kernelbot-data
收藏资源简介：
# KernelBot Competition Data This dataset contains GPU kernel submissions from the KernelBot competition platform. Submissions are optimized GPU kernels written for specific hardware targets. ## Data Files ### AMD MI300 Submissions | File | Description | |------|-------------| | `submissions.parquet` | All AMD competition submissions | | `successful_submissions.parquet` | AMD submissions that passed correctness tests | | `deduplicated_submissions.parquet` | AMD submissions deduplicated by (user, code) | | `deduplicated_successful_submissions.parquet` | Deduplicated passing AMD submissions | **AMD Problems:** fp8-gemm, moe (mixture of experts), mla-decode, all2all, gemm+reducescatter, allgather+gemm, mxfp4-mm, moe-mxfp4, mixed-mla ### AMD 1.1M Competition | File | Size | Description | |------|------|-------------| | `amd_1_1m_competition_submissions.parquet` | ~699 MB | Deduplicated submissions with code for `amd-mxfp4-mm` (763), `amd-moe-mxfp4` (764), and `amd-mixed-mla` (765) | ### Trimul | File | Size | Description | |------|------|-------------| | `trimul_submissions.parquet` | ~120 MB | Deduplicated submissions with code for `trimul` (leaderboard 496) | `trimul` is a separate mixed-GPU problem and is not grouped with the AMD competition exports. ### Helion B200_Nebius | File | Size | Description | |------|------|-------------| | `helion_b200_nebius_submissions.parquet` | ~4 MB | Deduplicated submissions with code for `causal_conv1d` (766), `fp8_quant` (767), `gated_deltanet_chunk_fwd_h` (768), `gated_deltanet_chunk_fwd_o` (769), and `gated_deltanet_recompute_w_u` (770) | **Measurement note:** these problems were run on `B200_Nebius`, and the measurements for this problem set are brittle. Treat leaderboard scores from this export with extra caution. ### NVIDIA Blackwell NVFP4 Submissions | File | Size | Description | |------|------|-------------| | `nvidia_nvfp4_submissions.parquet` | ~1.4 GB | NVFP4 submissions deduplicated by (user, code), with full code content | **NVFP4 Problems:** gemv (leaderboard 595), gemm (597), dual_gemm (598), modal_dual_gemm (697), group_gemm (730) **Note on Dual GEMM:** There are two variants of the dual_gemm problem. Midway through the competition, on-prem hardware measurements became unreliable, so a second leaderboard was created on Modal infrastructure. The Modal measurements (leaderboard 697, `modal_nvfp4_dual_gemm`) are more trustworthy. **Note:** Scores are execution time in seconds. **Lower is better.** ### PMPP v2 Submissions | File | Size | Description | |------|------|-------------| | `pmpp_v2_submissions.parquet` | ~28 MB | All PMPP v2 submissions with full code content | **PMPP v2 Problems:** conv2d_v2 (537), grayscale_v2 (538), histogram_v2 (539), matmul_v2 (540), prefixsum_v2 (541), sort_v2 (542), vectoradd_v2 (543), vectorsum_v2 (544) ### Linear Algebra Submissions | File | Size | Description | |------|------|-------------| | `linalg_submissions.parquet` | ~552 MB | Deduplicated submissions with code for `qr_v2` (leaderboard 774) | This export includes only `qr_v2`; the earlier `qr` leaderboard (773) is not included. ## Helper Scripts - `analyze_submissions.py` - Python functions for analyzing submissions - `skills.md` - Documentation for data processing workflows ### Quick Start ```python from analyze_submissions import load_submissions, top_contestants, author_progression # Load NVIDIA NVFP4 data df = load_submissions() # Get top 20 for a problem leaders = top_contestants(df, problem_name='nvfp4_gemm', n=20) # See a user's progression over time progression = author_progression(df, user_name='username', problem_name='nvfp4_gemm') ``` ## Learn More - Competition platform: [gpumode.com](https://gpumode.com) - Reference kernels and problem specs: [github.com/gpu-mode/reference-kernels](https://github.com/gpu-mode/reference-kernels) ## License This dataset is licensed under the [June 9 Researcher Reciprocity License](LICENSE). You are free to use, share, analyze, transform, and redistribute the material for research, education, benchmarking, publication, commercial analysis, and other lawful purposes, provided you give appropriate credit. This license adapts the Open RAIL-D structure and adds one specific use restriction: training, fine-tuning, distillation, synthetic-data generation for training, embedding for training, or otherwise using this dataset to improve an AI model or AI service requires Researcher Reciprocity. > If you train on it, you let us generate. Covered AI model and service providers may not use this dataset while imposing terms that prevent GPU Mode, dataset contributors, or authorized researchers from generating outputs, evaluating models, benchmarking, publishing research, or exploring their own research ideas on materially equal terms to ordinary users. **Attribution:** Please cite GPU Mode and link to this dataset. For academic papers, use the citation below. ## Citation If you use this dataset in your work, please cite: ```bibtex @inproceedings{ kernelbot2025, title={KernelBot: A Competition Platform for Writing Heterogeneous {GPU} Code}, author={Alex L Zhang and Matej Sirovatka and Erik Schultheis and Benjamin Horowitz and Mark Saroufim}, booktitle={Championing Open-source DEvelopment in ML Workshop @ ICML25}, year={2025}, url={https://openreview.net/forum?id=bq9U4dmuyJ} } ```
本数据集源自首届与第二届AMD 10万美元算力核（kernel）竞赛，包含约11万个针对FP8通用矩阵乘法（fp8-gemm）、混合专家模型（MoE）、矩阵注意力（MLA）、全对全通信（all2all）、gemm+归约散射（reducescatter）以及全收集+通用矩阵乘法（allgather+gemm）的优化内核，可在MI300架构GPU上高效运行。更多详情可访问gpumode.com/v2/news。 若需查看我们已举办及正在开展的全部算力核竞赛列表，可前往https://github.com/gpu-mode/reference-kernels，该页面同时包含参考内核及其输入形状与分布的详细说明。 我们计划后续新增针对Blackwell架构的NVFP4优化内核。 若您在研究工作中使用本数据集，请引用如下文献： bibtex @inproceedings{ zhang2025kernelbot, title={KernelBot: A Competition Platform for Writing Heterogeneous {GPU} Code}, author={Alex L Zhang and Matej Sirovatka and Erik Schultheis and Benjamin Horowitz and Mark Saroufim}, booktitle={Championing Open-source DEvelopment in ML Workshop @ ICML25}, year={2025}, url={https://openreview.net/forum?id=bq9U4dmuyJ} }
