AI 基礎設施
ROCm 10 Brings Agentic Kernel Optimization to General Availability, but the Upgrade Also Introduces ABI Breakage and Known Issues
AMD has released ROCm 10 and made ROCm.AI—integrating Hyperloom, AMD Skills, and a unified CLI—generally available. The company claims average inference and training performance gains of 3.3× and 2.4× on the same hardware, but those figures cover the entire optimization workflow and should not be interpreted as gains from simply upgrading the runtime.
AMD released ROCm 10 on August 27. The key change is not merely the major version bump, but ROCm.AI’s move from preview to general availability. This development layer consists of Hyperloom, AMD Skills, and the ROCm CLI. Hyperloom first uses tracing and roofline analysis to identify bottlenecks in host code, memory scheduling, or GPU kernels, then searches for, implements, and validates optimizations. It currently supports vLLM and SGLang workloads and can generate modifications using HIP, Triton, or FlyDSL. Skills brings AMD’s debugging, deployment, and performance workflows into Codex, Claude Code, and Cursor.
The ROCm CLI, which remains in technical preview, can create and switch among multiple managed ROCm environments on Windows and Linux. It can also run model-serving, diagnostic, and rollback operations. AMD claims that “systems configured with ROCm.AI” deliver average inference and training performance improvements of 3.3× and 2.4×, respectively, compared with ROCm 7 on the same hardware. These are system-level results combining kernel, memory, and scheduling optimizations; they do not mean existing applications will automatically receive equivalent acceleration after a package upgrade.
The underlying Core SDK has moved to the TheRock build system and expanded support for Instinct virtualization, HIP APIs, profiling, and communication libraries. The upgrade risks are also concrete: AMD SMI’s SONAME has been raised to `libamd_smi.so.27`, and some data fields and APIs are incompatible, requiring monitoring applications that depend on the old ABI to be recompiled. AMD also lists potential training-throughput regressions on MI350X, possible GPU resets with PyTorch on certain Radeon GPUs, potential failures in SGLang’s AITER backend, and possible crashes when running vLLM or ComfyUI on Ryzen AI. Engineering teams should rerun correctness, throughput, and stability tests using their actual models and sequence lengths before deciding whether agent-generated optimizations are suitable for merging into production branches.