Add pipeline tag and links to paper, project page and code
#1
by nielsr HF Staff - opened
README.md
CHANGED
|
@@ -1,19 +1,20 @@
|
|
| 1 |
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
base_model:
|
| 4 |
-
|
| 5 |
-
|
|
|
|
| 6 |
tags:
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
|
|
|
| 17 |
---
|
| 18 |
|
| 19 |
<div align="center">
|
|
@@ -27,6 +28,8 @@ MMFlash draft model for speculative decoding of [`Qwen/Qwen3.5-9B`](https://hugg
|
|
| 27 |
|
| 28 |
MMFlash builds on [DFlash](https://github.com/z-lab/dflash) block-diffusion drafting: the draft reads hidden states from five layers of the target and proposes a block of 16 tokens in one forward pass, and the target verifies the block in one forward pass. MMFlash keeps the DFlash architecture and is trained on multimodal data with a loss weighted by per-token visual-grounding (visual-KL) scores.
|
| 29 |
|
|
|
|
|
|
|
| 30 |
<div align="center">
|
| 31 |
<video autoplay muted loop playsinline controls width="100%">
|
| 32 |
<source src="https://huggingface.co/FSCCS/Qwen3.5-9B-MMFlash/resolve/main/assets/demo.mp4" type="video/mp4">
|
|
@@ -107,4 +110,11 @@ MMFlash at 1, 8 and 32 requests in flight, on a single NVIDIA H200. Each cell is
|
|
| 107 |
|
| 108 |
## Citation
|
| 109 |
|
| 110 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
|
|
|
| 2 |
base_model:
|
| 3 |
+
- Qwen/Qwen3.5-9B
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
pipeline_tag: image-text-to-text
|
| 6 |
tags:
|
| 7 |
+
- mmflash
|
| 8 |
+
- dflash
|
| 9 |
+
- speculative-decoding
|
| 10 |
+
- speculative-decoding-draft
|
| 11 |
+
- block-diffusion
|
| 12 |
+
- draft-model
|
| 13 |
+
- multimodal
|
| 14 |
+
- sglang
|
| 15 |
+
- qwen
|
| 16 |
+
- qwen3.5
|
| 17 |
+
inference: false
|
| 18 |
---
|
| 19 |
|
| 20 |
<div align="center">
|
|
|
|
| 28 |
|
| 29 |
MMFlash builds on [DFlash](https://github.com/z-lab/dflash) block-diffusion drafting: the draft reads hidden states from five layers of the target and proposes a block of 16 tokens in one forward pass, and the target verifies the block in one forward pass. MMFlash keeps the DFlash architecture and is trained on multimodal data with a loss weighted by per-token visual-grounding (visual-KL) scores.
|
| 30 |
|
| 31 |
+
This checkpoint is presented in the paper [MMFlash: Faster Speculative Decoding for Vision-Language Models](https://huggingface.co/papers/2611.18482). Check out the [project page](https://fscdc.github.io/MMFlash/) and the [GitHub repository](https://github.com/fscdc/MMFlash) for code, training scripts and evaluation details.
|
| 32 |
+
|
| 33 |
<div align="center">
|
| 34 |
<video autoplay muted loop playsinline controls width="100%">
|
| 35 |
<source src="https://huggingface.co/FSCCS/Qwen3.5-9B-MMFlash/resolve/main/assets/demo.mp4" type="video/mp4">
|
|
|
|
| 110 |
|
| 111 |
## Citation
|
| 112 |
|
| 113 |
+
```bibtex
|
| 114 |
+
@article{feng2026mmflash,
|
| 115 |
+
title = {MMFlash: Faster Speculative Decoding for Vision-Language Models},
|
| 116 |
+
author = {Feng, Sicheng and Chen, Zigeng and Wang, Song and Fang, Gongfan and Wang, Xinchao},
|
| 117 |
+
journal = {arXiv preprint arXiv:2611.18482},
|
| 118 |
+
year = {2026}
|
| 119 |
+
}
|
| 120 |
+
```
|