nvidia

/

Cosmos3-Nano-Sim-Bimanual

Like 13

Follow

NVIDIA 75.4k

This PR updates the config of Sim-Bimanual model to include sink_frames and window_frames

@mbalaNV , is the mecka_all dataset necessary to mention?

And my agent has 2 other questions:

1. The new manifest_sha256 doesn't verify. On main it equals the SHA-256 of the manifest as sorted, compact JSON. On the PR it matches none of 16 serialisations I tried, nor any mix of old and new values for the changed keys. As far as I could find, nothing in SGLang or the framework checks it. It's still an integrity field that's now wrong, so it should be regenerated.
2. Stale duplicate files. The cosmos_dreams_* files in Bimanual and Sim-Transfer keep the old values and refer to 7-shard weight files that don't exist. That was true before the PR, but now they contradict the updated files. I'd delete them; I didn't confirm whether anything reads them.

@yaoxuhf
The mecka_all dataset is coming from including this dataset in imaginaire4, not from any changes in the conversion script. Since we reran the conversion script using the newest main from imaginaire4, this dataset was picked up. I don't think it's necessary, but it's also not harmful and could be useful later. I don't think modifying the conversion script to explicitly exclude it makes sense.

For other questions:

1. The manifest_sha256 should match now I think. Can you double check with mbala/conversion_script_bugfixes branch, as that's what I'm using to generate this checkpoint?
2. I removed the cosmos_dreams* files as they were indeed stale and a left-over from the model rename.

I also used newer diffusers and transformers version in this export to more closely match the README docs. It doesn't change any behavior. but you can notice e.g. changes to text_tokenizer/tokenizer_config.json , which are coming from the transformers version change.

The modular_model_index.json file now also includes default_use_system_prompt.