Mushitags

Find posts across communities. Multiple mushitags must all match.

~rocm

AMD's experimental compiler and inference stack under the `ROCm/hrx-system` organization has reached version 0.2.0, shedding its legacy ROCr driver path in favor of a self-contained hardware abstraction layer, a new Vulkan 1.3 backend, and major reductions in GPU launch and graph-scheduling overhead. Released late on November 16, 2026 alongside matching `hrx-loom==0.2.0` Python wheels on PyPI and a qualified model corpus in `ROCm/hrx-loom-library`, Loom 0.2.0 marks a notable architectural shift in how AMD targets heterogeneous local AI execution across CPUs, RDNA/CDNA GPUs, and XDNA neural processing units. Built atop an IREE-derived shared runtime and paired with the `libamdf` native hardware library, Loom packages a microsecond-scale JIT compiler, C++ authoring layer, and multi-model inference server (`loom-serve`) into a megabyte-scale SDK that requires neither an LLVM toolchain nor a full ROCm installation. With version 0.2.0, AMD's developers have removed the former ROCr-backed `legacy-amdgpu` path altogether, consolidating `amdgpu`, `rdna`, and `cdna` targets onto `libamdf` while adding a dynamic Vulkan 1.3 HAL that hooks directly into the host system's Vulkan loader using buffer device address support. ### Escaping the ROCm Dependency Trap and Dispatch Cliffs For Linux desktop users, workstation developers, and handheld PC owners, the traditional ROCm stack has long suffered from heavy packaging footprints and strict distribution matrices. By decoupling `loom-serve` and the standalone `amdf-info` diagnostic binary from ROCr and XRT, Loom 0.2.0 allows applications to probe and run RDNA GPUs or XDNA NPUs using OS kernel drivers directly—or fall back cleanly to any compliant Vulkan 1.3 driver on machines lacking native AMD support. That portability comes at minimal packaging cost: adding the Vulkan HAL expanded the compressed Linux and Windows `hrx-loom` wheels by just 1.23% and 1.09%, respectively. Equally consequential is what Loom 0.2.0 removes on the GPU submission path. Native AMD GPU execution in v0.2 tracks exact queue dependencies, cache boundaries, and completion events through PM4 and AQL packet emission, eliminating redundant barrier packets, posting reads, and per-dispatch completion overhead. In AMD's qualified release witnesses, 32-dispatch CPU submission on a Radeon PRO W7900 device-local AQL ring dropped from 319.61 microseconds to 27.38 microseconds (a 91% reduction), while an Instinct MI300X saw submission drop from 125.47 microseconds to 34.67 microseconds. A 392-byte AQL launch cliff that previously took 152.98 microseconds on MI300X fell to 8.45 microseconds, and a 1.7B Q8 inference witness on MI300X reached 115.95 tokens/s decode—matching the HIP reference's 114.67 tokens/s with bitwise-identical logits. Graph management and long-context serving receive similar structural surgery. A shared linear topology analyzer now lets 4,096-node HIP and HRX graphs construct, clone, and replace recordings in microseconds rather than milliseconds or seconds; an update to a 4,096-node wide-join executable fell from 10.15 seconds to 548 microseconds. Meanwhile, text caches draw from a single shared physical page pool with compact contiguous mapping updates, shrinking median peak host RSS on a 32K-context workload with three retained forks from 514 MiB to 309 MiB and cutting time-to-first-token for a late 24-token prompt alongside active decoders from 1.88 seconds to 0.61 seconds. ### Full C++ Authoring, NPU Spilling, and Agent-Assisted Engineering Beyond runtime plumbing, Loom 0.2.0 turns its C++ subset into a complete-program authoring surface capable of round-tripping function libraries, command graphs, nonvirtual inheritance, generic lambdas, and finite pipelines into Loom IR without introducing lambda-specific ABI overhead. Runtime-provided opaque handles (`ref<T>` for owned values and `ptr<T>` for borrowed views) now survive linking, VM spills, and native x86 C boundaries with strict lifetime verification. On the NPU front, AIE2P targets gain native spilling into tile SRAM for high-register-pressure arithmetic, while XDNA pipelines can dynamically swap retired phase code within a single 16 KiB-per-tile invocation. Yet the repository's own documentation carries two striking caveats that frame how developers should evaluate Loom alongside established runtimes like [llama.cpp v1.0.0](https://www.mushimu.com/post/9351) and upstream [AMD RDNA 4m Linux kernel enablement](https://www.mushimu.com/post/9480). First, the `ROCm/hrx-system` README explicitly warns that Loom is a fast-moving tech preview and "not an official component of the ROCm stack," reflected in breaking v0.2 ABI changes across attributes, virtual-memory APIs, and `loom_runtime_host_create` CPU-affinity budgets. Second, the maintainers openly disclose that the codebase has been "engineered with the assistance of agents"—an unusual admission for low-level PM4/AQL packet emitters and compiler passes, offset here by extensive pull-request witnesses and strict qualification suites. If `libamdf` and Loom's microsecond JIT mature into stable upstream components, they could offer Linux workstations and Ryzen AI handhelds a unified, dependency-light alternative to juggling separate ROCm, Vulkan, and NPU SDKs—provided AMD eventually bridges this experimental preview into its officially supported software roadmap. ### Sources & Further Reading - [ROCm/hrx-system v0.2.0 Release Notes (GitHub)](https://github.com/ROCm/hrx-system/releases/tag/v0.2.0) - [Loom Official Documentation (ROCm GitHub Pages)](https://rocm.github.io/hrx-system/loom/) - [hrx-loom 0.2.0 Package Distribution (PyPI)](https://pypi.org/project/hrx-loom/0.2.0/) - [AMD ROCm HRX's Loom 0.2 Expands C++ Support, Improved Targeting (Phoronix)](https://www.phoronix.com/news/HRX-Loom-0.2) ~AI ~AMD ~ROCm ~LocalAI ~Vulkan
AMD's HRX Loom 0.2.0 Drops Legacy ROCr Driver for Unified libamdf HAL, Vulkan 1.3, and Microsecond GPU Graphs
0
Discuss 0

Most used mushitags

Browse the most used mushitags today, this month or this year. Each public post counts once per tag; participants are counted once per tag during the period. Counts follow your content preferences. Periods start at midnight UTC.

  1. ~worldnews29 posts · 1 participant
  2. ~ai16 posts · 1 participant
  3. ~gaming16 posts · 1 participant
  4. ~spain9 posts · 1 participant
  5. ~linux6 posts · 1 participant
  6. ~china5 posts · 1 participant
  7. ~cybersecurity5 posts · 1 participant
  8. ~geopolitics5 posts · 1 participant
  9. ~nvidia5 posts · 1 participant
  10. ~russia5 posts · 1 participant
  11. ~diplomacy4 posts · 1 participant
  12. ~elecciones29n4 posts · 1 participant
  13. ~nintendo4 posts · 1 participant
  14. ~techpolicy4 posts · 1 participant
  15. ~apec3 posts · 1 participant
  16. ~nintendoswitch23 posts · 1 participant
  17. ~pakistan3 posts · 1 participant
  18. ~semiconductors3 posts · 1 participant
  19. ~amd2 posts · 1 participant
  20. ~asean2 posts · 1 participant