v0.7.3: Fix the bug in the code where ref_space_uid was mistakenly written as ref_mask in the cache computation.

What's changed

• Fix the bug in the code where ref_space_uid was mistakenly written as ref_mask in the cache computation. commit 855973d.
  Full Changelog: v0.7.2...v0.7.3

v0.7.2: Allow for the absence of pairing.a3m in inference.

What's changed

• If the directory specified by precomputed_msa_dir under the msa field in the inference file does not contain the pairing.a3m file, no error will be thrown; instead, only the non_pairing.a3m file will be used for inference. In previous versions, this would have caused an immediate error.

example.json

[{
    "sequences": [
        {
            "proteinChain": {
                "sequence": "MGSSHHHHHHSSGLVPRGSHMSGKIQHKAVVPAPSRIPLTLSEIEDLRRKGFNQTEIAELYGVTRQAVSWHKKTYGGRLTTRQIVQQNWPWDTRKPHDKSKAFQRLRDHGEYMRVGSFRTMSEDKKKRLLSWWKMLRDNDLVLEFDPSIEPYEGMAGGGFRYVPRDISDDDLLIRVNEHTQLTAEGELLWSWPDDIEELLSEP",
                "count": 1,
                "msa": {
                    "precomputed_msa_dir": "./examples/7r6r/msa/1",
                    "pairing_db": "uniref100"
                }
            }
        },
        {
            "dnaSequence": {
                "sequence": "TTTCGGTGGCTGTCAAGCGGG",
                "count": 1
            }
        },
        {
            "dnaSequence": {
                "sequence": "CCCGCTTGACAGCCACCGAAA",
                "count": 1
            }
        }
    ],
    "name": "7r6r"
}
]

v0.7.1: enforce fp32 and torch kernels for triangle attention and multiplicative on V100.

What's Changed

• Add a dtype parameter to the Protenix CLI for inference, enabling FP32 inference via the -d flag.
• For inference on V100 GPUs, certain configurations are forcibly adjusted — for example, BF16 precision and unsupported optimized kernels are disabled by default.

v0.7.0: add options for faster diffusion inference: shared variable caching, efficient bias fusion, and TF32 acceleration.

What's Changed

We’re excited to announce the open-source release of Protenix v0.7.0, supported by @yangyanpinghpc, featuring several performance optimizations for diffusion inference. This version introduces three new optional acceleration flags (enabled by default in inference stage) and improved support for batched inference:

• --enable_cache
  Precomputes and caches shared intermediate variables (pair_z, p_lm, c_l) across the N_sample and N_step dimensions.
• --enable_fusion
  Fuses bias transformations and normalization in the 24-layer diffusion transformer blocks at compile time.
• --enable_tf32
  Enables TF32 precision for matrix multiplications when using FP32 computation, trading slight numerical accuracy for speed.
• Batched Diffusion Support (N_sample > 1)
  Shares s_trunk and z_pair across the N_sample dimension during diffusion, reducing memory and compute overhead without affecting results.

You can run it using the following example command:
(Note: if not specified, --enable_cache, --enable_fusion, and --enable_tf32 default to true.)

protenix predict -i examples/example.json -o  ./test_outputs/cmd/output_mini -s 105,106 -n "protenix_mini_default_v0.5.0" --triatt_kernel "torch" --trimul_kernel "torch" --enable_cache true --enable_fusion true --enable_tf32 true

v0.6.3: support polymer–polymer bond input at inference.

What's Changed

• Polymer–polymer bond input at inference. Inference can now read user-specified polymer–polymer covalent bonds from JSON and incorporate them into features. This supports cyclic peptides formed by head-to-tail amide linkage or disulfide bonds.
• CIF output quality. Cleaned and optimized fields in the generated CIF files for better downstream compatibility.
• msa_pairing.py: remove an assertion with deprecated np.string_ for improved NumPy compatibility.
• Updated inference README. Clarifies how to specify polymer–polymer bonds in JSON, the supported cyclic-peptide cases, and current limitations.

v0.6.2: update cuequivariance to 0.6.1 and update constraint api

What's Changed

• minor modification by @OccupyMars2025 in #177
• add compatibility with colabfold mmseqs server api by @JinyuanSun in #178
• tests: Add test cases for installation and compatibility issues by @ShadNygren in #192
• fix: Resolve DeepSpeed/Pydantic compatibility issue (#182) by @ShadNygren in #193
• Fix #185: Enable consumer GPU support (RTX 3090/4090) with Triton fallback by @ShadNygren in #194
• minor modification: switch residue_index to token_index by @OccupyMars2025 in #195
• fix typo for get_atom_permutation_list function by @mrzzmrzz in #196
• update cuequivariance to 0.6.1
• update constraint api and Protenix web server

New Contributors

• @JinyuanSun made their first contribution in #178
• @ShadNygren made their first contribution in #192
• @mrzzmrzz made their first contribution in #196

Full Changelog: v0.6.1...v0.6.2

v0.6.1: Fixed ESM model loading compatibility with PyTorch 2.6 and later versions.

• Fixed ESM model loading compatibility with PyTorch 2.6 and later versions.
  Full Changelog: v0.6.0...v0.6.1

v0.6.0: Optimized kernels and upgraded dependencies for enhanced performance in PyTorch 2.4+

What's Changed

1. Optimized the custom LayerNorm kernel, further boosting end-to-end inference and training speed.
2. Integrated a custom Triton-based implementation of the TriangleAttention operator (triattention), improving computational efficiency.
3. Integrated the cuEquivariance operator from NVIDIA/cuEquivariance to accelerate equivariant operations, with notable efficiency gains in the TriangleAttention and TriangleMultiplication modules.
4. Upgraded the container image and dependencies to resolve efficiency bottlenecks in PyTorch 2.4 and later versions; Supported Biotite 1.2 and above.

New Contributors

• @chaitjo made their first contribution in #175

Full Changelog: v0.5.5...v0.6.0

v0.5.5: Fix inference cache directory for users without root permissions.

• Fix inference cache directory for users without root permissions.
  Full Changelog: v0.5.4...v0.5.5

v0.5.4: fix redundant scaling in F.scaled_dot_product_attention

What's Changed

• find all new chain starts by inspecting atom_array.hetero[c_start:c_stop] by @OccupyMars2025 in #157
• some minor fixes for pdb_to_cif() execution flow by @OccupyMars2025 in #158
• fix redundant scaling in F.scaled_dot_product_attention when use_memory_efficient_kernel is enabled

Full Changelog: v0.5.3...v0.5.4