New: 98,000+ skills indexed from GitHub repos with 100+ stars · every skill safety-checkedBrowse top skillsTop skills
Esc to closeBrowse all skills
Submit a skillBrowse skills
Submit a skillBrowse skills

Agent skill

Sparkdiffusion Finetune

by AlibabaResearch in AlibabaResearch/SparkDiffusion

Configure and launch SparkDiffusion sparse finetuning for Wan 2.1 or Wan 2.2.

Apache-2.0Auto-check passedAI & LLM Engineering

Install Sparkdiffusion Finetune

skills CLI
$ npx skills add AlibabaResearch/SparkDiffusion --skill sparkdiffusion-finetune -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install AlibabaResearch/SparkDiffusion sparkdiffusion-finetune --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/AlibabaResearch/SparkDiffusion.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.agents/skills/sparkdiffusion-finetune .claude/skills/sparkdiffusion-finetune && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
sparkdiffusion-finetune
GitHub stars
542
Token cost
~904 tokens
SKILL.md length
304 words
Files
2
Skills in repo
4
Repo updated
First seen
Licence
Apache-2.0

At a glance

Configure and launch SparkDiffusion sparse finetuning for Wan 2.1 or Wan 2.2.

  • Works in 6 steps: Work from the repository root and source… → Export the required external SLA… → Confirm the requested model, task,… → …
  • A user asks to train sparse attention parameters
  • SKILL.md covers Before Launching, Wan 2.1, Wan 2.2 and Outputs and Resume
  • Calls bash

What it does

Sparkdiffusion Finetune is an agent skill from AlibabaResearch/SparkDiffusion. Configure and launch SparkDiffusion sparse finetuning for Wan 2.1 or Wan 2.2. Use when a user asks to train sparse attention parameters, select Wan experts, resume finetuning, or validate a finetuning configuration.

Its SKILL.md is about 900 tokens, which your agent loads only when the skill is triggered. The skill folder holds 2 other files (for example `agents/openai.yaml`).

It sits in AI & LLM Engineering, covering Fine-tuning. The repository describes itself as: SparkDiffusion: Accelerating DiT video generation by 265× via joint Sparsity, Distillation, Quantization, etc. The licence is Apache-2.0.

When your agent uses it

  • A user asks to train sparse attention parameters
  • Select Wan experts
  • Resume finetuning
  • Validate a finetuning configuration

Example prompts

  • “/sparkdiffusion-finetune”

Requirements

  • Python 3

Workflow steps

6 steps, taken from the first numbered list in SKILL.md.

  1. Work from the repository root and source the shared environment
  2. Export the required external SLA checkout before starting Python or torchrun
  3. Confirm the requested model, task, resolution, pretrained checkpoint, dataset shard pattern, GPU count, and output root.
  4. Confirm that the checkpoint and dataset exist. Use repository-relative defaults or explicit environment variables; never insert paths from…
  5. Run launcher validation
  6. For a new setup, start with a short smoke run using MAX_ITER, SAVE_ITER, and a small batch size before a full run.

What it can do on your machine

Read from SKILL.md and the folder at commit 6149ac5. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Shell commands in SKILL.md call:

    • bash

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Sparkdiffusion Finetune loads about 904 tokens when it runs. Until then it costs about 60 tokens; SKILL.md has 304 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~60
When it runs · the whole SKILL.md, loaded when a task matches
~904

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from AlibabaResearch/SparkDiffusion at commit 6149ac5, republished under its Apache-2.0 licence (© AlibabaResearch). 304 words, ~904 tokens.

Download SKILL.mdSave it as .claude/skills/sparkdiffusion-finetune/SKILL.md (or your agent's skills folder). This skill also uses 1 other file; get the full folder from GitHub.
name
sparkdiffusion-finetune
description
Configure and launch SparkDiffusion sparse finetuning for Wan 2.1 or Wan 2.2. Use when a user asks to train sparse attention parameters, select Wan experts, resume finetuning, or validate a finetuning configuration.

SparkDiffusion Sparse Finetuning

This skill handles the sparse-finetuning stage only. Distillation is a separate workflow.

Before Launching

  1. Work from the repository root and source the shared environment:

    bash
    source scripts/env.sh

  2. Export the required external SLA checkout before starting Python or torchrun:

    bash
    export SLA_SRC=/absolute/path/to/SLA
test -d "${SLA_SRC}/sparse_linear_attention"

    All standard sparse-finetuning experiments use RoLa and require the SLA backward kernel. Do not use a machine-specific default.

  3. Confirm the requested model, task, resolution, pretrained checkpoint, dataset shard pattern, GPU count, and output root.

  4. Confirm that the checkpoint and dataset exist. Use repository-relative defaults or explicit environment variables; never insert paths from another machine.

  5. Run launcher validation:

    bash
    bash -n scripts/sparse_finetune/*.sh

  6. For a new setup, start with a short smoke run using MAX_ITER, SAVE_ITER, and a small batch size before a full run.

Wan 2.1

Use one model and one training process:

bash
SLA_SRC=/absolute/path/to/SLA \
TASK=t2v \
MODEL_SIZE=14b \
RESOLUTION=480p \
NUM_GPUS=2 \
MAX_ITER=20 \
bash scripts/sparse_finetune/run_finetune_2pt1.sh

Important overrides:

  • TASK=t2v|i2v
  • MODEL_SIZE=1pt3b|14b
  • RESOLUTION=480p|720p
  • MODEL_ROOT or PRETRAINED_CKPT
  • DATASET
  • NUM_GPUS, CP_SIZE, FSDP_SHARD_SIZE
  • MAX_ITER, BATCH_SIZE, LR, SAVE_ITER
  • EXPERIMENT when using a custom registered configuration

Wan 2.1 1.3B is T2V-only in the standard launcher. Wan 2.1 I2V requires the 14B model and its image encoder.

Wan 2.2

The standard launcher supports high-noise, low-noise, and joint expert paths:

bash
export SLA_SRC=/absolute/path/to/SLA
TASK=t2v RESOLUTION=480p EXPERT=high \
  bash scripts/sparse_finetune/run_finetune_2pt2.sh

TASK=t2v RESOLUTION=480p EXPERT=low \
  bash scripts/sparse_finetune/run_finetune_2pt2.sh

TASK=t2v RESOLUTION=480p EXPERT=joint \
  bash scripts/sparse_finetune/run_finetune_2pt2.sh

EXPERT=both launches high and low training sequentially. It does not mean joint training.

For the two-expert setup, keep the high and low pretrained paths explicit:

bash
SLA_SRC=/absolute/path/to/SLA \
PRETRAINED_CKPT_HIGH=pretrain_weights/Wan2.2-T2V-A14B-Diffusers/transformer \
PRETRAINED_CKPT_LOW=pretrain_weights/Wan2.2-T2V-A14B-Diffusers/transformer_2 \
TASK=t2v RESOLUTION=480p EXPERT=joint \
bash scripts/sparse_finetune/run_finetune_2pt2.sh

Outputs and Resume

Training writes distributed checkpoints under the configured job output root, normally:

text
outputs/rola/<job>/checkpoints/iter_XXXXXXXXX/
  model/
  optim/
  scheduler/
  trainer/

Treat DCP as the resumable training checkpoint. Do not delete optim, scheduler, or trainer when resuming. Use the repository's configured checkpoint.load_path and load_training_state rather than manually copying shards.

Before passing a finetuning result to distillation or inference, inspect whether the consumer expects a model directory, a PTH file, or a DCP model directory. If the requested workflow requires a final portable PTH export, use the checkpoint-conversion workflow instead of assuming the DCP directory is a standalone weight file.

© AlibabaResearch, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 1 other file in .agents/skills/sparkdiffusion-finetune of AlibabaResearch/SparkDiffusion.

  • SKILL.md
  • agents/openai.yaml

Open the folder on GitHubat commit 6149ac5

Compare with similar skills

Sparkdiffusion Finetune next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Sparkdiffusion Finetune compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Sparkdiffusion Finetune this skillAlibabaResearch/SparkDiffusion542—~904Automated safety check: PassApache-2.0
Sentence-Transformers Training Routerhuggingface/skills11k1 repos~2.6kAutomated safety check: PassApache-2.0
Train RlOpenPipe/ART11k—~2.4kAutomated safety check: PassApache-2.0
Qwopus27b Rl TrainingR6410418/Jackrong-llm-finetuning-guide1.7k—~830Automated safety check: PassApache-2.0
Dataset Evaluationawslabs/agent-plugins9161 repos~1.3kAutomated safety check: PassApache-2.0
Train SftOpenPipe/ART11k—~2.9kAutomated safety check: PassApache-2.0

Similar skills

  • Sentence-Transformers Training Router

    huggingface/skills

    Official

    Routes a sentence-transformers training task to the right model type and required reference docs and example scripts, covering bi-encoders, rerankers, sparse and multi-vector models.

    11k GitHub starsUsed in 1 repo~2.6k tokens
    AI & LLM EngineeringAuto-check passed

  • Train Rl

    OpenPipe/ART

    RL training reference for the ART framework. An agent skill from OpenPipe/ART.

    11k GitHub stars~2.4k tokensUpdated today
    AI & LLM EngineeringAuto-check passed

  • Qwopus27b Rl Training

    R6410418/Jackrong-llm-finetuning-guide

    Prepare, validate, launch-plan, monitor, resume, and stop configurable Qwopus 27B reinforcement-learning workflows for GRPO or GSPO.

    1.7k GitHub stars~830 tokensUpdated 3 mo ago
    AI & LLM EngineeringAuto-check passed

  • Dataset Evaluation

    awslabs/agent-plugins

    Official

    Validates dataset formatting and quality for SageMaker model fine-tuning (SFT, DPO, or RLVR).

    916 GitHub starsUsed in 1 repo~1.3k tokens
    AI & LLM EngineeringAuto-check passed

  • Train Sft

    OpenPipe/ART

    SFT training reference for the ART framework. An agent skill from OpenPipe/ART.

    11k GitHub stars~2.9k tokensUpdated today
    AI & LLM EngineeringAuto-check passed

  • Fine Tuning With Trl

    Orchestra-Research/AI-Research-SKILLs

    Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training.

    13k GitHub starsUsed in 6 repos~2.9k tokens
    AI & LLM EngineeringAuto-check passed

More from AlibabaResearch/SparkDiffusion

  • Sparkdiffusion Distill

    AlibabaResearch/SparkDiffusion

    Configure and launch SparkDiffusion few-step distillation for Wan 2.1 or Wan 2.2.

    542 GitHub stars~876 tokensUpdated yesterday
    Auto-check passed

  • Sparkdiffusion Inference

    AlibabaResearch/SparkDiffusion

    Run validated SparkDiffusion inference for Wan 2.1 or Wan 2.2 T2V/I2V models.

    542 GitHub stars~853 tokensUpdated yesterday
    Auto-check passed

  • Sparkdiffusion Setup

    AlibabaResearch/SparkDiffusion

    Prepare a SparkDiffusion checkout for training or inference.

    542 GitHub stars~732 tokensUpdated yesterday
    Auto-check passed

Categories

Questions about Sparkdiffusion Finetune

What does Sparkdiffusion Finetune do?

Configure and launch SparkDiffusion sparse finetuning for Wan 2.1 or Wan 2.2. Sparkdiffusion Finetune is an agent skill from AlibabaResearch/SparkDiffusion.2.

When should I use Sparkdiffusion Finetune?

Sparkdiffusion Finetune fits situations like: A user asks to train sparse attention parameters; select Wan experts; resume finetuning; validate a finetuning configuration.

How do I install Sparkdiffusion Finetune in Claude Code?

Run `npx skills add AlibabaResearch/SparkDiffusion --skill sparkdiffusion-finetune -a claude-code`. Or copy the skill folder (.agents/skills/sparkdiffusion-finetune in AlibabaResearch/SparkDiffusion) into .claude/skills/sparkdiffusion-finetune in your project. Claude Code loads it when a task matches its description.

How do I install Sparkdiffusion Finetune in Codex?

Run `npx skills add AlibabaResearch/SparkDiffusion --skill sparkdiffusion-finetune -a codex`. Or copy the skill folder (.agents/skills/sparkdiffusion-finetune in AlibabaResearch/SparkDiffusion) into .agents/skills/sparkdiffusion-finetune in your project. Codex loads it when a task matches its description.

Can I use Sparkdiffusion Finetune in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add AlibabaResearch/SparkDiffusion --skill sparkdiffusion-finetune -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/sparkdiffusion-finetune, .gemini/skills/sparkdiffusion-finetune, .github/skills/sparkdiffusion-finetune and .opencode/skills/sparkdiffusion-finetune in your project.

What does Sparkdiffusion Finetune need to run?

Going by SKILL.md and its folder, Sparkdiffusion Finetune needs the command-line tools its instructions call (bash). Our summary lists: Python 3.

Does Sparkdiffusion Finetune access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Sparkdiffusion Finetune safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Sparkdiffusion Finetune use?

Sparkdiffusion Finetune is published under the Apache-2.0 licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Sparkdiffusion Finetune use?

About 904 tokens (SKILL.md is roughly 3.6k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Sparkdiffusion Finetune?

Skills that share tags, products or a category with Sparkdiffusion Finetune: Sentence-Transformers Training Router (huggingface/skills, 11k stars), Train Rl (OpenPipe/ART, 11k stars), Qwopus27b Rl Training (R6410418/Jackrong-llm-finetuning-guide, 1.7k stars) and Dataset Evaluation (awslabs/agent-plugins, 916 stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Sparkdiffusion Finetune?

AlibabaResearch (a GitHub organization) maintains it in AlibabaResearch/SparkDiffusion, which has 542 GitHub stars. The repository holds 4 skills in this directory. The repository was last updated on October 9, 2026.

Source: AlibabaResearch/SparkDiffusion on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.