nvidia

/

Cosmos3-Super-Image2Video-4Step

like 25

Follow

NVIDIA 64.3k

Adds an SGLang Diffusion usage section to the model card, mirroring the structure of the nvidia/Cosmos3-Super-Image2Video card.

Contents:

• Install block and serve command.
• A note that the 64B transformer is ~128 GB in BF16 and that sequence and classifier-free-guidance parallelism divide activations rather than weights, so on GPUs with less memory the weights need --use-fsdp-inference (sharding) or --dit-layerwise-offload (layer streaming).
• The fixed 4-step distilled schedule: SGLang reads it from scheduler/scheduler_config.json, so num_inference_steps and flow_shift are ignored and guidance_scale is forced to 1.0.
• An example request, verified end-to-end on 4xH100.

Please do not merge yet - holding until Cosmos 3 support is merged into SGLang: sgl-project/sglang#31590.