sf export

Convert a checkpoint to HuggingFace format

sf export my-grpo \
  --checkpoint checkpoints/step_500 \
  --checkpoint-format nemo-dcp

Converts a training checkpoint into HuggingFace format, which is what the Playground, model deployments, and everything outside this platform can load.

--checkpoint takes the path as the artifact registry recorded it — read it off the job’s Artifacts tab, or from sf job status. --checkpoint-format must match what the framework wrote; getting it wrong fails early with the formats it does accept.

Push to the Hub

sf export my-grpo \
  --checkpoint checkpoints/step_500 --checkpoint-format nemo-dcp \
  --push-repo alice/qwen-support-v3

Needs a linked Hugging Face account — see Hugging Face.

Automatically, after training

sf submit my-grpo --profile h200:8 --then export

Exports the final checkpoint as soon as the run succeeds, so the model is loadable without anyone having to remember.

sf export

Convert a checkpoint to HuggingFace format

sf export <EXP> [options]

Argument	Description
EXP	Experiment name or path (required)

Option	Description
--checkpoint	Checkpoint path as recorded in the artifact registry (required)
--checkpoint-format	nemo-dcp | nemo-megatron | verl-fsdp | verl-megatron | huggingface (required)
--push-repo	Upload the converted weights to this HF Hub repo (user/name, needs HF_TOKEN)
--profile	Hardware profile (managed by the server registry; sf status lists the available values)
--allow-dirty	Allow uncommitted changes in the working tree
--dry-run	Only print the command that would be submitted