HSpec-Qwen27B (RoPE-fixed, 50k perfectblend)

H-Spec mamba_attn_hybrid speculative drafter (draft depth 5, block size 6, draft vocab 64k, KV-sharing on verifier full-attention layers 3/19/35/51/63) for the target Qwen/Qwen3.8-27B, trained with vllm-project/speculators (branch hspec_qwen_27B, commit 577a974).

This checkpoint replaces the earlier pre-RoPE-fix upload. The drafter's Q/K rotation now matches the target's partial rotary (head_dim=256, partial_rotary_factor=0.25 → rotary_dim=64, rope_theta=1e7, neox pairing j↔j+32) — required because the drafter attends to the target's already-rotated shared K prefix.

Training

• 49.5k samples from OpenPerfectBlend (Qwen3.8-27B regenerated), 1 epoch, 7×H100, hidden states + shared KV extracted online from the target.
• Full training command in train_command.txt.

Results (spec decode depth 5)

metric	pre-fix 50k	this checkpoint
val loss	0.860	0.735
val accept rate	0.405	0.511
val EAL	2.428	2.998
HumanEval acceptance length	2.2275	2.5590

HumanEval per-position acceptance (positions 0–4): 0.623 / 0.390 / 0.254 / 0.173 / 0.119 (101,865 drafts). Raw eval data: humaneval/acceptance.csv, humaneval/run_HumanEval.json.

Serving

vllm serve Qwen/Qwen3.8-27B \
  --spec-model MeganEFlynn/HSpec-Qwen27B \
  --spec-method mamba_attn_hybrid \
  --spec-tokens 5 \
  --max-model-len 8196 --attention-backend FLASHINFER