> This page is for version 0.2.0 (latest) (default). > For other versions, use one of these documentation indexes: > - 0.2.0 (latest) (default): https://docs.nvidia.com/bionemo/inference-runtime/latest/llms.txt > - 0.1.1: https://docs.nvidia.com/bionemo/inference-runtime/0.1.1/llms.txt > - 0.1.0: https://docs.nvidia.com/bionemo/inference-runtime/0.1.0/llms.txt > For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/bionemo/inference-runtime/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/bionemo/inference-runtime/_mcp/server. # 0.2.0 Released 2026-10-20. ## Key Features and Enhancements - **RFProteína.** BioIR runs RFProteína (RFD4) as a conditional sampler. One call returns a sequence and an all-atom structure from a length and optional conditioning. - **Design tasks.** One model covers unconditional generation, fold-class and secondary-structure steering, indexed and unindexed motif scaffolding, protein and small-molecule binders, nucleic-acid binders, partial redesign, inverse folding, and metric-guided beam search. - **Checkpoints.** Load the released velocity-network and autoencoder checkpoints and call `RFProteina`. Checkpoints that carry training metadata load. - **Sampling schedule.** The default is the released medium policy: 60 steps, 185 trunk passes, a denser tail on both modalities, and a decoder conditioned on the predicted sequence. - **Design replay.** Sampling records one trunk recycle and replays it, so a shape is captured once whatever depths that schedule uses. - **Motif coordinates.** Motif scaffolding reads internal-coordinate conditioning from the coordinate group. A motif supplied only as internal coordinates is no longer dropped. ## Performance - **Folding replay.** For inputs up to 1024 tokens, Boltz-2, OpenFold3, and Protenix record one trunk recycle, one diffusion step, and the confidence Pairformer, and replay them on later calls of the same shape. Boltz-1 records the diffusion step. OpenFold2 and AlphaFold2 record one trunk recycle. The diffusion step extends to 2048 tokens on GPUs with at least 64 GiB. Larger inputs run without replay. Token counts are padded so nearby lengths share one recording. A serving request does not rerun the first recording to check it. - **Stable replay inputs.** A replay copies only tensors that changed. Diffusion steps keep features, trunk outputs, and prepared biases in place. - **Conditioning reuse.** Noise-independent pair and single conditioning, atom-pair biases, and diffusion noise levels are computed once per rollout. Boltz deduplicates diffusion conditioning across samples. Protenix reuses denoiser conditioning across samples, layers, and steps. - **Triangle updates.** Boltz-2, Protenix, and OpenFold3 embed relative position in one pass. Below 512 tokens, triangle attention uses a narrower tile. - **Fused triangle multiplication.** With token padding on, Boltz, Protenix, and OpenFold3 run triangle multiplication in the MSA, template, and pairformer stacks as one chain: input normalization and gated projections, the contraction, then output normalization, projection, gate, and masked residual. OpenFold2 does this in the Evoformer and extra-MSA stacks. RFProteína does it in the velocity-network pair updates. Confidence heads, OpenFold2 templates, and token counts that are not multiples of 8 take the dual-product path, including a residual add. On Hopper the output-tile weights stay resident while rows stream past. On Blackwell that projection also feeds conditioned transitions. - **Fused attention output.** Triangle attention and pair-biased attention fuse the output gate, projection, and residual, including the diffusion adaptive output gate. A residual-free form leaves the add to the caller. Token transformers use it when they add the residual themselves. The sigmoid gate can carry the residual. - **Fused transitions.** A transition keeps its hidden activation on chip. A conditioned transition that does not take that path uses the dual-product SwiGLU gate, then a fused SwiGLU. Adaptive normalization fuses the norm and the sigmoid gate. - **Fused pair updates.** Pair-to-single averaging and single-to-pair outer products are fused. OpenFold2 runs the dense outer product in BF16. Pair-bias normalization, projection, and layout share one pass. - **Atom transformers.** OpenFold3 and Protenix run atom attention in BF16 inside the FP32 encoder and decoder, and gather each local atom window in one pass. Boltz gathers atoms by token instead of a one-hot product. OpenFold3 projects only the atom pairs it reads, applies the encoder activation inside the atom reduction, and drops the unused atom-index matrix. - **Protenix trunk.** Pairformer nodes whose weights are zero are skipped. Triangle-bias heads are projected in the same pass as LayerNorm. Confidence summaries stay on the host. pLDDT weights are not rebuilt for every atom. - **Longer OpenFold3 inputs.** OpenFold3 fits longer sequences with the same recycles, diffusion steps, samples, and confidence outputs. Static atom conditioning and pair conditioning stay on the device for the rollout. Protein conformers are prefetched. - **Structure math.** OpenFold2 fuses inference rotation products. Quaternion conversion is fused. Rigid alignment reuses its solver setup. Diffusion noise increments are prepared once per rollout and do not change the random draws. - **Less host time.** Deep MSA decoding, template binning, and the OpenFold2 and OpenFold3 tokenizers take less time. Boltz-2 no longer waits on the device at every diffusion step. A serial processor overlaps CPU stages with inference instead of starting a distributed runtime. - **Pipeline confidence.** Pipelines keep the confidence scores and maps the writer uses and omit unused heads. Distance bins and their projection are fused. Existing calls keep the same arguments, and the written structure and confidence scores stay in place. - **Design sampling.** RFProteína projects shared pair biases in one pass and conditions pair adaptive normalization once per design. The first pass does not recycle a zero state. Sampling does not copy results back to the host on every step. ## Fixed Issues - **Ligand atom names.** Atoms outside the standard chemical vocabulary keep their real names and elements when a pipeline writes them. - **SMILES ligands.** A SMILES ligand is no longer written as protein only, and it no longer fails at the writer with duplicate atom ids. - **PDB templates.** OpenFold2 and OpenFold3 accept protein templates in PDB as well as mmCIF. Author chain ids, insertion codes, and unresolved residues are preserved. - **Template chain ids.** PDB inputs use author chain ids. mmCIF inputs keep label chain ids. When the sequence is taken from observed residues, selenomethionine is mapped to methionine. - **Protenix MSA.** MSA rows are resampled on each recycle with a fixed depth, so a recorded recycle still replays. - **Protenix seed.** An omitted sampler seed follows the caller's random state. An explicit sampling seed still overrides it. - **Protenix features.** Stock unbatched features are accepted, and missing atom windows are built inside the forward pass. - **Protenix memory.** On GPUs with at most 24 GiB, activations are chunked above 1024 tokens instead of running out of memory. - **Revised structures.** A revised RCSB entry is taken from the wwPDB archive when the newest RCSB file is a later remediation. - **Downloaded archives.** Archives are checked before they are used. The full public history for this release is on [GitHub](https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime/commits/v0.2.0).