[Bug] There is a gap between the acceptance rates of training and inference

[xinanjiao]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
• 4. If the issue you raised is not a bug but a question, please raise a discussion at https://github.com/sgl-project/SpecForge/discussions/new/choose Otherwise, it will be closed.
• 5. Please use English, otherwise it will be closed.

Describe the bug

I trained a Dflash application using the latest training code for Qwen3.5-4B. The draft configuration for 4B is from the official documentation, while Qwen3.5-4B is a version I fine-tuned using internal data. The task was OCR. This training used 70,000 data points, achieving a 98% acceptance rate during training.

I used a version of VLLM that supports Dflash for inference, but the average acceptance rate was only 10%. I've already aligned the chat-template, and I haven't used <think> during either training or inference.

Are there any gaps I might be overlooking?

Reproduction

BUILD_DATASET_NUM_PROC=64
ATTENTION_BACKEND=${2:-flex_attention}
NUM_GPUS=4
# Use patched specforge (fixes circular import in original)
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun \
    --standalone \
    --nproc_per_node $NUM_GPUS \
    $ROOT_DIR/scripts/train_dflash.py \
    --target-model-path qwen3.5_4B\
    --draft-config-path Qwen3.5-4B-Dflash/config.json \
    --train-data-path data_filtered.jsonl\
    --output-dir deflash_outputs/qwen3.5-4b-dflash-opc \
    --num-epochs 10 \
    --batch-size 2 \
    --learning-rate 6e-4 \
    --warmup-ratio 0.04 \
    --max-grad-norm 1.0 \
    --max-length 4096 \
    --chat-template qwen3.5-nothink \
    --attention-backend $ATTENTION_BACKEND \
    --num-anchors 512 \
    --loss-decay-gamma 7.0 \
    --log-interval 50 \
    --save-interval 10000 \
    --target-model-backend hf \
    --block-size 16 \

Environment

specforge [latest]
vllm [0.19.1.rc.0] nightly

[xinanjiao]

Similarly, I tested the latest sglang version with dflash, and unfortunately, the performance was just as bad. However, when testing with the official qwen3.5 4B dflash, the performance was acceptable. Are there any other ways to improve the matching between training and inference?