K8s deployment argocd

.gitignore

@@ -6,6 +6,7 @@ data/
 ckpts/
 wandb/
 results/
+kubernetes/secrets.yaml
 
 
 

Dockerfile

@@ -1,8 +1,11 @@
-FROM pytorch/pytorch:2.4.0-cuda12.4-cudnn9-devel
+# CUDA 13.1 + cuDNN 9 — matches node driver 590.48.01 on gx16 (A30, compute 8.0)
+FROM pytorch/pytorch:2.7.0-cuda13.1-cudnn9-devel
 
 USER root
 
 ARG DEBIAN_FRONTEND=noninteractive
+ARG TORCH_VERSION=2.7.0
+ARG CUDA_TAG=cu131
 
 LABEL github_repo="https://github.com/SWivid/F5-TTS"
 
@@ -15,10 +18,17 @@ RUN set -x \
 
 WORKDIR /workspace
 
-RUN git clone https://github.com/SWivid/F5-TTS.git \
+RUN git clone https://github.com/aihpi/F5-TTS.git \
     && cd F5-TTS \
     && pip install -e .[eval]
 
+# Force-reinstall torch + torchaudio from the official CUDA 13.1 wheel index so they
+# share the same ABI regardless of what pip resolved above.
+RUN pip install --no-cache-dir --force-reinstall \
+    torch==${TORCH_VERSION} \
+    torchaudio==${TORCH_VERSION} \
+    --index-url https://download.pytorch.org/whl/${CUDA_TAG}
+
 ENV SHELL=/bin/bash
 
 WORKDIR /workspace/F5-TTS

Dockerfile.new

@@ -0,0 +1,46 @@
+# CUDA 13.1 + cuDNN 9 — matches node driver 590.48.01 on gx16 (A30, compute 8.0)
+FROM pytorch/pytorch:2.7.0-cuda13.1-cudnn9-runtime
+
+USER root
+ARG DEBIAN_FRONTEND=noninteractive
+ARG TORCH_VERSION=2.7.0
+ARG CUDA_TAG=cu131
+
+LABEL github_repo="https://github.com/SWivid/F5-TTS"
+
+# Install dependencies
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        git \
+        ffmpeg \
+        build-essential \
+    && rm -rf /var/lib/apt/lists/* && \
+    apt-get clean
+
+WORKDIR /workspace
+
+# Clone F5-TTS repo with submodules and apply BigVGAN patch
+RUN git clone https://github.com/aihpi/F5-TTS \
+    && cd F5-TTS \
+    && git submodule update --init --recursive \
+    && printf "import os\nimport sys\nsys.path.append(os.path.dirname(os.path.abspath(__file__)))\n" \
+       | cat - src/third_party/BigVGAN/bigvgan.py > temp && mv temp src/third_party/BigVGAN/bigvgan.py
+
+# Copy local source code (e.g., for custom modifications)
+COPY . /workspace/F5-TTS
+
+WORKDIR /workspace/F5-TTS
+
+# Install Python dependencies
+RUN pip install --no-cache-dir -e .[eval] \
+    && pip install --no-cache-dir -r src/third_party/BigVGAN/requirements.txt \
+    && rm -rf ~/.cache/pip
+
+# Force-reinstall torch + torchaudio from the official CUDA 13.1 wheel index so they
+# share the same ABI regardless of what pip resolved above.
+RUN pip install --no-cache-dir --force-reinstall \
+    torch==${TORCH_VERSION} \
+    torchaudio==${TORCH_VERSION} \
+    --index-url https://download.pytorch.org/whl/${CUDA_TAG}
+
+ENV SHELL=/bin/bash

Dockerfile.server

@@ -0,0 +1,44 @@
+# CUDA 13.1 + cuDNN 9 — matches node driver 590.48.01 on gx16 (A30, compute 8.0)
+FROM pytorch/pytorch:2.7.0-cuda13.1-cudnn9-devel
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Pinned together — torchaudio must share the same PyTorch ABI.
+# Update both atomically whenever the base image is bumped.
+ARG TORCH_VERSION=2.7.0
+ARG CUDA_TAG=cu131
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        git \
+        ffmpeg \
+        build-essential \
+    && rm -rf /var/lib/apt/lists/* && \
+    apt-get clean
+
+WORKDIR /app
+
+ARG GIT_REPOSITORY="https://github.com/aihpi/F5-TTS"
+# Hacky workaround, as bigvgan is not properly setup to be installed as a package:
+# https://github.com/aihpi/F5-TTS?tab=readme-ov-file#2-local-editable-if-also-do-training-finetuning
+ARG BIGVGAN_PATCH="import os\nimport sys\nsys.path.append(os.path.dirname(os.path.abspath(__file__)))\n"
+
+RUN git clone $GIT_REPOSITORY \
+    && cd F5-TTS \
+    && git submodule update --init --recursive \
+    && printf "$BIGVGAN_PATCH" | cat - src/third_party/BigVGAN/bigvgan.py > temp && mv temp src/third_party/BigVGAN/bigvgan.py \
+    && pip install -e .[eval] \
+    && pip install -r src/third_party/BigVGAN/requirements.txt
+
+# Force-reinstall torch + torchaudio from the official CUDA 13.1 wheel index so they
+# share the same ABI regardless of what pip resolved above.  This must be a separate
+# RUN so Docker cache invalidation is scoped to version bumps, not source changes.
+RUN pip install --no-cache-dir --force-reinstall \
+    torch==${TORCH_VERSION} \
+    torchaudio==${TORCH_VERSION} \
+    --index-url https://download.pytorch.org/whl/${CUDA_TAG}
+
+# for prototyping
+COPY src/f5_tts/demo/ /app/F5-TTS/src/f5_tts/demo/
+
+CMD ["python3", "-m", "uvicorn", "f5_tts.demo.server:app", "--host", "0.0.0.0", "--port", "8000"]

argocd/f5-tts-application.yaml

@@ -0,0 +1,20 @@
+apiVersion: argoproj.io/v1alpha1
+kind: Application
+metadata:
+  name: f5-tts
+  namespace: argocd
+spec:
+  project: default
+  source:
+    repoURL: https://github.com/aihpi/F5-TTS.git
+    targetRevision: main
+    path: kubernetes
+  destination:
+    server: https://kubernetes.default.svc
+    namespace: f5-tts
+  syncPolicy:
+    automated:
+      prune: true
+      selfHeal: true
+    syncOptions:
+      - CreateNamespace=true

compose.yaml

@@ -0,0 +1,32 @@
+services:
+  f5tts_api_de:
+    build: .
+    command: python -m uvicorn f5_tts.api_server:app --host 0.0.0.0 --port 8000
+    ports:
+      - "8004:8000"
+    environment:
+      - MODEL_FILEPATH="/var/tmp/model_12600.pt"
+    volumes:
+      - /raid/shared/models/F5TTS/model_12600.pt:/var/tmp/model_12600.pt:ro
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [ gpu ]
+    restart: unless-stopped
+
+  f5tts_api_en:
+    build: .
+    command: python -m uvicorn f5_tts.api_server:app --host 0.0.0.0 --port 8000
+    ports:
+      - "8005:8000"
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    restart: unless-stopped

kubernetes/deployments.yaml

@@ -0,0 +1,46 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: f5tts-api
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: f5tts-api
+  template:
+    metadata:
+      labels:
+        app: f5tts-api
+    spec:
+      imagePullSecrets:
+        - name: ghcr-credentials
+      nodeSelector:
+        accelerator: a30
+      containers:
+        - name: api
+          image: ghcr.io/aihpi/f5-tts-optim:latest
+          imagePullPolicy: IfNotPresent
+          command:
+            - python3
+            - -m
+            - uvicorn
+            - f5_tts.demo.server:app
+            - --host
+            - 0.0.0.0
+            - --port
+            - "8000"
+          env:
+            - name: MODEL_FILEPATH
+              value: /var/tmp/model_430000.safetensors
+          ports:
+            - containerPort: 8000
+          resources:
+            limits:
+              nvidia.com/gpu: "1"
+          volumeMounts:
+            - name: model-volume
+              mountPath: /var/tmp/
+      volumes:
+        - name: model-volume
+          persistentVolumeClaim:
+            claimName: f5tts-model-pvc

kubernetes/ingresses.yaml

@@ -0,0 +1,19 @@
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: f5tts-api-ingress
+  annotations:
+    nginx.ingress.kubernetes.io/rewrite-target: /$2
+    nginx.ingress.kubernetes.io/use-regex: "true"
+spec:
+  rules:
+    - host: tts.aisc.hpi.de
+      http:
+        paths:
+          - path: /api(/|$)(.*)
+            pathType: Prefix
+            backend:
+              service:
+                name: f5tts-api
+                port:
+                  number: 8004

kubernetes/kustomization.yaml

@@ -0,0 +1,9 @@
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+namespace: f5-tts
+resources:
+  - namespace.yaml
+  - deployments.yaml
+  - services.yaml
+  - ingresses.yaml
+  - persistentvolumeclaims.yaml

kubernetes/namespace.yaml

@@ -0,0 +1,4 @@
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: f5-tts

kubernetes/persistentvolumeclaims.yaml

@@ -0,0 +1,11 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: f5tts-model-pvc
+spec:
+  accessModes:
+    - ReadWriteOnce
+  storageClassName: nfs-k8s-general
+  resources:
+    requests:
+      storage: 10Gi

kubernetes/secrets.example.yaml

@@ -0,0 +1,20 @@
+# Example secrets only. Do not commit real credentials here.
+apiVersion: v1
+kind: Secret
+metadata:
+  name: ghcr-credentials
+  namespace: f5-tts
+type: kubernetes.io/dockerconfigjson
+stringData:
+  .dockerconfigjson: |
+    {"auths":{"ghcr.io":{"username":"<ghcr-username>","password":"<ghcr-token>","email":"<email@example.com>","auth":"<base64(username:token)>"}}}
+---
+apiVersion: v1
+kind: Secret
+metadata:
+  name: f5tts-ui-env
+  namespace: f5-tts
+type: Opaque
+stringData:
+  API_USER: "<replace-me>"
+  API_PASSWORD: "<replace-me>"

kubernetes/services.yaml

@@ -0,0 +1,13 @@
+apiVersion: v1
+kind: Service
+metadata:
+  name: f5tts-api
+spec:
+  type: NodePort
+  selector:
+    app: f5tts-api
+  ports:
+    - port: 8004
+      targetPort: 8000
+      protocol: TCP
+      nodePort: 31493

kubernetes/ui-repo-secret.yaml

@@ -0,0 +1,13 @@
+apiVersion: v1
+kind: Secret
+metadata:
+  name: github-private-repo
+  namespace: argocd
+  labels:
+    argocd.argoproj.io/secret-type: repository    # Required label
+type: Opaque
+stringData:
+  type: git
+  url: https://github.com/aihpi/demo-tts.git
+  username: kalyanikumar-shunmugasundaram
+  password: <your-fine-grained-token>

pyproject.toml

@@ -30,8 +30,8 @@ dependencies = [
     "safetensors",
     "soundfile",
     "tomli",
-    "torch>=2.0.0",
-    "torchaudio>=2.0.0",
+    "torch>=2.7.0,<3.0.0",
+    "torchaudio>=2.7.0,<3.0.0",
     "torchdiffeq",
     "tqdm>=4.65.0",
     "transformers",
@@ -49,6 +49,8 @@ eval = [
     "modelscope",
     "zhconv",
     "zhon",
+    "loguru",
+    "DeepFilterNet",
 ]
 
 [project.urls]

src/f5_tts/demo/constants.py

@@ -0,0 +1,32 @@
+# Define model constants
+MODEL_CHECKPOINTS = {
+    # German vocos checkpoints
+    "kisz-german-vocos": "hf://aihpi/F5-TTS-German/F5TTS_Base/model_365000.safetensors",
+    # German bigvgan checkpoints
+    "kisz-german-bigvgan": "hf://aihpi/F5-TTS-German/F5TTS_Base_bigvgan/model_430000.safetensors",
+    # English checkpoints
+    "original-english-vocos": "hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.pt",
+    "original-english-bigvgan": "hf://SWivid/F5-TTS/F5TTS_Base_bigvgan/model_1250000.pt"
+}
+
+MODEL_METADATA = {
+    # German vocos checkpoints
+    "kisz-german-vocos": {
+        "vocoder_name": "vocos",
+        "language": "de"
+    },
+    # German bigvgan checkpoints
+    "kisz-german-bigvgan": {
+        "vocoder_name": "bigvgan",
+        "language": "de"
+    },
+    # English checkpoints
+    "original-english-vocos": {
+        "vocoder_name": "vocos",
+        "language": "en"
+    },
+    "original-english-bigvgan": {
+        "vocoder_name": "bigvgan",
+        "language": "en"
+    }
+}