# AMD ROCm 10.2.0 ships vLLM 0.32.0, SGLang 0.5.21, and Hyperloom for Radeon AI Pro

> **Key takeaway:** ROCm 10.2.0 refreshes its AI serving stack with vLLM 0.32.0 and SGLang 0.5.21 on Python 3.14, adds Ryzen support for 18 new processors, and extends Hyperloom tuning to single-node Radeon AI Pro R9700 GPUs.

**Published:** 2026-11-19T02:02:34.587653+00:00 
**Source:** AMD ROCm Official Blog 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/amd-rocm-1020-ships-vllm-0320-sglang-0521-and-hyperloom-for-.html 

## Summary
AMD has released ROCm 10.2.0, continuing its six-week cadence with framework updates and tooling improvements aimed at stability rather than disruptive changes. The release brings serving builds to vLLM 0.32.0 and SGLang 0.5.21, both shipping in Docker images and Python wheels with Python 3.14 support, while prebuilt AITER kernels now cover Instinct MI300X through MI355X and compile automatically on Radeon and Ryzen AI hardware. PyTorch 2.15 wheels arrive alongside JAX 0.11.2, MIGraphX 2.19, and ONNX Runtime 1.29, with a notable packaging change: TheRock-built PyTorch now installs ROCm as a separate dependency rather than bundling it.

On the deployment side, ROCm InferaSim adds configuration projection for serving setups, while Hyperloom gains power-budget targeting and early access for single-node Radeon AI Pro R9700 deployments. RCCL receives additional tuning for eight-GPU single-node Radeon PCIe servers validated with four-way tensor-parallel vLLM. ROCm CLI v0.3.0 now defaults to installing ROCm 10.2.0, and AMD Skills expands to 12 projects with three new additions for FlyDSL kernel work and CUDA-to-ROCm porting via Moat.

## Framework and tooling changes in ROCm 10.2.0
- **Serving stack refresh:** vLLM moves to 0.32.0 and SGLang to 0.5.21, both available in Docker images and Python wheels, with Python 3.14 support across the board. AITER integration is now automatic on Radeon GPUs and Ryzen AI processors without manual installation.
- **PyTorch packaging change:** PyTorch 2.15 wheels built by TheRock now install ROCm as a separate dependency rather than bundling it internally, which affects dependency management for existing deployments.
- **Hyperloom expansion:** The autonomous inference optimization agent adds power-budget controls and early access support for single-node Radeon AI Pro R9700 GPUs, extending beyond Instinct data-center hardware.
- **hipFile and Infinity Storage:** Filesystem support expands to logical volumes (LVMs) and software RAID (mdraid), building on the ROCm 10.1 storage foundation.

## Why it matters
The ROCm 10.2.0 release strengthens AMD's position as the primary open-source alternative to NVIDIA's CUDA ecosystem for AI/ML workloads, with particular significance for developers seeking to deploy on consumer and workstation hardware rather than datacenter-only GPUs. The automatic AITER compilation on Radeon and Ryzen AI, combined with 18 newly supported Ryzen processors, lowers barriers for local AI development on laptops and desktops, while the Hyperloom extension to Radeon AI Pro R9700 signals AMD's intent to compete in workstation inference optimization.