stepnivlk/Qwen3.6-27B-DSpark

Qwen3.6-27B DSpark drafter — vLLM rope-config fix

Fork of satgeze/Qwen3.6-27B-DSpark (revision 0642d218) with one metadata fix and bit-identical weights: config.json's rope fields are corrected to full-width RoPE (partial_rotary_factor: 1.0, rope_type: default, mrope fields removed).

Why

The original config inherits partial_rotary_factor: 0.25 and mrope fields from the target model's text config, but the DeepSpec training code never reads them — its Qwen3 attention applies RoPE across the full head dim (Qwen3RotaryEmbedding + rotate_half from dense modeling_qwen3).

Engines that build the drafter's rotary embedding from the config — notably vLLM's {"method": "dspark"} — honor the field, rotate only 64 of 256 head dims, and produce position-degenerate drafts:

Serving setup (target: unsloth/Qwen3.6-27B-NVFP4, block 15)	Draft acceptance
vLLM ≥ 0.25 + original config	~1.4% (net slowdown)
vLLM ≥ 0.25 + this config	~12.6% (≈2.9 tokens/step)
(sibling 0.8B head, dense bf16 target, same fix)	6–8% → 25.1%

llama.cpp (--spec-type draft-dspark) is unaffected — its GGUF conversion already uses full-width rope. For GGUF files, use the original repo.

Usage (vLLM ≥ 0.25)

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","num_speculative_tokens":15}'

Notes: num_speculative_tokens: 15 is the trained block size (see the original model card — lower values discard most of the speedup). Raise --max-num-batched-tokens if vLLM reports a negative max_num_scheduled_tokens (15 draft slots per sequence add up).

Provenance

• Weights: unchanged from satgeze/Qwen3.6-27B-DSpark @ 0642d2182fb557ac4efee5c77b22ecb5447137ea (model.safetensors, sha-identical). All training credit to the original author; head trained with DeepSpec against Qwen3.6-27B, warm-started from the z-lab DFlash head.
• Root-cause analysis: differential debugging of vLLM's DSpark integration against the DeepSpec reference implementation (2026-07-30). A config fix has been proposed to the upstream repo; this fork exists so vLLM users have a working head in the meantime.