DeepSeek 把 DSpark 开源自带进 vLLM，跑 DeepSeek-V4 实测单卡 250 token/s，比 MTP 快 12-42%，想搞投机解码的可以试试。
DeepSeek 的 DSpark 投机解码技术现已原生集成到 vLLM 推理框架中。DSpark 是一种半自回归草稿模型，通过非因果滑动窗口注意力并行生成多个 token，单次验证即可保持输出一致，减少解码步数。在 NVIDIA 8×B300 GPU 上，DeepSeek-V4-Pro-DSpark 在 batch size 1 时达到约 250 tokens/s，平均接受长度约 5，且在不同草稿深度下比 MTP 方法接受率高 12-42%。vLLM 通过复用 SparseMLA 后端、捕获完整草稿主干和采样循环到单一 CUDA graph，并支持前缀缓存和 FP8 KV cache。
🚀 @deepseek_ai's DSpark speculative decoding now r…
🚀 @deepseek_ai's DSpark speculative decoding now runs natively in vLLM!
What it is: a semi-autoregressive drafter that proposes several tokens in parallel with non-causal sliding-window attention, then verifies them in a single pass. Output stays identical, decoding takes fewer steps.
How vLLM runs it: it reuses the existing SparseMLA backends instead of custom attention kernels, captures the full draft backbone and sampling loop in one CUDA graph, and works with prefix caching and FP8 KV cache.
Performance on DeepSeek-V4-Pro-DSpark (verified on NVIDIA 8×B300 GPUs): - ~250 tokens/s at batch size 1 - average acceptance length ~5 - 12-42% higher acceptance than MTP across draft depths
Run with vLLM nightly today: vllm serve deepseek-ai/DeepSeek-V4-Pro-DSpark -tp 8 --trust-remote-code --kv-cache-dtype fp8 --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'
DSpark Core PR: https://t.co/Yyo6rYX4jG
Thanks @deepseek_ai for open-sourcing DSpark, and to @NVIDIAAI and the vLLM community for landing it! 🙏