AMD ROCm 10.1.0: PyTorch 2.14, vLLM 0.29, Radeon AI PRO R9600 support

AMD published the ROCm Core SDK 10.1.0 release notes on 2026-10-05. They apply to Linux and Windows. The release adds support for the Radeon AI PRO R9600 GPU. It also adds Ubuntu 26.04.1 (kernel 7.0 GA) for Instinct, Radeon and Ryzen GPUs. Supported frameworks move to PyTorch 2.14.0, JAX 0.11.1, vLLM 0.29.0, SGLang 0.5.18 and MIGraphX 2.18. These replace PyTorch 2.11.0, JAX 0.10.0, vLLM 0.27.0, SGLang 0.5.15 and MIGraphX 2.17. Support for TensorFlow 2.19.1 is dropped. The AI ecosystem table lists vLLM 0.29.0 for gfx950, gfx942, gfx1200, gfx1201, gfx1100-gfx1102 and gfx1150-gfx1152, on Linux only, with Python 3.14 and "requires PyTorch 2.13.0". The hardware table maps the consumer targets to RX 9000 and AI PRO R9000 cards (gfx1200/1201), RX 7000 cards (gfx1100-1102), Strix Halo and Gorgon Halo (gfx1151), and Strix Point, Krackan Point and Gorgon Point APUs (gfx1150, gfx1152). The gfx1153 Gorgon Point parts are not in the vLLM row. The PyTorch row lists 2.14.0 for Linux and Windows, and 2.13.0 and 2.12.0 for Instinct GPUs only, so it conflicts with the vLLM row for Radeon and Ryzen. hipFile adds an async fastpath for AMD Infinity Storage that moves data from storage to the GPU without host bounce-buffer copies, plus batch I/O and stats APIs. For Windows, the notes list the Adrenalin Edition 26.10.41.05 driver. AMD's driver notes, also dated October 5, list it for Windows 10 and 11, but ROCm's OS table lists only Windows 11 25H2 for Radeon and Ryzen. WSL2 support is a technical preview. ROCm SMI is no longer built or installed in the standard build; AMD points to `amd-smi`. GPU resets or crashes in PyTorch training and fine-tuning with Llama-Factory or Unsloth on some Radeon GPUs, such as the RX 9070 Series and AI PRO R9700, are listed as resolved.