logbook.json

{
  "schema_version": 1,
  "title": "Reproduction: FlashOptim: Optimizers for Memory-Efficient Training",
  "emoji": "🎯",
  "space_id": "Umong/repro-flashoptim-optimizers-for-memory-efficient-training",
  "paper": {
    "arxiv_id": "2602.23349"
  },
  "tags": [
    "icml2026-repro",
    "paper-Wfe1iJocjF"
  ],
  "updated_at": "2026-07-18T18:31:20+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: FlashOptim: Optimizers for Memory-Efficient Training",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-flashoptim-combines-improved-master-weight-splitting-with-companded-8-bit-optimizer-state-quantization-to-reduce-parameter-associated-optimizer-memory-section-3",
        "title": "Claim 1: FlashOptim combines improved master-weight splitting with companded 8-bit optimizer-state quantization to reduce parameter-associated optimizer memory (Section 3).",
        "file": "pages/claim-1-flashoptim-combines-improved-master-weight-splitting-with-companded-8-bit-optimizer-state-quantization-to-reduce-parameter-associated-optimizer-memory-section-3/page.md",
        "children": []
      },
      {
        "slug": "claim-2-flashoptim-reduces-adamw-training-memory-from-16-to-7-bytes-per-parameter-or-5-bytes-with-gradient-release-table-1",
        "title": "Claim 2: FlashOptim reduces AdamW training memory from 16 to 7 bytes per parameter, or 5 bytes with gradient release (Table 1).",
        "file": "pages/claim-2-flashoptim-reduces-adamw-training-memory-from-16-to-7-bytes-per-parameter-or-5-bytes-with-gradient-release-table-1/page.md",
        "children": []
      },
      {
        "slug": "claim-3-for-llama-3-1-8b-finetuning-flashoptim-reduces-peak-memory-from-175-gib-to-113-gib-by-compressing-parameters-and-optimizer-states-figure-1-table-4",
        "title": "Claim 3: For Llama-3.1-8B finetuning, FlashOptim reduces peak memory from 175 GiB to 113 GiB by compressing parameters and optimizer states (Figure 1; Table 4).",
        "file": "pages/claim-3-for-llama-3-1-8b-finetuning-flashoptim-reduces-peak-memory-from-175-gib-to-113-gib-by-compressing-parameters-and-optimizer-states-figure-1-table-4/page.md",
        "children": []
      },
      {
        "slug": "claim-4-flashoptim-variants-match-reference-optimizer-training-loss-trajectories-in-gpt-2-pretraining-and-resnet-50-image-classification-figure-2",
        "title": "Claim 4: FlashOptim variants match reference optimizer training-loss trajectories in GPT-2 pretraining and ResNet-50 image classification (Figure 2).",
        "file": "pages/claim-4-flashoptim-variants-match-reference-optimizer-training-loss-trajectories-in-gpt-2-pretraining-and-resnet-50-image-classification-figure-2/page.md",
        "children": []
      },
      {
        "slug": "claim-5-flashoptim-matches-reference-scores-on-resnet-50-validation-accuracy-llama-3-1-8b-gsm8k-finetuning-and-gpt-2-in-context-learning-benchmarks-table-2-table-3",
        "title": "Claim 5: FlashOptim matches reference scores on ResNet-50 validation accuracy, Llama-3.1-8B GSM8K finetuning, and GPT-2 in-context learning benchmarks (Table 2; Table 3).",
        "file": "pages/claim-5-flashoptim-matches-reference-scores-on-resnet-50-validation-accuracy-llama-3-1-8b-gsm8k-finetuning-and-gpt-2-in-context-learning-benchmarks-table-2-table-3/page.md",
        "children": []
      },
      {
        "slug": "claim-6-ulp-based-weight-splitting-lowers-fp32-reconstruction-error-and-companding-prevents-quantized-adamw-training-divergence-compared-with-linear-optimizer-state-quantization-figure-3-figure-5",
        "title": "Claim 6: ULP-based weight splitting lowers FP32 reconstruction error and companding prevents quantized AdamW training divergence compared with linear optimizer-state quantization (Figure 3; Figure 5).",
        "file": "pages/claim-6-ulp-based-weight-splitting-lowers-fp32-reconstruction-error-and-companding-prevents-quantized-adamw-training-divergence-compared-with-linear-optimizer-state-quantization-figure-3-figure-5/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 4816,
  "revision": "1784399480745309385"
}