--- pipeline_tag: text-generation library_name: transformers base_model: - Qwen/Qwen3.5-9B license: apache-2.0 inference: false tags: - dflash - speculative-decoding - speculative-decoding-draft - block-diffusion - draft-model - diffusion-language-model - efficiency - qwen - qwen3 - qwen3.5 - sglang --- # Qwen3.5-9B-DFlash [Paper](https://arxiv.org/abs/2602.06036) | [Github](https://github.com/z-lab/dflash) | [Blog](https://z-lab.ai/projects/dflash) This DFlash draft model is a joint retrain from [Z-Lab](https://z-lab.ai) and [Modal](https://modal.com), trained with 40k sequence length and sliding-window attention for improved long-context performance. It is mirrored across the following Hugging Face repositories: - [`z-lab/Qwen3.5-9B-DFlash`](https://huggingface.co/z-lab/Qwen3.5-9B-DFlash) - [`modal-labs/Qwen3.5-9B-DFlash`](https://huggingface.co/modal-labs/Qwen3.5-9B-DFlash) This repository contains a DFlash draft model for `Qwen/Qwen3.5-9B`. It is not a standalone language model. It is intended to be paired with the target model in a speculative decoding server. DFlash uses a lightweight block diffusion draft model to propose multiple tokens in parallel. The target model verifies those proposals, improving serving throughput while preserving the target model's output distribution. ## Quick Start This model should be used with an inference server that supports DFlash speculative decoding. An example SGLang deployment is: ```bash export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1 python -m sglang.launch_server \ --model-path Qwen/Qwen3.5-9B \ --trust-remote-code \ --speculative-algorithm DFLASH \ --speculative-draft-model-path modal-labs/Qwen3.5-9B-DFlash \ --speculative-dflash-block-size 8 \ --speculative-draft-attention-backend fa4 \ --attention-backend trtllm_mha \ --linear-attn-prefill-backend flashinfer \ --linear-attn-decode-backend flashinfer \ --mamba-scheduler-strategy extra_buffer \ --tp-size 1 \ --max-running-requests 32 \ --cuda-graph-max-bs-decode 32 \ --cuda-graph-backend-prefill tc_piecewise \ --enable-flashinfer-allreduce-fusion \ --mem-fraction-static 0.8 \ --host 0.0.0.0 \ --port 30000 ``` Block size `8` is the recommended default for higher-concurrency serving. Block size `16` gives longer accept lengths and strong concurrency-1 throughput in most workloads. ## Benchmark Results We benchmarked DFlash against the autoregressive baseline and Qwen's built-in MTP draft path. DFlash reaches up to `5.01x` speedup at concurrency 1 and `2.58x` at concurrency 32. Across the benchmark suite, DFlash delivers higher throughput than MTP at every matched setting where both completed. ### Setup - Runtime: SGLang on 1x NVIDIA B200 GPU, tensor parallel size 1, `bfloat16` - Backends: `trtllm_mha` target attention, `fa4` DFlash draft attention, `flashinfer` linear-attention prefill and decode - Workloads: GSM8K, MATH500, HumanEval, MBPP, and MT-Bench with the Qwen chat template - Decoding: greedy, thinking enabled, max output length 4096 tokens - Measurement: 5 independent runs per configuration at concurrency 1 and 32 with continuous batching - Throughput: generated output tokens / wall-clock benchmark time, including prefill and scheduling - Accept length: `completion_tokens / spec_verify_ct` per generation turn, averaged across generation turns ### Throughput and Speedup Each cell is `output tok/s (speedup)`. Bold marks the fastest speculative configuration in each row. #### Concurrency 1 | Workload | Baseline | MTP steps=3 | DFlash block=4 | MTP steps=7 | DFlash block=8 | MTP steps=15 | DFlash block=16 | | --- | --- | --- | --- | --- | --- | --- | --- | | gsm8k | 245.2 (1.00x) | 537.7 (2.19x) | 609.8 (2.49x) | 573.4 (2.34x) | 890.7 (3.63x) | 435.5 (1.78x) | **1027.3 (4.19x)** | | math500 | 244.6 (1.00x) | 558.3 (2.28x) | 636.3 (2.60x) | 617.3 (2.52x) | 987.6 (4.04x) | 485.3 (1.98x) | **1225.7 (5.01x)** | | humaneval | 243.4 (1.00x) | 537.6 (2.21x) | 633.2 (2.60x) | 586.9 (2.41x) | 959.5 (3.94x) | 447.6 (1.84x) | **1195.8 (4.91x)** | | mbpp | 244.7 (1.00x) | 526.4 (2.15x) | 624.4 (2.55x) | 543.6 (2.22x) | 935.7 (3.82x) | 403.6 (1.65x) | **1092.6 (4.46x)** | | mt-bench | 243.7 (1.00x) | 501.7 (2.06x) | 560.2 (2.30x) | 494.6 (2.03x) | 757.5 (3.11x) | 368.0 (1.51x) | **834.3 (3.42x)** | #### Concurrency 32 | Workload | Baseline | MTP steps=3 | DFlash block=4 | MTP steps=7 | DFlash block=8 | MTP steps=15 | DFlash block=16 | | --- | --- | --- | --- | --- | --- | --- | --- | | gsm8k | 5837.6 (1.00x) | 10421.0 (1.79x) | 11882.9 (2.04x) | 10132.5 (1.74x) | **13718.0 (2.35x)** | 8332.8 (1.43x) | 11019.7 (1.89x) | | math500 | 5885.7 (1.00x) | 11124.9 (1.89x) | 12534.0 (2.13x) | 11213.3 (1.91x) | **15198.3 (2.58x)** | 7785.0 (1.32x) | 13227.4 (2.25x) | | humaneval | 5513.0 (1.00x) | 9645.9 (1.75x) | 11882.7 (2.16x) | 9701.0 (1.76x) | **14229.9 (2.58x)** | 6901.7 (1.25x) | 12406.1 (2.25x) | | mbpp | 5538.8 (1.00x) | 9116.4 (1.65x) | 11561.2 (2.09x) | 8701.6 (1.57x) | **13460.3 (2.43x)** | 6220.5 (1.12x) | 11338.9 (2.05x) | | mt-bench | 5491.7 (1.00x) | 9135.0 (1.66x) | 10072.2 (1.83x) | 8436.9 (1.54x) | **10718.2 (1.95x)** | 5917.8 (1.08x) | 8495.7 (1.55x) | ### Accept Length Mean accept length at concurrency 1. Bold marks the higher value in each matched MTP/DFlash pair. | Workload | MTP steps=3 | DFlash block=4 | MTP steps=7 | DFlash block=8 | MTP steps=15 | DFlash block=16 | | --- | --- | --- | --- | --- | --- | --- | | gsm8k | **3.464** | 3.452 | 5.276 | **5.400** | 6.388 | **6.949** | | math500 | 3.541 | **3.555** | 5.466 | **5.757** | 6.728 | **7.721** | | humaneval | 3.493 | **3.571** | 5.326 | **5.798** | 6.399 | **7.927** | | mbpp | 3.338 | **3.454** | 4.790 | **5.376** | 5.508 | **6.820** | | mt-bench | **3.229** | 3.193 | 4.551 | **4.606** | 5.438 | **5.716** | ## Citation If you find DFlash useful, please cite the original paper: ```bibtex @article{chen2026dflash, title = {{DFlash: Block Diffusion for Flash Speculative Decoding}}, author = {Chen, Jian and Liang, Yesheng and Liu, Zhijian}, journal = {arXiv preprint arXiv:2602.06036}, year = {2026} } ```