Qwen3.6-35B-A3B-Magic-Prompt-FP8-DSpark

Qwen3.6-35B-A3B-Magic-Prompt DSpark drafter

DSpark speculative-decoding draft for fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8, magic prompt expander for ideogram v4.

<img src="https://cdn-uploads.huggingface.co/production/uploads/6a31acdf1c726944a6fcc23f/r11KbqlM6hSJ_50Hm2boc.png" alt="image" style="max-width: 600px; width: 100%; height: auto;">

DSpark = DFlash block-diffusion backbone + EAGLE-style Markov head (low-rank bigram logit bias) + confidence head (per-position accept-rate predictor). Objective: 0.1*CE + 0.9*L1(draft,target) + 1.0*confidence_BCE. Dataset used consisted of 250K samples of magic prompt expansions, regenerated using the target model.

• Base / target: fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8
• Init: warm-started from z-lab/Qwen3.6-35B-A3B-DFlash (Markov/confidence heads fresh)
• Schedule: 5 epochs, lr 5e-4 → 5e-5, block_size 15
• Eval: simulated_acc_len 3.335, acc 0.665, loss 0.995 (val set)

Serving

We recommend serving with SGLang using dogacel/sglang:v0.5.14-dspark-cu130 image.

B200:

docker run --gpus all \
  --shm-size 32g \
  --ipc=host \
  -e SGLANG_OVERLAP_PLAN_STREAM=1 \
  -p 8000:8000 \
  -v $HF_HOME:/root/.cache/huggingface \
  dogacel/sglang:v0.5.14-dspark-cu130 \
  python3 -m sglang.launch_server \
    --model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8 \
    --served-model-name qwen36-35b-magic-fp8 \
    --trust-remote-code \
    --speculative-algorithm DSPARK \
    --speculative-draft-model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8-DSpark \
    --speculative-dflash-block-size 8 \
    --linear-attn-prefill-backend flashinfer \
    --linear-attn-decode-backend flashinfer \
    --mamba-radix-cache-strategy extra_buffer \
    --cuda-graph-backend-prefill tc_piecewise \
    --tp-size 1 \
    --max-running-requests 32 \
    --cuda-graph-max-bs-decode 32 \
    --weight-loader-disable-mmap \
    --host 0.0.0.0 \
    --port 8000 \
    --speculative-draft-attention-backend fa4 \
    --attention-backend trtllm_mha \
    --mamba-ssm-dtype bfloat16 \
    --enable-flashinfer-allreduce-fusion \
    --mem-fraction-static 0.6

H100:

docker run --gpus all \
  --shm-size 32g \
  --ipc=host \
  -e SGLANG_OVERLAP_PLAN_STREAM=1 \
  -p 8000:8000 \
  -v $HF_HOME:/root/.cache/huggingface \
  dogacel/sglang:v0.5.14-dspark-cu130 \
  python3 -m sglang.launch_server \
    --model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8 \
    --served-model-name qwen36-35b-magic-fp8 \
    --trust-remote-code \
    --speculative-algorithm DSPARK \
    --speculative-draft-model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8-DSpark \
    --speculative-dflash-block-size 8 \
    --linear-attn-prefill-backend flashinfer \
    --linear-attn-decode-backend flashinfer \
    --mamba-radix-cache-strategy extra_buffer \
    --cuda-graph-backend-prefill tc_piecewise \
    --tp-size 1 \
    --max-running-requests 32 \
    --cuda-graph-max-bs-decode 32 \
    --weight-loader-disable-mmap \
    --host 0.0.0.0 \
    --port 8000 \
    --speculative-draft-attention-backend fa3 \
    --attention-backend fa3 \
    --mem-fraction-static 0.8