hf-family-workflows
Operate Hugging Face remote-code and Gradio-style workflows for MOSS-TTS family models.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- hf-family-workflows
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Operate Hugging Face remote-code and Gradio-style workflows for MOSS-TTS family models.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md10.4 KB
- 参考文档references/hf-generation-workflows.md16.0 KB
- 参考文档references/model-packaging.md9.4 KB
- 参考文档references/troubleshooting.md10.6 KB
- 脚本scripts/normalize_tts_text.py10.5 KB
03 SKILL.md 正文
hf-family-workflows
Use this sub-skill when the task is to run, adapt, or debug the Hugging Face AutoProcessor/AutoModel remote-code workflows or the Gradio-style UI parameters for the MOSS-TTS family.
Best-fit requests
- MOSS-TTS-v1.5 or MOSS-TTS 1.0 direct TTS, voice cloning, continuation, duration control, language tags, Pinyin, IPA, or
[pause X.Ys]prompt control. - MOSS-TTSD-v1.0 multi-speaker dialogue generation with
[S1]...[S5]tags and per-speaker reference audio. - MOSS-VoiceGenerator text + natural-language voice-instruction prompts.
- MOSS-SoundEffect v1 text-to-sound prompts through the Delay-family Hugging Face API.
- Choosing between Delay and Local architectures for non-streaming Hugging Face generation.
- Selecting
attn_implementation, dtype/device, model ID, codec path, sampling parameters, or Gradio launch flags. - Normalizing prompt text before sending it to
build_user_message. - Diagnosing import, package exposure, model download/cache, remote-code, reference-audio, or mixed checkpoint/code failures.
Route elsewhere
- Torch-free / low-memory llama.cpp, GGUF, ONNX, or TensorRT backend:
../llama-cpp-backend/SKILL.md. - MOSS-TTS-Local-Transformer-v1.5 realtime streaming app or streaming decode internals:
../local-v15-streaming/SKILL.md. - MOSS-TTS-Realtime voice-agent, FastAPI, SSE, or multi-turn streaming sessions:
../realtime-voice-agent/SKILL.md. - Fine-tuning, data JSONL construction,
audio_codes, training launchers, or post-finetune checks:../finetuning-data-prep/SKILL.md. - MOSS-SoundEffect-v2.0 DiT/Flow-Matching pipeline:
../soundeffect-v2/SKILL.md.
Operating map
- Start with
references/hf-generation-workflows.mdfor concrete generation recipes, model IDs, Gradio-style flags, and prompt patterns. - Use
references/api-reference.mdwhen you need exactbuild_user_message, conversation shape, processor modes, audio-code, and hyperparameter contracts. - Use
references/model-packaging.mdbefore changing model/code/codec packaging, using fused checkpoint layouts, or diagnosing package exposure. - Use
references/troubleshooting.mdfor known failure signatures and fixes. - Use
scripts/normalize_tts_text.pyfor safe stdlib-only prompt cleanup; it can run from any current working directory.
Quick safeguards
- Always pass
trust_remote_code=Truefor Hugging FaceAutoProcessorandAutoModelloads. - Prefer CUDA +
bfloat16for 8B models. CPU can load only in very constrained smoke tests and is usually impractical for generation. - Keep model code, config,
n_vq, tokenizer, and audio codec matched. Mixed Delay/Local/TTSD/VoiceGenerator code and checkpoints often produce shape errors or gibberish. - For v1.5 multilingual prompts, set the language field when known; for continuation, prepend the reference transcript to the text.
- Do not use this skill as proof that model download or full generation succeeded in the target runtime; those are runtime prerequisites to verify where the model will actually run.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).