vLLM Integrates DeepSeek DSpark for Faster Inference

The open-source model serving engine vLLM has merged DeepSeek's DSpark speculative decoding framework into its main branch under the pull request #47808. Accessible via the "enable_adaptive_verification" flag, this integration introduces adaptive verification to solve a long-standing challenge in speculative decoding. Traditionally, developers had to manually tune a static draft token length, balancing speed at low concurrency against wasted compute at high concurrency.

DSpark bypasses this trade-off by using a confidence head to score the survival probability of each draft token. Instead of committing to a fixed draft length, vLLM dynamically allocates verification slots to the highest-scoring tokens across an entire batch. On the DeepSeek-V4-Pro-0813 model, the acceptance rate decays steeply, with the first token of a seven-token draft surviving verification over 70 percent of the time, while the seventh token survives less than 10 percent.

With this update, a single configuration of "num_speculative_tokens: 7" maintains the Pareto frontier across all concurrencies from 1 to 256. This benchmark was achieved using eight B300 GPUs. However, the feature currently has several limitations. It requires NVIDIA SM100 hardware, specifically B300 GPUs, and does not support LoRA, pipeline parallelism, output logprobs, or eager mode.

For practitioners running DeepSeek-V4 at scale, this integration eliminates the need to constantly benchmark traffic shapes and re-tune static parameters when workloads shift. Because DSpark is fully integrated into vLLM as a free, open-source feature, users do not need to download a separate draft model to deploy it. This simplifies the deployment pipeline while maximizing hardware utilization under fluctuating user demands.