README: drop the training-data details
Browse filesCo-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LrUG2MPyVgRWHb7y6iKMf4
README.md
CHANGED
|
@@ -20,7 +20,7 @@ tags:
|
|
| 20 |
|
| 21 |
MMFlash draft model for speculative decoding of [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) on image and text inputs. It is not a standalone language model; it is paired with the target model in a speculative-decoding server (SGLang `DFLASH`).
|
| 22 |
|
| 23 |
-
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
|
| 24 |
|
| 25 |
## Model details
|
| 26 |
|
|
@@ -32,7 +32,6 @@ MMFlash builds on [DFlash](https://github.com/z-lab/dflash) block-diffusion draf
|
|
| 32 |
| Block size | 16 |
|
| 33 |
| Target hidden states | layers 1, 8, 15, 22, 29 (of 32) |
|
| 34 |
| Mask token id | 248070 |
|
| 35 |
-
| Training data | 1M samples of [`mvp-lab/LLaVA-OneVision-1.5-Instruct-Data`](https://huggingface.co/datasets/mvp-lab/LLaVA-OneVision-1.5-Instruct-Data), responses regenerated by Qwen3.5-4B, sequences up to 4096 tokens |
|
| 36 |
|
| 37 |
## Usage
|
| 38 |
|
|
@@ -93,7 +92,7 @@ Each cell is **τ / speedup**: τ is the average acceptance length (tokens accep
|
|
| 93 |
| DFlash (16) | 4.87 / 2.56× | 4.95 / 2.66× | 3.71 / 2.04× | 4.92 / 2.48× | 5.07 / 2.31× | 4.66 / 2.22× | 4.47 / 2.37× | 4.67 / 2.38× |
|
| 94 |
| **MMFlash (16)** | **5.23 / 2.74×** | **5.39 / 2.86×** | **3.96 / 2.18×** | **5.25 / 2.58×** | **5.75 / 2.56×** | **5.23 / 2.37×** | **4.69 / 2.49×** | **5.07 / 2.54×** |
|
| 95 |
|
| 96 |
-
Values in parentheses are the number of draft steps for MTP, the draft tree size for EAGLE-3, and the block size for DFlash and MMFlash. MTP is Qwen3.5-4B's built-in multi-token-prediction head. DFlash-z is the public [`z-lab/Qwen3.5-4B-DFlash`](https://huggingface.co/z-lab/Qwen3.5-4B-DFlash) draft. DFlash is our DFlash draft trained on the same
|
| 97 |
|
| 98 |
### Concurrency (temperature 0)
|
| 99 |
|
|
|
|
| 20 |
|
| 21 |
MMFlash draft model for speculative decoding of [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) on image and text inputs. It is not a standalone language model; it is paired with the target model in a speculative-decoding server (SGLang `DFLASH`).
|
| 22 |
|
| 23 |
+
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.
|
| 24 |
|
| 25 |
## Model details
|
| 26 |
|
|
|
|
| 32 |
| Block size | 16 |
|
| 33 |
| Target hidden states | layers 1, 8, 15, 22, 29 (of 32) |
|
| 34 |
| Mask token id | 248070 |
|
|
|
|
| 35 |
|
| 36 |
## Usage
|
| 37 |
|
|
|
|
| 92 |
| DFlash (16) | 4.87 / 2.56× | 4.95 / 2.66× | 3.71 / 2.04× | 4.92 / 2.48× | 5.07 / 2.31× | 4.66 / 2.22× | 4.47 / 2.37× | 4.67 / 2.38× |
|
| 93 |
| **MMFlash (16)** | **5.23 / 2.74×** | **5.39 / 2.86×** | **3.96 / 2.18×** | **5.25 / 2.58×** | **5.75 / 2.56×** | **5.23 / 2.37×** | **4.69 / 2.49×** | **5.07 / 2.54×** |
|
| 94 |
|
| 95 |
+
Values in parentheses are the number of draft steps for MTP, the draft tree size for EAGLE-3, and the block size for DFlash and MMFlash. MTP is Qwen3.5-4B's built-in multi-token-prediction head. DFlash-z is the public [`z-lab/Qwen3.5-4B-DFlash`](https://huggingface.co/z-lab/Qwen3.5-4B-DFlash) draft. DFlash is our DFlash draft trained on the same data without the visual-KL weighting.
|
| 96 |
|
| 97 |
### Concurrency (temperature 0)
|
| 98 |
|