| # Default configs for inference through generate.py | |
| # The defaults are geared more towards ease-of-use by designers, so running with | |
| # the existing analysis scripts will necessitate overrides | |
| defaults: | |
| - _self_ | |
| - inference_experiment: inference_on_user_inputs # Default for design | |
| - recycle_strategy: thinking_medium | |
| # NOTE: The experiment configs can override values specified in this config | |
| run_name_: inference | |
| verbose: false # Set true for routine warnings and detailed diagnostic output. | |
| hf_ckpt_path: nvidia/RFD4-Proteina # Optional local model bundle or HF repo; overrides ckpt_path/name when set. | |
| hf_ckpt_revision: a06e2509c7033b010fc43614e9fd840303588502 # Pin a Hub commit for reproducible model/AE loading. | |
| out_dir: null # If null, defaults to ./runs/<run_name>. For production, set to an external | |
| # path (e.g. /shared/runs/experiment_name). SLURM scripts set this | |
| # automatically. All outputs (samples, metrics) go under this directory. | |
| # Current SOTA; generation and regression tooling resolve the immutable checkpoint from here. | |
| # Regenerate the inference baselines when advancing it; use CLI overrides for one-off checkpoints. | |
| ckpt_name: rfd4-proteina-flow-release-2026-10-19.ckpt | |
| ckpt_path: ${oc.env:DATA_PATH}/flow_checkpoints | |
| adapter_path: null # Local PEFT adapter directory; load onto the exact base/AE pair used for fine-tuning. | |
| # Only applies when a pre-oracle design filter is active (i.e. its tag gate matched). | |
| # Determines whether to delete structure files for designs that fail this filter | |
| delete_failing_pre_oracle: True | |
| autoencoder_ckpt_path: ${oc.env:DATA_PATH}/ae_checkpoints/rfd4-proteina-ae-release-2026-10-19.ckpt | |
| num_replicates: 1 # Number of times to replicate each example | |
| # Token counts to pad batches to for torch.compile. Multiples of 16; null disables. | |
| compile_lengths: [128, 160, 192, 224, 256, 288, 320, 352, 384, 448, 512, 640, 768, 896, 1024] | |
| # Network settings that may change after training without touching the weights; null keeps the checkpoint's | |
| # value. Example: nn.torch_compile_mode=reduce-overhead nn.tri_mult_backend=cuequivariance | |
| nn: | |
| torch_compile_mode: null | |
| torch_compile_fullgraph: null | |
| tri_mult_backend: null | |
| attention_type: null | |
| precision: bf16-mixed | |
| batch_size: 1 | |
| subset_to_examples: null # If specified, only run inference on these example indices | |
| prevalidate_inputs: false # Whether to sanity-check each unique input once prior to generation (fail fast on bad inputs). | |
| seed: null # Will set a random seed if null | |
| torch_deterministic: false # Locks cuDNN/cuBLAS to deterministic algorithms; significantly reduces throughput | |
| skip_existing: true # If true, skip examples whose output structure files already exist in the output directory | |
| save_chain_of_interest: false # If true, write per-design chain-of-interest CIFs (only needed for foldseek clustering). | |
| compress_outputs: false # Compress persistent structure outputs, including refolds and CIF trajectories. | |
| metrics_tags: [] # If specified, determines which metrics are computed during inference | |
| search: | |
| checkpoints: null # Set checkpoint policies to enable beam search. | |
| save_states: false | |
| resume_from: null | |
| resume_seed: null | |
| num_workers: 4 # Number of dataloader workers | |
| hardware: | |
| accelerator: gpu | |
| ngpus_per_node_: 1 | |
| nnodes_: 1 | |
| # Cluster job settings (used by launch_infer_from_local.sh) | |
| ncpus_: 24 # Number of CPUs per task | |
| gen_njobs: 1 # Number of parallel generation jobs per config | |
| eval_njobs: 1 # Number of parallel evaluation jobs per config | |
| job_id: null # Set by launch_generate.sh; used to disambiguate per-job output files | |
| hydra: | |
| run: | |
| dir: ./hydra_outputs | |