Omega running Qwen3.8-27B on an AMD Instinct MI300X

Technical brief for the UXL Open Source Working Group. On 13 November 2026 a program written in Omega, my own language, ran the Qwen3.8-27B-FP8 model on an AMD Instinct MI300X with no AMD user-space runtime in its process: no HIP, no ROCr, no PyTorch. It talks to the Linux AMD compute interface (/dev/kfd) directly and runs 18 GPU kernels whose code objects my own Omega tooling wrote. Median decode time was 26.7 ms per token, measured twice that day; the second measurement followed only the public instructions. In a later session on 14 November 2026, a repaired launcher and new kernels took the same model to 16.226 ms per token; that optimized build has not landed in the public clone yet, so it carries a different evidence status from the published 13 November run. A paired run against AMD's HIP runtime on the published kernels put my launcher 0.08 ms per token behind, with dispatch cost equal on both sides. The brief carries the measurements, the correctness checks, an honest account of what the run does not show, and reproduce steps at three levels of effort: a five-minute evidence check needing no GPU, a from-source rebuild of the launcher and all 18 kernels, and a full run on an MI300X beginning with a small vector-add example.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24236688
Primary Topic
Parallel Computing and Optimization Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Omega running Qwen3.8-27B on an AMD Instinct MI300X

Drake Stapleton
Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
article

Omega running Qwen3.8-27B on an AMD Instinct MI300X

Drake Stapleton
article en

Abstract

Technical brief for the UXL Open Source Working Group. On 13 November 2026 a program written in Omega, my own language, ran the Qwen3.8-27B-FP8 model on an AMD Instinct MI300X with no AMD user-space runtime in its process: no HIP, no ROCr, no PyTorch. It talks to the Linux AMD compute interface (/dev/kfd) directly and runs 18 GPU kernels whose code objects my own Omega tooling wrote. Median decode time was 26.7 ms per token, measured twice that day; the second measurement followed only the public instructions. In a later session on 14 November 2026, a repaired launcher and new kernels took the same model to 16.226 ms per token; that optimized build has not landed in the public clone yet, so it carries a different evidence status from the published 13 November run. A paired run against AMD's HIP runtime on the published kernels put my launcher 0.08 ms per token behind, with dispatch cost equal on both sides. The brief carries the measurements, the correctness checks, an honest account of what the run does not show, and reproduce steps at three levels of effort: a five-minute evidence check needing no GPU, a from-source rebuild of the launcher and all 18 kernels, and a full run on an MI300X beginning with a small vector-add example.

Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon