AMD's hrx-loom stack: agent-written kernels beat llama.cpp across 51 LLMs on Strix Halo
jfowers_amd · reddit · 2026-11-14
An AMD AI Group employee showcased hrx-loom, an experimental end-to-end stack (compiler, runtime, serving tools) that lets anyone write, optimize, serve, and share model kernels without modifying the inference engine or server. Kernels are written in .loom files, a vector programming model reminiscent of MLIR that is complex for humans but easy for agents.
Key results:
- The author built an agent skill that generates kernels for any LLM's Q4KM GGUF, consistently faster than llama.cpp Vulkan and ROCm on Strix Halo across decode and prefill metrics
- Ran the skill over the top-51 trending Hugging Face models fitting on Strix Halo (19 model families): 100% agent-ported successfully, 449 measurements show median prefill speedups of 1.44×–1.50× and decode gains of 1.14×–1.19×
- Tool calling, speculative decoding, and image input all work; model programs are published to a Hugging Face collection
- Usable on Linux/Windows via loom-fetch/loom-serve from PyPI, or port your own models with the /loom skill
The team is working on an hrx-loom backend for llama.cpp. The SDK has no dependencies, builds from source in a minute, and targets x86, AMD GPU/NPU, and generic Vulkan devices.
More from coding & agent
- Nursa's accountant and retired parole officer built real tools with AI — every · 2026-11-14
- We don't need more chatbots: an agent harness lead's field notes — threepointone · 2026-11-14
- Nexus launches an AI context engine that turns deployment knowledge into agent-queryable graphs — rohanpaul_ai · 2026-11-14
- Developer now runs 100% of his agents locally on a Mac Mini, up from 25% earlier this year — josh_bickett · 2026-11-14
- Mediabunny ships 400 fixes from an LLM-assisted bug hunt across its 20M-download codebase — Vjeux · 2026-11-14
- Perplexity Teammates screenshot shows agents on Datadog duty, reading PRs overnight — testingcatalog · 2026-11-14