Umong's picture
Update logbook: Repro - FlashOptim: Memory Efficient Optimizers for Large-Scale Training
03a178a verified
Raw
History Blame Contribute Delete
2.14 kB
{
"schema_version": 1,
"title": "Repro - FlashOptim: Memory Efficient Optimizers for Large-Scale Training",
"emoji": "🎯",
"space_id": "Umong/repro-flashoptim-seven-byte-adamw",
"paper": {
"arxiv_id": "2602.23349"
},
"tags": [
"icml2026-repro",
"paper-Wfe1iJocjF"
],
"updated_at": "2026-07-17T12:48:58+00:00",
"root": {
"slug": "index",
"title": "Repro - FlashOptim: Memory Efficient Optimizers for Large-Scale Training",
"file": "pages/index.md",
"children": [
{
"slug": "claim-1-flashoptim-implements-compressed-master-weights-and-companded-8-bit-states",
"title": "Claim 1: FlashOptim implements compressed master weights and companded 8-bit states",
"file": "pages/claim-1-flashoptim-implements-compressed-master-weights-and-companded-8-bit-states/page.md",
"children": []
},
{
"slug": "claim-2-flashadamw-uses-the-stated-7-bytes-per-parameter-or-5-with-gradient-release",
"title": "Claim 2: FlashAdamW uses the stated 7 bytes per parameter or 5 with gradient release",
"file": "pages/claim-2-flashadamw-uses-the-stated-7-bytes-per-parameter-or-5-with-gradient-release/page.md",
"children": []
},
{
"slug": "claim-3-ulp-splitting-and-companding-reduce-numerical-error",
"title": "Claim 3: ULP splitting and companding reduce numerical error",
"file": "pages/claim-3-ulp-splitting-and-companding-reduce-numerical-error/page.md",
"children": []
},
{
"slug": "sources-protocol-and-integrity",
"title": "Sources, protocol, and integrity",
"file": "pages/sources-protocol-and-integrity/page.md",
"children": []
},
{
"slug": "conclusion",
"title": "Conclusion",
"file": "pages/conclusion/page.md",
"children": []
},
{
"slug": "flashoptim-63497-repro",
"title": "flashoptim-63497-repro",
"file": "pages/flashoptim-63497-repro/page.md",
"children": []
}
]
},
"agent_view_tokens": 2390,
"revision": "1784292538956810574"
}