stokiz/dots-tts-runtime-cuda

dots.tts runtime payload

This private Kaggle dataset is generated by Phorcys.Tools.DotsTtsRuntimeUploader for PHRunner.Kaggle.Service.DotsTTS.

Runtime flavor: LinuxCuda Python tag: python3.12 Generated UTC: 2026-09-10T00:49:22.3335893+00:00

The dataset intentionally contains runtime artifacts, not the GGUF model repository by default. Keep the model files in a separate private Kaggle dataset, for example kaggle-pool-account/dots-tts-v1-5-gguf.

Top-level runtime directories are uploaded as .phzip files by default. This keeps the Kaggle dataset mutation deterministic with --dir-mode skip and avoids Kaggle auto-extracting ordinary .zip directories during dataset processing. The service kernel extracts the small runtime payload into /kaggle/working/phorcys_dots_runtime/<dataset-name> during first activation. Experimental expanded-directory upload is available with --stage-opaque-archives false, but the .phzip mode remains the safer default for runtime because the virtual environment and vendor tree contain nested directory structures.

Kaggle-side activation example:

import sys
from pathlib import Path
# Try the current Kaggle direct mount first.
runtime_root = Path('/kaggle/input/dots-tts-runtime')
if not runtime_root.exists():
    # Newer Kaggle notebooks may expose private datasets under /kaggle/input/datasets/<owner>/<slug>.
    runtime_root = Path('/kaggle/input/datasets/kaggle-pool-account/dots-tts-runtime')
sys.path.insert(0, str(runtime_root))
import runtime_bootstrap
runtime_bootstrap.activate(runtime_root)

Run the bundled runtime Python against <runtime-root>/verify_runtime.py inside a Kaggle notebook when diagnosing import failures. The verifier imports torch, transformers, dots_tts.runtime, lingua, lingua.lingua, tn.chinese.normalizer, tn.english.normalizer, kaggle, and canonical PostgreSQL relay client modules. It also injects the same no-op tn.* fallback used by the runner, so a payload that lacks either the official package, ABI-compatible native dependencies, relay dependencies, Qwen2 transformer compatibility, pyOpenSSL/cryptography compatibility, optional text-normalizer compatibility, or portable torch fails before production use instead of failing only after a relay job starts. Linux Kaggle payloads include managed CPython 3.12 in python-runtime.phzip, direct-import site-packages in dot-venv.phzip, and the offline wheelhouse in wheelhouse.phzip. The generated service runner re-executes under this mounted Python by default. It should not use Kaggle worker Python, Kaggle-native torch, Kaggle-native torchaudio, or Kaggle-native ffmpeg. The only host component that cannot be portably shipped is the kernel-mode NVIDIA driver exposed by Kaggle/Colab to the container.