Qwen3.5 122B DFlash2 Drafter (block 12)

DFlash2 speculative-decoding drafter for Qwen/Qwen3.5-122B-A10B-FP8, trained for long-context software-engineering and agentic traffic and tuned for sampling at T=1. The target model is unchanged. The drafter only proposes tokens and the target verifies every one, so correctness is unchanged.

• Architecture: DFlash2DraftModel, block size 12, 6 layers, 8 target taps [1,7,14,20,26,32,39,45]
• Draft window: 8192 (sliding_window in config.json)
• Weights: BF16 (model.safetensors, ~1.86 GB, ~930M parameters)
• Warm start: z-lab's DFlash checkpoint for Qwen3.5-122B-A10B, with the DFlash2 path selector and depthwise convolution taps added
• Training data: 47,440 requests drawn from 31,214 open-source SWE-agent trajectories. For each request the Qwen3.5-122B target wrote the final assistant turn itself at T=1, and the drafter trained on the target's hidden states over that history and generation
• Training: one pass, loss DPAL-CE-TV, LR 1e-4 with warmup and cosine decay, checkpoint at step 5100 of 5,616 planned

Results

MLCommons agentic SWE traces, all 15,777 turns, T=1, batch size 1, H200 TP2, stock SGLang v0.5.21 with the target in FP8:

The gain over MTP-7 holds in every prompt-length bin from under 32K to over 128K tokens. The training set contains no MLCommons data (0 repo hits, 0 text hits).

Background and the full set of tables: We trained a diffusion drafter for Qwen3.5 122B.

Serving

Download the target and this drafter, then start SGLang with two GPUs:

hf download Qwen/Qwen3.5-122B-A10B-FP8 --local-dir /data/models/Qwen3.5-122B-A10B-FP8
hf download coral-bricks-ai/qwen35-122b-dflash2-drafter --local-dir /data/models/qwen35-122b-dflash2-drafter

docker run -d --name qwen35-122b-drafter \
  --gpus all --ipc=host --shm-size 32g --network host \
  -v /data:/data \
  lmsysorg/sglang:v0.5.21-cu130 \
  python3 -m sglang.launch_server \
    --model-path /data/models/Qwen3.5-122B-A10B-FP8 \
    --served-model-name qwen3.5-122b-a10b \
    --host 0.0.0.0 --port 30000 --tp-size 2 \
    --mem-fraction-static 0.85 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --attention-backend fa3 \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path /data/models/qwen35-122b-dflash2-drafter \
    --speculative-draft-attention-backend fa3 \
    --speculative-draft-model-quantization unquant \
    --speculative-dflash-block-size 12

The block size is set on the server (--speculative-dflash-block-size 12), and the sliding window is read from the drafter's config.json.

Files

• model.safetensors: BF16 weights
• config.json: DFlash2 config (block 12, taps, mask token 248077, window 8192)