pith. sign in

arxiv: 2606.09312 · v1 · pith:5AJH5D37new · submitted 2026-06-08 · 💻 cs.LG · cs.PL

Toward Compiler World Models: Learning Latent Dynamics for Efficient Tensor Program Search

Haolin Pan , Lianghong Huang , Xvlin Zhou , Mingjie Xing , Yanjun Wu This is my paper

Pith reviewed 2026-06-27 17:31 UTC · model grok-4.3

classification 💻 cs.LG cs.PL
keywords tensor program optimizationauto-schedulinglatent dynamicsworld modelscompiler optimizationTVMmachine learning systemscost modeling
0
0 comments X

The pith

A latent dynamics model that simulates scheduling actions in continuous space ranks tensor program candidates more accurately than static cost models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that treating schedule search as action-conditioned latent dynamics lets an evaluator roll out trajectories from the starting program without mutating ASTs or re-encoding code for every candidate. This matters because existing auto-schedulers judge each schedule as an isolated snapshot and therefore miss how actions depend on one another. If the claim holds, search can reach better programs with far fewer hardware measurements while still respecting hardware and action features. The approach is implemented inside TVM AutoScheduler and reports concrete gains on representative subgraphs and full models.

Core claim

The central claim is that modeling schedule evaluation as action-conditioned latent dynamics over program states, using a lightweight transition model to roll out actions in continuous latent space, produces a final dynamic representation that, when combined with action and hardware features, ranks candidates more effectively than static evaluators.

What carries the argument

Action-conditioned latent dynamics transition model that rolls out scheduling actions from the initial program state in continuous space.

If this is right

  • Under a fixed 64-trial budget the method reduces representative-subgraph latency by 1.37× on GPU and 1.54× on CPU compared with Ansor.
  • It reaches performance within 2.2 % geometric mean of an Ansor run that used 10 000 trials while using only one-tenth the measurements.
  • Full-model inference speed improves by 4.61× over PyTorch and 3.67× over PyTorch with cuDNN geometric mean.
  • The evaluator avoids repeated code generation and measurement for each candidate by simulating trajectories in latent space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same latent-dynamics pattern could be tried on other compiler passes that also build long action sequences, such as register allocation or loop nest transformations.
  • If the transition model proves accurate, it opens the possibility of gradient-based search inside the latent space rather than discrete trial-and-error.
  • Hardware-specific features are still concatenated at the end, so the method may still need retraining when the target device changes.

Load-bearing premise

The lightweight transition model operating in continuous latent space can faithfully predict the performance effects of scheduling actions without actual AST changes or repeated executions.

What would settle it

Measure the correlation between the latent model's predicted ranking of 1000 schedules and their actual measured latencies on the target hardware; if the ranking correlation falls below a usable threshold the claim fails.

Figures

Figures reproduced from arXiv: 2606.09312 by Haolin Pan, Lianghong Huang, Mingjie Xing, Xvlin Zhou, Yanjun Wu.

Figure 1
Figure 1. Figure 1: Common deep learning compiler pipeline. Our work focuses on candidate evaluation. schedules are modeled through action-conditioned latent state evolution. • We construct a TVM/TenSet-based state-prediction dataset for learning compiler state transitions. Built from tuning logs and aligned TensorIR states, the dataset organizes pre-schedule states, scheduling-action sequences, intermediate states, and post-… view at source ↗
Figure 2
Figure 2. Figure 2: Dynamic latent-state schedule evaluation. The proposed model represents scheduling as action-conditioned latent state evolution, reducing syntax sensitivity and resolv￾ing the referential ambiguity of action-only models. how large it is, or what its memory properties are. This is because the physical properties and existence of 𝑘 are strictly defined within the intermediate program state 𝑠1 that existed im… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our framework. Given an initial TensorIR state and a candidate scheduling-action sequence, the framework evaluates the candidate through the terminal-state representation induced by that sequence. Starting from the initial program state, it models action-conditioned state evolution in representation space and uses the predicted terminal-state representation together with action and hardware fea… view at source ↗
Figure 4
Figure 4. Figure 4: End-to-end latency across seven models. PyTorch￾opt denotes cuDNN-enabled PyTorch execution. The last group reports the geometric mean across the seven models. 4.2 End-to-End Model Performance We first evaluate the end-to-end performance of the opti￾mized models. This experiment compares our method with PyTorch, PyTorch-opt, TensorRT, and Ansor on the seven benchmark models described in Section 4.1. PyTorc… view at source ↗
Figure 5
Figure 5. Figure 5: Model-level weighted-latency speedup over Ansor. Each entry reports Ansor/Ours, so values larger than one indicate that our method obtains lower weighted latency. The last row reports the geometric mean across the seven models. We therefore use the following experiments to isolate the search quality of our method against Ansor under matched TVM tuning settings. 4.3 Model-Level Search Performance The end-to… view at source ↗
Figure 6
Figure 6. Figure 6: GPU representative-subgraph speedup over time. Each panel shows one representative subgraph, and the y-axis reports Ansor/Ours. The time-budget analysis starts from 200 seconds [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: shows the corresponding CPU trial-budget results. The same-budget comparison again favors our method: it achieves a 1.54× geometric-mean speedup at 64 trials and re￾mains slightly faster at 1024 trials with a 1.01× speedup. The cross-budget setting further shows that Ours-16 already out￾performs Ansor-64 by 1.25×, while Ours-64 nearly reaches Ansor-1024. The consistent trends on both GPU and CPU indicate t… view at source ↗
Figure 8
Figure 8. Figure 8: Sample efficiency compared with large-budget Ansor on GPU representative subgraphs. (a) Geometric-mean ratios against Ansor-1024, Ansor-2048, Ansor-4096, Ansor￾8192, and Ansor-10240. (b) Per-subgraph ratios for Ours-1024 against Ansor-10240, sorted by speedup. 8192, and 10240 trials on the 22 GPU representative sub￾graphs. The reported ratio is Ansor/Ours. Therefore, values larger than one indicate that ou… view at source ↗
read the original abstract

Tensor program optimization is essential for modern machine learning systems, but its search space is enormous. Existing auto-schedulers reduce measurement cost with learned cost models, yet they usually evaluate each candidate as a static code snapshot, ignoring the schedule trajectory that produced it. This makes them insensitive to action dependencies and vulnerable to superficial code variations. We propose a \emph{world-model-inspired} evaluator that models schedule evaluation as action-conditioned latent dynamics over program states. Starting from the initial program, it rolls out scheduling actions in a continuous latent space with a lightweight transition model, avoiding expensive AST mutation and repeated code encoding. The final dynamic representation is combined with action and hardware features to rank candidates. Implemented in TVM AutoScheduler, our method improves representative-subgraph latency over Ansor by 1.37$\times$ on GPU and 1.54$\times$ on CPU under the same 64-trial budget. It also matches Ansor-10K within 2.2% geometric mean using 10$\times$ fewer measurements, and accelerates full-model inference over PyTorch/PyTorch-opt(cuDNN) by 4.61$\times$/3.67$\times$ geometric mean.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The paper claims that modeling tensor program scheduling evaluation as action-conditioned latent dynamics in a continuous latent space, using a lightweight transition model starting from the initial program, allows efficient candidate ranking without repeated AST mutations or code encodings. Implemented in TVM AutoScheduler, the approach reports 1.37× GPU and 1.54× CPU improvements in representative-subgraph latency over Ansor under a 64-trial budget, matches Ansor-10K performance within 2.2% geometric mean using 10× fewer measurements, and yields 4.61×/3.67× geometric mean speedups over PyTorch/PyTorch-opt(cuDNN) for full-model inference.

Significance. If the results hold, this work could meaningfully advance auto-scheduling by incorporating dynamics modeling to capture action dependencies, potentially reducing measurement costs in tensor program optimization for ML systems. The real-hardware evaluation and comparison to strong baselines like Ansor are strengths.

minor comments (1)
  1. The abstract supplies no model architecture details, training procedure, validation splits, error bars, or ablation studies; the full manuscript should ensure these are clearly presented in the experimental section to support the quantitative claims.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work, including recognition of the real-hardware evaluation and strong baseline comparisons. The recommendation for minor revision is noted. As no specific major comments were provided in the report, we have no point-by-point responses to address at this time.

Circularity Check

0 steps flagged

No significant circularity; empirical results validated externally on hardware

full rationale

The paper's central claims consist of measured latency improvements (1.37×/1.54× over Ansor under fixed trial budget, matching Ansor-10K with 10× fewer measurements) obtained by deploying a learned latent transition model inside TVM AutoScheduler and evaluating the resulting schedules on real GPU/CPU hardware. No equations, derivations, or self-citations are presented that reduce a claimed prediction or uniqueness result to fitted parameters or prior author work by construction. The method is a standard learned surrogate for ranking; its validity is established by direct external measurement rather than internal self-definition. This is the most common honest non-finding for applied ML papers with hardware benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities beyond the general learned transition model; the central claim therefore rests on the unstated assumption that the latent dynamics accurately proxy real execution cost.

pith-pipeline@v0.9.1-grok · 5752 in / 1227 out tokens · 30770 ms · 2026-06-27T17:31:12.256581+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 9 canonical work pages · 6 internal anchors

  1. [1]

    Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. 2016. {TensorFlow}: a system for {Large-Scale} machine learning. In12th USENIX symposium on operating systems design and implementation (OSDI 16). 265–283

    2016

  2. [2]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al . 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    Andrew Adams, Karima Ma, Luke Anderson, Riyadh Baghdadi, Tzu- Mao Li, Michaël Gharbi, Benoit Steiner, Steven Johnson, Kayvon Fata- halian, Frédo Durand, et al. 2019. Learning to optimize halide with tree search and random programs.ACM Transactions on Graphics (TOG)38, 4 (2019), 1–12

    2019

  4. [4]

    Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, et al. 2024. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM international conference on architectural support for programmin...

    2024

  5. [5]

    Anthropic. 2024. The Claude 3 Model Family: Opus, Sonnet, Haiku.https://www-cdn.anthropic.com/ de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/ModelCardClaude3. pdfModel card

    2024

  6. [6]

    Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, et al. 2026. Motus: A unified latent action world model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 35101–35113

    2026

  7. [7]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for model- ing multi-relational data.Advances in neural information processing systems26 (2013)

    2013

  8. [8]

    Christopher Burges, Robert Ragno, and Quoc Le. 2006. Learning to rank with nonsmooth cost functions.Advances in neural information processing systems19 (2006)

    2006

  9. [9]

    Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 785–794

    2016

  10. [10]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hin- ton. 2020. A simple framework for contrastive learning of visual rep- resentations. InInternational conference on machine learning. PmLR, 1597–1607

    2020

  11. [11]

    Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. 2015. Mxnet: A flexible and efficient machine learning library for heteroge- neous distributed systems.arXiv preprint arXiv:1512.01274(2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  12. [12]

    Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. 2018. {TVM}: An automated {End-to-End} optimizing compiler for deep learning. In13th USENIX symposium on operating systems design and implementation (OSDI 18). 578–594

    2018

  13. [13]

    Tianqi Chen, Lianmin Zheng, Eddie Yan, Ziheng Jiang, Thierry Moreau, Luis Ceze, Carlos Guestrin, and Arvind Krishnamurthy. 2018. Learning to optimize tensor programs.Advances in Neural Information Processing Systems31 (2018)

    2018

  14. [14]

    Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Co- hen, John Tran, Bryan Catanzaro, and Evan Shelhamer. 2014. cudnn: Efficient primitives for deep learning.arXiv preprint arXiv:1410.0759 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  15. [15]

    Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, et al. 2020. Codebert: A pre-trained model for programming and natural lan- guages. InFindings of the association for computational linguistics: EMNLP 2020. 1536–1547

    2020

  16. [16]

    2026.oneDNN Documentation

    Intel. 2026.oneDNN Documentation. Intel.https://www. intel.com/content/www/us/en/developer/tools/oneapi/onednn- documentation.htmlAccessed: 2026-04-12

    2026

  17. [17]

    Jaber Jaber and Osama Jaber. 2026. AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search.arXiv preprint arXiv:2603.21331(2026)

    work page arXiv 2026

  18. [18]

    Isu Jeong and Seulki Lee. 2025. Bayesian code diffusion for efficient automatic deep learning program optimization. In19th USENIX Sym- posium on Operating Systems Design and Implementation (OSDI 25). 295–311

    2025

  19. [19]

    Guoliang Ji, Shizhu He, Liheng Xu, Kang Liu, and Jun Zhao. 2015. Knowledge graph embedding via dynamic mapping matrix. InPro- ceedings of the 53rd annual meeting of the association for computational linguistics and the 7th international joint conference on natural language processing (volume 1: Long papers). 687–696

    2015

  20. [20]

    Chris Lattner, Mehdi Amini, Uday Bondhugula, Albert Cohen, Andy Davis, Jacques Pienaar, River Riddle, Tatiana Shpeisman, Nicolas Vasi- lache, and Oleksandr Zinenko. 2020. MLIR: A compiler infrastructure for the end of Moore’s law.arXiv preprint arXiv:2002.11054(2020)

    work page arXiv 2020

  21. [21]

    Yann LeCun et al. 2022. A path towards autonomous machine intelli- gence version 0.9. 2, 2022-06-27.Open Review62, 1 (2022), 1–62

    2022

  22. [22]

    Yankai Lin, Zhiyuan Liu, Maosong Sun, Yang Liu, and Xuan Zhu. 2015. Learning entity and relation embeddings for knowledge graph com- pletion. InProceedings of the AAAI conference on artificial intelligence, Vol. 29

    2015

  23. [23]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al . 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    Yizhi Liu, Yao Wang, Ruofei Yu, Mu Li, Vin Sharma, and Yida Wang

  25. [25]

    In2019 USENIX Annual Technical Conference (USENIX ATC 19)

    Optimizing {CNN} model inference on {CPUs }. In2019 USENIX Annual Technical Conference (USENIX ATC 19). 1025–1040

  26. [26]

    2026.LLVM Documentation

    LLVM Project. 2026.LLVM Documentation. LLVM Project.https: //llvm.org/docs/Accessed: 2026-04-12

    2026

  27. [27]

    Martin Paul Lücke, Oleksandr Zinenko, William S Moses, Michel Steuwer, and Albert Cohen. 2025. The MLIR transform dialect: Your compiler is more powerful than you think. InProceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimiza- tion. 241–254

    2025

  28. [28]

    Zixuan Ma, Haojie Wang, Jingze Xing, Shuhong Huang, Liyan Zheng, Chen Zhang, Huanqi Cao, Kezhao Huang, Mingshu Zhai, Shizhi Tang, et al. 2025. IntelliGen: Instruction-Level Auto-tuning for Tensor Pro- gram with Monotonic Memory Optimization. InProceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Opti- mization. 107–122

    2025

  29. [29]

    2026.CUTLASS: CUDA Templates for Linear Algebra Subrou- tines and Solvers

    NVIDIA. 2026.CUTLASS: CUDA Templates for Linear Algebra Subrou- tines and Solvers. NVIDIA.https://docs.nvidia.com/cutlass/Accessed: 2026-04-12

    2026

  30. [30]

    2026.NVIDIA CUDA Compiler Driver NVCC Documentation

    NVIDIA. 2026.NVIDIA CUDA Compiler Driver NVCC Documentation. NVIDIA.https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/ index.htmlVersion 13.1, Accessed: 2026-04-12. 13 Haoin Pan et al

    2026

  31. [31]

    2026.NVIDIA TensorRT Documentation

    NVIDIA. 2026.NVIDIA TensorRT Documentation. NVIDIA.https: //docs.nvidia.com/deeplearning/tensorrt/latest/Accessed: 2026-04-12

    2026

  32. [32]

    2026.ONNX Documentation

    ONNX Community. 2026.ONNX Documentation. ONNX.https: //onnx.ai/onnx/Version 1.22.0, Accessed: 2026-04-12

    2026

  33. [33]

    Liang Qiao, Jun Shi, Xiaoyu Hao, Xi Fang, Sen Zhang, Minfan Zhao, Ziqi Zhu, Junshi Chen, Hong An, Xulong Tang, et al. 2025. Pruner: A draft-then-verify exploration mechanism to accelerate tensor pro- gram tuning. InProceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2. 949–965

    2025

  34. [34]

    Qiuyi Qu, Yicheng Sui, Yufei Sun, Rui Chen, Xiaofei Zhang, Yuzhi Zhang, Haofeng Wang, and Ge Lan. 2026. A Two-Stage GPU Kernel Tuner Combining Semantic Refactoring and Search-Based Optimiza- tion.arXiv preprint arXiv:2601.12698(2026)

    work page arXiv 2026

  35. [35]

    Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Hao- ran Wei, et al . 2024. Qwen2.5 Technical Report.arXiv preprint arXiv:2412.15115(2024).https://arxiv.org/abs/2412.15115

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Jared Roesch, Steven Lyubomirsky, Logan Weber, Josh Pollock, Marisa Kirisame, Tianqi Chen, and Zachary Tatlock. 2018. Relay: A new ir for machine learning frameworks. InProceedings of the 2nd ACM SIG- PLAN international workshop on machine learning and programming languages. 58–68

    2018

  37. [37]

    Junru Shao, Xiyou Zhou, Siyuan Feng, Bohan Hou, Ruihang Lai, Hongyi Jin, Wuwei Lin, Masahiro Masuda, Cody Hao Yu, and Tianqi Chen. 2022. Tensor program optimization with probabilistic programs. Advances in Neural Information Processing Systems35 (2022), 35783– 35796

    2022

  38. [38]

    Tadahiro Taniguchi, Ryo Ueda, Tomoaki Nakamura, Masahiro Suzuki, and Akira Taniguchi. 2026. Generative emergent communication: Large language model is a collective world model.Advanced Robotics (2026), 1–26

    2026

  39. [39]

    Philippe Tillet, Hsiang-Tsung Kung, and David Cox. 2019. Triton: an intermediate language and compiler for tiled neural network computa- tions. InProceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. 10–19

    2019

  40. [40]

    Mohammed Tirichine, Nassim Ameur, Nazim Bendib, Iheb Nassim Aouadj, Djad Bouchama, Rafik Bouloudene, and Riyadh Baghdadi

  41. [41]

    In2026 IEEE/ACM International Symposium on Code Generation and Optimization (CGO)

    A Reinforcement Learning Environment for Automatic Code Optimization in the MLIR Compiler. In2026 IEEE/ACM International Symposium on Code Generation and Optimization (CGO). IEEE, 696– 710

  42. [42]

    Nicolas Vasilache, Oleksandr Zinenko, Theodoros Theodoridis, Priya Goyal, Zachary DeVito, William S Moses, Sven Verdoolaege, Andrew Adams, and Albert Cohen. 2018. Tensor comprehensions: Framework- agnostic high-performance machine learning abstractions.arXiv preprint arXiv:1802.04730(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  43. [43]

    Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. InPro- ceedings of the AAAI conference on artificial intelligence, Vol. 28

    2014

  44. [44]

    Yi Zhai, Sijia Yang, Keyu Pan, Renwei Zhang, Shuo Liu, Chao Liu, Zichun Ye, Jianmin Ji, Jie Zhao, Yu Zhang, et al. 2024. Enabling tensor language model to assist in generating {High-Performance} tensor programs for deep learning. In18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24). 289–305

    2024

  45. [45]

    Yi Zhai, Yu Zhang, Shuo Liu, Xiaomeng Chu, Jie Peng, Jianmin Ji, and Yanyong Zhang. 2023. Tlp: A deep learning-based cost model for tensor program tuning. InProceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2. 833–845

    2023

  46. [46]

    Hongbin Zhang, Mingjie Xing, Yanjun Wu, and Chen Zhao. 2023. Com- piler Technologies in Deep Learning Co-Design: A Survey.Intelligent Computing(2023)

    2023

  47. [47]

    Jie Zhao, Bojie Li, Wang Nie, Zhen Geng, Renwei Zhang, Xiong Gao, Bin Cheng, Chen Wu, Yun Cheng, Zheng Li, et al. 2021. AKG: auto- matic kernel generation for neural processing units using polyhedral transformations. InProceedings of the 42nd ACM SIGPLAN Interna- tional Conference on Programming Language Design and Implementa- tion. 1233–1248

    2021

  48. [48]

    Lianmin Zheng, Chengfan Jia, Minmin Sun, Zhao Wu, Cody Hao Yu, Ameer Haj-Ali, Yida Wang, Jun Yang, Danyang Zhuo, Koushik Sen, et al. 2020. Ansor: Generating {High-Performance} tensor programs for deep learning. In14th USENIX symposium on operating systems design and implementation (OSDI 20). 863–879

    2020

  49. [49]

    Lianmin Zheng, Ruochen Liu, Junru Shao, Tianqi Chen, Joseph E Gonzalez, Ion Stoica, and Ameer Haj Ali. 2021. Tenset: A large-scale program performance dataset for learned tensor compilers. InThirty- fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1)

    2021

  50. [50]

    Size Zheng, Yun Liang, Shuo Wang, Renze Chen, and Kaiwen Sheng

  51. [51]

    InPro- ceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems

    Flextensor: An automatic schedule exploration and optimization framework for tensor computation on heterogeneous system. InPro- ceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems. 859–873. 14