Protenix 0.70 release vs cuequivariance speed

[rakeshr10]

@zhangyuxuann I have a question related to your new release. Is the newest 0.70 release faster than the cuequivariance based version, are there any significant advantages compared to 0.6.3 and does the accelerations work only on certain kinds of GPUs like A100, H200 etc.

When you say batched inference does it mean simultaneously modeling multiple protein complexes (containing multiple protein, DNA etc sequences) on a single GPU. I could n't locate any examples related to this.

Also was interested to know if this release contains the new Protenix+ or Mini+ model (https://arxiv.org/pdf/2510.12842) shown to be lightweight and faster.

Regards
Rakesh

[yangyanpinghpc]

inference time：0.6.3 vs 0.7.0 https://github.com/bytedance/Protenix/blob/main/assets/inference_time_vs_ntoken.png
optimization：kernel fusion, subgraph optimization, and parallelization of diffusion N-sample computation, remove unnecessary torch.cuda.empty_cache
These acceleration optimizations are general-purpose and do not rely on specific GPU.
Batched inference optimization: Parallelized diffusion N-sample computation, where the parallel count is controlled by sample_diffusion_chunk_size. This parameter’s setting is updated from 1 to 5（default config）.