Research map / Architectures and efficiency / Pruning and efficient tuning
FFT-based Dynamic Subspace Selection for Low-Rank Adaptive Optimization of Large Language Models
Ionut-Vlad Modoranu, Mher Safaryan, Erik Schultheis, Max Ryabinin, Artem Chumachenko, Dan Alistarh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Power-Iteration/QR with DCT dynamic column selection (Trion) | Dion · C4 pretraining, Llama-350M rank 256 | validation perplexity: 15.64 -> 15.30 | Improved |
| replace Power-Iteration/QR with DCT dynamic column selection (Trion) | Dion · C4 pretraining, Llama-800M rank 256 | max memory (GB): 71.75 -> 67.45 | Improved |
| replace Power-Iteration/QR with DCT dynamic column selection (Trion) | Dion · C4 pretraining, Llama-800M rank 512 | runtime: 8h 48m -> 8h 3m | Improved |
| replace SVD with DCT projection | FRUGAL · C4 pretraining, Llama-800M | runtime: 9h 45m -> 7h 57m | Improved |
| replace SVD with DCT projection | FRUGAL · C4 pretraining, Llama-800M | validation perplexity: 14.02 -> 14.23 | Hurt |
| replace SVD with DCT projection | FIRA · C4 pretraining, Llama-800M | validation perplexity: 17.67 -> 17.30 | Improved |
| use DCT instead of RandPerm/Random projection (ablation) | FRUGAL · C4 pretraining, Llama-800M | validation perplexity: 15.18 -> 14.23 (RandPerm); 15.61 -> 14.23 (Random) | Improved |
| replace LDAdamW projection with DCT-AdamW (with ZeRO-style updates) | LDAdamW · C4 pretraining, Llama-800M 100 tokens/param | memory (GiB): 72.10 -> 57.82 | Improved |
| replace LDAdamW projection with DCT-AdamW | LDAdamW · C4 pretraining, Llama-800M 100 tokens/param | validation perplexity: 13.91 -> 13.69 | Improved |
| replace SVD with DCT projection | FRUGAL · GSM8K, Llama-2 7B rank 32 | accuracy (%): 33.81 -> 35.93 | Improved |
| replace block power-iteration with DCT projection | LDAdamW · GSM8K, Llama-2 7B rank 32 | accuracy (%): 31.38 -> 30.09 | Hurt |
| add error feedback to DCT-AdamW (ablation) | DCT-AdamW · GSM8K, Llama-2 7B rank 32 | accuracy (%): 30.09 -> 29.11 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GaLore — GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- From Low Rank Gradient Subspace Stabilization to Low-Rank Weights: Observations, Theories, and Applications (ICML 2025 · similarity 0.89)
- SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLM Training (NeurIPS 2025 · similarity 0.88)
- LDAdam: Adaptive Optimization from Low-Dimensional Gradient Statistics (ICLR 2025 · similarity 0.88)
- Breaking the Frozen Subspace: Importance Sampling for Low-Rank Optimization in LLM Pretraining (NeurIPS 2025 · similarity 0.88)
- Subspace Optimization for Large Language Models with Convergence Guarantees (ICML 2025 · similarity 0.88)
- LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
