Question: Inference strategies and configurations for long RNA sequences (>1000 nt)

Hi Protenix team,

Thanks for releasing this great work and making it open source.

I am currently using Protenix for RNA 3D structure prediction, and I encountered out-of-memory (OOM) issues when running inference on long RNA sequences (e.g., >1000 nucleotides) on a 24GB GPU.

From the logs:
- N_token ≈ 1200–1300
- N_atom ≈ 27k
- OOM occurs during the trunk / attention computation

I have a few questions regarding long-sequence inference:

### 1. Are there recommended inference configurations for long RNA sequences?
For example:
- Reducing diffusion steps or cycles
- Disabling or simplifying certain modules (e.g., template, MSA)
- Using different precision settings
- Any hidden or advanced flags in `runner/inference.py`

### 2. Does Protenix support any form of memory-efficient inference?
For example:
- Chunked computation
- Attention chunking / block attention
- Gradient checkpointing-like techniques (for inference)
- Offloading to CPU

### 3. Is multi-GPU inference supported for a single long sequence?
Currently, it seems that inference runs on a single GPU even when multiple GPUs are visible (world size = 1).
Are there plans or existing methods to enable model parallelism or distributed inference for large inputs?

### 4. Are there recommended workflows for long RNA sequences?
For example:
- Domain decomposition (splitting long RNA into segments)
- Hierarchical modeling
- Combining local predictions into a global structure

### 5. Is Protenix-Mini or any other lighter model recommended for long sequences?

---

Overall, I would really appreciate any guidance or best practices for handling long RNA sequences efficiently.

Thanks again for your work!

Start with runner/inference.py and the reported logs showing N_token, N_atom, and trunk/attention OOM on a 24GB GPU. Determine which long-sequence configurations, memory-efficient methods, multi-GPU options, and lighter-model workflows are actually supported; the issue is done only when a concrete, reproducible recommendation or implementation scope is established.

Written by the indexing model from the issue text.