AMD Mi300 GPUs have 192GB of VRAM *each*, or 1.5TB on a 8x GPU node 🤯 To compare, Nvidia H100 GPUs only have 80GB each. The absurd amount of memory is for your model weights, and for your long contexts (KV cache). MI300 GPUs available on the AMD Developer Cloud today at

Excited to be part of the keynote at PyTorch Conference this year! 🫶 Open-source is so important for the community 🙂

🙏Its a team effort !!! And we are committed to making ROCm delightful to use.

Congratulation Waleed Atallah and Mohamed Abdelfattah !!! Looking forward on what you are building on AMD. AI at AMD

Come join the SGlang x AMD meetup on Friday Aug 22 @ Shack15 in SF.

NEW: Run Hugging Face models locally on your AMD Ryzen AI and Radeon PCs, with Lemonade! 🍋 + 🤗 = 💻🧠 GG Ramine Roane Vaibhav (VB) Srivastav AI at AMD

Awesome!!! Run Hugging Face models locally via AMD Lemonade. AI at AMD

RCCL now also has both single node and multi-node public CIs. More perf tests on the way. Lots of good things ahead for RCCL.

Two papers accepted in EMNLP 2025, one on LLM agent as research assistant and another a Tic-Tac-Toe style games for benchmark reasoning ability, proud of AMD team!!! AI at AMD arxiv.org/abs/2501.04227 arxiv.org/abs/2506.10209

At AMD , we’re building LLMs, vision-language, and text-to-video models that run on open software and fully open. Watch here → youtu.be/HgmVn7_bcgQ Try these models: Instella-3B LLM: bit.ly/418KZkZ Nitro Diffusion (Text-to-Image): bit.ly/3UY7UMe

.Anush Elangovan & Dylan Patel making dreams come true for developers at the SGLang and AI at AMD meetup!

Thank you SGLang team for the great meetup.

Grok 2.5 Day 0 support on AMD MI300!!! AI at AMD linkedin.com/posts/andyluo7…

will get even better with pip install

I kinda love these new algos that can be used for *both* RL training and (training-free) test-time reasoning - in this case, reranking. YES

Got to sit down with my boy Jeff Tatarchuk to talk about automated kernel gen Check it out

🚀🚀🚀

Awesome!!!

The next AMD Dev challenge is online. Multi-GPU kernels. The last dev challenge got us more kernel tokens for training kernel LLMs than what was available on the internet. Open Science, Open Models, Open Kernels FTW.

scraping some PyTorch Nightly UT logs: (Not done yet.. so WIP) And yes our commitment to the Quality and Performance of PyTorch on ROCm is unconditional and if there are gaps, fixed.. they will be.