What changed
AMD's Quark team has trained an EAGLE3.1 draft model for the large mixture-of-experts (MoE) target model, amd/MiniMax-M3-MXFP4, entirely on AMD Instinct MI350X GPUs using a ROCm-native stack of ROCm, vLLM, and TorchSpec. The draft model is designed for speculative decoding, which allows the model to propose multiple tokens at once, verified by the target model in a single forward pass. The training pipeline includes on-policy data generation, hidden-state extraction, cold-start training with the training-time-test (TTT) objective, export to a vLLM-compatible checkpoint, and evaluation using the SPEED-Bench benchmark.
What this means in practice
The trained EAGLE3.1 draft model can be deployed with vLLM for speculative decoding, improving the efficiency of large language model inference. The model is trained from scratch and is aligned with the target model's internal representations, making its proposals easier for the verifier to accept. The acceptance length (AL) metric, which measures the average number of tokens emitted per target model verification step, is used to evaluate the draft's quality. An AL of 1.0 indicates no acceleration beyond the guaranteed token, while a higher AL means the target confirms more tokens per verification step.
Limitations and availability
The training process requires an 8-GPU AMD Instinct MI350X node, with some GPUs hosting vLLM target engines that produce hidden states and others running the FSDP2 draft trainer. AMD Quark 0.13 provides user-friendly YAML recipes and a CLI surface for EAGLE3 training, with the Qwen3-8B Quick Start being a fully GPU-validated, reproducible path. The EAGLE-3 Large-Model Best Recipe uses amd/MiniMax-M3-MXFP4 as an example target for adapting the workflow to large, quantized models. The supplied announcement does not specify the exact performance benchmarks or independent testing results.
