Qwen SWE Drafter (DFlash2, block 8)
DFlash2 speculative-decoding drafter for Qwen/Qwen3.6-35B-A3B-FP8, trained for software-engineering and agentic traffic. The target model is unchanged. The drafter only proposes tokens and the target verifies every one, so correctness is unchanged.
- Architecture:
DFlash2DraftModel, block size 8, 6 layers, 8 target taps[1,6,11,16,22,27,32,37] - Draft window: 8192
- Weights: BF16 (
model.safetensors, ~1.05 GB) - Warm start: intermediate Qwen3.6 drafter checkpoint step 500
- Training: one pass over 33,868 unseen turns (SWE code-write, agentic support, Open-SWE), soft-label weight 0.50, LR 1e-4, 1048 steps
Background and measurements: Speculative decoding for Qwen3.6 on CoralBricks.
Serving
Step-by-step instructions for serving this drafter with the target on one H100 — stock SGLang, CUDA 13 image plus a CUDA 12.9 fallback, required flags, sizing and troubleshooting — are in SERVE.md.
Files
model.safetensors— BF16 weightsconfig.json— DFlash2 config (block 8, taps, mask token 248077)SERVE.md— serving playbook