Research map / Architectures and efficiency / Pruning and efficient tuning
PRAC: Principal-Random Subspace for LLM Activation Compression and Memory-Efficient Training
Yanyi Li, Yimu Zhang, Cong Fang
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 |
|---|---|---|---|
| use PRAC instead of RSO (add principal subspace to random) | RSO · LLaMA-1B C4 pre-training | validation perplexity: 15.72 -> 15.41 | Improved |
| use PRAC instead of RSO | RSO · LLaMA-1B C4 pre-training | peak memory (GB): 79.32 -> 60.48 | Improved |
| use PRAC instead of RSO | RSO · GPT-2-124M OpenWebText | validation perplexity: 22.31 -> 19.95 | Improved |
| use PRAC instead of GaLore | GaLore · LLaMA-130M C4 | validation perplexity: 25.12 -> 24.66 | Improved |
| use PRAC instead of full-activation baseline | AdamW baseline · LLaMA-130M C4 | validation perplexity: 24.41 -> 24.66 | Hurt |
| use PRAC instead of full-activation baseline | AdamW baseline · LLaMA-130M C4 | total memory reduction: 21.26 -> 15.65 | Improved |
| use PRAC with larger batch (64 -> 96) enabled by memory savings | PRAC · LLaMA-1B training time | training time (h): 156 -> 117 | Improved |
| use PRAC at same batch size 64 | baseline · LLaMA-1B training time | training time (h): 156 -> 179 | Hurt |
| use PRAC instead of RSO for RoBERTa fine-tuning | RSO rank 4 · GLUE average | average score: 86.10 -> 86.38 | Improved |
| use PRAC instead of full fine-tuning | Full Fine-Tuning · GLUE average | average score: 86.28 -> 86.38 | Improved |
| add principal subspace to random-only projection (PRAC vs RAC) (ablation) | RAC · LLaMA pre-training loss | convergence | Improved |
| add random subspace to principal-only projection (PRAC vs PAC) (ablation) | PAC · LLaMA pre-training loss | convergence | Improved |
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 5 of 5 shared results
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
- Memory-Efficient LLM Training with Online Subspace Descent (NeurIPS 2024 · similarity 0.89)
- Breaking the Frozen Subspace: Importance Sampling for Low-Rank Optimization in LLM Pretraining (NeurIPS 2025 · similarity 0.88)
- Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples (NeurIPS 2025 · similarity 0.88)
- SLoPe: Double-Pruned Sparse Plus Lazy Low-Rank Adapter Pretraining of LLMs (ICLR 2025 · similarity 0.87)
- Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint? (NeurIPS 2025 · similarity 0.87)
- A Memory Efficient Randomized Subspace Optimization Method for Training Large Language Models (ICML 2025 · 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.
