Update: AMD ships ROCm.ai; vLLM, SGLang, PyTorch run day-0 on MI455X
The Day-2 developer story from Advancing AI: your coding agent now reads AMD natively, an MIT kernel-tuning agent ships, and you still reach it cloud-first.
Copy markdown
vLLM, SGLang and PyTorch run day-0 on MI455X
AMD says PyTorch, Hugging Face, vLLM and SGLang are already enabled on the new MI455X — the serving stacks you'd actually deploy work on day one, not months after launch. That's the software gap that historically made Instinct a non-starter for inference.
Your coding agent now reads ROCm natively
ROCm.ai, a new AI-driven dev platform, teaches Claude Code, Codex and Cursor to reason about AMD hardware and the ROCm stack natively — so agent-written GPU code can target Instinct without you nursing the toolchain by hand.
Apex: point an agent at your AMD kernels, MIT-licensed
AMD open-sourced Apex, a pipeline that aims Claude Code, Codex or Cursor at your GPU kernels through five MCP servers, then profiles, rewrites and grades them for speedup on MI355X/MI300X/MI250X. One command runs it end-to-end, and you can export the runs as RL/SFT training data.
The catch: still cloud-first
MI455X lands in production systems through 2026, so for now you'll touch all this via cloud instances and the AMD Developer Cloud, not a card in your desktop. Local Ryzen AI Max PRO 400 dev boxes are promised “later this year.”