Downloads · 30 days
10
13% of all-time downloads
DSSD-Llama3-8B
FlorianVal/DSSD-Llama3-8B
DSSD-Llama3-8B is a machine learning model from FlorianVal. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for transformers.
This repository contains the trained auxiliary early-exit heads for meta-llama/Meta-Llama-3-8B. These heads are used for Dynamic Self-Speculative Decoding (DSSD), enabling faster inference with guaranteed identical ou…
All-time downloads
80
Public
Repo size
3.2 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt3.2 GB · 100%
At a glance
- Library
- transformers
- Access
- Public
- Created
- Jan 7, 2026
- Updated
- Jan 8, 2026
- SHA
- 9e7cdb7f
Base models
Spaces
Model card
From the Hugging Face model README
DSSD Auxiliary Heads for Llama 3 8B
This repository contains the trained auxiliary early-exit heads for meta-llama/Meta-Llama-3-8B. These heads are used for Dynamic Self-Speculative Decoding (DSSD), enabling faster inference with guaranteed identical output to the full model.
🚀 Usage
These heads are designed to be used with the DSSD Demo.
- GitHub Repository: FlorianVal/DSSD_demo
- Live Demo: HuggingFace Space
📦 Contents
aux_heads.pt: State dictionary of the trained auxiliary heads (Layers 8, 16, 24).config.json: Configuration file specifying head layers and parameters.calibration.json: Calibrated thresholds for different accuracy levels.
🔧 How it Works
The auxiliary heads allow the model to exit early at intermediate layers if confidence thresholds are met. If the drafted tokens are not verified by the full model, they are corrected, ensuring 100% output match.
- Library
- transformers
- Created
- Jan 7, 2026
- Updated
- Jan 8, 2026