turbo-diffusion
Use TurboDiffusion for accelerated Wan video generation, interactive serving, checkpoint conversion, CUDA acceleration backends, and TurboT2AV extension planning.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- turbo-diffusion
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use TurboDiffusion for accelerated Wan video generation, interactive serving, checkpoint conversion, CUDA acceleration backends, and TurboT2AV extension planning.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/installation-and-backends.md4.2 KB
- 参考文档references/model-and-asset-catalog.md4.6 KB
- 参考文档references/repo-provenance.md3.6 KB
- 参考文档references/repo-routing-metadata.json516 B
- 参考文档references/troubleshooting.md4.2 KB
- 脚本scripts/check_turbodiffusion_env.py5.4 KB
03 SKILL.md 正文
TurboDiffusion
Use this repo skill when a task involves TurboDiffusion, TurboWan, Wan2.1/Wan2.2 accelerated video generation, SLA/SageSLA attention, rCM/SLA checkpoint conversion or training, the turbodiffusion-serve TUI, or TurboT2AV audio-video generation.
TurboDiffusion is a CUDA-first package. Most real generation workflows need a compatible NVIDIA GPU, model checkpoints, Wan VAE/text encoder files, and source-layout imports. Do not start downloads, long generation, or training unless the user explicitly supplied the needed assets and budget.
First checks
- Read repo provenance when deciding whether this skill is current for a checkout.
- Read installation and backend notes before installing, compiling, or diagnosing CUDA/SageSLA/custom-op problems.
- Run scripts/check_turbodiffusion_env.py when a user asks whether their current environment can import TurboDiffusion, see CUDA, load custom ops, or use optional SageSLA.
- Read model and asset catalog before building commands that mention VAE/text encoder/DiT checkpoints, high/low I2V checkpoints, prompt files, or TurboT2AV assets.
- Use cross-cutting troubleshooting for install/import, source-layout, missing checkpoint, CUDA, and optional dependency failures.
Sub-skill routing
| User request | Read |
|---|---|
| Build a one-shot text-to-video or image-to-video command; choose Wan2.1/Wan2.2 model flags; validate prompt/image/checkpoint arguments; avoid running models while planning. | video-inference |
| Launch or debug the interactive terminal server; decide T2V vs I2V TUI flags; understand slash commands and runtime-adjustable parameters. | interactive-serving |
| Prepare SLA/rCM training commands, checkpoint conversion, checkpoint merging, safetensors-to-pth conversion, quantized checkpoint export, or data/checkpoint layout checks. | training-and-checkpoints |
Diagnose CUDA extension build/import, turbo_diffusion_ops, INT8/FastNorm, attention_type=sla vs sagesla, SpargeAttn, flash-attn, or source-layout import quirks. |
acceleration-backends |
| Plan TurboT2AV text-to-audio-video generation, LTX-2/Pixi setup, HF/Gemma checkpoint requirements, or TileLang/SageSLA acceleration choices. | turbot2av-extension |
Operating rules
- Prefer the bundled command builders in sub-skill
scripts/for command construction. They render commands and validate obvious mistakes without downloading weights, starting training, or running generation. - For source checkouts, TurboDiffusion scripts and even the installed TUI entry point may require adding the package source directory to
PYTHONPATHso top-level imports such asimaginaire,rcm,serve, andmodify_modelresolve. Capture this as a public source-layout requirement, not as a local path. - Treat
attention_type=sageslaas optional-backend dependent: install SpargeAttn first or fall back tosla/originalwhen SageSLA is not available. - Match
--quant_linearto quantized checkpoints. Do not add it to unquantized checkpoints on large-memory GPUs unless the user explicitly wants the quantized path. - Do not run full T2V/I2V/TurboT2AV generation unless the user provides the checkpoints, VAE/text encoder assets, output location, GPU budget, and permission to spend time/GPU memory.
- Do not run full rCM/SLA training unless the user provides checkpoint/data roots, distributed/GPU budget, logging policy, and any credentials such as WandB or Hugging Face tokens.
Minimal environment smoke
Use this public smoke from an installed/source-layout environment before deeper work:
python - <<'PY'
import torch
print('torch', torch.__version__, torch.version.cuda, torch.cuda.is_available())
try:
import turbo_diffusion_ops
print('custom ops: ok')
except Exception as exc:
print('custom ops:', type(exc).__name__, exc)
PY
For a fuller check, run the bundled environment diagnostic. If it reports missing imaginaire or rcm, use the source-layout PYTHONPATH guidance in installation and backend notes.
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).