mpashkovskii/dspark-qwen3-8b

DSpark Speculator for Qwen/Qwen3-8B

Trained using speculators with the DSpark algorithm (DFlash parallel backbone + Markov head + confidence head).

Usage

# Direct deployment (speculators config auto-detected)
vllm serve mpashkovskii/dspark-qwen3-8b

# Or explicit speculative config
vllm serve Qwen/Qwen3-8B \
  --speculative-config '{"method": "dspark", "model": "mpashkovskii/dspark-qwen3-8b"}'

Training Configuration

Parameter	Value
Algorithm	DSpark (DFlash + Markov head + confidence head)
Verifier	Qwen/Qwen3-8B
Block Size	3
Draft Layers	5
Hidden Size	4096
Markov Rank	256
Markov Head Type	vanilla
Confidence Head	True
Target Layer IDs	[2, 10, 18, 26, 34]
CE Loss α	0.1
TV Loss α	0.9
Confidence Loss α	1.0

Evaluation Results

Metric	Value
Full Acceptance Rate	0.6738
Per-Position Acceptance Rate	0.6602
Total Loss	0.0161
CE Loss	0.0258
TV Loss	0.0060
Confidence Loss	0.3979

Per-Position Acceptance

Position	Acceptance Rate
1	0.7494
2	0.5981