Azalia Mirhoseini
@Azaliamirh
Founder @RicursiveAI, Asst. Prof. of CS at Stanford. Prev: DeepMind, Anthropic, Brain. Co-Creator of MoEs, AlphaChip, Test Time Scaling.
加入 May 2013
626 正在關注 20.3K 粉絲
Check out this new synthetic data / RL approach for writing AMD HIP (a low-resource language) kernels!
Correctness went from 6% to 60% on KernelBench L2!
LLMs are good at CUDA because the internet is full of it. But a model that gives you highly optimized CUDA may still struggle to write compilable HIP.
We built a synthetic data pipeline with multi-agent search and post-trained a 14B open-source model with SFT + GRPO RL, leading to substantially better HIP compilation + correctness rates on AMD MI350X GPUs.
顯示更多
