => [stage-1 8/10] WORKDIR /home/ubuntu/speaches 0.1s => [stage-1 9/10] COPY --chown=ubuntu . . 0.1s => ERROR [stage-1 10/10] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/ 0.2s

[stage-1 10/10] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/uv pip install /ct2-wheel.whl --python .venv: 0.167 error: Required uv version ~=0.8.14 does not match the running version 0.10.4

2 warnings found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 51)
UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 57) Dockerfile.rocm-ct2:53

52 | 53 | >>> RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \ 54 | >>> /bin/uv pip install /ct2-wheel.whl --python .venv

55

ERROR: failed to build: failed to solve: process "/bin/sh -c /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/uv pip install /ct2-wheel.whl --python .venv" did not complete successfully: exit code: 2 folzo@folzo-Corp:~/speaches/src$ "

Excellent progrès — CTranslate2 compile et la wheel est générée ! L'erreur finale est simple : uv version 0.10.4 (latest) est incompatible avec le pyproject.toml de speaches qui requiert ~=0.8.14. Il suffit de fixer la version de uv :

nano ~/speaches/src/Dockerfile.rocm-ct2

Remplace la ligne COPY --from=ghcr.io/astral-sh/uv:latest par la version exacte utilisée lors du premier build :

COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv

C'est la même version que dans le Dockerfile original de speaches (visible au tout début quand on avait buildé speaches:rocm). Relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Les étapes 1 à 9 seront toutes en cache — seule l'étape 10 sera re-exécutée. Ce sera rapide.

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

[+] Building 9.7s (20/20) FINISHED docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.12kB 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:0.8.22 0.4s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => [stage-0 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => CACHED FROM ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a 0.0s => => resolve ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a0 0.0s => CACHED [stage-0 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git c 0.0s => CACHED [stage-0 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate 0.0s => CACHED [stage-0 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cma 0.0s => CACHED [stage-0 5/5] RUN cd /ct2/python && pip3 install --break-system-packages -r install_require 0.0s => CACHED [stage-1 2/10] COPY --from=0 /ct2-wheel.whl /ct2-wheel.whl 0.0s => CACHED [stage-1 3/10] COPY --from=0 /usr/local/lib/libctranslate2* /usr/local/lib/ 0.0s => CACHED [stage-1 4/10] RUN ldconfig 0.0s => CACHED [stage-1 5/10] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y pyt 0.0s => [stage-1 6/10] COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv 0.1s => [stage-1 7/10] RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true 0.2s => [stage-1 8/10] WORKDIR /home/ubuntu/speaches 0.1s => [stage-1 9/10] COPY --chown=ubuntu . . 0.1s => ERROR [stage-1 10/10] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/ 8.7s

[stage-1 10/10] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/uv pip install /ct2-wheel.whl --python .venv: 0.229 Using CPython 3.12.3 interpreter at: /usr/bin/python3 0.229 Creating virtual environment at: .venv 0.267 Building speaches @ file:///home/ubuntu/speaches 0.329 Downloading pydantic-core (1.9MiB) 0.329 Downloading tiktoken (1.1MiB) 0.329 Downloading faster-whisper (1.1MiB) 0.329 Downloading cryptography (4.0MiB) 0.329 Downloading gradio (60.6MiB) 0.330 Downloading espeakng-loader (9.6MiB) 0.330 Downloading clldutils (1.6MiB) 0.330 Downloading av (33.9MiB) 0.331 Downloading soundfile (1.3MiB) 0.331 Downloading hf-xet (3.0MiB) 0.331 Downloading hf-transfer (3.4MiB) 0.331 Downloading setuptools (1.2MiB) 0.332 Downloading babel (9.1MiB) 0.332 Downloading lxml (4.7MiB) 0.332 Downloading scikit-learn (9.0MiB) 0.332 Downloading tokenizers (2.8MiB) 0.333 Downloading ruff (12.7MiB) 0.333 Downloading piper-tts (13.2MiB) 0.334 Downloading grpcio (5.6MiB) 0.334 Downloading pygments (1.1MiB) 0.334 Downloading scipy (34.0MiB) 0.334 Downloading numpy (15.9MiB) 0.335 Downloading pandas (12.1MiB) 0.340 Downloading sympy (5.9MiB) 0.340 Downloading brotli (2.8MiB) 0.340 Downloading pylibsrtp (2.1MiB) 0.341 Downloading ctranslate2 (37.0MiB) 0.341 Downloading pillow (4.3MiB) 0.342 Downloading onnxruntime-gpu (270.1MiB) 0.867 Downloading faster-whisper 0.891 Downloading tiktoken 0.907 Built speaches @ file:///home/ubuntu/speaches 0.911 Downloading soundfile 1.080 Downloading clldutils 1.126 Downloading pygments 1.156 Building pylatexenc==2.10 1.171 Downloading pydantic-core 1.202 Downloading pylibsrtp 1.308 Downloading setuptools 1.347 Downloading tokenizers 1.350 Downloading brotli 1.367 Downloading hf-xet 1.442 Downloading hf-transfer 1.598 Downloading pillow 1.598 Downloading cryptography 1.652 Downloading lxml 1.776 Downloading grpcio 1.967 Downloading sympy 1.969 Built pylatexenc==2.10 2.170 Downloading scikit-learn 2.193 Downloading babel 2.273 Downloading espeakng-loader 2.457 Downloading ruff 2.460 Downloading pandas 2.487 Downloading piper-tts 2.629 Downloading numpy 3.430 Downloading av 3.453 Downloading scipy 3.527 Downloading ctranslate2 4.019 Downloading gradio 6.029 Downloading onnxruntime-gpu 6.029 Prepared 163 packages in 5.77s 6.161 Installed 163 packages in 132ms 8.120 Bytecode compiled 10021 files in 1.95s 8.120 + aiofiles==23.2.1 8.120 + aioice==0.10.1 8.120 + aiortc==1.13.0 8.120 + aiostream==0.7.0 8.120 + annotated-doc==0.0.3 8.120 + annotated-types==0.7.0 8.120 + anyio==4.9.0 8.120 + asgiref==3.8.1 8.120 + attrs==24.3.0 8.120 + av==14.4.0 8.120 + babel==2.16.0 8.120 + brotli==1.1.0 8.120 + cachetools==6.1.0 8.120 + certifi==2024.8.30 8.120 + cffi==1.17.1 8.120 + charset-normalizer==3.3.2 8.120 + click==8.1.7 8.120 + clldutils==3.21.0 8.120 + colorama==0.4.6 8.120 + coloredlogs==15.0.1 8.120 + colorlog==6.9.0 8.120 + cryptography==44.0.1 8.121 + csvw==3.5.1 8.121 + ctranslate2==4.5.0 8.121 + deprecated==1.2.14 8.121 + distro==1.9.0 8.121 + dlinfo==1.2.1 8.121 + dnspython==2.7.0 8.121 + einops==0.8.1 8.121 + espeakng-loader==0.2.4 8.121 + fastapi==0.121.1 8.121 + faster-whisper==1.1.1 8.121 + ffmpy==0.4.0 8.121 + filelock==3.16.1 8.121 + flatbuffers==24.3.25 8.121 + fsspec==2024.9.0 8.121 + google-crc32c==1.6.0 8.121 + googleapis-common-protos==1.65.0 8.121 + gradio==5.49.1 8.121 + gradio-client==1.13.3 8.121 + groovy==0.1.2 8.121 + grpcio==1.67.1 8.121 + h11==0.14.0 8.121 + hf-transfer==0.1.8 8.121 + hf-xet==1.1.5 8.121 + httpcore==1.0.5 8.121 + httpx==0.28.1 8.121 + httpx-sse==0.4.1 8.121 + httpx-ws==0.7.2 8.121 + huggingface-hub==0.35.3 8.121 + humanfriendly==10.0 8.121 + idna==3.10 8.121 + ifaddr==0.2.0 8.121 + importlib-metadata==8.4.0 8.121 + isodate==0.7.2 8.121 + jinja2==3.1.4 8.121 + jiter==0.11.1 8.121 + joblib==1.4.2 8.121 + jsonschema==4.23.0 8.121 + jsonschema-specifications==2024.10.1 8.121 + kaldi-native-fbank==1.22.3 8.121 + kokoro-onnx==0.4.5 8.121 + language-tags==1.2.0 8.121 + lxml==5.3.0 8.121 + markdown==3.7 8.121 + markdown-it-py==3.0.0 8.122 + markupsafe==2.1.5 8.122 + mdurl==0.1.2 8.122 + mpmath==1.3.0 8.122 + numpy==2.3.1 8.122 + onnx-asr==0.7.0 8.122 + onnx-diarization==0.1.0 8.122 + onnx-dl==0.1.0 8.122 + onnxruntime-gpu==1.22.0 8.122 + openai==2.7.1 8.122 + opentelemetry-api==1.29.0 8.122 + opentelemetry-distro==0.50b0 8.122 + opentelemetry-exporter-otlp==1.29.0 8.122 + opentelemetry-exporter-otlp-proto-common==1.29.0 8.122 + opentelemetry-exporter-otlp-proto-grpc==1.29.0 8.122 + opentelemetry-exporter-otlp-proto-http==1.29.0 8.122 + opentelemetry-instrumentation==0.50b0 8.122 + opentelemetry-instrumentation-asgi==0.50b0 8.122 + opentelemetry-instrumentation-asyncio==0.50b0 8.122 + opentelemetry-instrumentation-fastapi==0.50b0 8.122 + opentelemetry-instrumentation-grpc==0.50b0 8.122 + opentelemetry-instrumentation-httpx==0.50b0 8.122 + opentelemetry-instrumentation-logging==0.50b0 8.122 + opentelemetry-instrumentation-openai==0.37.1 8.122 + opentelemetry-instrumentation-openai-v2==2.1b0 8.122 + opentelemetry-instrumentation-requests==0.50b0 8.122 + opentelemetry-instrumentation-threading==0.50b0 8.122 + opentelemetry-instrumentation-urllib==0.50b0 8.122 + opentelemetry-instrumentation-urllib3==0.50b0 8.122 + opentelemetry-proto==1.29.0 8.122 + opentelemetry-sdk==1.29.0 8.122 + opentelemetry-semantic-conventions==0.50b0 8.122 + opentelemetry-semantic-conventions-ai==0.4.2 8.122 + opentelemetry-util-http==0.50b0 8.122 + orjson==3.10.7 8.122 + packaging==24.1 8.122 + pandas==2.2.3 8.123 + phonemizer-fork==3.3.1 8.123 + pillow==10.4.0 8.123 + piper-tts==1.3.0 8.123 + protobuf==5.29.3 8.123 + pyannote-core==6.0.1 8.123 + pycparser==2.22 8.123 + pydantic==2.11.7 8.123 + pydantic-core==2.33.2 8.123 + pydantic-settings==2.10.1 8.123 + pydub==0.25.1 8.123 + pyee==13.0.0 8.123 + pygments==2.18.0 8.123 + pylatexenc==2.10 8.123 + pylibsrtp==0.11.0 8.123 + pyopenssl==25.0.0 8.123 + pyparsing==3.2.1 8.123 + python-dateutil==2.9.0.post0 8.123 + python-dotenv==1.0.1 8.123 + python-multipart==0.0.20 8.123 + pytz==2024.2 8.123 + pyyaml==6.0.2 8.123 + rdflib==7.1.2 8.123 + referencing==0.35.1 8.123 + regex==2024.9.11 8.123 + requests==2.32.3 8.123 + rfc3986==1.5.0 8.123 + rich==13.8.1 8.123 + rpds-py==0.22.3 8.123 + ruff==0.14.4 8.123 + safehttpx==0.1.6 8.123 + scikit-learn==1.7.2 8.124 + scipy==1.16.3 8.124 + segments==2.2.1 8.124 + semantic-version==2.10.0 8.124 + setuptools==75.2.0 8.124 + shellingham==1.5.4 8.124 + six==1.16.0 8.124 + sniffio==1.3.1 8.124 + sortedcontainers==2.4.0 8.124 + sounddevice==0.5.2 8.124 + soundfile==0.13.1 8.124 + speaches==0.1.0 (from file:///home/ubuntu/speaches) 8.124 + starlette==0.41.3 8.124 + sympy==1.13.3 8.124 + tabulate==0.9.0 8.124 + threadpoolctl==3.6.0 8.124 + tiktoken==0.8.0 8.124 + tokenizers==0.20.1 8.124 + tomlkit==0.12.0 8.124 + tqdm==4.66.5 8.124 + typer==0.12.5 8.124 + typing-extensions==4.12.2 8.124 + typing-inspection==0.4.1 8.124 + tzdata==2024.2 8.124 + uritemplate==4.1.1 8.124 + urllib3==2.2.3 8.124 + uvicorn==0.35.0 8.124 + websockets==14.2 8.124 + wrapt==1.16.0 8.124 + wsproto==1.2.0 8.124 + zipp==3.20.2 8.186 error: The wheel filename "ct2-wheel.whl" is invalid: Must have a Python tag

2 warnings found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 51)
UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 57) Dockerfile.rocm-ct2:53

52 | 53 | >>> RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \ 54 | >>> /bin/uv pip install /ct2-wheel.whl --python .venv

55

Presque ! Tout s'installe parfaitement, dernière erreur mineure : uv refuse d'installer la wheel renommée en ct2-wheel.whl car le nom ne respecte pas le format PEP 427 (nom-version-pytag-abitag-platformtag.whl). Il faut copier la wheel avec son vrai nom :

nano ~/speaches/src/Dockerfile.rocm-ct2

Remplace ces lignes dans le stage-0 :

    cp dist/*.whl /ct2-wheel.whl

Par :

    cp dist/*.whl /

Et dans le stage-1, remplace :

COPY --from=0 /ct2-wheel.whl /ct2-wheel.whl

Par :

COPY --from=0 /ctranslate2*.whl /

Et remplace la ligne d'installation :

    /bin/uv pip install /ct2-wheel.whl --python .venv

Par :

    /bin/uv pip install /ctranslate2*.whl --python .venv

Relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Tout sera en cache sauf la dernière étape — ce sera rapide.

trop compliquer peux tu me regener le fichier en entier stp, je vais faire des erreurs sinon

Voici le Dockerfile.rocm-ct2 complet et corrigé :

FROM rocm/dev-ubuntu-24.04:6.3.4-complete AS builder

# ── Dépendances build
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev python3-pip \
    && rm -rf /var/lib/apt/lists/*

# ── Clone CTranslate2-rocm
RUN git clone --recurse-submodules \
    https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

# ── Compile CTranslate2-rocm pour gfx1101 (RX 7800 XT)
RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH=/opt/rocm-6.3.4 \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang \
        -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DWITH_MKL=OFF \
        -DOPENMP_RUNTIME=COMP \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=OFF \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

# ── Build wheel Python CTranslate2-rocm
RUN cd /ct2/python && \
    pip3 install --break-system-packages -r install_requirements.txt && \
    python3 setup.py bdist_wheel && \
    cp dist/*.whl /

# ── Image finale
FROM rocm/dev-ubuntu-24.04:6.3.4-complete

COPY --from=builder /ctranslate2*.whl /
COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/
RUN ldconfig

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    python3-pip ffmpeg \
    && rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv

RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true
WORKDIR /home/ubuntu/speaches
COPY --chown=ubuntu . .

RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \
    /bin/uv pip install /ctranslate2*.whl --python .venv

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0
ENV LD_LIBRARY_PATH=/usr/local/lib:/opt/rocm-6.3.4/lib:$LD_LIBRARY_PATH

EXPOSE 8000
CMD [".venv/bin/uvicorn", "speaches.main:create_app", \
     "--host", "0.0.0.0", "--port", "8000", "--factory"]

Copie-colle ce contenu dans nano :

nano ~/speaches/src/Dockerfile.rocm-ct2

Puis Ctrl+X, Y, Entrée pour sauvegarder. Lance le build :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Les étapes compilées seront en cache — seule la dernière étape uv pip install sera re-exécutée.

du coup c'est fini ? :

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 152.6s (21/21) FINISHED docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.16kB 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:0.8.22 0.4s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => CACHED FROM ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a 0.0s => => resolve ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a0 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => CACHED [builder 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => CACHED [builder 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git c 0.0s => CACHED [builder 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate 0.0s => CACHED [builder 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cma 0.0s => [builder 5/5] RUN cd /ct2/python && pip3 install --break-system-packages -r install_requirements. 17.6s => [stage-1 2/10] COPY --from=builder /ctranslate2*.whl / 0.4s => [stage-1 3/10] COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/ 0.1s => [stage-1 4/10] RUN ldconfig 0.3s => [stage-1 5/10] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y python3-p 64.7s => [stage-1 6/10] COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv 0.2s => [stage-1 7/10] RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true 0.2s => [stage-1 8/10] WORKDIR /home/ubuntu/speaches 0.1s => [stage-1 9/10] COPY --chown=ubuntu . . 0.1s => [stage-1 10/10] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/uv pip 8.9s => exporting to image 59.0s => => exporting layers 44.6s => => exporting manifest sha256:1dd298d1b3822fc9d2fc6d28438b441358d73977b3121fa3ba2304efd2840daa 0.0s => => exporting config sha256:ade63f546f7be00555c12f03171b1b2246fb3e484aa0ea1322f6b947fc6b66a9 0.0s => => exporting attestation manifest sha256:0bbb942dd7e7b8c230c5f28c77d786e67e6dd271d36f5ee34549b066c988e 0.1s => => exporting manifest list sha256:d29e09ad6ae6d74d1cd2fedf7715a91aaee05119366623cf165515b78f44ba1f 0.0s => => naming to docker.io/library/speaches:rocm-ct2 0.0s => => unpacking to docker.io/library/speaches:rocm-ct2 14.1s

2 warnings found (use docker --debug to expand):

UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 58)
CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 52) folzo@folzo-Corp:~/speaches/src$

Build réussi ! 🎉 speaches:rocm-ct2 est prêt. Maintenant il faut tester que CTranslate2 voit bien le GPU, puis relancer le stack.

1. Vérifie que ROCm est détecté

docker run --rm \
  --device /dev/kfd \
  --device /dev/dri \
  -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \
  speaches:rocm-ct2 \
  .venv/bin/python3 -c "
import ctranslate2
print('Version CT2:', ctranslate2.__version__)
print('ROCm compute types:', ctranslate2.get_supported_compute_types('cuda'))
"

Si tu vois float16, int8_float16 etc. → le GPU est reconnu ✅

2. Mets à jour docker-compose.yml

nano ~/speaches/docker-compose.yml

Change ces lignes :

image: speaches:rocm-ct2          # ← nouvelle image
environment:
  - WHISPER__MODEL=Systran/faster-whisper-large-v3
  - WHISPER__LANGUAGE=fr
  - WHISPER__DEVICE=cuda           # ← GPU
  - WHISPER__COMPUTE_TYPE=float16  # ← float16 sur GPU

3. Relance le stack

cd ~/speaches
docker compose down && docker compose up -d
docker logs -f speaches-stt

Tu devrais voir le modèle se charger sur le GPU au lieu du CPU.

folzo@folzo-Corp:~/speaches/src$ docker run --rm \

--device /dev/kfd \ --device /dev/dri \ -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \ speaches:rocm-ct2 \ .venv/bin/python3 -c " import ctranslate2 print('Version CT2:', ctranslate2.version) print('ROCm compute types:', ctranslate2.get_supported_compute_types('cuda')) " Traceback (most recent call last):

File "<string>", line 2, in <module>

File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/ctranslate2/__init__.py", line 21, in from ctranslate2._ext import ( ImportError: libomp.so: cannot open shared object file: No such file or directory folzo@folzo-Corp:~/speaches/src$

libomp.so n'est pas dans l'image finale — elle vient de l'image ROCm du builder mais n'a pas été copiée. Fix rapide dans le Dockerfile :

nano ~/speaches/src/Dockerfile.rocm-ct2

Dans le stage-1, ajoute libomp-dev dans le apt-get install :

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    python3-pip ffmpeg libomp-dev \
    && rm -rf /var/lib/apt/lists/*

Ou encore mieux, copie directement libomp.so depuis le builder (plus léger) en ajoutant après la ligne COPY --from=builder /usr/local/lib/libctranslate2* :

COPY --from=builder /opt/rocm-6.3.4/lib/llvm/lib/libomp.so /usr/local/lib/

Relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Tout sera en cache sauf les étapes après le changement — moins de 2 minutes.

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

[+] Building 62.1s (9/20) docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.14kB 0.0s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:0.8.22 0.6s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => CACHED [builder 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => FROM ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a0f12edd 0.0s => => resolve ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a0 0.0s => [builder 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y python3-pip 61.1s => ERROR [builder 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate2 0.3s

[builder 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2: 0.249 /bin/sh: 1: git: not found

2 warnings found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 52)
UndefinedVar: Usage of undefined variable '$LD_LIBRARY_PATH' (line 58) Dockerfile.rocm-ct2:9

8 | # ── Clone CTranslate2-rocm 9 | >>> RUN git clone --recurse-submodules \ 10 | >>> https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

11

ERROR: failed to build: failed to solve: process "/bin/sh -c git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2" did not complete successfully: exit code: 127 folzo@folzo-Corp:~/speaches/src$

Le cache a été invalidé et git a disparu du apt-get install du builder (tu as probablement modifié le mauvais bloc). Voici le fichier complet corrigé, régénère-le entièrement :

cat > ~/speaches/src/Dockerfile.rocm-ct2 << 'EOF'
FROM rocm/dev-ubuntu-24.04:6.3.4-complete AS builder

# ── Dépendances build
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev python3-pip \
    && rm -rf /var/lib/apt/lists/*

# ── Clone CTranslate2-rocm
RUN git clone --recurse-submodules \
    https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

# ── Compile CTranslate2-rocm pour gfx1101 (RX 7800 XT)
RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH=/opt/rocm-6.3.4 \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang \
        -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DWITH_MKL=OFF \
        -DOPENMP_RUNTIME=COMP \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=OFF \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

# ── Build wheel Python
RUN cd /ct2/python && \
    pip3 install --break-system-packages -r install_requirements.txt && \
    python3 setup.py bdist_wheel && \
    cp dist/*.whl /

# ── Image finale
FROM rocm/dev-ubuntu-24.04:6.3.4-complete

COPY --from=builder /ctranslate2*.whl /
COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/
COPY --from=builder /opt/rocm-6.3.4/lib/llvm/lib/libomp.so /usr/local/lib/
RUN ldconfig

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    python3-pip ffmpeg \
    && rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv

RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true
WORKDIR /home/ubuntu/speaches
COPY --chown=ubuntu . .

RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \
    /bin/uv pip install /ctranslate2*.whl --python .venv

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0
ENV LD_LIBRARY_PATH=/usr/local/lib:/opt/rocm-6.3.4/lib

EXPOSE 8000
CMD [".venv/bin/uvicorn", "speaches.main:create_app", \
     "--host", "0.0.0.0", "--port", "8000", "--factory"]
EOF

Cette fois j'utilise cat > pour éviter toute erreur de copie-colle dans nano. Puis relance :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

normalement c'est bon :

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 129.1s (22/22) FINISHED docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.20kB 0.0s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:0.8.22 0.4s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => CACHED FROM ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a 0.0s => => resolve ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a0 0.0s => [builder 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => CACHED [builder 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git c 0.0s => CACHED [builder 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate 0.0s => CACHED [builder 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cma 0.0s => CACHED [builder 5/5] RUN cd /ct2/python && pip3 install --break-system-packages -r install_require 0.0s => CACHED [stage-1 2/11] COPY --from=builder /ctranslate2*.whl / 0.0s => CACHED [stage-1 3/11] COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/ 0.0s => [stage-1 4/11] COPY --from=builder /opt/rocm-6.3.4/lib/llvm/lib/libomp.so /usr/local/lib/ 0.1s => [stage-1 5/11] RUN ldconfig 0.2s => [stage-1 6/11] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y python3-p 59.3s => [stage-1 7/11] COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv 0.2s => [stage-1 8/11] RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true 0.2s => [stage-1 9/11] WORKDIR /home/ubuntu/speaches 0.1s => [stage-1 10/11] COPY --chown=ubuntu . . 0.1s => [stage-1 11/11] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/uv pip 8.9s => exporting to image 59.3s => => exporting layers 45.1s => => exporting manifest sha256:e2ebfb6743d5668e26b1f16ee476cf585c1382d6f5b7906ce155a2042da931d9 0.0s => => exporting config sha256:927bb261eddcd487aa27358112308714d91c5b831845eff8292e68f17e946002 0.0s => => exporting attestation manifest sha256:dd40ff2a2458b3c692d6fee034693b7d78ea33227ba6527f7b0fb708d4c32 0.1s => => exporting manifest list sha256:08366edde66c5a5f761ae602bbc2f0c7298e65b26bece2e5210694747ece4e0a 0.0s => => naming to docker.io/library/speaches:rocm-ct2 0.0s => => unpacking to docker.io/library/speaches:rocm-ct2 13.9s

1 warning found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 53) folzo@folzo-Corp:~/speaches/src$

Build parfait, plus aucune erreur ! Teste maintenant que le GPU est bien détecté :

docker run --rm \
  --device /dev/kfd \
  --device /dev/dri \
  -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \
  speaches:rocm-ct2 \
  .venv/bin/python3 -c "
import ctranslate2
print('Version CT2:', ctranslate2.__version__)
print('ROCm compute types:', ctranslate2.get_supported_compute_types('cuda'))
"

Si tu vois float16 dans la liste → lance le stack :

cd ~/speaches
docker compose down && docker compose up -d
docker logs -f speaches-stt

Et surveille dans les logs que le modèle charge sur cuda au lieu de cpu. Tu peux aussi vérifier la VRAM en temps réel depuis un autre terminal :

watch -n1 rocm-smi

La transcription devrait être 5 à 10x plus rapide qu'en CPU. 🚀

log apres un test :

return self.model.encode(features, to_cpu=to_cpu)
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Conv1D on GPU currently requires the cuDNN library which is not integrated in this build

2026-02-21 18:07:44,815:DEBUG:speaches.dependencies:audio_file_dependency:89:Decoding audio file: f0ab6f70-dd33-4d48-a580-8ef47f9bb348.mp3, content_type: None, header: Headers({'content-disposition': 'form-data; name="file"; filename="f0ab6f70-dd33-4d48-a580-8ef47f9bb348.mp3"'}), size: 35565 2026-02-21 18:07:44,904:DEBUG:speaches.dependencies:audio_file_dependency:103:Decoded 4.38s of audio in 0.08933s (RTF: 0.02039520114276996) 2026-02-21 18:07:44,904:DEBUG:asyncio:init:64:Using selector: EpollSelector 2026-02-21 18:07:44,908:DEBUG:speaches.executors.shared.base_model_manager:load_model:138:silero_vad_v5 model already loaded 2026-02-21 18:07:44,908:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for silero_vad_v5, self.ref_count=1 2026-02-21 18:07:44,917:DEBUG:speaches.executors.silero_vad_v5:call:138:VAD model inference took 0.0091s 2026-02-21 18:07:44,918:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for silero_vad_v5, self.ref_count=0 2026-02-21 18:07:44,918:INFO:speaches.executors.shared.base_model_manager:_decrement_ref:98:Model silero_vad_v5 is idle, not unloading 2026-02-21 18:07:44,918:DEBUG:speaches.executors.silero_vad_v5:get_speech_timestamps:307:VAD processing took 0.0099s for 4.38s audio 2026-02-21 18:07:44,918:DEBUG:speaches.executors.shared.base_model_manager:load_model:138:Systran/faster-whisper-large-v3 model already loaded 2026-02-21 18:07:44,918:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:81:Model was set to expire in 300s, cancelling 2026-02-21 18:07:44,918:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for Systran/faster-whisper-large-v3, self.ref_count=1 2026-02-21 18:07:44,918:INFO:faster_whisper:transcribe:391:Processing audio with duration 00:04.380 2026-02-21 18:07:44,918:INFO:faster_whisper:transcribe:428:VAD filter removed 00:02.332 of audio 2026-02-21 18:07:44,935:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for Systran/faster-whisper-large-v3, self.ref_count=0 2026-02-21 18:07:44,935:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:91:Model Systran/faster-whisper-large-v3 is idle, scheduling offload in 300s INFO: 192.168.1.27:54094 - "POST /v1/audio/transcriptions HTTP/1.1" 500 Internal Server Error ERROR: Exception in ASGI application Traceback (most recent call last): File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/uvicorn/protocols/http/h11_impl.py", line 403, in run_asgi result = await app( # type: ignore[func-returns-value] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/uvicorn/middleware/proxy_headers.py", line 60, in call return await self.app(scope, receive, send) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/fastapi/applications.py", line 1134, in call await super().call(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/applications.py", line 113, in call await self.middleware_stack(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/middleware/errors.py", line 187, in call raise exc File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/middleware/errors.py", line 165, in call await self.app(scope, receive, _send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/middleware/exceptions.py", line 62, in call await wrap_app_handling_exceptions(self.app, conn)(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app raise exc File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app await app(scope, receive, sender) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/fastapi/middleware/asyncexitstack.py", line 18, in call await self.app(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/routing.py", line 715, in call await self.middleware_stack(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/routing.py", line 735, in app await route.handle(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/routing.py", line 288, in handle await self.app(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/fastapi/routing.py", line 125, in app await wrap_app_handling_exceptions(app, request)(scope, receive, send) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app raise exc File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app await app(scope, receive, sender) File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/fastapi/routing.py", line 111, in app response = await f(request) ^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/fastapi/routing.py", line 391, in app raw_response = await run_endpoint_function( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/fastapi/routing.py", line 292, in run_endpoint_function return await run_in_threadpool(dependant.call, **values) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/concurrency.py", line 39, in run_in_threadpool return await anyio.to_thread.run_sync(func, *args) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/anyio/to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/anyio/_backends/_asyncio.py", line 2470, in run_sync_in_worker_thread return await future ^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/anyio/_backends/_asyncio.py", line 967, in run result = context.run(func, *args) ^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/speaches/routers/stt.py", line 171, in transcribe_file res = transcription_executor.model_manager.handle_transcription_request(transcription_request) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/speaches/executors/whisper.py", line 231, in handle_transcription_request return self.handle_non_streaming_transcription_request(request, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/speaches/tracing.py", line 31, in wrapper return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/speaches/executors/whisper.py", line 161, in handle_non_streaming_transcription_request segments, transcription_info = whisper_model.transcribe( ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/faster_whisper/transcribe.py", line 451, in transcribe ) = self.model.detect_language( ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/faster_whisper/transcribe.py", line 1777, in detect_language encoder_output = self.encode( ^^^^^^^^^^^^ File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/faster_whisper/transcribe.py", line 1358, in encode return self.model.encode(features, to_cpu=to_cpu) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ RuntimeError: Conv1D on GPU currently requires the cuDNN library which is not integrated in this build

L'erreur est claire : CTranslate2-rocm a été compilé sans le support cuDNN (-DWITH_CUDNN=OFF) mais Whisper utilise Conv1D qui en a besoin sur GPU. Il faut activer WITH_CUDNN=ON dans le build.

Mais sur ROCm, cuDNN s'appelle MIOpen. Il faut compiler avec -DWITH_CUDNN=ON ET les libs MIOpen disponibles dans l'image ROCm complète :

cat > ~/speaches/src/Dockerfile.rocm-ct2 << 'EOF'
FROM rocm/dev-ubuntu-24.04:6.3.4-complete AS builder

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev python3-pip \
    && rm -rf /var/lib/apt/lists/*

RUN git clone --recurse-submodules \
    https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH=/opt/rocm-6.3.4 \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang \
        -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DWITH_MKL=OFF \
        -DOPENMP_RUNTIME=COMP \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON \
        -DCUDnn_INCLUDE_DIR=/opt/rocm-6.3.4/include \
        -DCUDnn_LIBRARY=/opt/rocm-6.3.4/lib/libMIOpen.so \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

RUN cd /ct2/python && \
    pip3 install --break-system-packages -r install_requirements.txt && \
    python3 setup.py bdist_wheel && \
    cp dist/*.whl /

FROM rocm/dev-ubuntu-24.04:6.3.4-complete

COPY --from=builder /ctranslate2*.whl /
COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/
COPY --from=builder /opt/rocm-6.3.4/lib/llvm/lib/libomp.so /usr/local/lib/
RUN ldconfig

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    python3-pip ffmpeg \
    && rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv

RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true
WORKDIR /home/ubuntu/speaches
COPY --chown=ubuntu . .

RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \
    /bin/uv pip install /ctranslate2*.whl --python .venv

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0
ENV LD_LIBRARY_PATH=/usr/local/lib:/opt/rocm-6.3.4/lib

EXPOSE 8000
CMD [".venv/bin/uvicorn", "speaches.main:create_app", \
     "--host", "0.0.0.0", "--port", "8000", "--factory"]
EOF

Le seul changement est dans le bloc cmake :

-DWITH_CUDNN=ON — réactivé
-DCUDnn_INCLUDE_DIR et -DCUDnn_LIBRARY pointent vers MIOpen de ROCm

Relance le build — les étapes 1 à 3 seront en cache, seule la compilation cmake repart :

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

folzo@folzo-Corp:~/speaches$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

[+] Building 0.1s (1/1) FINISHED docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2B 0.0s ERROR: failed to build: failed to solve: failed to read dockerfile: open Dockerfile.rocm-ct2: no such file or directory folzo@folzo-Corp:~/speaches$

Tu es dans ~/speaches mais le Dockerfile est dans ~/speaches/src. Soit tu te déplaces, soit tu précises le chemin :

docker build -f ~/speaches/src/Dockerfile.rocm-ct2 -t speaches:rocm-ct2 ~/speaches/src/

Ou alors recrée le fichier directement au bon endroit :

cat > ~/speaches/src/Dockerfile.rocm-ct2 << 'EOF'
FROM rocm/dev-ubuntu-24.04:6.3.4-complete AS builder

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    git cmake ninja-build python3-dev python3-pip \
    && rm -rf /var/lib/apt/lists/*

RUN git clone --recurse-submodules \
    https://github.com/arlo-phoenix/CTranslate2-rocm.git /ct2

RUN cd /ct2 && \
    PYTORCH_ROCM_ARCH=gfx1101 \
    HIP_PATH=/opt/rocm-6.3.4 \
    cmake -S . -B build \
        -DCMAKE_CXX_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DCMAKE_C_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang \
        -DCMAKE_HIP_COMPILER=/opt/rocm-6.3.4/lib/llvm/bin/clang++ \
        -DWITH_MKL=OFF \
        -DOPENMP_RUNTIME=COMP \
        -DWITH_HIP=ON \
        -DCMAKE_HIP_ARCHITECTURES=gfx1101 \
        -DBUILD_TESTS=OFF \
        -DWITH_CUDNN=ON \
        -DCUDnn_INCLUDE_DIR=/opt/rocm-6.3.4/include \
        -DCUDnn_LIBRARY=/opt/rocm-6.3.4/lib/libMIOpen.so \
        -GNinja && \
    ninja -C build -j$(nproc) && \
    cmake --install build --prefix /usr/local && \
    ldconfig

RUN cd /ct2/python && \
    pip3 install --break-system-packages -r install_requirements.txt && \
    python3 setup.py bdist_wheel && \
    cp dist/*.whl /

FROM rocm/dev-ubuntu-24.04:6.3.4-complete

COPY --from=builder /ctranslate2*.whl /
COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/
COPY --from=builder /opt/rocm-6.3.4/lib/llvm/lib/libomp.so /usr/local/lib/
RUN ldconfig

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    python3-pip ffmpeg \
    && rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv

RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true
WORKDIR /home/ubuntu/speaches
COPY --chown=ubuntu . .

RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && \
    /bin/uv pip install /ctranslate2*.whl --python .venv

ENV HSA_OVERRIDE_GFX_VERSION=11.0.0
ENV LD_LIBRARY_PATH=/usr/local/lib:/opt/rocm-6.3.4/lib

EXPOSE 8000
CMD [".venv/bin/uvicorn", "speaches.main:create_app", \
     "--host", "0.0.0.0", "--port", "8000", "--factory"]
EOF

Puis :

cd ~/speaches/src
docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

normalement il a fini : folzo@folzo-Corp:~/speaches$ cd ~/speaches/src

docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 . [+] Building 192.6s (22/22) FINISHED docker:default => [internal] load build definition from Dockerfile.rocm-ct2 0.0s => => transferring dockerfile: 2.14kB 0.0s => [internal] load metadata for docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete 0.0s => [internal] load metadata for ghcr.io/astral-sh/uv:0.8.22 0.4s => [internal] load .dockerignore 0.0s => => transferring context: 110B 0.0s => CACHED [builder 1/5] FROM docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32 0.0s => => resolve docker.io/rocm/dev-ubuntu-24.04:6.3.4-complete@sha256:76e99e263ef6ce69ba5d32905623c801fff3f 0.0s => CACHED FROM ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a 0.0s => => resolve ghcr.io/astral-sh/uv:0.8.22@sha256:9874eb7afe5ca16c363fe80b294fe700e460df29a55532bbfea234a0 0.0s => [internal] load build context 0.0s => => transferring context: 4.06kB 0.0s => CACHED [builder 2/5] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y git c 0.0s => CACHED [builder 3/5] RUN git clone --recurse-submodules https://github.com/arlo-phoenix/CTranslate 0.0s => [builder 4/5] RUN cd /ct2 && PYTORCH_ROCM_ARCH=gfx1101 HIP_PATH=/opt/rocm-6.3.4 cmake -S 50.1s => [builder 5/5] RUN cd /ct2/python && pip3 install --break-system-packages -r install_requirements. 16.9s => [stage-1 2/11] COPY --from=builder /ctranslate2*.whl / 0.4s => [stage-1 3/11] COPY --from=builder /usr/local/lib/libctranslate2* /usr/local/lib/ 0.1s => [stage-1 4/11] COPY --from=builder /opt/rocm-6.3.4/lib/llvm/lib/libomp.so /usr/local/lib/ 0.1s => [stage-1 5/11] RUN ldconfig 0.2s => [stage-1 6/11] RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y python3-p 56.2s => [stage-1 7/11] COPY --from=ghcr.io/astral-sh/uv:0.8.22 /uv /bin/uv 0.2s => [stage-1 8/11] RUN useradd --create-home --shell /bin/bash --uid 1000 ubuntu || true 0.2s => [stage-1 9/11] WORKDIR /home/ubuntu/speaches 0.1s => [stage-1 10/11] COPY --chown=ubuntu . . 0.1s => [stage-1 11/11] RUN /bin/uv sync --frozen --compile-bytecode --no-dev --no-editable && /bin/uv pip 8.8s => exporting to image 58.3s => => exporting layers 44.2s => => exporting manifest sha256:b20e7cd9203afb6585754225862fa50e5bcb86841e0f50146c7aa30e63a4b0e8 0.0s => => exporting config sha256:9256afc3cfda94cba45c1f98afd0e1ac2e7d1bfa5dd224cd08a504bb53e209f1 0.0s => => exporting attestation manifest sha256:42e2c94beb748e1721b3e660d0ce28d71b5e3da2a5356619a80a6cbde2548 0.1s => => exporting manifest list sha256:aa096042ec20418b0f5ea4def6ab3432b5d411daf55a08015ea11672ed93a3af 0.0s => => naming to docker.io/library/speaches:rocm-ct2 0.0s => => unpacking to docker.io/library/speaches:rocm-ct2 13.9s

1 warning found (use docker --debug to expand):

CopyIgnoredFile: Attempting to Copy file "." that is excluded by .dockerignore (line 50) folzo@folzo-Corp:~/speaches/src$ .

Build terminé sans erreur ! Teste d'abord le GPU avant de relancer le stack :

docker run --rm \
  --device /dev/kfd \
  --device /dev/dri \
  -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \
  speaches:rocm-ct2 \
  .venv/bin/python3 -c "
import ctranslate2
print('Version CT2:', ctranslate2.__version__)
print('ROCm compute types:', ctranslate2.get_supported_compute_types('cuda'))
"

Si tu vois float16 → lance le stack :

cd ~/speaches
docker compose down && docker compose up -d
docker logs -f speaches-stt

Et vérifie la VRAM dans un autre terminal :

watch -n1 rocm-smi

Tu devrais voir la VRAM augmenter quand le modèle large-v3 se charge (~3GB), et le GPU s'activer lors d'une transcription.

autre probleme le serveur effi ubunutu ne repond plus comment connaitre son etat ?

Pour diagnostiquer un serveur Ubuntu qui ne répond plus, voici les étapes par ordre de priorité :

Vérifier la connectivité réseau

ping -c 4 <IP_serveur>
traceroute <IP_serveur>

Tenter une connexion SSH

ssh ubuntu@<IP_serveur> -v   # -v pour voir où ça bloque

Si SSH timeout, essaie avec un timeout court :

ssh -o ConnectTimeout=5 ubuntu@<IP_serveur>

Via ton réseau local (si même LAN)

# Scanner les ports ouverts
nmap -Pn <IP_serveur>

# Vérifier si la machine répond ARP (même si ping bloqué)
arp -n | grep <IP_serveur>

Si tu as accès au hyperviseur / Proxmox

# Lister les VMs
qm list

# Console directe sans réseau
qm terminal <VMID>

Si c'est une machine physique accessible

Accès console direct ou IPMI/iDRAC si disponible :

ipmitool -H <IP_IPMI> -U admin -P <pass> chassis status
ipmitool -H <IP_IPMI> -U admin -P <pass> sol activate  # console série

Quel est ton setup pour ce serveur ? (VM Proxmox, machine physique, accès réseau local ou distant ?) → je peux affiner le diagnostic.

le ping repond :

folzo@folzo-Corp:~/speaches$ ping -c 4 192.168.1.27 PING 192.168.1.27 (192.168.1.27) 56(84) bytes of data. 64 bytes from 192.168.1.27: icmp_seq=1 ttl=64 time=2.81 ms 64 bytes from 192.168.1.27: icmp_seq=2 ttl=64 time=2.69 ms 64 bytes from 192.168.1.27: icmp_seq=3 ttl=64 time=2.68 ms 64 bytes from 192.168.1.27: icmp_seq=4 ttl=64 time=2.77 ms

--- 192.168.1.27 ping statistics --- 4 packets transmitted, 4 received, 0% packet loss, time 3006ms rtt min/avg/max/mdev = 2.682/2.738/2.808/0.054 ms folzo@folzo-Corp:/speaches$ traceroute 192.168.1.27 La commande « traceroute » n'a pas été trouvée, mais peut être installée avec : sudo apt install inetutils-traceroute # version 2:2.6-1ubuntu3.1, or sudo apt install traceroute # version 1:2.1.6-1 folzo@folzo-Corp:/speaches$

Masintenant comment voir son etat de charge et l'etat des docker qu'elle a ?

Le ping répond donc la machine est vivante — probablement SSH ou un service qui rame. Voici comment tout voir en une fois :

Connexion SSH + état instantané

ssh ubuntu@192.168.1.27

Une fois connecté, charge système :

# CPU / RAM / load average en temps réel
htop

# ou version simple sans htop
top -b -n 1 | head -20

# Charge en une ligne
uptime

État des conteneurs Docker

# Tous les conteneurs (running + stopped)
docker ps -a

# Stats CPU/RAM en temps réel de tous les conteneurs
docker stats

# Logs d'un conteneur spécifique
docker logs --tail 50 -f <nom_conteneur>

Vérifier si c'est le disque qui sature (cause fréquente de freeze)

# Espace disque
df -h

# I/O disque en temps réel
iostat -x 1 5

# Processus qui écrivent/lisent le plus
iotop -o

Si SSH ne répond toujours pas

Le ping répond mais SSH peut être surchargé ou crashé :

# Depuis ta machine, tenter sur le port SSH
nc -zv 192.168.1.27 22

# Si le port est fermé → service SSH down, faut aller en console physique

Qu'est-ce que tu vois avec ssh ubuntu@192.168.1.27 ? Timeout, connexion refusée, ou ça se connecte ? ²¹²²²³

⁂

j'ai fait un reboot au server mais du coup il y a un probleme de connection avec openwebui : Comment faire pour trouver la panne ?

L'erreur 500: Open WebUI: Server Connection Error signifie qu'Open WebUI tourne mais ne peut pas joindre le backend (Ollama ou l'API configurée). Après un reboot, c'est souvent un conteneur qui n'a pas redémarré.

1. Vérifier l'état des conteneurs

ssh ubuntu@192.168.1.27
docker ps -a

Cherche les conteneurs avec le status Exited ou Restarting — ce sont les coupables.

2. Redémarrer les conteneurs arrêtés

# Si tu utilises docker compose
cd ~/ton-projet-openwebui
docker compose up -d

# Ou forcer le redémarrage de tout
docker compose restart

3. Voir les logs du conteneur en erreur

# Open WebUI
docker logs --tail 50 open-webui

# Ollama (ou le backend LLM)
docker logs --tail 50 ollama

4. Vérifier que les conteneurs ont bien `restart: unless-stopped`

docker inspect open-webui | grep RestartPolicy
docker inspect ollama | grep RestartPolicy

Si la policy est no, les conteneurs ne redémarrent pas automatiquement après un reboot. Corrige dans ton docker-compose.yml :

services:
  ollama:
    restart: unless-stopped
  open-webui:
    restart: unless-stopped

Colle le résultat de docker ps -a et je t'indique exactement quel conteneur relancer. ²⁴

⁂

effi@effi-server:~$ htop

effi@effi-server:$ docker ps -a CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES 0e449d33d747 ghcr.io/open-webui/open-webui:main "bash start.sh" 5 hours ago Up 3 minutes (healthy) 0.0.0.0:3000->8080/tcp, [::]:3000->8080/tcp open-webui 95fde8885273 fedirz/faster-whisper-server:latest-cpu "uv run uvicorn --fa…" 6 hours ago Up 3 minutes 0.0.0.0:8200->8000/tcp, [::]:8200->8000/tcp whisper 68865e9ef0fb ghcr.io/matatonic/openedai-speech "/bin/sh -c 'bash st…" 6 hours ago Up 3 minutes 0.0.0.0:8100->8000/tcp, [::]:8100->8000/tcp openedai-speech 90b176cf78e9 ghcr.io/open-webui/mcpo:main "sh /mcpo-patch.sh" 8 hours ago Up 3 minutes 0.0.0.0:8000->8000/tcp, [::]:8000->8000/tcp mcpo ae9d09e8e441 ghcr.io/open-webui/pipelines:main "bash start.sh" 2 days ago Up 3 minutes 0.0.0.0:9099->9099/tcp, [::]:9099->9099/tcp pipelines 330376345ad2 couchdb:3.3 "tini -- /docker-ent…" 2 days ago Up 3 minutes 4369/tcp, 9100/tcp, 0.0.0.0:5984->5984/tcp, [::]:5984->5984/tcp couchdb b6a9ea6eec2b qdrant/qdrant:latest "./entrypoint.sh" 2 days ago Up 3 minutes 0.0.0.0:6333-6334->6333-6334/tcp, [::]:6333-6334->6333-6334/tcp qdrant c2ae7a572b0c docker.n8n.io/n8nio/n8n:latest "tini -- /docker-ent…" 2 days ago Up 3 minutes 0.0.0.0:5678->5678/tcp, [::]:5678->5678/tcp n8n 8a7b2eb13e12 effi-wol-proxy "python app.py" 2 days ago Up 3 minutes wol-proxy 87661e0e585b hello-world "/hello" 2 days ago Exited (0) 2 days ago naughty_mcnulty effi@effi-server:$ # Open WebUI docker logs --tail 50 open-webui

Ollama (ou le backend LLM)

docker logs --tail 50 ollama 2026-02-21 21:36:30.589 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/a83a7d5d-f320-45ff-9c16-f597a238105d HTTP/1.1" 200 2026-02-21 21:36:41.016 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:36:41.017 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:36:41.027 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /ollama/api/version HTTP/1.1" 500 2026-02-21 21:36:41.030 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/models HTTP/1.1" 200 2026-02-21 21:36:52.016 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:36:52.020 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /ollama/api/version HTTP/1.1" 500 2026-02-21 21:36:53.976 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 2026-02-21 21:37:01.205 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/immutable/chunks/DxrC7fln.js HTTP/1.1" 304 2026-02-21 21:37:01.205 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/immutable/nodes/8.DSQK34AA.js HTTP/1.1" 304 2026-02-21 21:37:01.205 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/immutable/nodes/24.b41V1Uwu.js HTTP/1.1" 304 2026-02-21 21:37:01.209 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/immutable/nodes/5.BXafm3JG.js HTTP/1.1" 304 2026-02-21 21:37:01.209 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/immutable/chunks/BtUDCm8d.js HTTP/1.1" 304 2026-02-21 21:37:03.455 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /c/63dc352c-72a8-466d-925f-76fa60a9a3cb HTTP/1.1" 304 2026-02-21 21:37:03.536 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /manifest.json HTTP/1.1" 200 2026-02-21 21:37:03.573 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/config HTTP/1.1" 200 2026-02-21 21:37:03.616 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/auths/ HTTP/1.1" 200 2026-02-21 21:37:03.658 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/version HTTP/1.1" 200 2026-02-21 21:37:03.660 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/config HTTP/1.1" 200 2026-02-21 21:37:03.700 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/users/user/settings HTTP/1.1" 200 2026-02-21 21:37:03.726 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/configs/banners HTTP/1.1" 200 2026-02-21 21:37:03.727 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/users/c09565eb-ce13-4606-a2a5-2905e3ddb8c9/profile/image HTTP/1.1" 200 2026-02-21 21:37:03.747 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/folders/ HTTP/1.1" 200 2026-02-21 21:37:03.753 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/all/tags HTTP/1.1" 200 2026-02-21 21:37:03.760 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/users/user/settings HTTP/1.1" 200 2026-02-21 21:37:03.777 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/tools/ HTTP/1.1" 200 2026-02-21 21:37:03.792 | INFO | open_webui.routers.openai:get_all_models:490 - get_all_models() 2026-02-21 21:37:03.792 | INFO | open_webui.routers.ollama:get_all_models:322 - get_all_models() 2026-02-21 21:37:03.798 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/pinned HTTP/1.1" 200 2026-02-21 21:37:03.815 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/?page=1 HTTP/1.1" 200 2026-02-21 21:37:03.839 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "POST /api/v1/tasks/active/chats HTTP/1.1" 200 2026-02-21 21:37:03.949 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/?page=2 HTTP/1.1" 200 2026-02-21 21:37:05.079 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/cf5e869f-8a5b-4521-9b39-bc6ba3a7d88c HTTP/1.1" 200 2026-02-21 21:37:05.090 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/88493c4a-d6c7-4962-af7d-993484ccfcbb HTTP/1.1" 200 2026-02-21 21:37:14.015 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:37:25.016 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:37:25.025 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/models HTTP/1.1" 200 2026-02-21 21:37:25.072 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/63dc352c-72a8-466d-925f-76fa60a9a3cb HTTP/1.1" 200 2026-02-21 21:37:25.106 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/chats/63dc352c-72a8-466d-925f-76fa60a9a3cb/tags HTTP/1.1" 200 2026-02-21 21:37:25.137 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/tasks/chat/63dc352c-72a8-466d-925f-76fa60a9a3cb HTTP/1.1" 200 2026-02-21 21:37:25.212 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/v1/tools/ HTTP/1.1" 200 2026-02-21 21:37:48.164 | INFO | open_webui.routers.openai:get_all_models:490 - get_all_models() 2026-02-21 21:37:48.164 | INFO | open_webui.routers.ollama:get_all_models:322 - get_all_models() 2026-02-21 21:37:59.016 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:37:59.017 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:37:59.021 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /ollama/api/version HTTP/1.1" 500 2026-02-21 21:38:03.522 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 2026-02-21 21:38:10.016 | ERROR | open_webui.routers.ollama:send_get_request:104 - Connection error: 2026-02-21 21:38:10.025 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /api/models HTTP/1.1" 200 2026-02-21 21:39:03.523 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - 192.168.1.254:0 - "GET /_app/version.json HTTP/1.1" 200 Error response from daemon: No such container: ollama effi@effi-server:~$

Le problème est clair : le conteneur ollama n'existe pas (No such container: ollama). Open WebUI essaie de s'y connecter en permanence et échoue.

Vérifier où tourne Ollama

Ollama n'est peut-être pas dans ce docker-compose, mais installé directement sur le serveur :

# Service systemd ?
systemctl status ollama

# Processus actif ?
ps aux | grep ollama

# Port 11434 ouvert ?
ss -tlnp | grep 11434

Si Ollama est en systemd (installé natif)

# Le démarrer
sudo systemctl start ollama
sudo systemctl enable ollama  # pour les prochains reboots

# Vérifier
curl http://localhost:11434/api/version

Si Ollama devrait être un conteneur Docker

Retrouve son docker-compose :

find ~ -name "docker-compose*.yml" | xargs grep -l ollama 2>/dev/null

Puis le relancer :

cd /dossier/trouvé
docker compose up -d ollama

Vérifier l'URL configurée dans Open WebUI

Dans tous les cas, vérifie que l'URL Ollama configurée dans Open WebUI correspond à la réalité :

Si Ollama est en systemd natif → l'URL doit être http://host.docker.internal:11434 ou http://172.17.0.1:11434
Si Ollama est en conteneur Docker → l'URL doit être http://ollama:11434 (même réseau Docker)

Colle le résultat de systemctl status ollama et ss -tlnp | grep 11434.

openweb ui a un probleme, il est tres lent repond pas ce n'est pas normal

Vérifions les ressources consommées en temps réel :

# Charge globale CPU/RAM
free -h && uptime

# Top des processus les plus gourmands
ps aux --sort=-%cpu | head -20

# Stats temps réel de tous les conteneurs
docker stats --no-stream

Et spécifiquement open-webui :

# Logs en temps réel pour voir ce qui bloque
docker logs -f --tail 100 open-webui

# Inspecter si le conteneur est vraiment healthy
docker inspect open-webui | grep -A 10 Health

Causes fréquentes après un reboot

RAM saturée — trop de conteneurs qui redémarrent en même temps, Ollama qui charge un gros modèle en mémoire
Disque I/O saturé — les conteneurs écrivent tous leurs données au démarrage
Open WebUI qui tourne des migrations DB au premier démarrage après update

# Vérifier I/O disque
iostat -x 1 3

# Espace disque
df -h

Colle le résultat de docker stats --no-stream — ça donnera immédiatement quel conteneur bouffe les ressources.

effi@effi-server:~$ # Charge globale CPU/RAM

free -h && uptime

Top des processus les plus gourmands

ps aux --sort=-%cpu | head -20

Stats temps réel de tous les conteneurs

docker stats --no-stream total used free shared buff/cache available Mem: 7.6Gi 2.4Gi 3.1Gi 12Mi 2.5Gi 5.2Gi Swap: 4.0Gi 0B 4.0Gi 21:46:52 up 11 min, 1 user, load average: 0.22, 0.22, 0.14 USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND effi 3051 200 0.0 12336 5380 pts/0 R+ 21:46 0:00 ps aux --sort=-%cpu root 1456 4.5 10.0 2857648 801764 ? Ssl 21:35 0:32 /usr/local/bin/python3 -m uvicorn open_webui.main:app --host 0.0.0.0 --port 8080 --forwarded-allow-ips * --workers 1 root 2012 3.6 11.3 4802776 905912 ? SLl 21:35 0:25 python speech.py effi 1684 2.7 2.6 22355548 208828 ? Sl 21:35 0:19 node /usr/local/bin/n8n root 1950 2.1 2.2 598904 179932 ? Sl 21:35 0:15 /root/faster-whisper-server/.venv/bin/python /root/faster-whisper-server/.venv/bin/uvicorn --factory faster_whisper_server.main:create_app 5984 1918 1.6 0.8 2947568 64928 ? Sl 21:35 0:11 /opt/couchdb/bin/../erts-12.3.2.15/bin/beam.smp -SDio 16 -zdbbl 32768 -sbwt none -sbwtdcpu none -sbwtdio none -Bd -- -root /opt/couchdb/bin/.. -progname couchdb -- -home /opt/couchdb -- -boot /opt/couchdb/bin/../releases/3.3.3/couchdb -kernel inet_dist_listen_min 9100 -kernel inet_dist_listen_max 9100 -kernel error_logger silent -sasl sasl_error_logger false -kernel prevent_overlapping_partitions false -noshell -noinput -- -config /opt/couchdb/bin/../releases/3.3.3/sys.config -- effi 2068 0.7 1.3 21695696 104172 ? Sl 21:35 0:05 node --disallow-code-generation-from-strings --disable-proto=delete /usr/local/lib/node_modules/n8n/node_modules/.pnpm/@n8n+task-runner@file+packages+@n8n+task-runner_@opentelemetry+api@1.9.0_@opentelemetry_c2ea36637674b99f45b42fa0a37672ba/node_modules/@n8n/task-runner/dist/start.js root 1463 0.7 0.9 87440 73360 ? Ss 21:35 0:05 /app/.venv/bin/python3 /app/.venv/bin/mcpo --config /app/config/config.json root 1917 0.5 0.8 254576 67796 ? Sl 21:35 0:03 /usr/local/bin/python /usr/local/bin/uvicorn main:app --host 0.0.0.0 --port 9099 --forwarded-allow-ips * --loop auto root 878 0.4 1.1 3389440 92432 ? Ssl 21:35 0:02 /usr/bin/dockerd -H fd:// --containerd=/run/containerd/containerd.sock root 1838 0.3 0.0 0 0 ? Z 21:35 0:02 [mcpo] root 740 0.2 0.6 2388572 54624 ? Ssl 21:34 0:01 /usr/bin/containerd root 1500 0.2 0.4 536376 36360 ? Ssl 21:35 0:01 python app.py root 2043 0.2 0.0 0 0 ? Z 21:35 0:01 [python] root 1757 0.2 1.7 648280 140876 ? Sl 21:35 0:01 ./qdrant root 2107 0.2 0.7 212724 57184 ? Sl 21:35 0:01 /root/.cache/uv/archive-v0/vZnPVCcyBNrvSAUg9sMQn/bin/python /root/.cache/uv/archive-v0/vZnPVCcyBNrvSAUg9sMQn/bin/mcp-obsidian root 1 0.1 0.1 22256 13512 ? Ss 21:34 0:01 /sbin/init root 324 0.0 0.2 66856 17348 ? S<s 21:34 0:00 /usr/lib/systemd/systemd-journald root 3032 0.0 0.0 4948 3032 ? S 21:46 0:00 jq -ne input.status == true CONTAINER ID NAME CPU % MEM USAGE / LIMIT MEM % NET I/O BLOCK I/O PIDS 0e449d33d747 open-webui 0.01% 972.9MiB / 7.595GiB 12.51% 411kB / 255kB 1.12GB / 1.9MB 20 95fde8885273 whisper 1.19% 248.8MiB / 7.595GiB 3.20% 7.06kB / 1.69kB 331MB / 0B 9 68865e9ef0fb openedai-speech 0.37% 1.06GiB / 7.595GiB 13.95% 228kB / 5.69kB 1.35GB / 434kB 9 90b176cf78e9 mcpo 0.39% 177.1MiB / 7.595GiB 2.28% 20.1kB / 85.5kB 234MB / 35.7MB 8 ae9d09e8e441 pipelines 0.29% 84.11MiB / 7.595GiB 1.08% 7.43kB / 4.75kB 121MB / 20.5kB 7 330376345ad2 couchdb 0.33% 64.22MiB / 7.595GiB 0.83% 1.86kB / 126B 59.1MB / 0B 34 b6a9ea6eec2b qdrant 0.28% 137MiB / 7.595GiB 1.76% 8.89kB / 4.61kB 102MB / 90.1kB 24 c2ae7a572b0c n8n 2.79% 295.4MiB / 7.595GiB 3.80% 26.7kB / 27.2kB 381MB / 111MB 20 8a7b2eb13e12 wol-proxy 0.02% 42.49MiB / 7.595GiB 0.55% 0B / 0B 53.7MB / 0B 2 effi@effi-server:~$ # Vérifier I/O disque iostat -x 1 3

Espace disque

df -h Linux 6.8.0-100-generic (effi-server) 02/21/2026 x86_64 (4 CPU)

avg-cpu: %user %nice %system %iowait %steal %idle 3.83 0.00 1.42 0.26 0.00 94.49

Device r/s rkB/s rrqm/s %rrqm r_await rareq-sz w/s wkB/s wrqm/s %wrqm w_await wareq-sz d/s dkB/s drqm/s %drqm d_await dareq-sz f/s f_await aqu-sz %util dm-0 111.08 2885.50 0.00 0.00 0.41 25.98 9.91 145.75 0.00 0.00 0.82 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 2.14 loop0 0.01 0.02 0.00 0.00 0.00 1.27 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme0n1 86.12 2904.90 27.06 23.91 0.29 33.73 5.47 145.92 5.23 48.90 1.21 26.70 0.00 0.00 0.00 0.00 0.00 0.00 0.70 1.04 0.03 1.20

avg-cpu: %user %nice %system %iowait %steal %idle 3.53 0.00 0.76 0.00 0.00 95.72

Device r/s rkB/s rrqm/s %rrqm r_await rareq-sz w/s wkB/s wrqm/s %wrqm w_await wareq-sz d/s dkB/s drqm/s %drqm d_await dareq-sz f/s f_await aqu-sz %util dm-0 0.00 0.00 0.00 0.00 0.00 0.00 7.00 28.00 0.00 0.00 0.00 4.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 loop0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme0n1 0.00 0.00 0.00 0.00 0.00 0.00 7.00 28.00 0.00 0.00 0.57 4.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

avg-cpu: %user %nice %system %iowait %steal %idle 0.51 0.00 0.51 0.00 0.00 98.99

Device r/s rkB/s rrqm/s %rrqm r_await rareq-sz w/s wkB/s wrqm/s %wrqm w_await wareq-sz d/s dkB/s drqm/s %drqm d_await dareq-sz f/s f_await aqu-sz %util dm-0 0.00 0.00 0.00 0.00 0.00 0.00 2.00 8.00 0.00 0.00 1.00 4.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.10 loop0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 nvme0n1 0.00 0.00 0.00 0.00 0.00 0.00 1.00 8.00 1.00 50.00 2.00 8.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.10

Filesystem Size Used Avail Use% Mounted on tmpfs 778M 2.0M 776M 1% /run efivarfs 128K 60K 64K 49% /sys/firmware/efi/efivars /dev/mapper/ubuntu--vg-ubuntu--lv 98G 38G 56G 41% / tmpfs 3.8G 0 3.8G 0% /dev/shm tmpfs 5.0M 0 5.0M 0% /run/lock /dev/nvme0n1p2 2.0G 103M 1.7G 6% /boot /dev/nvme0n1p1 1.1G 6.2M 1.1G 1% /boot/efi tmpfs 778M 12K 778M 1% /run/user/1000 effi@effi-server:~$

Les ressources sont saines (CPU idle à 95%, RAM OK, disque OK). Le vrai problème est double :

1. Ollama est absent → Open WebUI tente de le joindre toutes les 11 secondes, ça génère des timeouts qui ralentissent l'UI.

2. Deux processus zombies (mcpo et python) qui indiquent des crashs silencieux.

Corriger Ollama en priorité

Ollama tourne en systemd natif sur ce serveur (pas en Docker) :

systemctl status ollama
sudo systemctl start ollama
curl http://localhost:11434/api/version

Si pas installé du tout :

curl -fsSL https://ollama.com/install.sh | sh
sudo systemctl enable --now ollama

Vérifier l'URL Ollama dans Open WebUI

Une fois Ollama démarré, vérifie que Open WebUI pointe vers la bonne adresse. Dans l'interface → Admin Panel → Settings → Connections, l'URL Ollama doit être :

http://host.docker.internal:11434

Ou depuis le serveur, vérifie la variable d'env du conteneur :

docker inspect open-webui | grep -i ollama

Nettoyer les zombies mcpo/python

# Redémarrer mcpo qui a un zombie
docker restart mcpo

# Vérifier
docker ps | grep mcpo
ps aux | grep defunct

Commence par systemctl status ollama — c'est la cause racine de tout le problème.

ce n'est pas ollama mais un wol proxy qui tourne sur ce serveur, celui-ci a pour objectif de pinguer le pc ou se trouve ollama et de lancer un wol si besoin

Ah voilà ! Le PC avec Ollama est éteint ou inaccessible. Le wol-proxy est censé le réveiller mais Open WebUI continue de recevoir des erreurs de connexion en attendant.

Vérifier l'état du PC Ollama

# Le PC répond-il ?
ping -c 4 <IP_PC_Ollama>

# Le port Ollama est-il accessible ?
nc -zv <IP_PC_Ollama> 11434

Vérifier les logs du wol-proxy

docker logs --tail 50 -f wol-proxy

Il devrait montrer s'il a tenté d'envoyer le magic packet et si le PC a répondu.

Forcer manuellement un WOL

Si le proxy n'a pas déclenché le réveil :

# Depuis le serveur effi
docker exec wol-proxy python -c "import app; app.wake()"

# Ou avec wakeonlan directement
sudo apt install wakeonlan
wakeonlan <MAC_adresse_PC_Ollama>

Si le PC est allumé mais Ollama ne répond pas

# Tester directement l'API Ollama
curl http://<IP_PC_Ollama>:11434/api/version

# Vérifier si le service Ollama a redémarré automatiquement sur le PC
ssh user@<IP_PC_Ollama> "systemctl status ollama"

Quelle IP a le PC avec Ollama ? Et les logs du wol-proxy montrent quoi ?

c'est bon j'ai trouvé, l'interface reseau filaire de ubuntu avais planté j'ai fait off on et s'est revenu. C'est souvant sur ubuntu ?

Oui, ça arrive sur Ubuntu, surtout dans certains contextes. Les causes les plus fréquentes :

Causes connues

Driver réseau qui se plante — particulièrement avec certaines cartes Intel (i225/i226) qui ont des bugs firmware bien documentés sous Linux
NetworkManager qui déraille après un reboot si des services démarrent dans le mauvais ordre
Carte réseau en power saving agressif — le kernel met la carte en veille et elle ne se réveille pas correctement
Mise à jour de kernel qui change le comportement du driver sans redémarrage propre

Vérifier si c'est récurrent

# Voir les erreurs réseau dans les logs système
journalctl -b -1 | grep -i "eth\|ens\|enp\|network\|link"

# Identifier ta carte réseau
lspci | grep -i ethernet

Fix préventif — désactiver le power saving

# Identifier l'interface
ip link show

# Désactiver le power management (remplace enp3s0 par ton interface)
sudo ethtool -s enp3s0 wol g
echo 'ACTION=="add", SUBSYSTEM=="net", NAME=="enp3s0", RUN+="/usr/sbin/ethtool -s enp3s0 wol g"' \
  | sudo tee /etc/udev/rules.d/99-disable-net-powersave.rules

Fix NetworkManager au boot

sudo systemctl restart NetworkManager
# S'assurer qu'il démarre après les dépendances réseau
sudo systemctl enable NetworkManager-wait-online.service

Quelle est ta carte réseau (lspci | grep -i ethernet) ? Si c'est une Intel i225/i226, il y a un fix kernel spécifique connu.

ca arrive souvent ?

s/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:17:11 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/191711.196852:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (04188B92404DAAE85867E49F8A16C6AF) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:19:37 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/191937.416298:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:19:37 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/191937.416312:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (9E3E98679BE446CA24EEF9489FF30765) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:20:08 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192008.076312:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:20:08 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192008.076327:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (48DDE1311F8337B23E8734D061A9B52E) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:21:47 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192147.497511:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:21:47 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192147.497523:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (AE474897B0B31097DDE2F35E47F1DB67) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:23:01 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192301.462650:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:23:01 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192301.462662:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (2A5C1470A844C4030024E37347990544) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:25:14 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192514.804885:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:25:14 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/192514.804896:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (C30E435F582773C0428A92B6F0F71498) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:30:05 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193005.010784:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:30:05 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193005.010795:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (28F5E7DD906FAE47DB9D35EC79A5F138) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:31:17 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193117.054763:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:31:17 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193117.054779:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (448E3CF3870F56D436A7611F42C9A287) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:31:57 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193157.559701:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:31:57 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193157.559715:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (5D2EA3F1077E40D3914FECED42A18849) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:33:35 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193335.526589:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:33:35 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193335.526601:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (6EEF8EE784D95BD6F2DB10FDB2129A9E) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:34:15 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193415.826222:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:34:15 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193415.826238:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (82B2DBAFB5108032EF0936D4C13D5E38) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:35:56 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193556.246806:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:35:56 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193556.246818:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (7520A764A864F58DB58B10852F915706) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:36:27 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193627.781403:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:36:27 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193627.781419:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (A4E2D28C17C7F0C915226D5774D15E52) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:36:58 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193658.183636:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:36:58 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193658.183648:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (FF29E2BDE203E228149EA523DDD33307) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:38:57 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193857.379396:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:38:57 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193857.379408:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (363870A292CDAD221740C19586226856) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:39:30 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193930.516099:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:39:30 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/193930.516112:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (691F76D8B89B69576B7BE1C2E76BA78C) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:42:21 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194221.881442:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:42:21 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194221.881454:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (82806B4A4937FA7B5FBA16C17E661CA2) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:47:17 folzo-Corp systemd[1]: Starting update-notifier-motd.service - Check to see whether there is a new version of Ubuntu available... févr. 19 19:47:17 folzo-Corp systemd[1]: Finished update-notifier-motd.service - Check to see whether there is a new version of Ubuntu available. févr. 19 19:47:21 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194721.882753:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:47:21 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194721.882769:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (E0B233411021C83672A77E000A95D4C5) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:49:03 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194903.449104:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:49:03 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194903.449117:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (C5FA9A66450D5A0D3DCA383B82FED8F9) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:49:33 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194933.793632:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:49:33 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/194933.793647:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (4B89E2D1521A6DAA82677F0597C3AA09) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 19:50:54 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/195054.113204:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 19:50:54 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/195054.113216:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (577EB19E7D1BC2898D3E41843F5693B5) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 20:49:03 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/204903.453338:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 20:49:03 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/204903.453352:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (43D85A5A7A9ED714E6B2324A4268E3D4) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 21:31:30 folzo-Corp NetworkManager[1702]: [1771533090.9087] dhcp4 (wlp5s0): state changed new lease, address=192.168.1.12 févr. 19 21:31:30 folzo-Corp dbus-daemon[1610]: [system] Activating via systemd: service name='org.freedesktop.nm_dispatcher' unit='dbus-org.freedesktop.nm-dispatcher.service' requested by ':1.10' (uid=0 pid=1702 comm="/usr/sbin/NetworkManager --no-daemon" label="unconfined") févr. 19 21:31:30 folzo-Corp systemd[1]: Starting NetworkManager-dispatcher.service - Network Manager Script Dispatcher Service... févr. 19 21:31:30 folzo-Corp systemd[1]: Started NetworkManager-dispatcher.service - Network Manager Script Dispatcher Service. févr. 19 21:31:40 folzo-Corp systemd[1]: NetworkManager-dispatcher.service: Deactivated successfully. févr. 19 21:49:03 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/214903.449550:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 21:49:03 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/214903.449561:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (0798AA464C584278E24FF480D31072E6) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 21:57:40 folzo-Corp NetworkManager[1702]: [1771534660.6455] dhcp4 (enp6s0): state changed new lease, address=192.168.1.11 févr. 19 21:57:40 folzo-Corp NetworkManager[1702]: [1771534660.6460] device (enp6s0): conflict detected for IP address 192.168.1.11 with host 5C:87:9C:65:95:23 févr. 19 21:57:40 folzo-Corp dbus-daemon[1610]: [system] Activating via systemd: service name='org.freedesktop.nm_dispatcher' unit='dbus-org.freedesktop.nm-dispatcher.service' requested by ':1.10' (uid=0 pid=1702 comm="/usr/sbin/NetworkManager --no-daemon" label="unconfined") févr. 19 21:57:40 folzo-Corp systemd[1]: Starting NetworkManager-dispatcher.service - Network Manager Script Dispatcher Service... févr. 19 21:57:40 folzo-Corp systemd[1]: Started NetworkManager-dispatcher.service - Network Manager Script Dispatcher Service. févr. 19 21:57:50 folzo-Corp systemd[1]: NetworkManager-dispatcher.service: Deactivated successfully. févr. 19 21:58:26 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/215826.649279:ERROR:chromium/services/network/restricted_cookie_manager.cc:1148] site_for_cookies from renderer='SiteForCookies: {site=null; schemefully_same=false}' from browser='SiteForCookies: {site=https://perplexity.ai; schemefully_same=true}'; févr. 19 21:58:26 folzo-Corp vivaldi_vivaldi-stable.desktop[8143]: [8143:8156:0219/215826.649291:ERROR:chromium/services/network/restricted_cookie_manager.cc:1157] top_frame_origin from renderer='null [internally: (E1C5D4E93809A6F2E3C3F761B0CF68A4) anonymous]' from browser='https://www.perplexity.ai'; févr. 19 21:58:27 folzo-Corp vivaldi_vivaldi-stable.desktop[7944]: [7944:7944:0219/215827.217266:ERROR:CONSOLE:0] "Unchecked runtime.lastError: The browser is shutting down.", source: chrome-extension://mpognobbkildjkofajifpdfhcoklimli/main.html (0) févr. 19 21:58:27 folzo-Corp vivaldi_vivaldi-stable.desktop[7944]: at _handleInProgressSearch (chrome-extension://mpognobbkildjkofajifpdfhcoklimli/bundle.js:1:132491)", source: chrome-extension://mpognobbkildjkofajifpdfhcoklimli/main.html (0) févr. 19 21:58:27 folzo-Corp vivaldi_vivaldi-stable.desktop[7944]: [7944:7944:0219/215827.217310:ERROR:CONSOLE:0] "Unchecked runtime.lastError: The browser is shutting down.", source: chrome-extension://mpognobbkildjkofajifpdfhcoklimli/main.html (0) févr. 19 21:58:27 folzo-Corp vivaldi_vivaldi-stable.desktop[7944]: [7944:7944:0219/215827.217330:ERROR:CONSOLE:0] "Unchecked runtime.lastError: The browser is shutting down.", source: chrome-extension://mpognobbkildjkofajifpdfhcoklimli/main.html (0) févr. 19 21:58:27 folzo-Corp vivaldi_vivaldi-stable.desktop[7944]: at chrome-extension://mpognobbkildjkofajifpdfhcoklimli/bundle.js:1:134020", source: chrome-extension://mpognobbkildjkofajifpdfhcoklimli/main.html (0) févr. 19 21:58:27 folzo-Corp vivaldi_vivaldi-stable.desktop[7944]: [7944:7944:0219/215827.217350:ERROR:CONSOLE:0] "Unchecked runtime.lastError: The browser is shutting down.", source: chrome-extension://mpognobbkildjkofajifpdfhcoklimli/main.html (0) févr. 19 21:58:39 folzo-Corp systemd[5000]: app-gnome-gnome\x2dnetwork\x2dpanel-274879.scope: Consumed 2.322s CPU time, 87.2M memory peak. févr. 19 21:58:51 folzo-Corp com.heroicgameslauncher.hgl.desktop[1810576]: which: no wine in (/app/bin:/app/bin:/app/utils/bin:/usr/bin:/usr/lib/extensions/vulkan/MangoHud/bin:/usr/lib/extensions/vulkan/gamescope/bin:/usr/lib/extensions/vulkan/OBSVkCapture/bin:/app/bin/heroic/resources/app.asar.unpacked/build/bin/linux) févr. 19 21:58:54 folzo-Corp com.heroicgameslauncher.hgl.desktop[1810576]: which: no wine in (/app/bin:/app/bin:/app/utils/bin:/usr/bin:/usr/lib/extensions/vulkan/MangoHud/bin:/usr/lib/extensions/vulkan/gamescope/bin:/usr/lib/extensions/vulkan/OBSVkCapture/bin:/app/bin/heroic/resources/app.asar.unpacked/build/bin/linux) févr. 19 21:59:05 folzo-Corp systemd[1]: Stopped target nss-lookup.target - Host and Network Name Lookups. févr. 19 21:59:05 folzo-Corp systemd[1]: Stopped update-notifier-motd.timer - Check to see whether there is a new version of Ubuntu available. févr. 19 21:59:05 folzo-Corp systemd[5000]: Stopped target org.gnome.SettingsDaemon.ScreensaverProxy.target - GNOME FreeDesktop screensaver target. févr. 19 21:59:05 folzo-Corp NetworkManager[1702]: [1771534745.6229] modem-manager: ModemManager no longer available févr. 19 21:59:05 folzo-Corp systemd[1]: Stopping networkd-dispatcher.service - Dispatcher daemon for systemd-networkd... févr. 19 21:59:05 folzo-Corp systemd[1]: networkd-dispatcher.service: Deactivated successfully. févr. 19 21:59:05 folzo-Corp systemd[1]: Stopped networkd-dispatcher.service - Dispatcher daemon for systemd-networkd. févr. 19 21:59:05 folzo-Corp dbus-daemon[1610]: [system] Rejected send message, 1 matched rules; type="method_call", sender=":1.3" (uid=0 pid=1608 comm="/usr/libexec/bluetooth/bluetoothd" label="unconfined") interface="org.bluez.GattProfile1" member="Release" error name="(unset)" requested_reply="0" destination=":1.86" (uid=1000 pid=5046 comm="/usr/bin/wireplumber" label="unconfined") févr. 19 21:59:05 folzo-Corp systemd[1]: Stopped target network-online.target - Network is Online. févr. 19 21:59:05 folzo-Corp systemd[1]: NetworkManager-wait-online.service: Deactivated successfully. févr. 19 21:59:05 folzo-Corp systemd[1]: Stopped NetworkManager-wait-online.service - Network Manager Wait Online. févr. 19 21:59:05 folzo-Corp avahi-daemon[1607]: Leaving mDNS multicast group on interface enp6s0.IPv4 with address 192.168.1.11. févr. 19 21:59:05 folzo-Corp dbus-daemon[5026]: [session uid=1000 pid=5026 pidfd=5] Activating via systemd: service name='org.gtk.vfs.Daemon' unit='gvfs-daemon.service' requested by ':1.269' (uid=1000 pid=1812087 comm="gjs /usr/share/gnome-shell/extensions/ding@rasters" label="desktop-icons-ng (unconfined)") févr. 19 21:59:06 folzo-Corp systemd[5000]: Stopping ssh-agent.service - OpenSSH Agent... févr. 19 21:59:06 folzo-Corp systemd[5000]: Stopped ssh-agent.service - OpenSSH Agent. févr. 19 21:59:06 folzo-Corp systemd[5000]: Closed dirmngr.socket - GnuPG network certificate management daemon. févr. 19 21:59:06 folzo-Corp systemd[5000]: Stopping ssh-agent.socket - OpenSSH Agent socket... févr. 19 21:59:06 folzo-Corp systemd[5000]: Closed ssh-agent.socket - OpenSSH Agent socket. févr. 19 21:59:08 folzo-Corp systemd[1]: Stopped target network.target - Network. févr. 19 21:59:08 folzo-Corp systemd[1]: Stopping NetworkManager.service - Network Manager... févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7567] caught SIGTERM, shutting down normally. févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7571] manager: NetworkManager state is now CONNECTED_GLOBAL févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activating via systemd: service name='org.freedesktop.nm_dispatcher' unit='dbus-org.freedesktop.nm-dispatcher.service' requested by ':1.10' (uid=0 pid=1702 comm="/usr/sbin/NetworkManager --no-daemon" label="unconfined") févr. 19 21:59:08 folzo-Corp systemd[1]: Requested transaction contradicts existing jobs: Transaction for NetworkManager-dispatcher.service/start is destructive (integritysetup.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activation via systemd failed for unit 'dbus-org.freedesktop.nm-dispatcher.service': Transaction for NetworkManager-dispatcher.service/start is destructive (integritysetup.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7691] dhcp4 (enp6s0): canceled DHCP transaction févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7692] dhcp4 (enp6s0): activation: beginning transaction (timeout in 45 seconds) févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7692] dhcp4 (enp6s0): state changed no lease févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7827] device (wlp5s0): state change: activated -> deactivating (reason 'unmanaged', managed-type: 'full') févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activating via systemd: service name='org.freedesktop.nm_dispatcher' unit='dbus-org.freedesktop.nm-dispatcher.service' requested by ':1.10' (uid=0 pid=1702 comm="/usr/sbin/NetworkManager --no-daemon" label="unconfined") févr. 19 21:59:08 folzo-Corp systemd[1]: Requested transaction contradicts existing jobs: Transaction for NetworkManager-dispatcher.service/start is destructive (cryptsetup.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activation via systemd failed for unit 'dbus-org.freedesktop.nm-dispatcher.service': Transaction for NetworkManager-dispatcher.service/start is destructive (cryptsetup.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7841] dispatcher: (43) failed (after 0.001 sec): Transaction for NetworkManager-dispatcher.service/start is destructive (cryptsetup.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7842] device (wlp5s0): state change: deactivating -> unmanaged (reason 'unmanaged-quitting', managed-type: 'full') févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7842] dhcp4 (wlp5s0): canceled DHCP transaction févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7843] dhcp4 (wlp5s0): activation: beginning transaction (timeout in 45 seconds) févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.7843] dhcp4 (wlp5s0): state changed no lease févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activating via systemd: service name='org.freedesktop.nm_dispatcher' unit='dbus-org.freedesktop.nm-dispatcher.service' requested by ':1.10' (uid=0 pid=1702 comm="/usr/sbin/NetworkManager --no-daemon" label="unconfined") févr. 19 21:59:08 folzo-Corp systemd[1]: Requested transaction contradicts existing jobs: Transaction for NetworkManager-dispatcher.service/start is destructive (swap.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activation via systemd failed for unit 'dbus-org.freedesktop.nm-dispatcher.service': Transaction for NetworkManager-dispatcher.service/start is destructive (swap.target has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.8447] manager: NetworkManager state is now CONNECTED_LOCAL févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.8448] device (p2p-dev-wlp5s0): state change: disconnected -> unmanaged (reason 'unmanaged-link-not-init', managed-type: 'removed') févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activating via systemd: service name='org.freedesktop.nm_dispatcher' unit='dbus-org.freedesktop.nm-dispatcher.service' requested by ':1.10' (uid=0 pid=1702 comm="/usr/sbin/NetworkManager --no-daemon" label="unconfined") févr. 19 21:59:08 folzo-Corp systemd[1]: Requested transaction contradicts existing jobs: Transaction for NetworkManager-dispatcher.service/start is destructive (systemd-tmpfiles-setup-dev-early.service has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp dbus-daemon[1610]: [system] Activation via systemd failed for unit 'dbus-org.freedesktop.nm-dispatcher.service': Transaction for NetworkManager-dispatcher.service/start is destructive (systemd-tmpfiles-setup-dev-early.service has 'stop' job queued, but 'start' is included in transaction). févr. 19 21:59:08 folzo-Corp NetworkManager[1702]: [1771534748.8749] exiting (success) févr. 19 21:59:08 folzo-Corp systemd[1]: NetworkManager.service: Deactivated successfully. févr. 19 21:59:08 folzo-Corp systemd[1]: Stopped NetworkManager.service - Network Manager. févr. 19 21:59:08 folzo-Corp systemd[1]: NetworkManager.service: Consumed 1min 1.494s CPU time, 31.4M memory peak. févr. 19 21:59:08 folzo-Corp systemd[1]: Stopped target network-pre.target - Preparation for Network. févr. 19 21:59:08 folzo-Corp systemd[1]: Stopping systemd-resolved.service - Network Name Resolution... févr. 19 21:59:08 folzo-Corp systemd[1]: Stopped systemd-resolved.service - Network Name Resolution. 06:00.0 Ethernet controller: Realtek Semiconductor Co., Ltd. RTL8111/8168/8211/8411 PCI Express Gigabit Ethernet Controller (rev 26) folzo@folzo-Corp:~/speaches$

Ces logs viennent de ta machine locale (folzo-Corp), pas du serveur effi. Deux choses distinctes à noter :

Les erreurs `restricted_cookie_manager` → sans gravité

Ce sont des logs Chromium/Vivaldi totalement bénins — un iframe ou web worker sur perplexity.ai avec un contexte cookie null. Ça n'affecte rien, c'est du bruit de debug Chromium. À ignorer.

Le vrai problème repéré : conflit IP

conflict detected for IP address 192.168.1.11 with host 5C:87:9C:65:95:23

Ta machine folzo-Corp et une autre machine sur le réseau se battaient pour la même IP 192.168.1.11. C'est probablement la cause du plantage réseau — le DHCP a attribué la même adresse deux fois.

Fix : IP statiques ou réservations DHCP

Sur ton routeur/box, assigne des IP fixes par MAC address pour éviter les conflits :

Machine	IP fixe suggérée
folzo-Corp (`enp6s0`)	`192.168.1.11` réservée
effi-server	`192.168.1.27` réservée
PC Ollama	IP réservée

Ta carte réseau : Realtek RTL8111

Bonne nouvelle, c'est un driver très stable (r8169 sous Linux). Pas de bug connu de power saving agressif contrairement aux Intel i225. Le plantage vient donc bien du conflit DHCP, pas du driver.

# Vérifier l'IP actuelle et la MAC pour la réservation
ip addr show enp6s0 | grep -E "inet |ether"

ok autre probleme avec la reconnaissance vocal, ca ne marche pas : j'ai une erreur : "error: 400: [error : 500 : Error transcibing chunk : external : 520 server error: Internal server Error for url : https://192.168.1.11:8200/v1/audio/transcriptions]]

L'erreur vient en cascade : Open WebUI → appelle le serveur STT sur 192.168.1.11:8200 → qui répond 500. C'est le conteneur whisper (fedirz/faster-whisper-server) sur le serveur effi qui plante.

Vérifier les logs du whisper

ssh ubuntu@192.168.1.27
docker logs --tail 50 -f whisper

Causes probables

1. Le modèle n'est pas encore chargé — au démarrage il télécharge/charge le modèle, ça peut prendre quelques minutes :

docker logs whisper 2>&1 | grep -i "model\|load\|error\|cuda\|cpu"

2. Problème de mémoire RAM — openedai-speech prend déjà 1.06GB, whisper en prend ~500MB en plus :

free -h

3. Le modèle configuré n'existe pas — vérifier quelle variable d'env est définie :

docker inspect whisper | grep -A 20 '"Env"'

Tester directement le endpoint

# Depuis effi-server, tester l'API whisper directement
curl -X POST http://localhost:8200/v1/audio/transcriptions \
  -H "Authorization: Bearer doesntmatter" \
  -F "file=@/tmp/test.mp3" \
  -F "model=Systran/faster-whisper-large-v3"

Colle le docker logs --tail 50 whisper — l'erreur 500 interne sera visible dedans.

effi@effi-server:~$ curl -X POST http://192.168.1.11:8200/v1/audio/transcriptions -H "Authorization: Bearer doesntmatter" -F "file=@/tmp/test.mp3" -F "model=Systran/faster-whisper-large-v3"

curl: (26) Failed to open/read local data from file/application effi@effi-server:~$

folzo@folzo-Corp:/speaches$ docker ps CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES d19885b93bcd speaches:rocm-ct2 ".venv/bin/uvicorn s…" 33 minutes ago Up 33 minutes 0.0.0.0:8200->8000/tcp, [::]:8200->8000/tcp speaches-stt folzo@folzo-Corp:/speaches$ docker compose down && docker compose up -d [+] down 3/3 ✔ Container speaches-init Removed 0.1s ✔ Container speaches-stt Removed 1.3s ✔ Network speaches_default Removed 0.1s [+] up 3/3 ✔ Network speaches_default Created 0.0s ✔ Container speaches-stt Created 0.1s ✔ Container speaches-init Created 0.1s folzo@folzo-Corp:~/speaches$

Attention, il y a une confusion importante : le conteneur whisper qui reçoit les requêtes tourne sur folzo-Corp (toi), pas sur effi-server !

folzo-Corp → speaches:rocm-ct2 sur le port 8200 ← c'est celui-là qui est appelé
effi-server → fedirz/faster-whisper-server sur le port 8200 ← celui-là n'est pas utilisé

Vérifier les logs de speaches sur folzo-Corp

docker logs --tail 50 -f speaches-stt

Tester avec un vrai fichier audio

# Créer un fichier de test (si tu n'en as pas)
ffmpeg -f lavfi -i "sine=frequency=440:duration=3" /tmp/test.wav

# Tester l'API
curl -X POST http://localhost:8200/v1/audio/transcriptions \
  -H "Authorization: Bearer doesntmatter" \
  -F "file=@/tmp/test.wav" \
  -F "model=Systran/faster-whisper-large-v3"

Le problème MIOpen/cuDNN est peut-être encore là

Le build avec -DWITH_CUDNN=ON a bien compilé, mais il faut vérifier si MIOpen est bien détecté au runtime :

docker logs speaches-stt 2>&1 | grep -iE "cudnn|miopen|conv1d|error|gpu|hip"

Colle le résultat — si tu revois Conv1D on GPU currently requires the cuDNN library, le lien vers libMIOpen.so ne fonctionne pas au runtime.

folzo@folzo-Corp:~/speaches$ docker logs --tail 50 -f speaches-stt

2026-02-21 22:02:10,196:DEBUG:speaches.main:create_app:102:Config: stt_model_ttl=300 tts_model_ttl=300 vad_model_ttl=-1 api_key=None log_level='debug' host='0.0.0.0' port=8000 allow_origins=None enable_ui=True whisper=WhisperConfig(inference_device='auto', device_index=0, compute_type='float16', cpu_threads=0, num_workers=1) loopback_host_url=None chat_completion_base_url='http://localhost:11434/v1' chat_completion_api_key=SecretStr('**********') unstable_ort_opts=OrtOptions(exclude_providers=['TensorrtExecutionProvider'], provider_priority={'CUDAExecutionProvider': 100}, provider_opts={}) otel_exporter_otlp_endpoint=None otel_service_name='speaches' preload_models=[] 2026-02-21 22:02:11,122:DEBUG:urllib3.connectionpool:_new_conn:1051:Starting new HTTPS connection (1): huggingface.co:443 2026-02-21 22:02:11,153:DEBUG:urllib3.connectionpool:_new_conn:1051:Starting new HTTPS connection (1): huggingface.co:443 2026-02-21 22:02:11,298:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "HEAD /api/telemetry/https%3A/api.gradio.app/gradio-initiated-analytics HTTP/11" 200 0 2026-02-21 22:02:11,383:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "HEAD /api/telemetry/https%3A/api.gradio.app/gradio-initiated-analytics HTTP/11" 200 0 new 2026-02-21 22:02:11,395:INFO:speaches.main:create_app:202:

To view the gradio web ui of speaches open your browser and visit:

http://localhost:8000/

Tester l'API

curl -X POST http://localhost:8200/v1/audio/transcriptions \ -H "Authorization: Bearer doesntmatter" \ -F "file=@/tmp/test.wav" \ -F "model=Systran/faster-whisper-large-v3" La commande « ffmpeg » n'a pas été trouvée, mais peut être installée avec : sudo snap install ffmpeg # version 4.3.1, or sudo apt install ffmpeg # version 7:7.1.1-1ubuntu4.2 Voir « snap info ffmpeg » pour des versions additionelles. Internal Server Errorfolzo@folzo-Corp:/speaches$ docker logs speaches-stt 2>&1 | grep -iE "cudnn|miopen|conv1d|error|gpu|hip" docker logs speaches-stt 2>&1 | grep -iE "cudnn|miopen|conv1d|error|gpu|hip" 2026-02-21 22:04:47.797231021 [E:onnxruntime:Default, provider_bridge_ort.cc:2195 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1778 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcublasLt.so.12: cannot open shared object file: No such file or directory 2026-02-21 22:04:47.797249742 [W:onnxruntime:Default, onnxruntime_pybind_state.cc:1055 CreateExecutionProviderInstance] Failed to create CUDAExecutionProvider. Require cuDNN 9.* and CUDA 12.. Please install all dependencies as mentioned in the GPU requirements page (https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements), make sure they're in the PATH, and that your GPU is supported. 2026-02-21 22:04:47.801851599 [E:onnxruntime:Default, provider_bridge_ort.cc:2195 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1778 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcublasLt.so.12: cannot open shared object file: No such file or directory 2026-02-21 22:04:47.801861709 [W:onnxruntime:Default, onnxruntime_pybind_state.cc:1055 CreateExecutionProviderInstance] Failed to create CUDAExecutionProvider. Require cuDNN 9. and CUDA 12.*. Please install all dependencies as mentioned in the GPU requirements page (https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements), make sure they're in the PATH, and that your GPU is supported. INFO: 172.18.0.1:55960 - "POST /v1/audio/transcriptions HTTP/1.1" 500 Internal Server Error ERROR: Exception in ASGI application File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/middleware/errors.py", line 187, in call File "/home/ubuntu/speaches/.venv/lib/python3.12/site-packages/starlette/middleware/errors.py", line 165, in call RuntimeError: parallel_for failed: hipErrorInvalidDeviceFunction: invalid device function folzo@folzo-Corp:/speaches$

Deux problèmes distincts dans ces logs :

Problème 1 : ONNX Runtime cherche CUDA/cuBLAS

Les erreurs libcublasLt.so.12 et CUDAExecutionProvider viennent du modèle Silero VAD qui tourne via ONNX Runtime — il cherche des libs CUDA qui n'existent pas sur ROCm. C'est un warning non bloquant, il fallback en CPU automatiquement.

Problème 2 : `hipErrorInvalidDeviceFunction` ← cause réelle du crash

RuntimeError: parallel_for failed: hipErrorInvalidDeviceFunction: invalid device function

CTranslate2-ROCm a été compilé pour gfx1101 mais ta GPU répond à un kernel HIP incompatible. Vérifie quelle architecture est réellement exposée :

# Sur folzo-Corp, vérifier la GPU vue par ROCm
docker run --rm \
  --device /dev/kfd \
  --device /dev/dri \
  -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \
  speaches:rocm-ct2 \
  /opt/rocm-6.3.4/bin/rocminfo | grep -E "gfx|Name"

Fix probable : forcer le fallback CPU pour CT2

En attendant, force speaches à utiliser le CPU dans ton docker-compose.yml :

services:
  speaches-stt:
    environment:
      - WHISPER__INFERENCE_DEVICE=cpu
      - WHISPER__COMPUTE_TYPE=int8

Puis :

docker compose down && docker compose up -d
docker logs -f speaches-stt

Le CPU sera plus lent (~2-3x) mais fonctionnel. On réglera le GPU après avoir confirmé l'architecture exacte avec rocminfo.

folzo@folzo-Corp:~/speaches$ # Sur folzo-Corp, vérifier la GPU vue par ROCm

docker run --rm \ --device /dev/kfd \ --device /dev/dri \ -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \ speaches:rocm-ct2 \ /opt/rocm-6.3.4/bin/rocminfo | grep -E "gfx|Name" Name: AMD Ryzen 7 5800X3D 8-Core Processor Marketing Name: AMD Ryzen 7 5800X3D 8-Core Processor Vendor Name: CPU Name: gfx1100 Marketing Name: AMD Radeon RX 7800 XT Vendor Name: AMD Name: amdgcn-amd-amdhsa--gfx1100 folzo@folzo-Corp:~/speaches$

Voilà le problème ! Ta GPU est gfx1100 (RX 7800 XT) mais le build a été compilé pour gfx1101. Ce sont deux architectures différentes et les kernels HIP ne sont pas compatibles entre elles.

Rebuild avec gfx1100

Il faut changer 3 occurrences de gfx1101 → gfx1100 dans le Dockerfile :

sed -i 's/gfx1101/gfx1100/g' ~/speaches/src/Dockerfile.rocm-ct2

# Vérifier
grep gfx ~/speaches/src/Dockerfile.rocm-ct2

Tu dois voir :

-DCMAKE_HIP_ARCHITECTURES=gfx1100
PYTORCH_ROCM_ARCH=gfx1100
HSA_OVERRIDE_GFX_VERSION=11.0.0  ← celui-là reste inchangé

Puis rebuild — les étapes 1-3 seront en cache, seule la compilation cmake repart (~50s) :

cd ~/speaches/src
docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Ensuite relance le stack :

cd ~/speaches
docker compose down && docker compose up -d
docker logs -f speaches-stt

cd ~/speaches

docker compose down && docker compose up -d docker logs -f speaches-stt

Tu as juste collé les commandes sans les lancer ou sans coller le résultat. Lance-les :

cd ~/speaches
docker compose down && docker compose up -d
docker logs -f speaches-stt

Et colle ce que tu vois dans le terminal — notamment si tu vois encore hipErrorInvalidDeviceFunction ou au contraire GPU qui se charge correctement.

folzo@folzo-Corp:~/speaches/src$ cd ~/speaches

docker compose down && docker compose up -d docker logs -f speaches-stt [+] down 3/3 ✔ Container speaches-init Removed 0.0s ✔ Container speaches-stt Removed 2.3s ✔ Network speaches_default Removed 0.1s [+] up 3/3 ✔ Network speaches_default Created 0.0s ✔ Container speaches-stt Created 0.1s ✔ Container speaches-init Created 0.1s 2026-02-21 22:11:04,355:DEBUG:speaches.main:create_app:102:Config: stt_model_ttl=300 tts_model_ttl=300 vad_model_ttl=-1 api_key=None log_level='debug' host='0.0.0.0' port=8000 allow_origins=None enable_ui=True whisper=WhisperConfig(inference_device='auto', device_index=0, compute_type='float16', cpu_threads=0, num_workers=1) loopback_host_url=None chat_completion_base_url='http://localhost:11434/v1' chat_completion_api_key=SecretStr('**********') unstable_ort_opts=OrtOptions(exclude_providers=['TensorrtExecutionProvider'], provider_priority={'CUDAExecutionProvider': 100}, provider_opts={}) otel_exporter_otlp_endpoint=None otel_service_name='speaches' preload_models=[] 2026-02-21 22:11:05,251:DEBUG:urllib3.connectionpool:_new_conn:1051:Starting new HTTPS connection (1): huggingface.co:443 2026-02-21 22:11:05,276:DEBUG:urllib3.connectionpool:_new_conn:1051:Starting new HTTPS connection (1): huggingface.co:443 2026-02-21 22:11:05,490:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "HEAD /api/telemetry/https%3A/api.gradio.app/gradio-initiated-analytics HTTP/11" 200 0 2026-02-21 22:11:05,510:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "HEAD /api/telemetry/https%3A/api.gradio.app/gradio-initiated-analytics HTTP/11" 200 0 new 2026-02-21 22:11:05,522:INFO:speaches.main:create_app:202:

To view the gradio web ui of speaches open your browser and visit:

http://localhost:8000/

INFO: Started server process [1] INFO: Waiting for application startup. INFO: Application startup complete. INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit) 2026-02-21 22:11:05,959:INFO:httpx:_send_single_request:1025:HTTP Request: GET https://api.gradio.app/pkg-version "HTTP/1.1 200 OK" 2026-02-21 22:11:05,959:INFO:httpx:_send_single_request:1025:HTTP Request: GET https://api.gradio.app/pkg-version "HTTP/1.1 200 OK" 2026-02-21 22:11:06,061:INFO:httpx:_send_single_request:1025:HTTP Request: GET https://api.gradio.app/pkg-version "HTTP/1.1 200 OK" 2026-02-21 22:11:06,061:INFO:httpx:_send_single_request:1025:HTTP Request: GET https://api.gradio.app/pkg-version "HTTP/1.1 200 OK" 2026-02-21 22:11:07,377:DEBUG:urllib3.connectionpool:_new_conn:1051:Starting new HTTPS connection (1): huggingface.co:443 2026-02-21 22:11:07,550:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "GET /api/models?filter=ctranslate2&filter=automatic-speech-recognition&cardData=True HTTP/11" 200 242732 2026-02-21 22:11:07,725:WARNING:huggingface_hub.repocard_data:init:368:Invalid model-index. Not loading eval results into CardData. Invalid model-index. Not loading eval results into CardData. 2026-02-21 22:11:07,837:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "GET /api/models/Systran/faster-whisper-large-v3/revision/main HTTP/11" 200 4572 Fetching 6 files: 100%|██████████| 6/6 [00:00<00:00, 139810.13it/s] INFO: 172.18.0.3:53654 - "POST /v1/models/Systran/faster-whisper-large-v3 HTTP/1.1" 200 OK 2026-02-21 22:11:25,034:DEBUG:speaches.dependencies:audio_file_dependency:89:Decoding audio file: 00aab224-be8c-4045-9158-bdfd2b9f24c0.mp3, content_type: None, header: Headers({'content-disposition': 'form-data; name="file"; filename="00aab224-be8c-4045-9158-bdfd2b9f24c0.mp3"'}), size: 43629 2026-02-21 22:11:25,109:DEBUG:speaches.dependencies:audio_file_dependency:103:Decoded 5.4s of audio in 0.07514s (RTF: 0.013914737591457432) 2026-02-21 22:11:25,110:DEBUG:asyncio:init:64:Using selector: EpollSelector 2026-02-21 22:11:25,118:DEBUG:speaches.executors.shared.base_model_manager:_load:72:Loading model silero_vad_v5 2026-02-21 22:11:25,119:DEBUG:speaches.executors.shared.base_model_manager:get_ort_providers_with_options:23:Available ONNX Runtime providers: ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'] 2026-02-21 22:11:25,119:DEBUG:speaches.executors.shared.base_model_manager:get_ort_providers_with_options:33:Using ONNX Runtime providers: [('CUDAExecutionProvider', {}), ('CPUExecutionProvider', {})] 2026-02-21 22:11:25.120366595 [E:onnxruntime:Default, provider_bridge_ort.cc:2195 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1778 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcublasLt.so.12: cannot open shared object file: No such file or directory

2026-02-21 22:11:25.120378835 [W:onnxruntime:Default, onnxruntime_pybind_state.cc:1055 CreateExecutionProviderInstance] Failed to create CUDAExecutionProvider. Require cuDNN 9.* and CUDA 12.*. Please install all dependencies as mentioned in the GPU requirements page (https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements), make sure they're in the PATH, and that your GPU is supported. 2026-02-21 22:11:25.123420246 [E:onnxruntime:Default, provider_bridge_ort.cc:2195 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1778 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcublasLt.so.12: cannot open shared object file: No such file or directory

2026-02-21 22:11:25.123429467 [W:onnxruntime:Default, onnxruntime_pybind_state.cc:1055 CreateExecutionProviderInstance] Failed to create CUDAExecutionProvider. Require cuDNN 9.* and CUDA 12.*. Please install all dependencies as mentioned in the GPU requirements page (https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements), make sure they're in the PATH, and that your GPU is supported. 2026-02-21 22:11:25,124:INFO:speaches.executors.shared.base_model_manager:_load:75:Model silero_vad_v5 loaded in 0.01s 2026-02-21 22:11:25,124:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for silero_vad_v5, self.ref_count=1 2026-02-21 22:11:25,148:DEBUG:speaches.executors.silero_vad_v5:call:138:VAD model inference took 0.0240s 2026-02-21 22:11:25,149:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for silero_vad_v5, self.ref_count=0 2026-02-21 22:11:25,149:INFO:speaches.executors.shared.base_model_manager:_decrement_ref:98:Model silero_vad_v5 is idle, not unloading 2026-02-21 22:11:25,149:DEBUG:speaches.executors.silero_vad_v5:get_speech_timestamps:307:VAD processing took 0.0307s for 5.40s audio 2026-02-21 22:11:25,149:DEBUG:speaches.executors.shared.base_model_manager:_load:72:Loading model Systran/faster-whisper-large-v3 2026-02-21 22:11:25,257:DEBUG:urllib3.connectionpool:_make_request:546:https://huggingface.co:443 "GET /api/models/Systran/faster-whisper-large-v3/revision/main HTTP/11" 200 4572 2026-02-21 22:11:33,945:INFO:speaches.executors.shared.base_model_manager:_load:75:Model Systran/faster-whisper-large-v3 loaded in 8.80s 2026-02-21 22:11:33,945:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for Systran/faster-whisper-large-v3, self.ref_count=1 2026-02-21 22:11:33,945:INFO:faster_whisper:transcribe:391:Processing audio with duration 00:05.400 2026-02-21 22:11:33,945:INFO:faster_whisper:transcribe:428:VAD filter removed 00:02.424 of audio 2026-02-21 22:11:38,128:INFO:faster_whisper:transcribe:463:Detected language 'fr' with probability 0.99 2026-02-21 22:11:38,525:INFO:speaches.executors.whisper:handle_non_streaming_transcription_request:181:Transcribed 5.4 seconds of audio in 13.375739333001547 seconds 2026-02-21 22:11:38,525:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for Systran/faster-whisper-large-v3, self.ref_count=0 2026-02-21 22:11:38,525:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:91:Model Systran/faster-whisper-large-v3 is idle, scheduling offload in 300s 2026-02-21 22:11:38,525:ERROR:speaches.routers.stt:transcription_response_to_http_response:105:Unexpected streaming transcription response type: <class 'openai.types.audio.transcription.Transcription'> INFO: 192.168.1.27:39786 - "POST /v1/audio/transcriptions HTTP/1.1" 200 OK ^Cfolzo@folzo-Corp:~/speaches$

La transcription fonctionne ! 200 OK et langue détectée fr avec probabilité 0.99. Mais deux points à noter :

✅ Ce qui marche

Transcription réussie en 200 OK
Langue française détectée à 99%
Modèle large-v3 chargé correctement

⚠️ GPU pas encore utilisé

Le modèle tourne encore en CPU — 13.37s pour 5.4s d'audio (RTF > 1), ce qui confirme qu'il n'utilise pas la RX 7800 XT. Le rebuild avec gfx1100 n'a pas encore été fait.

🔴 Erreur mineure à corriger

Unexpected streaming transcription response type: Transcription

C'est un bug speaches côté format de réponse, non bloquant pour l'instant.

Lance le rebuild maintenant pour activer le GPU :

cd ~/speaches/src
# Vérifier que gfx1101 a bien été remplacé
grep gfx Dockerfile.rocm-ct2

# Si encore gfx1101, corriger
sed -i 's/gfx1101/gfx1100/g' Dockerfile.rocm-ct2

# Rebuild
docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

Avec gfx1100 correct, la transcription devrait passer de ~13s à <1s grâce au GPU.

ok sur le serveur de effi j'ai le podcast speech qui tourne, il prend toute la puissance du processeur, il faut que je transfert celui-ci sur le pc d ejue

Bonne idée — openedai-speech est gourmand en CPU et ton PC de jeu a une RX 7800 XT qui peut accélérer la synthèse vocale.

Récupérer la config actuelle sur effi

ssh ubuntu@192.168.1.27
docker inspect openedai-speech | grep -A 30 '"Env"'
cat $(docker inspect openedai-speech | grep -o '"[^"]*docker-compose[^"]*"' | head -1)

Ou directement trouver le docker-compose :

find ~ -name "docker-compose*.yml" | xargs grep -l openedai-speech 2>/dev/null

Arrêter sur effi

# Sur effi-server
cd /dossier/du/compose
docker compose stop openedai-speech
docker compose rm openedai-speech

Déployer sur folzo-Corp

Crée un dossier dédié sur ta machine :

mkdir ~/openedai-speech && cd ~/openedai-speech

Colle le docker-compose.yml d'effi et adapte si besoin pour ROCm :

services:
  openedai-speech:
    image: ghcr.io/matatonic/openedai-speech
    restart: unless-stopped
    ports:
      - "8100:8000"
    volumes:
      - ./voices:/app/voices
      - ./config:/app/config
    devices:
      - /dev/kfd
      - /dev/dri
    environment:
      - HSA_OVERRIDE_GFX_VERSION=11.0.0

Colle le docker-compose actuel d'effi pour openedai-speech — je l'adapte exactement pour folzo-Corp avec ROCm.

effi@effi-server:~$ find ~ -name "docker-compose*.yml" | xargs grep -l openedai-speech 2>/dev/null

/home/effi/effi/docker-compose.yml effi@effi-server:~~$ cd /home/effi effi@effi-server:~~$ cd /home/effi/ effi@effi-server:~$ pwd /home/effi

cat ~/effi/docker-compose.yml

Colle le contenu — je prépare la config adaptée pour folzo-Corp avec ROCm.

effi@effi-server:~$ cat ~/effi/docker-compose.yml

═══════════════════════════════════════════════════════════

EFFI SERVER - Stack IA Locale

Machine : effi-server (192.168.1.27)

═══════════════════════════════════════════════════════════

services:

─────────────────────────────────────────────────────────

1. OPEN WEBUI - Interface Chat Effi

→ http://192.168.1.27:3000

─────────────────────────────────────────────────────────

open-webui: image: ghcr.io/open-webui/open-webui:main container_name: open-webui restart: unless-stopped extra_hosts:

"host.docker.internal:host-gateway" ports:
"3000:8080" environment:
OLLAMA_BASE_URL=http://host.docker.internal:3011
WEBUI_SECRET_KEY=${WEBUI_SECRET_KEY}
VECTOR_DB=qdrant
QDRANT_URI=http://qdrant:6333
DEFAULT_LOCALE=fr-FR
ENABLE_MEMORIES=true
AUDIO_STT_ENGINE=openai
AUDIO_STT_OPENAI_API_BASE_URL=http://192.168.1.11:8200/v1
AUDIO_STT_OPENAI_API_KEY=anything
AUDIO_STT_MODEL=Systran/faster-whisper-large-v3 volumes:
open-webui-data:/app/backend/data depends_on:
qdrant
wol-proxy networks:
effi-net

─────────────────────────────────────────────────────────

2. WOL PROXY - Réveil PC Gaming avant appel Ollama

→ Proxy :3011 → PC Gaming :11434

─────────────────────────────────────────────────────────

wol-proxy: build: ./wol-proxy container_name: wol-proxy restart: unless-stopped network_mode: host # obligatoire pour broadcast WOL environment:

TARGET_HOST=${OLLAMA_HOST} - TARGET_PORT=11434 - LISTEN_PORT=3011 - MAC_ADDRESS=${PC_GAMING_MAC}
LISTEN_PORT=3011
WAKE_TIMEOUT=60
CHECK_INTERVAL=5

─────────────────────────────────────────────────────────

3. MCPO - Proxy outils MCP (Obsidian, etc.)

→ http://192.168.1.27:8000

─────────────────────────────────────────────────────────

mcpo: image: ghcr.io/open-webui/mcpo:main container_name: mcpo restart: unless-stopped ports:

"8000:8000" volumes:
./mcpo-config:/app/config:ro
./mcpo-patch.sh:/mcpo-patch.sh:ro entrypoint: ["sh", "/mcpo-patch.sh"]

─────────────────────────────────────────────────────────

4. QDRANT - Vector Store (RAG sémantique)

→ http://192.168.1.27:6333

─────────────────────────────────────────────────────────

qdrant: image: qdrant/qdrant:latest container_name: qdrant restart: unless-stopped ports:

"6333:6333"
"6334:6334" volumes:
./qdrant-storage:/qdrant/storage:z networks:
effi-net

─────────────────────────────────────────────────────────

5. COUCHDB - Backend Obsidian LiveSync

→ http://192.168.1.27:5984

─────────────────────────────────────────────────────────

couchdb: image: couchdb:3.3 container_name: couchdb restart: unless-stopped ports:

"5984:5984" environment:
COUCHDB_USER=${COUCHDB_USER} - COUCHDB_PASSWORD=${COUCHDB_PASSWORD} volumes:
./couchdb-data:/opt/couchdb/data networks:
effi-net

─────────────────────────────────────────────────────────

6. N8N - Automatisation autonome Effi

→ http://192.168.1.27:5678

─────────────────────────────────────────────────────────

n8n: image: docker.n8n.io/n8nio/n8n:latest container_name: n8n restart: unless-stopped ports:

"5678:5678" environment:
N8N_HOST=0.0.0.0
N8N_PORT=5678
N8N_PROTOCOL=http
WEBHOOK_URL=http://192.168.1.27:5678
N8N_BASIC_AUTH_ACTIVE=true
N8N_BASIC_AUTH_USER=${N8N_USER} - N8N_BASIC_AUTH_PASSWORD=${N8N_PASSWORD}
GENERIC_TIMEZONE=Europe/Paris
TZ=Europe/Paris volumes:
./n8n-data:/home/node/.n8n networks:
effi-net

─────────────────────────────────────────────────────────

7. PIPELINES - Middleware Open WebUI (mem0 memory)

→ http://192.168.1.27:9099

─────────────────────────────────────────────────────────

pipelines: image: ghcr.io/open-webui/pipelines:main container_name: pipelines restart: unless-stopped ports:

"9099:9099" volumes:
pipelines-data:/app/pipelines networks:
effi-net

─────────────────────────────────────────────────────────

8. WHISPER - Speech To Text (STT)

→ http://192.168.1.27:8200

─────────────────────────────────────────────────────────

whisper: image: fedirz/faster-whisper-server:latest-cpu container_name: whisper restart: unless-stopped ports:

"8200:8000" volumes:
whisper-cache:/root/.cache/huggingface environment:
WHISPER__MODEL=guillaumekln/faster-whisper-tiny
WHISPER__LANGUAGE=fr
HF_HUB_OFFLINE=0

─────────────────────────────────────────────────────────

9. OPENEDAI-SPEECH - Text To Speech (TTS)

→ http://192.168.1.27:8100

─────────────────────────────────────────────────────────

openedai-speech: image: ghcr.io/matatonic/openedai-speech container_name: openedai-speech restart: unless-stopped ports:

"8100:8000" volumes:
./tts-voices:/app/voices
./tts-config:/app/config

networks: effi-net: driver: bridge

volumes: open-webui-data: pipelines-data: whisper-cache:

Parfait, j'ai tout ce qu'il faut. Voici le plan :

Sur effi-server — retirer openedai-speech

ssh ubuntu@192.168.1.27
cd ~/effi

# Arrêter et supprimer le conteneur
docker compose stop openedai-speech
docker compose rm -f openedai-speech

# Copier les données voices/config vers folzo-Corp
scp -r ~/effi/tts-voices folzo@192.168.1.11:~/openedai-speech/
scp -r ~/effi/tts-config folzo@192.168.1.11:~/openedai-speech/

Sur folzo-Corp — créer le service

mkdir -p ~/openedai-speech
cat > ~/openedai-speech/docker-compose.yml << 'EOF'
services:
  openedai-speech:
    image: ghcr.io/matatonic/openedai-speech
    container_name: openedai-speech
    restart: unless-stopped
    ports:
      - "8100:8000"
    volumes:
      - ./tts-voices:/app/voices
      - ./tts-config:/app/config
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
      - render
    environment:
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      - HIP_VISIBLE_DEVICES=0
EOF

cd ~/openedai-speech
docker compose up -d
docker logs -f openedai-speech

Mettre à jour Open WebUI sur effi

Dans ~/effi/docker-compose.yml, changer l'URL TTS de 192.168.1.27:8100 → 192.168.1.11:8100 :

# Sur effi-server
grep -n "8100\|TTS\|speech" ~/effi/docker-compose.yml

Si l'URL est configurée dans l'interface Open WebUI plutôt que dans le compose, va dans Admin Panel → Settings → Audio et change l'URL TTS vers http://192.168.1.11:8100.

effi@effi-server:~$

cd ~/effi

Arrêter et supprimer le conteneur

docker compose stop openedai-speech docker compose rm -f openedai-speech

Copier les données voices/config vers folzo-Corp

scp -r ~~/effi/tts-voices folzo@192.168.1.11:~~/openedai-speech/ scp -r ~~/effi/tts-config folzo@192.168.1.11:~~/openedai-speech/ [+] stop 1/1 ✔ Container openedai-speech Stopped 10.7s Going to remove openedai-speech [+] remove 1/1 ✔ Container openedai-speech Removed 0.1s ssh: connect to host 192.168.1.11 port 22: Connection refused scp: Connection closed ssh: connect to host 192.168.1.11 port 22: Connection refused scp: Connection closed effi@effi-server:~/effi$

SSH n'est pas installé/activé sur folzo-Corp. Deux options :

Option 1 — Activer SSH sur folzo-Corp

# Sur folzo-Corp
sudo apt install openssh-server
sudo systemctl enable --now ssh
sudo systemctl status ssh

Puis relancer le scp depuis effi :

scp -r ~/effi/tts-voices folzo@192.168.1.11:~/openedai-speech/
scp -r ~/effi/tts-config folzo@192.168.1.11:~/openedai-speech/

Option 2 — Copier dans l'autre sens (depuis folzo-Corp)

Si tu préfères ne pas ouvrir SSH sur folzo-Corp, tire les fichiers depuis effi vers toi :

# Sur folzo-Corp
mkdir -p ~/openedai-speech
scp -r ubuntu@192.168.1.27:~/effi/tts-voices ~/openedai-speech/
scp -r ubuntu@192.168.1.27:~/effi/tts-config ~/openedai-speech/

Les dossiers tts-voices et tts-config sont-ils vides ou contiennent-ils des fichiers importants ? Si c'est vide/par défaut, tu peux aussi juste créer les dossiers vides sur folzo-Corp et laisser openedai-speech les recréer au démarrage :

mkdir -p ~/openedai-speech/tts-voices ~/openedai-speech/tts-config

folzo@folzo-Corp:~/speaches/src$ # Sur folzo-Corp

mkdir -p ~~/openedai-speech scp -r effi@192.168.1.27:~~/effi/tts-voices ~~/openedai-speech/ scp -r effi@192.168.1.27:~~/effi/tts-config /openedai-speech/ effi@192.168.1.27's password: en_GB-northern_english_male-medium.onnx.json 100% 4847 4.8MB/s 00:00 onyx.wav 100% 372KB 47.0MB/s 00:00 config.json 100% 4368 5.4MB/s 00:00 vocab.json 100% 353KB 57.0MB/s 00:00 speakers_xtts.pth 100% 7573KB 105.5MB/s 00:00 model.pth 100% 1781MB 111.2MB/s 00:16 hash.md5 100% 32 47.0KB/s 00:00 en_US-libritts_r-medium.onnx.json 100% 20KB 17.5MB/s 00:00 version.txt 100% 1 1.7KB/s 00:00 alloy.wav 100% 300KB 56.3MB/s 00:00 echo.wav 100% 386KB 59.5MB/s 00:00 en_GB-northern_english_male-medium.onnx 100% 60MB 111.4MB/s 00:00 en_US-libritts_r-medium.onnx 100% 75MB 111.5MB/s 00:00 nova.wav 100% 400KB 74.5MB/s 00:00 voices.json 100% 204KB 49.4MB/s 00:00 shimmer.wav 100% 365KB 56.9MB/s 00:00 fable.wav 100% 393KB 57.1MB/s 00:00 effi@192.168.1.27's password: pre_process_map.yaml 100% 503 985.3KB/s 00:00 voice_to_speaker.yaml 100% 1536 3.2MB/s 00:00 folzo@folzo-Corp:/speaches/src$ mkdir -p ~/openedai-speech/tts-voices ~~/openedai-speech/tts-config folzo@folzo-Corp:~~/speaches/src$

Parfait, les voix/config sont bien copiés. Il reste à lancer le service TTS sur ton PC de jeu.

1. Préparer la structure sur folzo-Corp

Tu as déjà les dossiers, vérifie juste :

ls ~/openedai-speech
ls ~/openedai-speech/tts-voices
ls ~/openedai-speech/tts-config

2. Créer un docker-compose pour openedai-speech (avec ROCm)

Sur folzo-Corp :

cat > ~/openedai-speech/docker-compose.yml << 'EOF'
services:
  openedai-speech:
    image: ghcr.io/matatonic/openedai-speech
    container_name: openedai-speech
    restart: unless-stopped
    ports:
      - "8100:8000"
    volumes:
      - ./tts-voices:/app/voices
      - ./tts-config:/app/config
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
      - render
    environment:
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      - HIP_VISIBLE_DEVICES=0
EOF

Puis lance :

cd ~/openedai-speech
docker compose up -d
docker logs -f openedai-speech

Tu dois voir qu’il écoute sur 0.0.0.0:8000.

3. Pointer Open WebUI vers le nouveau TTS

Comme l’URL de TTS est désormais sur ton PC de jeu :

Hôte : 192.168.1.11
Port : 8100
Base URL TTS : http://192.168.1.11:8100

Dans Open WebUI (sur effi) → Admin / Settings / Audio (ou TTS) → mets cette URL.

Ensuite, teste une génération vocale depuis Open WebUI; si ça ne répond pas, donne-moi les 20 dernières lignes de docker logs openedai-speech. ²⁵²⁶²⁷²⁸²⁹³⁰³¹³²³³³⁴

⁂

folzo@folzo-Corp:~/speaches/src$ cd ~/openedai-speech

docker compose up -d docker logs -f openedai-speech [+] up 16/16 ✔ Image ghcr.io/matatonic/openedai-speech Pulled 83.3s ✔ Network openedai-speech_default Created 0.0s ✔ Container openedai-speech Created 0.5s Error response from daemon: Unable to find group render: no matching entries in group file folzo@folzo-Corp:~/openedai-speech$

Le conteneur démarre, mais l’option group_add: render casse car le groupe n’existe pas sur ta machine.

1. Voir les groupes GPU disponibles

getent group | grep -E "video|render"

Tu verras sûrement video, mais pas render.

2. Corriger le docker-compose

Édite ~/openedai-speech/docker-compose.yml et retire render :

services:
  openedai-speech:
    image: ghcr.io/matatonic/openedai-speech
    container_name: openedai-speech
    restart: unless-stopped
    ports:
      - "8100:8000"
    volumes:
      - ./tts-voices:/app/voices
      - ./tts-config:/app/config
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
    environment:
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      - HIP_VISIBLE_DEVICES=0

Puis relance :

docker compose down
docker compose up -d
docker logs -f openedai-speech

Colle les 20 dernières lignes des logs si tu veux qu’on vérifie qu’il écoute bien sur 0.0.0.0:8000.

folzo@folzo-Corp:~/openedai-speech$ getent group | grep -E "video|render"

video:x:44:ollama render:x:991:ollama folzo@folzo-Corp:~/openedai-speech$

Docker a râlé parce que ces groupes n’existent pas dans le conteneur, même s’ils existent sur l’hôte. Il faut passer les GID numériques plutôt que les noms.

1. Récupérer les GID

Tu les as déjà :

video:x:44:ollama → GID 44
render:x:991:ollama → GID 991

2. Modifier `docker-compose.yml`

Remplace group_add par les GID :

services:
  openedai-speech:
    image: ghcr.io/matatonic/openedai-speech
    container_name: openedai-speech
    restart: unless-stopped
    ports:
      - "8100:8000"
    volumes:
      - ./tts-voices:/app/voices
      - ./tts-config:/app/config
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - "44"
      - "991"
    environment:
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      - HIP_VISIBLE_DEVICES=0

3. Relancer le service

cd ~/openedai-speech
docker compose down
docker compose up -d
docker logs -f openedai-speech

Si tout est bon, tu devrais voir un log du style « Running on http://0.0.0.0:8000 ».

📋 Résumé Complet Thread - Optimisé Markdown

🎯 Objectif Initial

Déployer speaches (STT Whisper ROCm) sur folzo-Corp (RX 7800 XT) pour libérer effi-server (CPU only).

🔧 Problèmes Résolus

Problème	Cause	Solution
Whisper cuDNN crash	CTranslate2 compilé sans MIOpen	Dockerfile avec `-DWITH_CUDNN=ON -DCUDnn_LIBRARY=libMIOpen.so` ³⁵
hipErrorInvalidDeviceFunction	gfx1101 au lieu de gfx1100	`sed -i 's/gfx1101/gfx1100/g' Dockerfile.rocm-ct2`
OpenWebUI lent (500)	Ollama inaccessible	WOL proxy déclenche PC gaming (192.168.1.11:3011→11434)
Interface réseau Ubuntu plantée	Conflit DHCP 192.168.1.11	Réservations MAC routeur + `ethtool wol g`
TTS CPU 100% effi-server	openedai-speech gourmand	Migré vers folzo-Corp (8100) avec ROCm gfx1100

🏗️ Architecture Finale

effi-server (192.168.1.27) ← OpenWebUI:3000
├── STT → folzo-Corp:8200 (speaches:rocm-ct2)
├── TTS → folzo-Corp:8100 (openedai-speech ROCm)
└── WOL → PC Gaming Ollama (192.168.1.11:11434)

✅ Status Services

Service	Hôte	Port	GPU	Statut
speaches-stt	folzo-Corp	8200	RX 7800 XT (gfx1100)	✅ CPU → GPU rebuild pending
openedai-speech	folzo-Corp	8100	RX 7800 XT	✅ Démarré (group_add: "44","991")
OpenWebUI	effi-server	3000	CPU	✅ Pointé vers folzo-Corp STT/TTS

📁 Configs Clés

# speaches ROCm (folzo-Corp)
~/speaches/src/Dockerfile.rocm-ct2  # gfx1100 + MIOpen
docker compose up speaches-stt

# TTS ROCm (folzo-Corp)  
~/openedai-speech/docker-compose.yml # group_add: ["44","991"]

🎛️ Prompt Nouveau Chat

**Contexte IA Locale Migrée GPU**

J'ai migré mon stack IA de effi-server (CPU) vers folzo-Corp (RX 7800 XT gfx1100 ROCm 6.3.4):
- speaches-stt:8200 → Whisper large-v3 (rebuild gfx1100 pending)
- openedai-speech:8100 → TTS ROCm (group_add 44,991 ✅)
- OpenWebUI:3000 (effi) → pointe folzo STT/TTS + WOL Ollama

**Prochaine étape**: Finaliser rebuild speaches gfx1100 et optimiser pipeline vocal complet.

**Question précise**: [INSÉRER ICI]

Copie-colle le prompt avec ta question suivante ! 🚀

⁂

📋 Résumé Complet du Thread

🖥️ Infrastructure

Machine	IP	Rôle	GPU/CPU
folzo-Corp	192.168.1.11	PC gaming / STT+TTS	AMD RX 7800 XT (gfx1100)
effi-server	192.168.1.27	Serveur IA principal	CPU Intel 4 cores / 7.6GB RAM
PC Gaming Ollama	variable	LLM Ollama	GPU dédié

🎯 Objectif du Thread

Faire tourner Whisper (STT) sur la RX 7800 XT de folzo-Corp via ROCm, en construisant CTranslate2-ROCm depuis les sources (speaches), puis migrer le TTS (openedai-speech) d'effi-server vers folzo-Corp pour libérer le CPU.

🔨 Étapes Réalisées

1. Build CTranslate2-ROCm pour speaches

Problème initial : RuntimeError: Conv1D on GPU currently requires the cuDNN library which is not integrated in this build

Cause : CTranslate2-ROCm compilé sans support MIOpen (équivalent ROCm de cuDNN)

Solution : Dockerfile custom ~/speaches/src/Dockerfile.rocm-ct2 avec :

-DWITH_CUDNN=ON
-DCUDnn_INCLUDE_DIR=/opt/rocm-6.3.4/include
-DCUDnn_LIBRARY=/opt/rocm-6.3.4/lib/libMIOpen.so

Image base : rocm/dev-ubuntu-24.04:6.3.4-complete Source CTranslate2 : github.com/arlo-phoenix/CTranslate2-rocm

2. Erreur Architecture GPU Incorrecte

Problème : RuntimeError: parallel_for failed: hipErrorInvalidDeviceFunction: invalid device function

Cause : Build compilé pour gfx1101 mais la RX 7800 XT est gfx1100

Diagnostic :

docker run --device /dev/kfd --device /dev/dri \
  -e HSA_OVERRIDE_GFX_VERSION=11.0.0 speaches:rocm-ct2 \
  /opt/rocm-6.3.4/bin/rocminfo | grep -E "gfx|Name"
# → gfx1100 (AMD Radeon RX 7800 XT)

Fix :

sed -i 's/gfx1101/gfx1100/g' ~/speaches/src/Dockerfile.rocm-ct2
cd ~/speaches/src
docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

3. Transcription fonctionnelle (CPU en attendant rebuild)

Après le premier build, speaches fonctionne en CPU (fallback) :

200 OK sur /v1/audio/transcriptions
Langue fr détectée à 99%
RTF > 1 (13s pour 5.4s audio) → confirme CPU, pas GPU
ONNX Runtime cherche libcublasLt.so.12 (CUDA) → warning non bloquant, fallback CPU automatique

Rebuild avec gfx1100 en cours pour activer l'accélération GPU (~10x plus rapide attendu).

4. Problème réseau effi-server

Symptômes : Interface filaire effi-server plantée après reboot, OpenWebUI inaccessible

Diagnostic :

conflict detected for IP address 192.168.1.11 with host 5C:87:9C:65:95:23

Cause : Conflit DHCP — deux machines avec la même IP 192.168.1.11

Fix court terme : nmcli device disconnect enp6s0 && nmcli device connect enp6s0

Fix permanent recommandé : Réservations DHCP par MAC sur le routeur + ethtool wol g pour la carte Realtek RTL8111

5. OpenWebUI lent / erreurs 500

Symptômes : 500: Open WebUI: Server Connection Error, lenteur UI

Diagnostic :

ERROR: open_webui.routers.ollama: Connection error

Cause : Ollama inaccessible — le PC gaming était éteint. Le wol-proxy sur effi-server assure le Wake-on-LAN mais le PC n'avait pas répondu à temps.

Architecture WOL :

OpenWebUI → port 3011 effi → wol-proxy → WOL magic packet → PC Gaming
                                       → proxy TCP vers 11434

6. Deux processus zombies sur effi-server

Détectés via ps aux :

[mcpo] <defunct>
[python] <defunct>

Fix :

docker restart mcpo

7. Migration openedai-speech → folzo-Corp

Raison : openedai-speech consommait ~100% CPU sur effi-server (1.06GB RAM, CPU intensif)

Étapes :

# Arrêt sur effi
docker compose stop openedai-speech && docker compose rm -f openedai-speech

# Copie des données depuis folzo-Corp
scp -r effi@192.168.1.27:~/effi/tts-voices ~/openedai-speech/
scp -r effi@192.168.1.27:~/effi/tts-config ~/openedai-speech/

Contenu transféré : modèles XTTS (~1.8GB), voix WAV (alloy/echo/nova/shimmer/fable/onyx), modèles ONNX piper, configs YAML

Problème group_add : Unable to find group render: no matching entries in group file

Fix : Utiliser GID numériques au lieu des noms :

group_add:
  - "44"    # video
  - "991"   # render

📁 Fichiers Clés

folzo-Corp:
├── ~/speaches/
│   ├── src/Dockerfile.rocm-ct2     # Build CTranslate2 gfx1100 + MIOpen
│   └── docker-compose.yml          # speaches-stt:8200
└── ~/openedai-speech/
    ├── docker-compose.yml           # TTS ROCm gfx1100
    ├── tts-voices/                  # Voix XTTS + ONNX
    └── tts-config/                  # voice_to_speaker.yaml etc.

effi-server:
└── ~/effi/docker-compose.yml        # OpenWebUI + WOL + n8n + qdrant...

✅ État Final des Services

Service	Hôte	Port	État	Notes
speaches-stt	folzo-Corp	8200	✅ Fonctionne	CPU actuellement, rebuild gfx1100 pending
openedai-speech	folzo-Corp	8100	✅ Démarré	ROCm gfx1100, group_add 44+991
OpenWebUI	effi-server	3000	✅ Fonctionne	Pointé STT+TTS vers 192.168.1.11
wol-proxy	effi-server	3011	✅ Fonctionne	Proxy Ollama + WOL PC gaming
n8n	effi-server	5678	✅ Running
qdrant	effi-server	6333	✅ Running	Vector store RAG
pipelines	effi-server	9099	✅ Running	mem0 memory
couchdb	effi-server	5984	✅ Running	Obsidian LiveSync
mcpo	effi-server	8000	⚠️ Zombie détecté	Redémarré

🔜 Prochaines Étapes

Finaliser rebuild speaches gfx1100 → activer GPU sur Whisper large-v3
Tester TTS ROCm → vérifier que openedai-speech utilise bien la RX 7800 XT
Mettre à jour l'URL TTS dans OpenWebUI Admin → Audio → http://192.168.1.11:8100
Réservations DHCP sur le routeur pour fixer les IPs de façon permanente
Tester pipeline vocal complet : micro → STT folzo → OpenWebUI effi → LLM Ollama → TTS folzo → speaker

🎛️ Prompt Nouveau Chat

## Contexte infrastructure IA locale

### Machines
- folzo-Corp (192.168.1.11) : PC gaming, AMD RX 7800 XT (gfx1100), ROCm 6.3.4, Fedora/Ubuntu
- effi-server (192.168.1.27) : Serveur IA, CPU Intel 4 cores, 7.6GB RAM, Ubuntu 24.04
- PC gaming Ollama : LLM via WOL proxy

### Stack déployée
- OpenWebUI:3000 → effi-server
- speaches STT:8200 → folzo-Corp (speaches:rocm-ct2, Whisper large-v3, gfx1100, build CTranslate2-ROCm custom avec MIOpen)
- openedai-speech TTS:8100 → folzo-Corp (ROCm, group_add: 44+991)
- wol-proxy:3011 → effi-server (proxy TCP+WOL vers Ollama PC gaming)
- n8n:5678, qdrant:6333, pipelines:9099, couchdb:5984, mcpo:8000 → effi-server

### Configs importantes
- Dockerfile ROCm CTranslate2 : ~/speaches/src/Dockerfile.rocm-ct2
  - Architecture : gfx1100 (RX 7800 XT)
  - MIOpen : -DWITH_CUDNN=ON -DCUDnn_LIBRARY=/opt/rocm-6.3.4/lib/libMIOpen.so
- TTS : ~/openedai-speech/docker-compose.yml (group_add: "44","991")
- HSA_OVERRIDE_GFX_VERSION=11.0.0

### TODO en cours
- Rebuild speaches avec gfx1100 (sed fait, build en cours)
- Vérifier GPU actif sur TTS openedai-speech
- Tester pipeline vocal bout en bout

### Question
[INSÉRER ICI]

et si tu devais donner un titre a ce thread ?

"Migration Stack IA ROCm : CTranslate2 gfx1100 + Pipeline Vocal GPU sur RX 7800 XT"

Ou en version plus courte :

"Speaches + OpenedAI-Speech ROCm — folzo-Corp RX 7800 XT"

227 KiB Raw Permalink Blame History Unescape Escape

effi@effi-server:~/effi$ docker logs open-webui --tail 20 -f

Le vrai problème : voix nova inexistante

^Ceffi@effi-server:~/effils ~/effi/tts-voices/s/

j'ai changé pour tts-1-hd, voici les logs :

Vérifie le mode vocal dans l'UI

voici les logs

Ce qui reste à ignorer (non bloquant)

Bilan final du projet Effi — Session d'aujourd'hui

tout marche mais ce n'est pas assez rapide il faut utiliser la carte graphique avec rocm

Plan : Rebuilder speaches avec CTranslate2-ROCm

avant d'aller plus loins voici les infos de rocm :

0 1 0x747e, 60998 51.0°C 50.0W N/A, N/A, 0 2737Mhz 96Mhz 14.9% auto 212.0W 72% 100%

Le vrai problème

Ce que je suspecte

folzo@folzo-Corp:~/speaches$ docker exec speaches-stt rocm-smi

0 1 0x747e, 60998 50.0°C 56.0W N/A, N/A, 0 2734Mhz 456Mhz 14.9% auto 212.0W 72% 100%

folzo@folzo-Corp:~/speaches$ docker exec -it speaches-stt bash

il y a une erreur ? le PIP ne semble pas marcher :

speaches) ubuntu@f268afe88f2b:~/speaches$ /home/ubuntu/speaches/.venv/bin/pip uninstall ctranslate2 -y

(speaches) ubuntu@f268afe88f2b:~/speaches$ /bin/uv pip list --python /home/ubuntu/speaches/.venv | grep ctranslate

j'ai modifier le fichier yml, mais la compilation n'a pas marché :

toujours un probleme :

folzo@folzo-Corp:~/speaches/src$ docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \

── Dépendances build

── Compile CTranslate2-rocm pour gfx1101 (RX 7800 XT)

── Build wheel Python CTranslate2-rocm

── Image finale speaches + CT2-rocm

folzo@folzo-Corp:~/speaches/src$ docker run --rm rocm/dev-ubuntu-24.04:6.3.4-complete \

encore une erreur :

on est allé plus loin mais il y aencor eune erreur :

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

trop compliquer peux tu me regener le fichier en entier stp, je vais faire des erreurs sinon

du coup c'est fini ? :

1. Vérifie que ROCm est détecté

2. Mets à jour docker-compose.yml

3. Relance le stack

folzo@folzo-Corp:~/speaches/src$ docker run --rm \

folzo@folzo-Corp:~/speaches/src$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

normalement c'est bon :

log apres un test :

folzo@folzo-Corp:~/speaches$ docker build -f Dockerfile.rocm-ct2 -t speaches:rocm-ct2 .

normalement il a fini : folzo@folzo-Corp:~/speaches$ cd ~/speaches/src

autre probleme le serveur effi ubunutu ne repond plus comment connaitre son etat ?

Vérifier la connectivité réseau

Tenter une connexion SSH

Via ton réseau local (si même LAN)

Si tu as accès au hyperviseur / Proxmox

Si c'est une machine physique accessible

le ping repond :

Connexion SSH + état instantané

État des conteneurs Docker

Vérifier si c'est le disque qui sature (cause fréquente de freeze)

Si SSH ne répond toujours pas

j'ai fait un reboot au server mais du coup il y a un probleme de connection avec openwebui : Comment faire pour trouver la panne ?

1. Vérifier l'état des conteneurs

2. Redémarrer les conteneurs arrêtés

3. Voir les logs du conteneur en erreur

4. Vérifier que les conteneurs ont bien restart: unless-stopped

effi@effi-server:~$ htop

Ollama (ou le backend LLM)

Vérifier où tourne Ollama

Si Ollama est en systemd (installé natif)

Si Ollama devrait être un conteneur Docker

Vérifier l'URL configurée dans Open WebUI

openweb ui a un probleme, il est tres lent repond pas ce n'est pas normal

Causes fréquentes après un reboot

effi@effi-server:~$ # Charge globale CPU/RAM

Top des processus les plus gourmands

Stats temps réel de tous les conteneurs

Espace disque

Corriger Ollama en priorité

Vérifier l'URL Ollama dans Open WebUI

Nettoyer les zombies mcpo/python

ce n'est pas ollama mais un wol proxy qui tourne sur ce serveur, celui-ci a pour objectif de pinguer le pc ou se trouve ollama et de lancer un wol si besoin

Vérifier l'état du PC Ollama

Vérifier les logs du wol-proxy

Forcer manuellement un WOL

Si le PC est allumé mais Ollama ne répond pas

c'est bon j'ai trouvé, l'interface reseau filaire de ubuntu avais planté j'ai fait off on et s'est revenu. C'est souvant sur ubuntu ?

227 KiB

Raw Permalink Blame History

Le vrai problème : voix `nova` inexistante

4. Vérifier que les conteneurs ont bien `restart: unless-stopped`

Les erreurs `restricted_cookie_manager` → sans gravité

Problème 2 : `hipErrorInvalidDeviceFunction` ← cause réelle du crash