Qwen3.6-27B DSpark draft head

DSpark speculative-decoding draft head for Avesed/Qwen3.6-27B-W4A16, trained on (Dataset )

Requires VLLM version 0.26 or later, or my fork for Ampere GPUs 0.4.1 or later

Serve this head with the vllm-ampere-optimized fork (ghcr.io/avesed/vllm-ampere-optimized:latest).

Update 2026-10-09: model.safetensors re-uploaded

About 20% of the rows of the frozen embed_tokens (a copy of the target's embedding) had been zeroed by a storage fault before the first upload. The embedding is now byte-identical to the target's, so vLLM shares the target's embedding instead of loading its own copy (~2.5 GB less GPU memory). Output was lossless either way (speculative decoding verifies every token); acceptance is equal or slightly higher (greedy +3% on our 12-prompt set). Every other tensor is unchanged.

Accepted length (num_speculative_tokens=7, greedy, 10 prompts/category — robust)

~2× decode speedup on English; multilingual output verified clean (0 degeneration). Chinese acceleration is modest (the draft head is English-centric); more Chinese on-policy data would raise it.

Serve (Ampere fork)

docker run --runtime nvidia --gpus all --shm-size 8g -p 8000:8000 \
  ghcr.io/avesed/vllm-ampere-optimized:latest \
  Avesed/Qwen3.6-27B-W4A16 --tensor-parallel-size 2 --trust-remote-code \
  --speculative-config '{"method":"dspark","model":"Avesed/Qwen3.6-27B-DSpark","num_speculative_tokens":7}'