Day 0 Developer Guide: Running the Latest Open Models from OpenAI on AMD AI Hardware — ROCm Blogs

[giscus[bot]]

Day 0 Developer Guide: Running the Latest Open Models from OpenAI on AMD AI Hardware — ROCm Blogs

Day 0 support across our AI hardware ecosystem from our flagship AMD InstinctTM MI355X and MI300X GPUs, AMD Radeon™ AI PRO R700 GPUs and AMD Ryzen™ AI Processors

https://rocm.blogs.amd.com/ecosystems-and-partners/openai-day-0/README.html

[ehartford]

What about mi210

[hongxiayang]

The mi300 docker also works for mi210. You can give it a try.

[rajwanur]

The link in the following line is broken in the article -

See how to run it today and learn more: How To Run OpenAI’s GPT-OSS 20B and 120B Models on AMD Ryzen™ AI Processors and Radeon™ Graphics Cards

[anshulgupta4-del]

https://www.amd.com/en/blogs/2025/how-to-run-openai-gpt-oss-20b-120b-models-on-amd-ryzen-ai-radeon.html

[brawlspecials]

This looks awesome. I like it. Thanks for this valuable information..

[medioxor]

begging you to support ROCm/vllm#633

[anshulgupta4-del]

“No need to beg Medioxor. We’ll support for our question— our approach is always developer-first. Could you please elaborate on your question?”

[medioxor]

Given that Strix Halo with 128gb of unified memory is the most affordable hardware to run and fine tune decently sized models I assume that a significant amount of developers will require vllm support. As of right now there is no vllm support and so using libraries such as Unsloth, Verifiers, etc for Reinforcement Learning is not possible. While it's great to be able to run models, without fine tuning capabilities it's hard to apply these models to specific use cases.

[anshulgupta4-del]

Thanks for your interest in ROCm on Radeon and Ryzen! Our current focus is building a highly optimized inference stack to support key AI workloads. While fine-tuning and training are not available yet, they’re part of our long-term roadmap—and we’ll keep you updated as those capabilities roll out.
Also, we’re excited to share that vLLM is already supported on our discrete Radeon GPUs. For Strix Halo, official ROCm support is coming in Q3 2025, with ongoing feature expansion to give developers more flexibility and performance.
Stay tuned—there’s more to come!

[hieule88]

How to run accuracy evaluation in MI300. I'm using docker image rocm/vllm-dev:open-mi300-08052025, try to use gpt_oss.evals as in https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html#accuracy-evaluation-panels but find no module evals

[hieule88]

solved

[bmsrank2-sketch]

The release of OpenAI’s open models with AMD’s seamless support shows how technology is evolving to give developers more flexibility and efficiency. Similarly, parle italian wordle is a free online word game site offering unlimited Wordle puzzles, multi-language support, and varied word-length modes for fun and learning. Both highlight how open access and diverse options can enhance creativity and user experience in different fields.