推論系統
Anthropic Releases 36 Biological Model Acceleration Packages, Incorporating Numerical Consistency Into Execution Modes
Developed with Claude’s assistance, the 36 packages use shared GPU kernels to accelerate biological models and reduce memory requirements. Anthropic distinguishes between identical-output and approximate acceleration, but does not commit to ongoing maintenance of the public release.
On September 17, Anthropic released 36 inference optimization packages for biological models, developed with Claude’s assistance and covering structure prediction, protein design, and genomic models. In structure prediction tests, Anthropic reported an average speedup of approximately 4× when small numerical differences were permitted, compared with approximately 1.6× when identical outputs were required. These two measures must be considered separately. [Research announcement](https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling)
The technical core is FlashPairformer, which provides GPU kernels for the triangle attention and triangle multiplication operations repeatedly used in molecular structure models. A shared execution layer selects implementations based on the device, data type, and tensor shape. Interfaces connect PyTorch, JAX, and different kernel backends, allowing multiple models to reuse the optimizations. [Shared kernel documentation](https://raw.githubusercontent.com/anthropics/uplifting-biomolecular-modeling/main/common/opt_core/README.md)
The packages separate behavior into four modes: `off` runs a pinned version of the upstream code; `exact` aims for identical outputs; `fast` accepts documented numerical differences; and `big` prioritizes reducing GPU memory requirements. This lets researchers determine their error tolerance before choosing a performance strategy, although the modes available vary by model. [Project overview](https://github.com/anthropics/uplifting-biomolecular-modeling)
For Boltz-2, for example, `exact` uses CUDA Graph and hoists loop-invariant computations out of loops, while `fast` additionally incorporates low-precision triangle operations. The `big` mode processes data in chunks and can distribute pair representations by row across multiple GPUs on the same host. However, the documentation notes that each invocation still incurs approximately 25 seconds of overhead for startup, weight loading, and feature processing, limiting end-to-end gains for small inputs. Bitwise identical results also cannot be maintained for certain input lengths on A100 GPUs; in those cases, `exact` reports that it is not enabled and exits with an error code. [Boltz-2 documentation](https://raw.githubusercontent.com/anthropics/uplifting-biomolecular-modeling/main/boltz2/README.md)
The release also includes a reproducibility workflow: each package lists upstream versions, dependencies, and weight checksums, and provides a container or virtual environment setup. Users can check whether a mode can be enabled before execution, while logs record the optimizations actually applied. This information helps distinguish differences caused by models or execution environments from failures in acceleration kernels, reducing the risk of misleading conclusions based on a single runtime figure. [Reproducibility and execution logs](https://github.com/anthropics/uplifting-biomolecular-modeling)
Deployment limitations are equally explicit: this is a reference release tied to pinned upstream versions. Anthropic states that it does not plan to provide ongoing updates and does not accept pull requests. Original code is licensed under Apache 2.0, while upstream code retains its respective licenses. For research teams, the immediately usable assets are the inspectable kernels and mode-switching mechanism. Teams should then separately measure cold-start time, steady-state inference, peak memory usage, and output error on their own data, and verify continued compatibility after upstream upgrades. [Versioning and maintenance notes](https://github.com/anthropics/uplifting-biomolecular-modeling)