{ "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" }