cosmos-gr1-sft-nano-state-20k-8gpu

cosmos-gr1-sft-nano-state-20k-8gpu

SFT checkpoint at iteration 20000 of NVIDIA Cosmos3-Nano (Omni-MoT World Foundation Model) fine-tuned on the GR-1 (Fourier GR-1, 44-DOF) LeRobot dataset using the built-in action path (action_gen / DomainAwareLinear / rectified_flow_action, gr1 embodiment, domain id 21).

Run: cosmos3_gr1_sft_nano_20k_8gpu_45ktoken_state_1000 — 8×B200, FSDP, 45k-token state config.

Contents

PyTorch Distributed Checkpoint (DCP) format, layout preserved for direct resume:

iter_000020000/
  model/      # 8 sharded .distcp + .metadata  (model weights only)
  scheduler/  # LR scheduler state
  trainer/    # trainer/step metadata
latest_checkpoint.txt   # -> iter_000020000

Optimizer state (optim/) is intentionally omitted — this checkpoint is for inference/eval and weight loading, not exact optimizer resume.

Usage

Download:

hf download periphanes/cosmos-gr1-sft-nano-state-20k-8gpu --local-dir ./ckpt

The DCP model/ directory is loaded by the cosmos-framework training/eval harness. To export to a single consolidated state dict, use torch.distributed.checkpoint (DCP) loading APIs or the framework's conversion utilities. The shard count (8) reflects the 8-way FSDP training layout but DCP loading is resharding-agnostic.

Notes

• Base model: nvidia/Cosmos3-Nano. License inherits from the base model — see NVIDIA's terms.
• Trained as part of long-horizon latent world-modeling research for VLA models.