Inference throughput (Residues embedded per second on one NVIDIA H20 for 2,048 held-out proteins (639,806 residues) in length-sorted batches of at most 16,384 tokens; HyPLM-Flash on its INT8 engine, where INT8 applies only to packed batches of at least 1,024 tokens) of HyPLM-Flash, INT8 inference engine (HyPLM report) on Throughput workload of 2,048 held-out proteins (HyPLM report)
131,000 Inference throughput (Residues embedded per second on one NVIDIA H20 for 2,048 held-out proteins (639,806 residues) in length-sorted batches of at most 16,384 tokens; HyPLM-Flash on its INT8 engine, where INT8 applies only to packed batches of at least 1,024 tokens)
Methods
- Tested configuration
- HyPLM-Flash, INT8 inference engine (HyPLM report)
- Protocol
- Embedding throughput on one NVIDIA H20 (HyPLM report)
- Dataset
- Throughput workload of 2,048 held-out proteins (HyPLM report)
- Procedure
- Embedding throughput on one NVIDIA H20 (HyPLM report)
- Evaluation
- INT8 on embedding-throughput-h20 (HyPLM report)
- Coverage
- Not reported scored / Not reported eligible
- Uncertainty
- Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
- Evidence
- Author-reported evaluation · source checkedHyPLM Technical Report (Hyra Team, 2026-10-12) · Section 3, sentence 'On a common workload of 2,048 held-out proteins, the engine embeds 131,000 residues per second on one H20' (page 4 of 13)
A source-checked result verifies the numerical transcription, not every model or protocol detail. Evaluation metadata: source checked. Source checked does not mean independently reproduced.
Reproduction
- Split
- Test split as printed in Table 4
- Adaptation
- Frozen BF16 representations; the encoder is not fine-tuned
- Scoring implementation
- As stated in the protocol
No execution recipe has been verified for this exact configuration and evaluation. A benchmark's general instructions may use different inputs, splits or model settings.
Reproducing this published result requires matching its model configuration, data, split and scorer. Source checking or a successful smoke test does not establish score reproduction.
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.
1 evidence row matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Reported result 131,000 Individual claims | HyPLM Technical Report (Hyra Team, 2026-10-12) Section 3, sentence 'On a common workload of 2,048 held-out proteins, the engine embeds 131,000 residues per second on one H20' (page 4 of 13) Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages) | source checked ["independent-cell-check","deterministic-table-parse","pdf-text-parse","source-hash-verification"] · 2026-11-06T00:00:00Z author reported Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted. Extraction artifact SHA-256: |
Sources and history
Release 2026-11-21-c8bea25e5abb · Record review: source checked
1 source record and release history
- HyPLM Technical Report (Hyra Team, 2026-10-12) · Original source · Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Technical metadata and extraction receipts
Stable ID: hyplm-flash-20261106-result-embedding-throughput-h20-hyplm-flash-int8
- limitations
- The README's precision table rounds this to 'about 130,000'. The BF16 mode's 90,000 residues per second is stated without a workload, so it is a claim, not a result.; INT8 applies only to packed batches of at least 1,024 tokens, so this rate holds for batched workloads and not for single short proteins, which run BF16.
- metric
- inference-throughput
- metric direction
- higher
- metric qualifier
- Residues embedded per second on one NVIDIA H20 for 2,048 held-out proteins (639,806 residues) in length-sorted batches of at most 16,384 tokens; HyPLM-Flash on its INT8 engine, where INT8 applies only to packed batches of at least 1,024 tokens
- missing metadata
- uncertainty: note: Section 4 says intervals are paired bootstrap intervals; none is printed here.; reason: unreported
- numeric value
- 131000
- printed value
- 131,000
- review
- artifact sha256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710; method: independent-cell-check; deterministic-table-parse; pdf-text-parse; source-hash-verification; method note: Independent review pass. All six pinned artifacts were re-downloaded at the pinned commit and their SHA-256 checked against the source records; pdftotext -layout reproduced the archived text layer byte for byte. npm run batch -- check-values was run over the whole batch (value-check.tsv). The 469 recorded cells of Tables 3, 5, 6 and 7 were then re-read by a reviewer script written for this pass, which takes each table's column order from its printed header line and splits every body row by printed cell, without using extract/parse.py; every printed value and column matched. The two arithmetic constraints were re-derived independently rather than taken from the extractor's checker: each Table 3 perplexity against exp of its NLL over the NLL's own rounding interval, and each Table 5 average against the mean of its five categories. Column assignment in Tables 6 and 7 is anchored outside those tables, by Table 3's NLL values and by Figure 3a's and Table 5's ProteinGym values, and the four printed paired differences reconcile with the Table 6 ProteinGym row. The 15 locators check-values cannot parse, four Figure 3a bar labels and eleven prose values, were read by hand against the text layer; in Figure 3a each bar label and its value share one line. extract/check_cells.py was rerun and reproduces cell-check.tsv.; retrieval url: https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyplm_flash/HyPLM_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
- Section 3, sentence 'On a common workload of 2,048 held-out proteins, the engine embeds 131,000 residues per second on one H20' (page 4 of 13)
- unit
- sample-per-second
- unit detail
- Residues per second; printed as '131,000'
Related records
- evaluation: INT8 on embedding-throughput-h20 (HyPLM report)