training-and-checkpoints
Operate TurboDiffusion rCM/SLA training setup, dry-run command construction, checkpoint conversion, model merge, quantized export, and data/checkpoint layout validation.
- License 许可
- Apache-2.0
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- training-and-checkpoints
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Operate TurboDiffusion rCM/SLA training setup, dry-run command construction, checkpoint conversion, model merge, quantized export, and data/checkpoint layout validation.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/checkpoint-workflows.md7.6 KB
- 参考文档references/cli-reference.md7.9 KB
- 参考文档references/training-and-data.md6.9 KB
- 参考文档references/troubleshooting.md8.2 KB
03 SKILL.md 正文
TurboDiffusion Training And Checkpoints
Use this sub-skill when the task is about TurboDiffusion rCM/SLA training setup, debug-only training command construction, Distributed Checkpoint (DCP) versus PyTorch .pth formats, merging SLA/rCM updates, converting sharded safetensors, or exporting modified/quantized Wan checkpoints.
Do not run full training or model conversion by default. Full training, dataset synthesis, checkpoint quantization, and inference-format conversion are GPU/model/data heavy and remain skip-expensive unless the user explicitly authorizes them and provides checkpoints, data, credentials policy, and backend readiness.
Route first
- Use references/training-and-data.md for rCM/SLA training prerequisites, debug/dry-run planning, data shard layout, checkpoint roots, and WANDB/environment decisions.
- Use references/checkpoint-workflows.md for DCP-to-
.pth,.pth-to-DCP, merge arithmetic, safetensors conversion, andmodify_modelexport semantics. - Use references/cli-reference.md for safe command templates and supported flags for public training/checkpoint utilities.
- Use references/troubleshooting.md when a user reports DCP/PTH confusion, missing roots/shards, training extras, WANDB issues, mismatched model names, or quantization/SLA failures.
Bundled safe helpers
These helpers only build commands or run a tiny CPU tensor arithmetic check; they do not download, train, execute model inference, or require the original checkout.
- scripts/build_training_dryrun_command.py: render a
torchrun ... -m scripts.train --dryruncommand with explicit config/checkpoint/data overrides. - scripts/build_modify_model_command.py: render a
modify_model.pycheckpoint export/quantization command for Wan2.1/Wan2.2 model profiles. - scripts/build_safetensors_to_pth_command.py: render a sharded safetensors-to-
.pthcommand with optional key prefixing. - scripts/tiny_merge_models_check.py: create temporary tiny tensors and verify the source merge formula
base + w * (diff_target - diff_base)on CPU.
Operating rules
- Start with layout validation: checkpoint root, teacher DCP, VAE, text encoder, negative embedding, output root, and WebDataset shard pattern.
- For training requests, prefer a dry-run/config-composition command first. Explain that actual training needs CUDA/multi-GPU readiness, training extras, model/data files, and a WANDB/offline logging decision.
- For DCP conversion, distinguish direction:
.pthteacher checkpoint must be converted to DCP before FSDP training; saved training DCP directories can later be converted back to.pth. - For model merge, warn that mismatched shapes/keys are not fatal in the source utility but require manual key-coverage inspection before treating the output as good.
- For quantized export, ensure the source checkpoint is already in the expected rCM/SLA state-dict shape, choose a valid Wan model name, and align
--quant_linearwith later inference/serving commands. - For source-layout use, mention the public package quirk: top-level imports such as
rcm,imaginaire,SLA,ops,scripts, andmodify_modeloften require settingPYTHONPATHto the package source directory.
Cross-skill routing
- Route full T2V/I2V use of converted checkpoints to the
video-inferencesub-skill. - Route interactive serving with converted checkpoints to the
interactive-servingsub-skill. - Route CUDA extension build, INT8/FastNorm op failures, plain SLA versus SageSLA backend decisions, or SpargeAttn installation issues to the
acceleration-backendssub-skill.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.