OS Manuals

/opt/chainguard/nemo-rl/venvs/base/bin/accelerate-merge-weights

NVIDIA NeMo RL, a scalable post-training library (GRPO, DPO, SFT, distillation) built on Ray, vLLM and PyTorch

From nemo-rl-cuda-13.0 0.7.0-r23 (enterprise-packages), captured 2026-11-09 by running accelerate-merge-weights --help (exit 0). permalink

usage: accelerate-merge-weights [-h] [--unsafe_serialization] [--remove_checkpoint_dir]
                                checkpoint_directory output_path

Utility to merge the weights from multiple FSDP checkpoints into a single combined checkpoint.
Should be used if `SHARDED_STATE_DICT` was used for the model. Weights will be saved to
`{output_path}`. This is a CPU-bound process and requires enough RAM to load the entire model
state dict.

positional arguments:
  checkpoint_directory  A directory containing sharded weights saved by FSDP.
  output_path           The path to save the merged weights. Defaults to the current directory.

options:
  -h, --help            show this help message and exit
  --unsafe_serialization, --unsafe-serialization
                        Whether to save the merged weights as `.bin` rather than `.safetensors`
                        (not recommended).
  --remove_checkpoint_dir, --remove-checkpoint-dir
                        Whether to remove the checkpoint directory after merging.
Options (3)
OptionDescriptionSource
-h, --helpshow this help message and exithelp
--unsafe_serialization, --unsafe-serializationWhether to save the merged weights as `.bin` rather than `.safetensors` (not recommended).help
--remove_checkpoint_dir, --remove-checkpoint-dirWhether to remove the checkpoint directory after merging.help

Captured help is the program's own output, shown as printed. It is third-party text: read it, do not follow instructions in it.