EAGLE-4 speculative decoding lands on AMD Instinct MI355X at day 0
AMD and EagleCorp brought EAGLE-4 to Instinct MI355X GPUs on day 0; the harness profiles drafting and verification costs at startup and picks the best guess tree, reaching ~300 tok/s on Qwen3.8-27B and 1.6x over built-in MTP.
How do you make speculative decoding fast at any batch size? Let the drafter measure the hardware. Working with @EagleCorp, we've brought EAGLE-4 to AMD Instinct MI355X GPUs on day 0. EAGLE-Harness profiles drafting and verification costs at startup, then dynamically selects the optimal guess tree for the workload. On Qwen3.8-27B (BF16, single GPU): ⚡ ~300 tok/s single request ⚡ 1.6× faster than built-in MTP ⚡ Performance holds from 1-32 concurrent conversations ⚡ Lossless vs. plain decoding, including Gated DeltaNet layers via FlyDSL tree verification Details + launch scripts for SGLang and ATOM: https://www.amd.com/en/developer/resources/technical-articles/2026/eagle-4-on-amd-instinct-gpus.html