Uno in vLLM: An Independent Implementation and Empirical Serving Study

Adapter-based speculative decoding gives a language model a parallel drafting path while retaining its original weights for verification. Turning that method into useful serving throughput requires the drafter to participate in the inference engine's cache, scheduler and execution lifecycle. We present a native implementation of Uno in vLLM and an independent replication of its serving benefit. On an NVIDIA H100 with Qwen3-8B in bfloat16, a controlled math workload measures 2.58 times plain-vLLM throughput at one concurrent request, 2.20 times at eight and 1.74 times at 32. Throughput remains within 3% of the SGLang reference at all three concurrency levels, reproducing its scaling pattern in a second engine. With production-oriented settings enabled, the vLLM implementation's reported mean throughput slightly exceeds the reference at eight, 24 and 32 streams, reaching 7,657 tokens per second against 7,505 at 32. We describe the execution contracts that make the integration effective: conditional adapter routing, complete draft key/value state, asynchronous scheduling, graph replay, overlapping adapter computation and seed-row reuse. A separate within-vLLM ablation reduces warm first-token latency by 25% by increasing the API-process count from one to two. Additional recorded configurations span NVIDIA Ampere, Hopper and Blackwell, alongside fused-drafting research. The work contributes a serving implementation, replicated acceleration, reusable engineering findings and a reproducible configuration for further Uno research.

This record includes the research paper and a companion archive containing the evidence inventory, selected measurement data, figure assets and figure-generation script.