AMD ROCm 10.1 Adds Async File I/O, NUMA Memory, and New CLI

AMD has released ROCm 10.1, a version that targets data-movement bottlenecks in AI and HPC workloads. The update introduces three enhancements to the hipFILE interface: an asynchronous fast-path backend, a batch I/O API, and multi-tier I/O statistics. The asynchronous fast-path runs hipFILE reads and writes directly on a HIP stream, bypassing host-memory staging. The batch I/O API submits multiple file requests together and dispatches them across worker threads. Multi-tier I/O statistics provide richer telemetry to ROCm profiling tools.

Memory and Profiling Updates

The HIP runtime now includes NUMA-aware host memory allocation, letting applications allocate memory on a specific CPU node. This enables zero-copy, NUMA-local pipelines for libraries such as RCCL. On the profiling side, ROCprofiler-SDK adds beta kernel replay, which re-executes each dispatch in place and restores device memory between passes, allowing full counter sets to be collected in a single run. AI framework profilers can now start and stop at any point during a run. The ROCm Compute Profiler roofline report is now an interactive HTML page with per-kernel isolation and axis switching. PC sampling now reports per-kernel sample and stall counts down to individual instructions. ROCm Systems Profiler supports Gorgon Point 1-3 APUs (gfx1150, gfx1152, gfx1153) on Linux, and ROCm Optiq visualizes rocprofv3 data alongside Systems and Compute Profiler traces.

New CLI and Tooling

ROCm CLI v0.2.0 is a single-binary tool for installing, configuring, and running local AI workloads. It supports automatic detection of flash-attn and amd-aiter wheels for vLLM on both ROCm 10.0 and 10.1. The install command includes a flag for fully non-interactive installs suitable for CI scripts. The CLI provides a full-screen TUI dashboard showing live GPU telemetry, model serving, and chat. It is available on Linux, Windows, and WSL2. AMD also added two new skills to its AMD Skills package for AI coding agents: quark-install validates AMD Quark with a matching PyTorch build and checks imports, kernels, and quark-cli; quark-torch-llm-ptq performs post-training quantization on PyTorch and Hugging Face LLMs, supporting FP8, INT4, and other formats. A new PC sampling agent skill guides AI assistants through rocprofv3 PC sampling and hotspot reporting.

Compiler and Library Changes

ROCm 10.1 moves the compiler stack to LLVM 24, with amdclang++ upgrading from LLVM 23. LLVM 24 brings reduced rebuild times and newer C++ language support. Incremental builds are faster by caching LTO partitions of HIP device code. Upgrading from ROCm 10.0 requires code review because the clang_major macro now reports 24. Composable Kernel adds more quantized matrix-multiply types and speeds up attention on Radeon GPUs. MIGraphX switches its backend compiler from rocMLIR to rocMLIRTriton, improving inference performance and compile speed. rocSPARSE can now perform triangular solves directly on ELLPACK (ELL) format matrices.

Platform and Monitoring

ROCm SMI is deprecated, with amd-smi replacing it as the unified monitoring tool for AMD GPUs, CPUs, and APUs. amd-smi offers a subcommand-driven CLI, a native C library, and bindings for Python, Rust, and Go. It reports container IDs for Docker and Kubernetes, enabling GPU process attribution in containers. WSL2 support enters tech preview, with ROCm packages installing directly in the guest without manual patches. GPU virtualization via paravirtualization allows PyTorch, HIP, and AI workloads to run inside WSL2. KVM SR-IOV support extends to Ubuntu 26.04 LTS for both host and guest, using driver version 9.3.0.K. The hipThreads library brings standard C++ threading concepts to heterogeneous device programming, mapping CPU threading APIs onto GPU execution on Linux and Windows for incremental acceleration.