DSpark
A speculative decoding algorithm released by DeepSeek, engineered specifically for the Multi-head Latent Attention (MLA) architecture with a dedicated draft strategy. It has been merged into the vLLM main branch and supports the Qwen3 and Gemma model series.
Timeline (last 90 days)
Meta provides two dedicated draft models for speculative decoding: DFlash V2 is designed for SGLang and is reportedly achieving a 3.7x throughput improvement on AMD MI355X hardware, accepting an average of 6.4 draft tokens per step; DSpark is designed for vLLM and proposes up to seven tokens per pass.
DSpark is a representative baseline method in the field of speculative decoding, belonging to one of the routes of joint optimization of the draft and the target model, and is used as a reference in JetSpec++'s comparison as the current state-of-the-art (SOTA) method.
JetSpec++ achieves a 1.54x wall-clock time improvement relative to DSpark.
ReaDy is built on top of DSpark's parallel backbone and sequence prediction head, using lightweight two-stage adaptation.
Compared with the DSpark baseline, ReaDy improves on all benchmarks, with throughput increasing by an average of 7.97% and accepted length increasing by an average of 8.50%.
MiniCPM-V 4.7 35B-A3B uses a combination of three techniques—Lookahead Expert Scheduling (LES), fine-grained layer-by-layer quantization, and DSpark speculative sampling—to accelerate on-device inference.
LFM2-VL achieves a 1.66x to 2.56x speedup at batch 1 via DSpark speculative decoding.
DFlash and DSpark have made targeted optimizations in draft generation efficiency and hardware adaptation, and may outperform classic solutions under specific models and workloads.
vLLM v0.28.0 introduces DFlash2 and DSpark confidence-scheduled verification in the speculative decoding domain.
DeepSeek's official DSpark is essentially speculative decoding/multi-token prediction technology, requiring a drafter model and two sets of KV Cache resident in high-speed memory to achieve optimal results.
1 more timeline events
All Facts (11)
Meta provides two dedicated draft models for speculative decoding: DFlash V2 is designed for SGLang and is reportedly achieving a 3.7x throughput improvement on AMD MI355X hardware, accepting an average of 6.4 draft tokens per step; DSpark is designed for vLLM and proposes up to seven tokens per pass.
50%UnverifiedDSpark is a representative baseline method in the field of speculative decoding, belonging to one of the routes of joint optimization of the draft and the target model, and is used as a reference in JetSpec++'s comparison as the current state-of-the-art (SOTA) method.
50%UnverifiedJetSpec++ achieves a 1.54x wall-clock time improvement relative to DSpark.
50%UnverifiedReaDy is built on top of DSpark's parallel backbone and sequence prediction head, using lightweight two-stage adaptation.
50%UnverifiedCompared with the DSpark baseline, ReaDy improves on all benchmarks, with throughput increasing by an average of 7.97% and accepted length increasing by an average of 8.50%.
50%UnverifiedMiniCPM-V 4.7 35B-A3B uses a combination of three techniques—Lookahead Expert Scheduling (LES), fine-grained layer-by-layer quantization, and DSpark speculative sampling—to accelerate on-device inference.
50%UnverifiedLFM2-VL achieves a 1.66x to 2.56x speedup at batch 1 via DSpark speculative decoding.
50%UnverifiedDFlash and DSpark have made targeted optimizations in draft generation efficiency and hardware adaptation, and may outperform classic solutions under specific models and workloads.
50%UnverifiedvLLM v0.28.0 introduces DFlash2 and DSpark confidence-scheduled verification in the speculative decoding domain.
50%UnverifiedDeepSeek's official DSpark is essentially speculative decoding/multi-token prediction technology, requiring a drafter model and two sets of KV Cache resident in high-speed memory to achieve optimal results.
50%UnverifiedThe Q8 main model occupies 162 GB, preventing the drafter from fitting into unified memory, rendering DSpark acceleration ineffective; after moving three layers of MoE experts back to the CPU, output only improved from 22.87 to 23.49 tokens/second (2.7% faster), while prefill dropped from 95.38 to 76.65.
50%