Research map / Architectures and efficiency / Pruning and efficient tuning
Mastering Sparse CUDA Generation through Pretrained Models and Deep Reinforcement Learning
Yaoyu Wang, Hankun Dai, Zhidong Yang, Junmin Xiao, Guangming Tan
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 |
|---|---|---|---|
| add CUDA-specific pretraining (ablation) | SparseRL · SuiteSparse SpMV | pass@1000: 40.75 -> 49.25 | Improved |
| add CUDA-specific pretraining (ablation) | SparseRL · SuiteSparse SpMV | average execution speed (GFLOPS): 95.22 -> 116.20 | Improved |
| add PPO stage after SFT (ablation) | SparseRL · SuiteSparse SpMV | pass@1000: 32.75 -> 49.25 | Improved |
| add SFT stage before PPO (ablation) | SparseRL · SuiteSparse SpMV | pass@1000: 15.25 -> 49.25 | Improved |
| use sinusoidal embedding instead of raw indices (ablation) | SparseRL+CodeT5 · SuiteSparse SpMV | pass@1000: 40.75 -> 48.75 | Improved |
| use sinusoidal embedding instead of Max-Min normalization (ablation) | SparseRL+CodeT5 · SuiteSparse SpMV | pass@1000: 43.25 -> 48.75 | Improved |
| use sinusoidal embedding instead of UniXcoder-style embedding (ablation) | SparseRL+CodeT5 · SuiteSparse SpMV | pass@1000: 43.00 -> 48.75 | Improved |
| add execution efficiency reward to correctness-only RL (ablation) | SparseRL · SuiteSparse SpMV | GFLOPS | Improved |
| raise efficiency scale r_eff from 0.5 to 1.0 (ablation) | SparseRL · SuiteSparse SpMV | average GFLOPS: 102.3 -> 116.2 | Improved |
| raise efficiency scale r_eff from 1.0 to 1.5 (ablation) | SparseRL · SuiteSparse SpMV | pass@1000 and GFLOPS: 49.25 -> 48.50 pass@1000; 116.2 -> 118.7 GFLOPS | Mixed |
| raise memory penalty r_penalty from 0.3 to 0.5 (ablation) | SparseRL · SuiteSparse SpMV | pass@1000: 49.25 -> 46.80 | Hurt |
| weight correctness more (alpha 0.5 -> 0.9) (ablation) | SparseRL · SuiteSparse SpMV | pass@1000 and GFLOPS: 49.25 -> 51.2 pass@1000; 116.2 -> 98.7 GFLOPS | Mixed |
Closest papers by meaning
- DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation (NeurIPS 2026 · similarity 0.84)
- Navigating Extremes: Dynamic Sparsity in Large Output Spaces (NeurIPS 2024 · similarity 0.83)
- StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning (ICML 2026 · similarity 0.83)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.83)
- Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.83)
- QiMeng-GEMM: Automatically Generating High-Performance Matrix Multiplication Code by Exploiting Large Language Models (AAAI 2025 · similarity 0.83)
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.
