Training peak GPU memory under progressive optimizations, sequence length 1000, batch size 16 (Saadat and Fellay 2026 Figure 3)
Peak GPU memory of a training step at sequence length 1000 and batch size 16, adding FlashAttention, activation checkpointing and ZeRO Stage-2 CPU offload in turn.
Overview
An explanatory profile has not yet been reviewed for this record.
Consult the linked sources for architecture or protocol details. Missing evidence is not evidence of a missing capability.
4 recorded evaluations, 4 metric rows. A comparison chart has not yet been validated for these results. The table retains the individual findings and their sources.
Results
Results are available, but no reviewed comparison panel is linked in this release.
All evaluations
4 evaluations · 4 results. Different protocols are not a single leaderboard.
Protocol: Training peak GPU memory under progressive optimizations, sequence length 1000, batch size 16 (Saadat and Fellay 2026 Figure 3) · Dataset: Synthetic uniform random DNA sequences, seed 42 (Saadat and Fellay 2026)
Sorted by Peak memory (Training, sequence length 1000, batch size 16, stage + Checkpointing (FlashAttention and activation checkpointing)) (lower is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.
| Tested configuration | Peak memory (Training, sequence length 1000, batch size 16, stage + Checkpointing (FlashAttention and activation checkpointing)) | Peak memory (Training, sequence length 1000, batch size 16, stage + DeepSpeed (FlashAttention, checkpointing and ZeRO Stage-2 CPU offload)) |
|---|---|---|
| GENA-LM 110M, training with FlashAttention and checkpointing (Saadat and Fellay 2026)Evaluation | 1.18 (best in column) | Not reported |
| NT 2.5B, training with FlashAttention and checkpointing (Saadat and Fellay 2026)Evaluation | 33.9 | Not reported |
| GENA-LM 110M, training with FlashAttention, checkpointing and ZeRO-2 offload (Saadat and Fellay 2026)Evaluation | Not reported | 2.49 (best in column) |
| NT 2.5B, training with FlashAttention, checkpointing and ZeRO-2 offload (Saadat and Fellay 2026)Evaluation | Not reported | 11.5 |
All 4 result rows with coverage, uncertainty and sources
Filter evaluations
Applied filters: All linked evaluations
Source checking is not independent reproduction. Release 2026-10-22-1b7cf027e4f6.
Methods and evaluation design
Procedure, tasks and evaluated configurations
Recorded evaluations
Each evaluation records what was tested and under which conditions.
Baseline coverage
Reference methods help show what a model adds beyond simple controls. We track a null control and a conventional method for each protocol.
0 of 2 active baseline roles have published Rewire measurements in this release. Measurements on a selected protocol do not establish coverage of an entire suite.
No execution recipe linked to this protocol. Recipe availability does not establish a completed evaluation.
- External evaluations
- 4
Literature evidence is not a Rewire measurement. Executed but unpublished runs and private review status are not included.
Null control
Proposed control: requires review
Unintegrated representation using protocol preprocessing
Protocol-specific applicability, permitted inputs, access, split, evaluator and execution requirements need review before implementation or execution.
This is a suggested selection rule, not a validated method or a measured score.
Conventional reference
Proposed control: requires review
Protocol-specific conventional integration method
Protocol-specific applicability, permitted inputs, access, split, evaluator and execution requirements need review before implementation or execution.
This is a suggested selection rule, not a validated method or a measured score.
Protocol coverage CSV (gzip) · Model evaluation matrix (gzip) · Source table (gzip) · Release and checksums (gzip)
Coverage is derived from release 2026-10-22-1b7cf027e4f6. Source citations describe the original records; they do not validate an unreviewed baseline proposal. No results have been generated by this audit.
Run instructions
No runnable recipe has been reviewed for this protocol. Dataset access, model requirements, licences and compute requirements must be checked against its sources before execution.
Strengths, limitations and unresolved questions
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.
0 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|
No evidence rows match these filters. Choose another scope or clear the search.
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-protocol-training-peak-memory
- areas
- dna-genomes
- contexts
- research
- protocol
- Measure peak allocated GPU memory (PyTorch allocator statistics) during training at sequence length 1000 and batch size 16, for the baseline and after cumulatively adding FlashAttention, activation checkpointing and ZeRO Stage-2 CPU offload (DeepSpeed), on an NVIDIA H100.
- metric
- peak-memory
- unit
- gigabyte
- memory measurement
- Peak allocated GPU memory from the framework allocator statistics during the measured region (Appendix A.3)
- version
- ICLR 2026 workshop paper, Section 3.3
- limitations
- Only the stage values quoted in Section 3.3 are printed; the other Figure 3 points are plotted without values.; The training input is not named. The paper lists synthetic sequences as its input for implementation measurements and the promoter and enhancer tasks as its only real data; the training ablation is linked to the synthetic data on that basis.; Whether the ablation trains with LoRA (the stated default for fine-tuning) or updates all weights is not stated. Section 3.3 attributes the NT 2.5B ZeRO saving to optimizer state dominating, which fits a full-weight update better than LoRA rank 16 on query, key and value projections, but the paper does not say.; Whether 'sequence length 1000' counts nucleotides or tokens is not stated; NT uses 6-mer tokens and GENA-LM uses BPE tokens, so the two families see different token counts for 1000 nucleotides.; Single hardware type (H100 94GB); offload benefit depends on CPU-GPU bandwidth, as the authors note.
- source locator
- Section 3.3 and Figure 3 (pages 5-6); Appendix A.3 'Peak GPU memory'
- missing metadata
- replicates: 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; ai-assisted-source-review; scope-review; method note: Descriptive fields, identity links and missing-metadata reasons checked against the PDF.; note: Limitations added in review: sequence-length unit and the LoRA-or-full-update question.
Related records
- uses data: Synthetic uniform random DNA sequences, seed 42 (Saadat and Fellay 2026)
- assessment: GENA-LM 110M training peak memory, stage + Checkpointing
- assessment: GENA-LM 110M training peak memory, stage + DeepSpeed
- assessment: NT 2.5B training peak memory, stage + Checkpointing
- assessment: NT 2.5B training peak memory, stage + DeepSpeed