AMD releases ROCm HRX Loom v0.1 compiler SDK for CPU, GPU and NPU targeting

AMD has published the first tagged release of Loom, a specialized compiler and intermediate representation stack within the ROCm HRX (HIP Runtime Extended) ecosystem. The v0.1 SDK provides tools for compiling kernels and asynchronous device programs, linking model programs, and specializing them for execution across AMD's heterogeneous compute portfolio including CPUs, discrete GPUs, and neural processing units.

The release packages a compiler, runtime, Python APIs, and command-line utilities including loom-fetch for installing model programs and assets, and loom-serve for HTTP request serving. Loom currently supports AMDGPU HSA emission for native GPU execution, SPIR-V/Vulkan for portable cross-vendor deployment, and has maturing x86 and WebAssembly backends. The software is available on GitHub and installable via Python's pip package manager as hrx-loom.

Architecture and compiler targets in Loom v0.1

• Compiler IR and specialization pipeline: Loom carries source facts, target facts, configuration, correctness cases, and compiler reports through a unified authoring and execution path, enabling specialization for multiple hardware targets from shared program representations.
• AMDGPU HSA emission: Native GPU execution targets the Heterogeneous System Architecture intermediate representation for direct AMD GPU deployment within the ROCm stack.
• SPIR-V/Vulkan portability layer: The compiler can target SPIR-V with Vulkan execution for portable deployment across conformant GPU implementations beyond AMD hardware.
• Maturing CPU and WebAssembly backends: x86 and WebAssembly support is present but noted as still developing, indicating AMD's intent for broad host and edge deployment scenarios.
• Python SDK and HTTP serving infrastructure: The hrx-loom package provides Python bindings and command-line tools including loom-fetch for model asset management and loom-serve for HTTP-based inference serving.

Why it matters

Loom represents AMD's most explicit open-source effort yet to unify programming across its CPU, GPU, and NPU product lines under a single compiler toolchain, reducing fragmentation for developers building agentic AI workloads. The SPIR-V/Vulkan target in particular offers a path to reduce vendor lock-in compared to AMD-specific intermediate representations alone, though the practical portability will depend on how completely Loom's runtime abstractions can be retargeted. For the Linux AI development community, pip-installable tooling with Python APIs lowers the barrier to experimentation with AMD's heterogeneous compute stack alongside existing PyTorch and ONNX Runtime workflows.