- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper presents ModelParts, an RL training data generation technique to generate training tasks for coding models to produce efficient kernel code on new hardware, where no existing kernel library is available. The dataset consists of different partitions of sets of ATen IR operations, captured from real model graphs during torch.compile. Starting from the torchinductor's own partitioning, the authors run a Metropolis–Hastings walk with shift, fuse, and split moves to sample alternative partitions, and each resulting part becomes a task. The model must write an NKI kernel covering the operations. The authors propose using a reward based on the roofline.
Using this dataset, the authors train Qwen3.8-27B with GRPO and compare against training on synthetic operator compositions (NKI-Agent-Ops), with and without profiler access during the agent loop. ModelParts training improves Pass@8 and transfers better than synthetically generated tasks (NKO Agent Ops).
Critical strengths
- Important problem to work on.
- Authors demonstrate decent improvements in coding performance using this approach.
Critical weaknesses
- It is not clear that RL enables efficient kernel code generation on new hardware. Knowledge of a new platform is learnt during pretraining.
- It is not clear that the coding model trained on ModelParts transfers to other models, since all evaluation tasks are drawn from the same five models used for training.
- ModelParts addresses the data generation problem for a "new platform", but NKI/Trainium is not a new platform: it already has an SDK and expert-written kernels, and the base model has likely seen NKI code during pretraining. The cold-start setting is therefore only emulated.
Questions for authors
I think the paper was motivates a real problem of writing efficient kernel code for brand new hardware. I think the paper falls short on a few accounts. Please see my comments below:
-
Premise of the paper. My main concern is the premise of the paper. The authors argue that current models would struggle with a new hardware release and therefore need RL training on generated tasks before a kernel ecosystem exists. Prior research, however, has shown that acquiring new knowledge happens during pretraining, and RL training only assists in sampling the model, i.e., it sharpens the distribution toward solutions the base model can already produce rather than expanding its capabilities [1]. The authors should thus narrow the motivation of the Model Parts.
-
More general claim In my experience, modern coding models work really well at producing low-level GPU code, and with a brief description of the new DSL and hardware, this is a transferable skill. So it is not well established whether a new accelerator requires RL training on generated tasks at all, rather than prompting a strong existing model with documentation and profiler feedback. The paper's own results support this (untrained Opus 5 with PITL achieves the best NKILib speedup).
Evaluation/Results
-
Learning transferable skills across models. Qwen trained with ModelParts using RL is able to achieve a good speedup because the tasks in dataset are generated by MH sampling of different operation sets from the same model graphs used for training. MH naturally explores peration sets with faster times using torchinductor. For a fair apples-to-apples comparison against NKI-Agent-Ops, the authors must evaluate Qwen+MP on a new model not present in the dataset. Another way to establish this is to omit one particular model from the ModelParts dataset, evaluate on tasks generated from the omitted model, and compare that to NAO.
-
NKI and Trainium are not new hardware The base model has access to NKI code during pretraining. To study this, the authors should report pass@k for larger k for the base and RL-trained models to show whether RL expands what the model can do or only improves sampling.
-
Comparison to prior work. Could similar gains be achieved simply by using an existing harness without any RL training? AccelOpt [2] targets the same problem on NKI through a test-time search. is it possible to ocmpare the performance of ModelParts tuned model against AccelOpt [2] on the same base model (Qwen3.8-27B)?
-
Opus performs worse than untrained Qwen on ModelParts tasks in Table 2. Why is that?
Minor: Typo change “amd” to “and" in Section 3.1 (line 179)
References: [1] Yue, Yang, et al. "Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?." Advances in Neural Information Processing Systems 38 (2026): 57654-57689.
[2] Zhang, Genghan, et al. "Accelopt: A self-improving llm agentic system for ai accelerator kernel optimization." Proceedings of Machine Learning and Systems 8 (2026): 541-568.
Thank you for the thoughtful review and the concrete experiment suggestions.
CW1 We agree that RL is unlikely to teach knowledge absent from pretraining, and we do not claim it does. Our claim concerns the training pipeline: ModelParts supplies a realistic, verifiable training distribution without an existing kernel ecosystem, and RL on it makes an open model write correct, faster NKI kernels far more reliably. With PITL, Pass@8 on held-out ModelParts tasks rises from 38.3 to 84.9, and the aggregate speedup on 66 NKILib kernels from 1.37× to 1.65×.
CW2 We have not yet evaluated tasks from a model outside the five and will add this. Two results already test transfer beyond them: without PITL, ModelParts training raises Pass@8 on NKI-Agent-Ops tasks, a different generator, from 23.2 to 32.2, nearly matching training on that distribution (32.9); and on NKILib, whose tasks start from expert kernels, it raises the 66 kernel speedup with PITL from 1.37× to 1.65× (faster on 57 of 66), versus 1.42× for NKI-Agent-Ops training.
CW3 Agreed: the base model has likely seen NKI code, so the cold start is emulated for the training pipeline (no NKI kernels, baselines or curated tasks), not for the model's prior knowledge, as §6 states. That exposure does not yield working kernels: before any feedback, 95.2% of base Qwen's and 94.4% of Opus 5's first submissions fail to compile. A newly released accelerator is the right test of the stronger claim; §6 lists it as future work.
Q1 We agree, and we narrowed the motivation and now cite Yue et al. Our results do not show RL adding knowledge absent from pretraining. They show RL turning partial NKI knowledge into working kernels: only 4.8% of the base policy's first submissions compile and pass, and Pass@8 with PITL rises from 38.3 to 84.9.
Q2 On NKILib, prompting a frontier model works well: Opus 5 with PITL is our strongest optimizer (1.97× on 66 kernels), as the revised abstract states. Writing kernels from scratch is different: on held-out ModelParts tasks, Opus 5 with PITL reaches 32.5 Pass@8, below the untrained 27B policy (38.3) and the trained one (84.9). Our claim concerns training open models. ModelParts tasks and the verifier can also improve prompted frontier agents.
ERQ1 Current transfer evidence favors ModelParts: it nearly matches NKI-Agent-Ops training on NKI-Agent-Ops tasks (32.2 vs 32.9 Pass@8 without PITL), while NKI-Agent-Ops training reaches 52.8 on ModelParts tasks vs 67.2. Note that the walk's energy is a roofline latency estimate (launch overhead plus per-part roofline time, §3.1), not TorchInductor timings; TorchInductor supplies only the initial partition.
ERQ2 Agreed; we will report pass@k up to k = 64 for the base and trained policies. Current numbers do not separate the two views: with PITL, the base policy already solves 38.3% of held-out ModelParts tasks within eight attempts, and training raises this to 84.9%.
ERQ3 Our untrained conditions are exactly this setting: the same harness (Harbor with Mini-SWE-Agent) and profiler access, without RL. Training raises Pass@8 on held-out ModelParts tasks from 38.3 to 84.9 and the 66-kernel NKILib speedup from 1.37× to 1.65×. AccelOpt is test-time search over an existing kernel, so it is complementary: a ModelParts-trained policy can serve as its proposer. We will report AccelOpt with base and trained Qwen3.8-27B on NKIBench at matched compute.
ERQ4 Opus 5 has no head start in writing NKI from scratch. First
check-solutionoutcomes on the held-out generated tasks of Table 2, with PITL:NKILib tasks instead start from a correct, compiling expert kernel, and frontier models had 128 steps there rather than 64, which is where Opus 5's skill shows. We replaced §5.1's "harder distribution" reading with this explanation and added the table as Appendix A.12.
Minor. Typo fixed.
We hope these analyses and narrower claims address your concerns. If they do, we would be grateful if you would consider raising your score.