rewirebio.iobenchmarks
Source

HyInv package README (Tencent Hunyuan, commit 0dcc533c)

GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a · Project source

Original source

What this source contributed

Records in this release that cite this source. A record can cite several sources.

Claims

Statements this source supports, each with where in the source it was checked.

  • HyInv report: architecture: Architecture A message-passing graph network with 1,957,573 parameters following the encoder-decoder design of ProteinMPNN: width 128, 4 encoder layers and 3 decoder layers, each residue connected to its 48 nearest neighbours with rotation- and translation-invariant distance and orientation features, random initialisation. The encoder runs once per backbone; the decoder predicts amino acid distributions at masked positions from the backbone and the observed sequence context. At inference the model needs neither a protein language model, a structure predictor nor an alignment. (Section 2.1 'Architecture'; Appendix C 'Configuration'; AI4Science/hyinv/README.md opening paragraph)
  • HyInv report: availability: Availability Public. The hyinv package, its command line and Python API and the FP32 weights (weights/model.safetensors, 7.8 MB, with config.json) ship in AI4Science/hyinv/ of the Tencent-Hunyuan/Hyra-results repository; the repository README states the 7.8 MB weights ship in the folder. Python 3.10+ and PyTorch, with safetensors and numpy; accelerated inference also needs triton. Runs on CPU and GPU. No hash or revision is published for the weights file. (AI4Science/hyinv/README.md 'Installation' and 'Layout'; repository README.md 'Model releases')
  • HyInv report: BF16 speed-up against the reference sampler: Reference sampler speedup bf16 Section 4 states HyInv designs 128 sequences for a 402-residue backbone in 50.6 ms on one NVIDIA H20, '52.1x faster than the reference implementation of the sampler and 57x faster than ProteinMPNN'; the abstract gives the same figure as 'about 52x faster than the reference sampler'. This is HyInv's optimised BF16 path against HyInv's own FP32 reference sampler, not against a baseline model, and it is the printed ratio of the two Table 1 times for the 402-residue request (2,633 and 50.6 ms). The 57x recorded as a result is the ratio against ProteinMPNN's released implementation. Both were timed with the supervised-stage weights at rewrite temperature 0.23, not the released checkpoint. (Section 4 'Efficient inference', first sentence; abstract)
  • HyInv report: exact FP32 speed-up against the reference sampler: Reference sampler speedup exact fp32 The abstract states 'an exact FP32 path that returns the reference sequences is 8.7x faster', against HyInv's own FP32 reference sampler. Table 1's printed times for the 402-residue request give the same ratio (2,633 and 303 ms). The 9.6x recorded as a result is the ratio against ProteinMPNN's released implementation, so the two figures describe different comparisons and neither is a speed-up of HyInv over a baseline at equal optimisation. (Abstract, final speed sentence)
  • HyInv report: inference speed conditions: Inference speed conditions Latencies were measured on one NVIDIA H20 (Hopper class, sm_90) of a node on which nothing else ran, with PyTorch 2.7.1 and CUDA 12.9. The timed region excludes model loading, CUDA graph capture and kernel compilation; a first call costs 67.7 ms to 834 ms per shape bucket in the report's accounting, and the package README puts cold-cache Triton compilation at about 25 s for the exact path and 14 s for the BF16 path, plus about 0.1 to 0.2 s to capture a graph for a new shape. The kernels were verified on an H20 only and have not been verified on other architectures. HyInv's timing rows used the supervised-stage weights and a rewrite temperature of 0.23, not the released checkpoint's default. Baselines ran with their released implementations in FP32 without graph capture or compilation, DPLM-2 3B in BF16 autocast, and unbatched where the released adapter handles one structure per call, so the report states the speed-ups are against those implementations and not against the best achievable speed of the baseline models. (Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference')
  • HyInv report: inputs: Inputs Protein backbone N, CA, C and O coordinates read from the ATOM records of the first model of a PDB file (MSE treated as MET), optionally a chain selection and a set of fixed residues given by PDB residue number. Each residue takes one slot and residues with missing backbone atoms keep their slot and are designed; residues are numbered 0..n-1 within each chain without filling gaps in the PDB numbering. mmCIF must be converted to PDB first. For multi-chain structures fixed positions must be qualified with a chain name. (AI4Science/hyinv/README.md 'Usage' and 'Notes'; Section 2.1)
  • HyInv report: licence: Licence MIT. The package LICENSE carries two copyright lines, 'Copyright (c) 2022 Justas Dauparas (ProteinMPNN)' and 'Copyright (c) 2026 HyInv contributors', because the model modules come from ProteinMPNN, which is also MIT. The repository's own top-level LICENSE is Apache-2.0, so the package licence and the repository licence differ. (AI4Science/hyinv/LICENSE; AI4Science/hyinv/README.md 'Citation and license'; repository README.md 'License')
  • HyInv report: numerical paths: Numerical paths Two accelerated paths. The exact FP32 path computes matrix products by a three-pass TF32 emulation of FP32 with erf GELU and returns the reference sampler's sequences for the same seeds (1,600 of 1,600 eager, 400 of 400 with graph capture, 352 of 352 in padded multi-target batches), with graph replay equal to eager execution bit for bit; the report calls only this path identical. The BF16 path uses BF16 operands with FP32 accumulation and tanh GELU and is a numerically different path: against the reference on 200 development backbones the largest absolute logit difference is 0.069, the mean KL divergence per sampled position 5.2e-5 (99th percentile 7.5e-4), the argmax agrees at 99.73% of positions and, with the same seeds, 49.4% of sequences are identical and 98.7% of positions agree. All sequences scored in Section 5 come from the reference sampler. (Appendix F 'The two numerical paths'; Section 4 'Quality'; AI4Science/hyinv/README.md 'Accelerated inference')
  • HyInv report: outputs: Outputs Designed amino-acid sequences over the 20 standard amino acids, written as FASTA by the command line or returned as a list of strings by the Python API; chains are separated by '/' in the output and fixed positions are kept unchanged. (AI4Science/hyinv/README.md 'Usage' and 'Notes')
  • HyInv report: readme speed discrepancy: Readme speed discrepancy The package README's own speed figures differ from the report's for the exact FP32 path: about 119 / 281 ms for 148 / 402 residues against the report's 130 / 303 ms. The README states the reason, that the report's exact path uses another set of kernels, and gives the BF16 and reference numbers as the report's (about 25 / 54 ms against 24.7 / 50.6 ms, and 999 / 2,633 ms). The report's values are the ones recorded as results here. (AI4Science/hyinv/README.md 'Accelerated inference', speed paragraph; Table 1)
  • HyInv report: reproducibility: Reproducibility Outputs are identical for the same seed, batch size (default 64, automatically tightened by length), device type and PyTorch version. Each batch has its own random stream (batch k uses seed+k), so changing the batch size changes the result, and the CPU and GPU random streams differ, so their sequences are not identical. (AI4Science/hyinv/README.md 'Notes')
  • HyInv report: sampler: Sampler Deployment samples in 16 decoder passes after one encoder pass: the positions are split at random into 8 fixed equal-size groups, 8 rounds of blind drafting at temperature 0.6 sample one group each from the all-masked state, then 8 rounds of rewriting at temperature 0.34 mask and resample one group at a time with all other positions visible. The group permutation is drawn from a SHA-256 hash of the seed and the chain identifier and is independent of structure, sequence and weights. Truncation at inference uses top-p 0.9 and min-p 0.1. Fixed residues are visible as the native residue from the first round and are redrawn unchanged at every stage. (Sections 3.1 and 3.2; Appendix E 'Procedure and groups' and 'Random streams, truncation and fixed residues'; AI4Science/hyinv/README.md opening paragraph)

Models, benchmarks and other records

Models (1)

Configurations (13)

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

21 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-adc15377fab2
Property and statementOriginal source and locationReview and provenance
Archive note
Archived in the batch as artifacts/hyinv-README.md.
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_note

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Archive sha256
b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_sha256

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Artifact SHA-256
b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Artifact url
https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/README.md
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_url

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Licence
MIT (AI4Science/hyinv/LICENSE)
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.licence

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Limitations
2 values
  • The README's speed figures differ from the technical report's for the exact FP32 path (about 119 / 281 ms against the report's 130 / 303 ms) because, as it states, the report's exact path uses another set of kernels. The report's values are the ones recorded as results.
  • The README states the weights are the structure-reward reinforcement-learning version but does not print an update number; the report's update 640 is what identifies the checkpoint, and no hash of the weights file is published.
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.limitations

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Media type
text/markdown
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.media_type

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Publication status
official_project_source
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.publication_status

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Retrieved at
2026-11-06T09:52:00Z
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieved_at

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Source type
repository README
Source metadata
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

No field-specific location recorded

Version: GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Retrieved: 2026-11-06T09:52:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.source_type

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

How to read the evidence

Sources and history

Release 2026-11-21-adc15377fab2 · Record review: source checked

README of the released hyinv package, with the checkpoint identity, the command line and Python interface, the input and output formats and the accelerated inference modes.

Release history

No supporting source is linked yet.

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: hyinv-20261106-source-model-readme

areas
proteins-complexes; protein-structure
contexts
research
archive note
Archived in the batch as artifacts/hyinv-README.md.
archive sha256
b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891
artifact sha256
b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891
artifact url
https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/README.md
licence
MIT (AI4Science/hyinv/LICENSE)
limitations
The README's speed figures differ from the technical report's for the exact FP32 path (about 119 / 281 ms against the report's 130 / 303 ms) because, as it states, the report's exact path uses another set of kernels. The report's values are the ones recorded as results.; The README states the weights are the structure-reward reinforcement-learning version but does not print an update number; the report's update 640 is what identifies the checkpoint, and no hash of the weights file is published.
media type
text/markdown
publication status
official_project_source
retrieved at
2026-11-06T09:52:00Z
source type
repository README
url
https://github.com/Tencent-Hunyuan/Hyra-results/blob/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/README.md
version
GitHub repository Tencent-Hunyuan/Hyra-results at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a
Related records

Suggest a correction