[imathur1]

Release esm 3.4.2: Batch API for folding, max-accuracy folding

Batch API (ForgeBatchClient)

• See cookbook/snippets/batch.md and cookbook/snippets/fold_max_accuracy.md.

Sync client

• The fold, generate, encode, decode and inverse-fold calls now request pickle responses (ACCEPT_ES_PICKLE). fold takes the sequence from the server response

Models

• ESMFold2:
  • fp8 inference for the ESMC backbone.
  • The pair bias is now cached for each fold.
  • The forward-only pair transition no longer writes the buffer that only backward reads.
  • set_kernel_backend("fast") now also reaches the MSA encoder's triangle updates.
  • Transformer Engine is removed from the ESMC tensor-parallel path.
• ESMC: a model loaded on CUDA and moved with .to("cpu") now falls back from flash-attention / Triton to PyTorch instead of crashing, which fixes the 3.x README quickstart on GPU machines.
• ProteinChain and MolecularComplex fixes, each with tests.

Packaging

• Version 3.4.2. cuequivariance and xformers are no longer base dependencies. Triton is pinned to the torch 2.11 version on Linux x86-64. The old fused / fold-cp / cueq* extras are removed, and a packaging test keeps the base install free of accelerator packages on other platforms.
• README: install options and sizes, how to install with CPU-only torch, and a Batch API section.

[github-actions]

Tests	Skipped	Failures	Errors	Time
525	3 💤	0 ❌	0 🔥	47.534s ⏱️