Known Issues¶

This page tracks user-impacting issues in the macOS/MLX port.

trp_cage Non-determinism Across Runs¶

In some configurations, trp_cage can show run-to-run variation beyond expected stochasticity.

Workarounds:

• Use a fixed --seed.
• Keep precision and diffusion settings constant between comparisons.
• Prefer relative comparisons on identical hardware/software environments.

No Apple Neural Engine (ANE) Backend¶

MLX in this project runs on CPU/GPU. ANE acceleration is not currently used.

Impact:

• Performance expectations should be based on GPU execution, not ANE offload.

Large Inputs Can Exhaust Unified Memory¶

Very large complexes can run out of memory, especially with high sample counts.

Workarounds:

• Reduce --num_samples.
• Reduce --diffusion_steps for exploratory runs.
• Use --precision float16 or --precision bfloat16 when appropriate.

Full Data Pipeline Requires External Databases¶

--run_data_pipeline requires correctly configured HMMER binaries and large sequence/template databases.

Symptoms:

• Validation errors for missing DB files.
• Jobs failing before inference starts.

Workarounds:

• Confirm AF3_DB_DIR or per-database environment variables.
• Run in sequence-only mode until pipeline dependencies are ready.

Reference Generation Workflows Are Linux-Oriented¶

Cross-platform parity reference generation (Docker/Apptainer) is a separate workflow from normal Mac inference.

Impact:

• Most users do not need Linux reference generation for everyday prediction use.

Weights Availability and Terms¶

Predictions require official AF3 weights and compliance with their terms of use.

Impact:

• Installation can complete without weights, but inference jobs will fail until weights are placed correctly.