strix-llm

strix-llm is a known-good local LLM inference configuration tool for the Framework Desktop (Ryzen AI MAX+ 395 / Strix Halo). Maintained by community member Alberto Migliorato (GitHub: AlbeMiglio), it provides a tested, reproducible setup for running large language models locally on Framework Desktop hardware using ROCm and llama.cpp¹.

Running a 70B model on Strix Halo is feasible — up to 96 GB of the 128 GB unified memory pool can be allocated to the Radeon 8060S iGPU, which is enough for Llama-70B at usable speeds. However, getting there requires hand-tuning BIOS VRAM/GTT splits, kernel parameters, ROCm versions, and llama.cpp build flags. strix-llm is the layer that makes that path reproducible and maintained, so you can go from a fresh Framework Desktop to a running model with minimal manual configuration¹.

What It Does

strix-llm is not another inference engine. It sits on top of ROCm and llama.cpp and does the integration work¹:

Command	Description
strix-llm doctor	Detects hardware and checks configuration (VRAM/GTT split, ROCm, kernel, llama.cpp build) against a known-good baseline; tells you exactly what to fix
strix-llm list	Shows curated model + quantization presets sized to the unified-memory budget
strix-llm run <preset>	Prints the exact llama.cpp command to launch a model with sane defaults
strix-llm bench --parse-file <log>	Measures tokens/s, max context, and power draw from a saved llama.cpp log

Without --model, run prints the command it would launch and where to get the model, so you see exactly what will happen before downloading tens of GB¹.

Installation

pipx install git+https://github.com/AlbeMiglio/strix-llm
# or: pip install git+https://github.com/AlbeMiglio/strix-llm

Usage

strix-llm doctor                          # check the machine and the local setup
strix-llm list                            # show curated model presets
strix-llm run llama-3.3-70b               # print the exact llama.cpp command to launch
strix-llm run llama-3.3-70b --model ~/models/llama-3.3-70b-q4.gguf
strix-llm bench --parse-file run.log      # tokens/sec from a saved llama.cpp log

Example doctor output on a configured Framework Desktop¹:

strix-llm doctor

  [ok] os: Linux
  [ok] cpu: AMD Ryzen AI Max+ 395 w/ Radeon 8060S
  [ok] memory: 119 GiB unified
  [ok] vram-split: 96 GiB allocated to the iGPU
  [ok] rocm: ROCm 6.2.0
  [ok] llama.cpp: /usr/bin/llama-server (ROCm build)

strix-llm doctor --json prints the same checks in machine-readable form¹.

Target Hardware

• Framework Desktop / Ryzen AI MAX+ 395 (Strix Halo), 128 GB unified memory
• Radeon 8060S iGPU, ROCm on Linux¹

Roadmap

• doctor: hardware, ROCm, and config detection with remediation
• Model + quant presets and one-command launch (list / run)
• Benchmark parsing of llama.cpp timings
• Known-good baseline configs (BIOS notes, kernel params, ROCm pinning)
• Published benchmark results measured on real hardware
• Quantization / KV-cache tuning for unified memory¹

Technical Details

Detail	Value
Language	Python
License	MIT
Stars	0
Commits	4
Maintainer	Alberto Migliorato (@AlbeMiglio)
Dependencies	No external dependencies (built test-first)
Status	Early work in progress
Last updated	June 15, 2026

Related Projects

Project	Description
Strix Halo Guide	Comprehensive local LLM setup and benchmark guide for Strix Halo
framework-rocm-containers	Reproducible ROCm containers for Framework Desktop (PyTorch/JAX)

Footnotes

1. albemiglio/strix-llm — GitHub ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸