arXiv subjects

Ali Tehrani

Publications and source records attributed to Ali Tehrani.

3 recordsLinked to original sources

arXiv2026

Fine-Tuning GPT-5 for GPU Kernel Generation

Developing efficient GPU kernels is essential for scaling modern AI systems, yet it remains a complex task due to intricate hardware architectures and the need for specialized optimization expertise. Although Large Language Models (LLMs) demonstrate strong capabilities in general sequential code generation, they face significant challenges in GPU code generation because of the scarcity of high-quality labeled training data, compiler biases when generating synthetic solutions, and limited generalization across hardware generations. This precludes supervised fine-tuning (SFT) as a scalable methodology for improving current LLMs. In contrast, reinforcement learning (RL) offers a data-efficient and adaptive alternative but requires access to relevant tools, careful selection of training problems, and a robust evaluation environment. We present Makora's environment and tools for reinforcement learning finetuning of frontier models and report our results from fine-tuning GPT-5 for Triton code generation. In the single-attempt setting, our fine-tuned model improves kernel correctness from 43.7% to 77.0% (+33.3 percentage points) and increases the fraction of problems outperforming TorchInductor from 14.8% to 21.8% (+7 percentage points) compared to baseline GPT-5, while exceeding prior state-of-the-art models on KernelBench. When integrated into a full coding agent, it is able to solve up to 97.4% of problems in an expanded KernelBench suite, outperforming the PyTorch TorchInductor compiler on 72.9% of problems with a geometric mean speedup of 2.12x. Our work demonstrates that targeted post-training with reinforcement learning can unlock LLM capabilities in highly specialized technical domains where traditional supervised learning is limited by data availability, opening new pathways for AI-assisted accelerator programming.

cs.DC↗

arXiv2020

Software Defect Prediction Based On Deep Learning Models: Performance Study

In recent years, defect prediction, one of the major software engineering problems, has been in the focus of researchers since it has a pivotal role in estimating software errors and faulty modules. Researchers with the goal of improving prediction accuracy have developed many models for software defect prediction. However, there are a number of critical conditions and theoretical problems in order to achieve better results. In this paper, two deep learning models, Stack Sparse Auto-Encoder (SSAE) and Deep Belief Network (DBN), are deployed to classify NASA datasets, which are unbalanced and have insufficient samples. According to the conducted experiment, the accuracy for the datasets with sufficient samples is enhanced and beside this SSAE model gains better results in comparison to DBN model in the majority of evaluation metrics.

cs.SE↗

arXiv2019

DeepRace: Finding Data Race Bugs via Deep Learning

With the proliferation of multi-core hardware, parallel programs have become ubiquitous. These programs have their own type of bugs known as concurrency bugs and among them, data race bugs have been mostly in the focus of researchers over the past decades. In fact, detecting data races is a very challenging and important task. There have been several research paths in this area with many sophisticated tools designed and utilized that focus on detecting data race at the file level. In this paper, we propose DeepRace, a novel approach toward detecting data races in the source code. We build a deep neural network model to find data races instead of creating a data race detector manually. Our model uses a one-layer convolutional neural network (CNN) with different window size to find data races method. Then we adopt the class activation map function with global average pooling to extract the weights of the last convolutional layer and backpropagate it with the input source code to extract the line of codes with a data race. Thus, the DeepRace model can detect the data race bugs on a file and line of code level. In addition, we noticed that DeepRace successfully detects several buggy lines of code at different locations of the file. We tested the model with OpenMP and POSIX source code datasets which consist of more than 5000 and 8000 source code files respectively. We were able to successfully classify buggy source code files and achieve accuracies ranging from 81% and 86%. We also measured the performance of detecting and visualizing the data race at the line of code levels and our model achieved promising results. We only had a small number of false positives and false, ranging from 1 to 10. Furthermore, we used the intersection of union to measure the accuracy of the buggy lines of code, our model achieved promising results of 66 percent.

cs.DC↗