# Direct deployment (speculators config auto-detected)
vllm serve olka-amd/dspark-gemma-4-31b-it

# Or explicit speculative config
vllm serve google/gemma-4-31B-it \
  --speculative-config '{"method": "dspark", "model": "olka-amd/dspark-gemma-4-31b-it"}'

Training Configuration

Parameter	Value
Algorithm	DSpark (DFlash + Markov head + confidence head)
Verifier	`google/gemma-4-31B-it`
Block Size	7
Draft Layers	5
Hidden Size	5376
Markov Rank	256
Markov Head Type	vanilla
Confidence Head	True
Target Layer IDs	[2, 16, 30, 44, 58]
CE Loss α	0.1
TV Loss α	0.9
Confidence Loss α	1.0

Evaluation Results

Metric	Value
Full Acceptance Rate	0.6934
Per-Position Acceptance Rate	0.6663
Total Loss	0.4404
CE Loss	0.7204
TV Loss	0.1543
Confidence Loss	0.2294

Per-Position Acceptance

Position	Acceptance Rate
0	0.8375
1	0.7722
2	0.7216
3	0.6798
4	0.6437
5	0.6129
6	0.5836