Inference backends

rfd4-generate runs the native model by default (inference_backend=native, compiled or eager according to compile_lengths). inference_backend=bioir runs the whole workflow through BioIR instead, with the same input and sampling options:

rfd4-generate inputs=data/benchmarks/monomer out_dir=./runs/bioir inference_backend=bioir

BioIR is built for CUDA 13. Install it with pip install "rfd4-proteina[metrics-cuda13,bioir]"; the cu13 Docker image ships it, and the cu12 and cpu images stay native. It runs the released checkpoints, including the 147-class autoencoder, wide recycle conditioning and the checkpoint’s LoRA adapters (folded into the base weights when the model loads).

BioIR runs feature representations, sampling and the autoencoder; input parsing, structure formatting, metrics and reward adapters stay native. It needs bionemo-ir 0.2.0 or later, which provides the RFProteina model (graph_status, reset_graphs, configure_graphs); older runtimes are rejected before anything is converted. It is available from rfd4-generate, not from training-time validation. These options are rejected when BioIR is selected, instead of being ignored: gradient or classifier-free guidance, trajectory export, LoRA adapters passed through adapter_path and native nn.* overrides.

Native beam search (search.checkpoints) runs through BioIR’s own beam search, scored with the native metric and oracle managers and the per-checkpoint metrics and rewards you configure. BioIR supports one branching width, rollout scoring, a last checkpoint at the end of the schedule and a scalar minimum for saved intermediates; use_clean_prediction, continuation_criteria, per-checkpoint n_branch, save_states and resume_from are rejected. By default BioIR draws its own noise, so results are not bitwise identical to the native backend; bioir.match_reference_rng=true draws in native’s order, which is what a one-to-one comparison needs. Under ODE sampling from identical noise, on the released checkpoints in bf16 (144 PPI designs), the median aligned CA RMSD between the backends is 0.17 A after one step, 2.2 A after four and 0.8 A after sixty: bf16 kernel differences move trajectories between basins after the first few steps and then neither grow nor heal. Compare designs by distribution, not one to one. The first velocity prediction from identical noise is the exception, and test_bioir_first_step_velocity_matches_native pins it.

BioIR takes no padding buckets: compile_lengths is ignored, and each request runs at its own length. CUDA graphs are captured per (batch, tokens) shape. The settings live under bioir: in configs/inference.yaml:

Key	Meaning	Default
bioir.cuda_graph	Capture CUDA graphs for the velocity trunk	true
bioir.cuda_graph_max_tokens	Requests above this many tokens run the trunk eagerly (graph buffers grow with the square of the length); null keeps BioIR’s limit	511
bioir.max_graphs	Graph slots; null reserves two (a full batch and its tail), an explicit value below two is rejected	null
bioir.attention_backend	BioIR attention kernel; null selects its default	null
bioir.match_reference_rng	Draw noise and feature permutations in the native backend’s order, so one seed gives both backends the same noise; for accuracy comparison, costs an extra draw per design	false

What BioIR refuses rather than approximates

The bridge raises instead of silently running a different algorithm. It rejects:

• an installed BioIR that lacks what the workflow needs, checked before anything is converted: the model API (graph_status, reset_graphs, configure_graphs), the config fields tail_steps, tail_step_multiplier, decoder sequence-conditioning and match_reference_rng, and the random-stream hooks (normal_draws, permutations); and RFProteinaVFStepConfig.tsr_scaled_score_drift, which stochastic VF-TSR sampling uses;

• a checkpoint whose residue vocabulary or recycle conditioning the installed BioIR cannot convert. The released 147-class AE and wide recycle conditioning convert on current BioIR, and are refused up front, by name, only on a runtime that predates them;

• guidance, trajectory export, non-beam search and search.continuation_policy other than stop.