Skip to content

Add examples of fintuning Cosmos3#218

Merged
foreverlms merged 6 commits into
mainfrom
finetune-cookbook
Jun 17, 2026
Merged

Add examples of fintuning Cosmos3#218
foreverlms merged 6 commits into
mainfrom
finetune-cookbook

Conversation

@foreverlms

@foreverlms foreverlms commented Jun 16, 2026

Copy link
Copy Markdown
Collaborator

No description provided.

@foreverlms foreverlms marked this pull request as draft June 16, 2026 14:19
lfengad
lfengad reviewed Jun 16, 2026
View reviewed changes
Comment thread cookbooks/cosmos3/finetune/README.md Outdated
Comment thread cookbooks/cosmos3/finetune/README.md Outdated
Comment thread cookbooks/cosmos3/reasoner/finetune/toml/sft_config/llava_ov.toml
Comment thread cookbooks/cosmos3/finetune/launch_sft_videophy2_nano.sh Outdated
@lfengad

lfengad commented Jun 16, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

https://github.com/QwenLM/Qwen3-VL/tree/main/qwen-vl-finetune we could mimic this to make the finetune part with a new top level folder?
Or we could follow #214 this PR, then add finetune folder as sibling of the basic_example folder in each case?

@foreverlms foreverlms marked this pull request as ready for review June 16, 2026 14:54
@Xuanmeng-Zhang
restructure cookbook into per-capability folders + address review
8abd6d8 
  - Place SFT recipes under their capability cookbooks: vision recipes ->
    cookbooks/cosmos3/generator/audiovisual/finetune/, reasoner recipes ->
    cookbooks/cosmos3/reasoner/finetune/ (sibling of existing inference content,
    forward-compatible with #214).
  - Surface the finetune cookbooks on the repo landing page (README Finetune).
  - Trim each README to the happy path; link advanced config + raw torchrun to the
    canonical framework docs (training.md, sft_config.md).
  - Add the recommended NGC PyTorch base image to Prerequisites.
  - Rewrite each launch_sft_*.sh as a simple, self-contained recipe: linear
    numbered steps (download -> convert -> train) with hardcoded paths, dropping
    the shared launcher helper and all env-var override knobs.
@Xuanmeng-Zhang

Xuanmeng-Zhang commented Jun 16, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

https://github.com/QwenLM/Qwen3-VL/tree/main/qwen-vl-finetune we could mimic this to make the finetune part with a new top level folder? Or we could follow #214 this PR, then add finetune folder as sibling of the basic_example folder in each case?

I followed with the #214-style placement. Rather than a separate top-level finetune/ folder, each recipe sits as a finetune/ sibling next to the existing inference content in its capability cookbook (generator/audiovisual/finetune/, reasoner/finetune/). This mirrors the cookbook layout and stays forward-compatible with #214. When #214 moves inference into basic_examples/, our finetune/ is already the sibling it expects, with no rename/conflict.

lfengad
lfengad approved these changes Jun 17, 2026
View reviewed changes

@lfengad lfengad left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, we need to make sure these recipes have passed locally.

@Xuanmeng-Zhang

Xuanmeng-Zhang commented Jun 17, 2026

Copy link
Copy Markdown
Collaborator

LGTM, we need to make sure these recipes have passed locally.

Got, will double check before merge.

@foreverlms

foreverlms commented Jun 17, 2026

Copy link
Copy Markdown
Collaborator Author

SFT-NANO

image

@foreverlms foreverlms merged commit 444d861 into main Jun 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Xuanmeng-Zhang Xuanmeng-Zhang Xuanmeng-Zhang left review comments

@lfengad lfengad lfengad approved these changes

@Dinghow Dinghow Dinghow approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@foreverlms @lfengad @Xuanmeng-Zhang @Dinghow