Skip to content

feat(cosmos3): Cosmos3-Nano DROID SFT performance optimization - #148

Open
xiaojunjie wants to merge 1 commit into
baidu-baige:masterfrom
xiaojunjie:cosmos3-perf-opt
Open

feat(cosmos3): Cosmos3-Nano DROID SFT performance optimization#148
xiaojunjie wants to merge 1 commit into
baidu-baige:masterfrom
xiaojunjie:cosmos3-perf-opt

Conversation

@xiaojunjie

Copy link
Copy Markdown
  • move the post-ColorJitter augmentation tail (view split, half-resolution shoulder views, concat_view layout, uint8 formatting, temporal truncation, resize + pad) into the data preprocessor so it can run on the model device via data.colorjitter_on_gpu; pipeline stage order and RNG draws unchanged
  • rename run_cosmos3_nano_droid_fsdp.sh -> run_cosmos3_nano_droid_fsdp_finetune.sh and drive VAE / dataset / checkpoint / tokenizer paths from env vars
  • forward extra trainer kwargs to models declaring only **kwargs
  • docs: refresh EN/ZH quick_start_cosmos3 (resource paths, key arguments, performance switches, correctness verification)

Verified against the official cosmos-framework launch_sft_action_policy_droid recipe: loss is bit-identical over 26 steps (max |delta| = 0).

- move the post-ColorJitter augmentation tail (view split, half-resolution
  shoulder views, concat_view layout, uint8 formatting, temporal truncation,
  resize + pad) into the data preprocessor so it can run on the model device
  via data.colorjitter_on_gpu; pipeline stage order and RNG draws unchanged
- rename run_cosmos3_nano_droid_fsdp.sh -> run_cosmos3_nano_droid_fsdp_finetune.sh
  and drive VAE / dataset / checkpoint / tokenizer paths from env vars
- forward extra trainer kwargs to models declaring only **kwargs
- docs: refresh EN/ZH quick_start_cosmos3 (resource paths, key arguments,
  performance switches, correctness verification)

Verified against the official cosmos-framework launch_sft_action_policy_droid
recipe: loss is bit-identical over 26 steps (max |delta| = 0).
@github-actions github-actions Bot added checkpoint documentation Improvements or additions to documentation model training labels Aug 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

checkpoint documentation Improvements or additions to documentation model training

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant