Hugging Face has begun testing the Transformers library on the AMD MI455X, the latest datacenter GPU for AI introduced at Advancing AI 2026. The MI455X features 432 GB of high-bandwidth memory, a significant jump from the 192 GB in the previous MI300 generation. The company aims to ensure its ecosystem works seamlessly across AMD hardware and fully utilizes each platform. This testing comes after Transformers already runs daily CI on MI300 infrastructure, with plans to expand coverage to Ryzen and Radeon systems. The collaboration includes working with the AITER team to bring optimized AMD kernels to the Hugging Face Hub for direct use in Transformers workflows. Hugging Face expressed excitement about the MI455X and used early access to conduct preliminary tests. The results highlight the potential of the MI455X for handling larger models and long-context workloads efficiently. The 432 GB memory capacity allows for more concurrent requests, with the MI455X supporting over three times more than the MI300. This was demonstrated through a capacity study using a 64 GB Qwen3-32B BF16 model, where the MI455X showed a significantly higher KV cache capacity. The company plans to follow up with speed comparisons as testing progresses. Hugging Face also mentioned the importance of the test suite for 24 key model architectures in assessing the health of the broader ecosystem. The MI455X achieved a 99.5% success rate across these models, similar to the MI300 and NVIDIA A10. This success required collaboration across the stack, including enabling a stable Flash Attention path, adding torchcodec support, and resolving output-comparison issues. The company intends to add MI455 to Transformers CI and continue expanding validation across more Hugging Face libraries. The AMD team provided early access and close collaboration, enabling these results to be available promptly.
Source: huggingface
