Peak memory (Embedding inference with 4-bit weights; the paper calls it parameter memory in Section 2.4 and peak memory in Figure 5) of NT 2.5B, 4-bit weights for embedding inference (Saadat and Fellay 2026) on Synthetic uniform random DNA sequences, seed 42 (Saadat and Fellay 2026)
5.6 Peak memory (Embedding inference with 4-bit weights; the paper calls it parameter memory in Section 2.4 and peak memory in Figure 5)
Methods
- Tested configuration
- NT 2.5B, 4-bit weights for embedding inference (Saadat and Fellay 2026)
- Protocol
- Peak memory under 8-bit and 4-bit quantization relative to BF16 (Saadat and Fellay 2026 Figure 5)
- Dataset
- Synthetic uniform random DNA sequences, seed 42 (Saadat and Fellay 2026)
- Procedure
- Peak memory under 8-bit and 4-bit quantization relative to BF16 (Saadat and Fellay 2026 Figure 5)
- Evaluation
- NT 2.5B 4-bit inference memory
- Coverage
- Not reported scored / Not reported eligible
- Uncertainty
- Not reported by the source
- Evidence
- Independent external evaluation · source checkedOptimizing genomic language models for efficient training, fine-tuning, and inference (Saadat and Fellay, ICLR 2026 workshop) · Section 3.5 'Memory' paragraph (page 7): 'For NT 2.5B, memory decreases from 6.7 GB (BF16) to 5.6 GB (4-bit).'
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
- Not applicable
- Adaptation
- None (pretrained weights)
- Scoring implementation
- Not reported
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 5.6 Individual claims | Optimizing genomic language models for efficient training, fine-tuning, and inference (Saadat and Fellay, ICLR 2026 workshop) Section 3.5 'Memory' paragraph (page 7): 'For NT 2.5B, memory decreases from 6.7 GB (BF16) to 5.6 GB (4-bit).' Version: OpenReview PDF headed 'Published as a workshop paper at ICLR 2026' (11 pages; PDF CreationDate 2026-03-09) | source checked ["source-hash-verification","pdf-text-parse","independent-cell-check","rendered-source-visual-check"] · 2026-10-15T22:00:00Z independent paper Audit detailsValue matches the printed sentence. Plotted bar differs slightly and the quantity is ambiguous (parameter or peak memory); see source_discrepancy. Kept as a result-level note. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Extraction artifact SHA-256: |
Sources and history
Release 2026-10-22-1b7cf027e4f6 · Record review: source checked
1 source record and release history
- Optimizing genomic language models for efficient training, fine-tuning, and inference (Saadat and Fellay, ICLR 2026 workshop) · Original source · OpenReview PDF headed 'Published as a workshop paper at ICLR 2026' (11 pages; PDF CreationDate 2026-03-09)
Technical metadata and extraction receipts
Stable ID: saadat-fellay-20261015-result-nt-2-5b-quantization-memory-4bit
- metric
- peak-memory
- metric direction
- lower
- unit
- gigabyte
- printed value
- 5.6
- numeric value
- 5.6
- source locator
- Section 3.5 'Memory' paragraph (page 7): 'For NT 2.5B, memory decreases from 6.7 GB (BF16) to 5.6 GB (4-bit).'
- metric qualifier
- Embedding inference with 4-bit weights; the paper calls it parameter memory in Section 2.4 and peak memory in Figure 5
- missing metadata
- uncertainty: reason: unreported
- review
- reviewer: claude; reviewer note: Claude review agent; no human review claimed. Independent Claude (Opus 5.5) review agent, separate from the extraction agent that built this batch. Automated review only. Transcription, not reproduction.; reviewed at: 2026-10-15T22:00:00Z; artifact sha256: 4bc2452110f280a7ba8baa4c7c28beac89769511019666bc77fc33af9b3bcd60; retrieval url: https://openreview.net/pdf?id=NUleWh1GFx; method: source-hash-verification; pdf-text-parse; independent-cell-check; rendered-source-visual-check; method note: Matched the pinned PDF SHA-256 (archived copy of the hand download; OpenReview blocks scripted retrieval). review/review.py re-parsed the PDF with pdftotext and asserted the source sentence, printed and numeric value, unit, direction, model label, protocol, data link and Figure 3 stage. The stage was also checked by eye against Figure 3, whose plotted points agree with the printed pairs.; note: Value matches the printed sentence. Plotted bar differs slightly and the quantity is ambiguous (parameter or peak memory); see source_discrepancy. Kept as a result-level note.
- hardware
- NVIDIA H100 (section 2.4); 'H100 94GB GPUs' (section 4)
- unit detail
- Printed as GB
- source discrepancy
- Section 2.4 says the quantization experiment measures parameter memory, and Section 3.5 opens 'Quantization reduces parameter memory' before giving these values, but the Figure 5 right panel is labelled 'Peak Memory (GB)'. A 1.1 GB drop is small for parameters alone (2.5 billion weights are about 5 GB in BF16 and about 1.25 GB at 4 bits; an extractor's estimate, not printed), so the values look like peak memory of a run whose input length, batch size and quantised layers ('a conservative strategy for NT models') are not stated. Recorded as peak memory. Independent review 2026-10-15: the Figure 5 right-panel bars for NT 2.5B also do not match the printed sentence. Against the panel's gridlines they stand at about 6.6 GB (BF16) and about 5.45 GB (4-bit), against 6.7 and 5.6 in the text (reviewer's reading of bar heights, not a printed value). The direction and size of the drop agree, so the printed values are kept with this note and no source-level concern is raised.
Related records
- evaluation: NT 2.5B 4-bit inference memory