{"paper":{"title":"GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"GaLore projects full gradients onto low-rank subspaces periodically, cutting optimizer memory by 65.5% while training every parameter of large language models.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Anima Anandkumar, Beidi Chen, Jiawei Zhao, Yuandong Tian, Zhangyang Wang, Zhenyu Zhang","submitted_at":"2024-03-06T07:29:57Z","abstract_excerpt":"Training Large Language Models (LLMs) presents significant memory challenges, predominantly due to the growing size of weights and optimizer states. Common memory-reduction approaches, such as low-rank adaptation (LoRA), add a trainable low-rank matrix to the frozen pre-trained weight in each layer, reducing trainable parameters and optimizer states. However, such approaches typically underperform training with full-rank weights in both pre-training and fine-tuning stages since they limit the parameter search to a low-rank subspace and alter the training dynamics, and further, may require full"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That periodically recomputed low-rank bases for gradient projection preserve optimization dynamics close enough to full-rank gradients that final model quality remains comparable across pre-training and fine-tuning regimes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GaLore performs full-parameter LLM training with up to 65.5% less optimizer memory by projecting gradients onto a low-rank subspace at each step, matching full-rank performance on LLaMA pre-training and RoBERTa fine-tuning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"GaLore projects full gradients onto low-rank subspaces periodically, cutting optimizer memory by 65.5% while training every parameter of large language models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"716137a95d8b5a93ac44d1d24eb53cd4af93ab812ef2053f7d65aaa71f64fa98"},"source":{"id":"2403.03507","kind":"arxiv","version":2},"verdict":{"id":"8a22c18f-eb03-4180-bb8b-a3ccd8e03981","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T23:47:37.157229Z","strongest_claim":"We demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.","one_line_summary":"GaLore performs full-parameter LLM training with up to 65.5% less optimizer memory by projecting gradients onto a low-rank subspace at each step, matching full-rank performance on LLaMA pre-training and RoBERTa fine-tuning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That periodically recomputed low-rank bases for gradient projection preserve optimization dynamics close enough to full-rank gradients that final model quality remains comparable across pre-training and fine-tuning regimes.","pith_extraction_headline":"GaLore projects full gradients onto low-rank subspaces periodically, cutting optimizer memory by 65.5% while training every parameter of large language models."},"references":{"count":55,"sample":[{"doi":"","year":2019,"title":"Memory efficient adaptive optimization","work_id":"852fa34d-c73a-4d07-8012-d983fe698557","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Belle: Be everyone's large language model engine","work_id":"84a8bc20-e93c-4601-8413-3f9c43212cae","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Continual learning in low-rank orthogonal subspaces","work_id":"c83f5d43-b1ee-4cf1-8c48-0c547920edd0","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Non- Convex Projected Gradient Descent for Generalized Low-Rank Tensor Regression","work_id":"ab8e2819-94e0-4e6d-bcfd-75ab25d5680b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Training Deep Nets with Sublinear Memory Cost","work_id":"f2c5c287-a500-40e4-a136-e7e3172db1d7","ref_index":5,"cited_arxiv_id":"1604.06174","is_internal_anchor":true}],"resolved_work":55,"snapshot_sha256":"fb16556210ca767e58199a9a7a212f2ab5f7dace7778461b526f01ce8ca68a16","internal_anchors":9},"formal_canon":{"evidence_count":2,"snapshot_sha256":"27962be8dc6346f405f988acb1bfa29e625c6b7e11ae694deef8035f87cf45c8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}