Add notebook

train_moshi_de_colab.ipynb

@@ -0,0 +1,262 @@
+{
+ "nbformat": 4,
+ "nbformat_minor": 0,
+ "metadata": {
+  "colab": {
+   "name": "train_moshi_de_colab.ipynb",
+   "provenance": [],
+   "machine_shape": "hm"
+  },
+  "kernelspec": {
+   "name": "python3",
+   "display_name": "Python 3"
+  },
+  "language_info": {
+   "name": "python"
+  },
+  "accelerator": "GPU"
+ },
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# German PersonaPlex — Colab Training\n",
+    "\n",
+    "LoRA fine-tune von Kyutai Moshi auf deutsche Podcast-Daten (Kuechenstudio, CC-BY).\n",
+    "\n",
+    "**Voraussetzungen:**\n",
+    "- Colab Pro+ (A100 40GB) empfohlen — L4 24GB geht auch mit kleinerer Batch-Size\n",
+    "- Runtime: GPU (A100 > L4 > V100)\n",
+    "- HuggingFace Token (Secret `HF_TOKEN` einrichten: links im Menue → Key-Icon)\n",
+    "\n",
+    "**Ablauf:** Runtime → Run all. Training dauert ~6-10h. Checkpoint wird auto auf HF gepusht."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": ["## 1. Setup"]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# GPU check\n",
+    "!nvidia-smi"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# HF Token aus Colab Secrets laden\n",
+    "from google.colab import userdata\n",
+    "import os\n",
+    "os.environ['HF_TOKEN'] = userdata.get('HF_TOKEN')\n",
+    "os.environ['HUGGINGFACE_HUB_TOKEN'] = os.environ['HF_TOKEN']\n",
+    "print('HF_TOKEN set:', os.environ['HF_TOKEN'][:10] + '...')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# moshi-finetune repo klonen und installieren\n",
+    "!git clone https://github.com/kyutai-labs/moshi-finetune.git\n",
+    "%cd moshi-finetune\n",
+    "!pip install -q -e ."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "!pip install -q huggingface_hub sphn soundfile"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": ["## 2. Dataset herunterladen"]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "from huggingface_hub import snapshot_download\n",
+    "\n",
+    "DATASET_REPO = 'playa555x/german-personaplex-10h'\n",
+    "dataset_path = snapshot_download(DATASET_REPO, repo_type='dataset')\n",
+    "print('Dataset at:', dataset_path)\n",
+    "\n",
+    "!ls -la {dataset_path}/kuechenstudio/ | head"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# Manifest fuer moshi-finetune vorbereiten — absolute Pfade\n",
+    "import json\n",
+    "from pathlib import Path\n",
+    "\n",
+    "src = Path(dataset_path) / 'manifest.jsonl'\n",
+    "manifest_entries = [json.loads(l) for l in src.read_text().splitlines() if l.strip()]\n",
+    "out = Path('/content/train_manifest.jsonl')\n",
+    "with out.open('w') as fh:\n",
+    "    for e in manifest_entries:\n",
+    "        e['path'] = str(Path(dataset_path) / e['path'])\n",
+    "        fh.write(json.dumps(e) + '\\n')\n",
+    "print('Manifest:', out, '-', len(manifest_entries), 'samples')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": ["## 3. Base Model laden"]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "BASE_MODEL = 'kyutai/moshiko-pytorch-bf16'\n",
+    "model_path = snapshot_download(BASE_MODEL)\n",
+    "print('Base model:', model_path)\n",
+    "!ls -la {model_path}"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": ["## 4. Training Config"]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "config_yaml = '''\n",
+    "# LoRA Fine-Tune Moshi -> Deutsch\n",
+    "run_dir: /content/checkpoints/moshi-de-lora-v1\n",
+    "\n",
+    "data:\n",
+    "  train_data: /content/train_manifest.jsonl\n",
+    "  eval_data: /content/train_manifest.jsonl\n",
+    "  duration_sec: 30\n",
+    "  shuffle: true\n",
+    "\n",
+    "model:\n",
+    "  moshi_paths:\n",
+    "    hf_repo_id: kyutai/moshiko-pytorch-bf16\n",
+    "  lora:\n",
+    "    enable: true\n",
+    "    rank: 64\n",
+    "    scaling: 2.0\n",
+    "\n",
+    "training:\n",
+    "  batch_size: 2\n",
+    "  gradient_accumulation: 4\n",
+    "  optim:\n",
+    "    lr: 1.0e-4\n",
+    "    weight_decay: 0.01\n",
+    "  max_steps: 2000\n",
+    "  eval_freq: 500\n",
+    "  save_freq: 500\n",
+    "  log_freq: 20\n",
+    "  seed: 42\n",
+    "\n",
+    "system:\n",
+    "  mixed_precision: bf16\n",
+    "  gradient_checkpointing: true\n",
+    "  num_workers: 2\n",
+    "'''\n",
+    "\n",
+    "config_path = '/content/moshi-de-lora.yaml'\n",
+    "with open(config_path, 'w') as fh:\n",
+    "    fh.write(config_yaml)\n",
+    "print('Config written:', config_path)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": ["## 5. Training starten"]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# Lizenz fuer Moshi akzeptieren falls noch nicht\n",
+    "!huggingface-cli login --token $HF_TOKEN --add-to-git-credential\n",
+    "\n",
+    "# Training\n",
+    "%cd /content/moshi-finetune\n",
+    "!torchrun --nproc-per-node=1 -m train /content/moshi-de-lora.yaml 2>&1 | tee /content/train.log"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": ["## 6. Checkpoint hochladen"]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "from huggingface_hub import HfApi, create_repo\n",
+    "import os\n",
+    "\n",
+    "CHECKPOINT_REPO = 'playa555x/moshi-de-lora-v1'\n",
+    "api = HfApi(token=os.environ['HF_TOKEN'])\n",
+    "create_repo(CHECKPOINT_REPO, private=True, exist_ok=True, token=os.environ['HF_TOKEN'])\n",
+    "\n",
+    "api.upload_folder(\n",
+    "    folder_path='/content/checkpoints/moshi-de-lora-v1',\n",
+    "    repo_id=CHECKPOINT_REPO,\n",
+    "    commit_message='LoRA fine-tune: Moshi -> Deutsch (Kuechenstudio 10h)',\n",
+    ")\n",
+    "print(f'Checkpoint uploaded: https://huggingface.co/{CHECKPOINT_REPO}')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 7. Quick-Test: Generate German Audio\n",
+    "\n",
+    "Nach dem Training: ein deutsches Sample erzeugen zum Anhoeren."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# Simple inference test — erzeuge 10s Audio als Reaktion auf deutschen Prompt\n",
+    "# Details je nach moshi-finetune API\n",
+    "!python -c \"print('Training complete. Download checkpoint for local inference.')\""
+   ]
+  }
+ ]
+}