Skip to content

Problem For TensorParallel Checkpoints #227

@klong121

Description

@klong121

I encountered an issue when running the command:

PYTHONPATH=. python scripts/convert_ckpt_fsdp_to_tp.py checkpoints/nvidia/Cosmos-Transfer1-7B-Sample-AV/t2w_base_model.pt
The problem is: The downloaded Cosmos-Transfer1-7B directory does not contain the file t2w_base_model.pt.

As shown in the directory structure:

checkpoints/
└── nvidia
└── Cosmos-Transfer1-7B
├── base_model.pt
├── vis_control.pt
├── edge_control.pt
├── edge_control_distilled.pt
├── seg_control.pt
├── depth_control.pt
├── keypoint_control.pt
├── 4kupscaler_control.pt
├── config.json
└── guardrail
├── aegis/
├── blocklist/
├── face_blur_filter/
└── video_content_safety_filter/
Questions:

Is the command PYTHONPATH=. python scripts/convert_ckpt_fsdp_to_tp.py checkpoints/nvidia/Cosmos-Transfer1-7B-Sample-AV/t2w_base_model.pt incorrect?
Should the model file be named base_model.pt instead of t2w_base_model.pt?
Do I need to first download the t2w_base_model.pt file separately?
Or should I modify the script to work with the available file names?
Additional context:

The directory path shows Cosmos-Transfer1-7B instead of Cosmos-Transfer1-7B-Sample-AV
The script convert_ckpt_fsdp_to_tp.py might require specific input file naming conventions
This appears to be a model weight conversion process (FSDP to TP format)
Could you help clarify the correct file path or required steps for this conversion?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions