AMD released DFlash on MI355X, saying it enables up to 5.02× faster single-request throughput for Qwen3.5-27B inference. It is the company's first speculative decoding update since the introduction of MTP in recent models.
AMD reported 396 tokens per second throughput for Qwen3.5-27B on MI355X, measured on HumanEval with concurrency 1. That compares with 245 tokens per second for MTP.
DFlash is built on ROCm and targets high-performance inference for large language models. Availability begins with vLLM 0.22.1rc1.dev43+g8c3cc98cf, initially for developers and researchers.
"Speculative decoding pairs a small draft model with the large target model," said Hang Yang, lead researcher. The drafter proposes several tokens, the target verifies them in one parallel forward pass, and every token that matches what the target would have produced is committed for free.
The announcement follows AMD's focus on improving AI inference efficiency. AMD said the results show that speculation and quantization stack, with an mxfp4 target adding another ~10–20% on top for free.
AMD did not say how DFlash performs on other models, and raised the open question of whether quantization degrades draft acceptance. The company said it will continue exploring ways to enhance speculative decoding.
Source: amd
