Research map / Architectures and efficiency / RNNs and normalization
POET-X: Memory-efficient LLM Training by Scaling Orthogonal Transformation
Zeju Qiu, Lixin Liu, Adrian Weller, Han Shi, Weiyang Liu
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 POET-X_fast instead of original POET | POET · Single-layer latency | fwd+bwd time (ms): 10.59ms to 1.38ms | Improved |
| Use gradient checkpointing (POET-X_mem) instead of standard autograd | POET-X · Single-layer latency | fwd+bwd time (ms): 10.59ms to 1.89ms | Improved |
| Add permutation reduction (merge 2 of 4 permutations into W) (ablation) | POET-X · Permutation benchmark, block 256, uncompiled, hidden 4096 | time (ms): 15.858 -> 12.017 (1.32x) | Improved |
| Add permutation reduction (merge 2 of 4 permutations into W) (ablation) | POET-X · Permutation benchmark, block 512, compiled, hidden 16384 | time (ms): 96.798 -> 86.380 (1.12x) | Improved |
| Use custom CUDA permutation operator instead of PyTorch-native (ablation) | POET-X · Permutation, hidden dim 2048 | time (ms): 1.621 -> 0.086 (18.75x) | Improved |
| Use custom CUDA permutation operator instead of PyTorch-native (ablation) | POET-X · Permutation, hidden dim 16384 | time (ms): 13.246 -> 0.946 (14.00x) | Improved |
| Use customized batch-parallel block-diagonal multiplication instead of PyTorch (ablation) | POET-X · Seq length 2048 | time (ms): 0.790 -> 0.334 (2.37x) | Improved |
| Store only half of skew-symmetric matrices in CNP (memory reduction) (ablation) | POET-X · Seq length 2048 | memory (MB): 280 -> 192 (31.43%) | Improved |
| Use Triton CNP kernels instead of PyTorch-native, N=64 (ablation) | POET-X · CNP benchmark N=64 | time (ms): 0.316 -> 0.107 (2.96x) | Improved |
| Use POET-X_b=512 instead of b=256 (larger block size) | POET-X · C4 Llama-3B val PPL | validation perplexity: 12.76 -> 12.05 | Improved |
| Use POET-XQ_b=512 instead of b=256 | POET-XQ · C4 Llama-3B quantized val PPL | validation perplexity: 16.21 -> 14.78 | Improved |
| Replace AdamW with POET-X b=512 | Llama-3B · C4 Lmax=256 | validation perplexity: 12.69 -> 12.05 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- GaLore — GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024) · better on 3 of 3 shared results
- POET — POET: Training Neural Networks on Tiny Devices with Integrated Rematerialization and Paging (ICML 2022) · 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
- Reparameterized LLM Training via Orthogonal Equivalence Transformation (NeurIPS 2025 · similarity 0.94)
- GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024 · similarity 0.88)
- Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation (NeurIPS 2026 · similarity 0.87)
- ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization (NeurIPS 2024 · similarity 0.87)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.87)
- GANQ: GPU-Adaptive Non-Uniform Quantization for 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.
