DSpark
Accelerate decodingSpeculative decodingReferenceDeepSeek's drafter: a parallel backbone proposes a whole block at once, a lightweight serial head glances one token back to stop the tail drifting, and a confidence head schedules how far to verify — so it doesn't pay to draft-and-discard past the acceptance knee.
Runs on 2× DGX Spark via vLLM --speculative-config '{"method":"dspark",...}'; the confidence-scheduled block is the right shape for GB10's bandwidth-bound decode, where a fixed large k wastes the shared memory bus. See models/ DSpark bring-up log.
- Objective
- Accelerate decoding
- Relieves
- Bandwidth
- Format
- Semi-autoregressive parallel drafter + confidence-scheduled verify
- Granularity
- Block draft (parallel backbone + serial Markov head)
- Lifecycle
- QAT / QAD · needs gradients
- Calibration
- Small calibration set
- Compression
- ~1.6–1.9× decode throughput (semi-autoregressive block draft)
- Quality
- 100% — verification keeps only tokens the target model would emit
- Hardware
- Trained DSpark drafter (baked into the repo or a companion speculator repo); confidence-scheduled verify kernel
- Runtimes
- vLLM
Open-sourced DSpark checkpoints plus DeepSpec, the training repo — 57–85% faster inference, lossless, on V4 / Qwen / GLM / Gemma.
Companion DSpark drafter repo for GLM-5.2 — the 'ships a speculator' pattern the optimizer detects.
End-to-end recipe to train and export a DSpark drafter for a target model.
In-flight llama.cpp support — DSpark reaching the local-inference stack beyond vLLM.
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.