hip

ROCmFPX's gated delta net prefill kernel was optimized to process 16-token chunks via RDNA3 WMMA instead of one token at a time, cutting latency from 772µs to 610µs per 512-token call and improving prompt-processing throughput by ~2-4% across configurations, with Strix Halo testing ongoing.

GNU Guix 正式将 AMD 的 HIP 与 ROCm 生态作为一等公民支持，使在 Guix 中进行 AMD GPU 计算栈的功能化打包与部署更加顺畅。

AMD 发布 ROCm 10.1，重点突破数据搬运瓶颈，通过 AMD Infinity Storage (hipFile)、HIP 运行时的 NUMA 感知内存分配以及 ROCm CLI 和 AMD Skills 等开发者工具提升 GPU 数据传输与训练推理效率。

The article introduces AMDKernelVault, an open dataset and training framework for AMD GPU kernel optimization, featuring large-scale HIP and Triton kernels and agent-driven pipelines for generating and validating kernels.

This tool provides a reproducible Windows setup to run CUDA-targeted applications on AMD GPUs using ZLUDA and AMD HIP/ROCm, validated with RX 9060 XT and LibTorch workloads.

This pull request introduces Flash Attention tuning optimizations for CUDA/HIP in the llama.cpp project, targeting gfx1201 hardware to enhance inference performance.

A fork of llama.cpp has been created and optimized for AMD GFX906 GPUs, improving performance on Mi50, Mi60, Radeon VII, and GCN HIP for machine learning and LLM applications.

A multi-agent synthetic data pipeline with SFT and GRPO RL post-training improves HIP compilation and correctness on AMD MI350X GPUs for a 14B open-source model.