rewirebio.iobenchmarks
Evidence claim

HyInv report: inference speed conditions

Inference speed conditions of HyInv

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.

Descriptive fact read from the pinned sources at the locator given. Not a measurement on a benchmark panel.

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.

12 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance
Claimed field
inference_speed_conditions
Context-only references
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.field

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

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

Inspected artifact

Claimed field
inference_speed_conditions
Context-only references
HyInv Technical Report (Hyra Team, 2026-10-14)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages)
Retrieved: 2026-11-06T09:52:00Z

not individually reviewed

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.field

Source artifact SHA-256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted.

Inspected artifact

Source location
Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'
Context-only references
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.source_locator

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

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

Inspected artifact

Source location
Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'
Context-only references
HyInv Technical Report (Hyra Team, 2026-10-14)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages)
Retrieved: 2026-11-06T09:52:00Z

not individually reviewed

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.source_locator

Source artifact SHA-256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted.

Inspected artifact

Claimed value
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.
Context-only references
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.value

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

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

Inspected artifact

Claimed value
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.
Context-only references
HyInv Technical Report (Hyra Team, 2026-10-14)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages)
Retrieved: 2026-11-06T09:52:00Z

not individually reviewed

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.value

Source artifact SHA-256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted.

Inspected artifact

Description
Descriptive fact read from the pinned sources at the locator given. Not a measurement on a benchmark panel.
Context-only references
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: description

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

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

Inspected artifact

Description
Descriptive fact read from the pinned sources at the locator given. Not a measurement on a benchmark panel.
Context-only references
HyInv Technical Report (Hyra Team, 2026-10-14)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages)
Retrieved: 2026-11-06T09:52:00Z

not individually reviewed

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: description

Source artifact SHA-256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted.

Inspected artifact

Relationship: subject
hyinv-20261106-model-hyinv
Context-only references
HyInv package README (Tencent Hunyuan, commit 0dcc533c)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: links:subject:hyinv-20261106-model-hyinv

Source artifact SHA-256: b44c6d0c4c7f73aaf920cf734670a8d4649df175c14a105afaf5e74172fd6891

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

Inspected artifact

Relationship: subject
hyinv-20261106-model-hyinv
Context-only references
HyInv Technical Report (Hyra Team, 2026-10-14)

Original source ↗

Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages)
Retrieved: 2026-11-06T09:52:00Z

not individually reviewed

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: links:subject:hyinv-20261106-model-hyinv

Source artifact SHA-256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted.

Inspected artifact

How to read the evidence

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

2 source records and release historyDownload this release (gzip)
Technical metadata and extraction receipts

Stable ID: hyinv-20261106-claim-hyinv-inference-speed-conditions

areas
proteins-complexes; protein-structure
contexts
research
field
inference_speed_conditions
review
artifact sha256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419; method: independent-cell-check; deterministic-table-parse; pdf-text-parse; source-hash-verification; method note: Independent review pass. The pinned PDF was re-downloaded at the pinned commit and its SHA-256 checked against the source record; pdftotext -layout reproduced the archived text layer byte for byte. npm run batch -- check-values was run over the whole batch against that text layer (value-check.tsv). The 170 cells of Tables 6, 7, 8 and 9 and the 290 cells of the Figure 14 and Figure 15 matrices were then re-read by a reviewer script that finds each row by its printed label and splits the printed cells by position, written for this pass without using the extractor's parser; every printed value and column matched. Column assignment in Figure 14 was confirmed against the labelled columns of Table 2 and against Section 5.2's prose figures; Figure 15's was confirmed by the coordinator reading the rotated column headers on the rendered page 17, which match Figure 14's order. Table 1, Figure 4a and the eleven prose values were read by hand against the text layer. Each claim was re-read against the passage its locator names.; retrieval url: https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/HyInv_Technical_Report.pdf; reviewed at: 2026-11-06T00:00:00Z; reviewer: claude; reviewer note: Claude research agent; no human review claimed. Reviewed by a different agent from the one that extracted the batch. Transcription checked against the pinned source; nothing was reproduced and no value was recomputed.
source locator
Appendix F 'Timing protocol' and 'Baseline timings'; Table 1 footnote; AI4Science/hyinv/README.md 'Accelerated inference'
value
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.
Related records

Suggest a correction