You can deploy this model and start earning money today!

stepnivlk/
Qwen3.6-27B-DSpark-adapted

Name: Qwen3.6-27B-DSpark-adapted
Availability: InStock

satgeze/Qwen3.6-27B-DSpark

finetune

apache-2.0

Views

No views yet

Qwen3.6-27B DSpark head — adapted to unsloth/Qwen3.6-27B-NVFP4

Warm-start finetune of satgeze/Qwen3.6-27B-DSpark (with the rope config fix) on 20M tokens of the quantized target's own generations: 21.3k mixed prompts (40% chat / 30% code-instruct / 30% math), responses regenerated by unsloth/Qwen3.6-27B-NVFP4 with thinking on (temp 0.7 / top-p 0.95), DeepSpec pipeline, 2 epochs, lr 2e-5, max_length 4096, embeddings/lm_head frozen.

Benchmarks

All measured on a DGX Spark (GB10), vLLM nightly, target unsloth/Qwen3.6-27B-NVFP4, temp 0.7 / top-p 0.95, real-text prompts, warm-only. Baseline (no speculative decoding) = 11.0 tok/s flat across all content. Full methodology and raw data in the source repo benchmarks.

Single-stream speedup over baseline, by workload

config	chat	instruct-code	raw-code	math	mean tok/s
no spec decode	1.00×	1.00×	0.99×	1.00×	11.0
MTP n=2 (target's built-in head)	1.97×	1.95×	1.96×	2.17×	22.1
MTP n=3	2.24×	2.15×	2.10×	2.42×	24.5
original DSpark head, n=15 (rope-fixed)	2.07×	1.91×	1.97×	3.52×	26.1
this head, n=15	2.25×	2.18×	2.58×	3.88×	30.0
this head, n=8	2.36×	2.26×	2.81×	3.57×	30.3

Draft acceptance with n=8: 31–57% by content (vs 12.6% for the unadapted head on this target).

Context-length behaviour (real-code review prompts)

prompt tokens	MTP n=3 acc / tok/s	this head n=8 acc / tok/s
2k	72% / ~25	56% / ~42
8k	84% / ~29	30% / ~26
16k	78% / ~26	18% / ~18
24k	84% / ~27	10% / ~13

The head was trained at max_length 4096; acceptance decays beyond that window (crossover with MTP ≈ 8k). At concurrency 8 (mixed pool) MTP also wins: 146 vs 123 tok/s aggregate.

Usage (vLLM ≥ 0.25)

bash

vllm serve unsloth/Qwen3.6-27B-NVFP4 \
  --max-num-seqs 8 --max-num-batched-tokens 8448 \
  --enable-prefix-caching --mamba-block-size 256 \
  --speculative-config '{"method":"dspark","model":"stepnivlk/Qwen3.6-27B-DSpark-adapted","num_speculative_tokens":8}'