Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -28,13 +28,6 @@ This implementation is based on the paper **"Fuzzy Speculative Decoding for a Tu
|
|
| 28 |
- `js`: Jensen-Shannon divergence
|
| 29 |
- `draft_tokens`: Absolute difference in draft token probabilities
|
| 30 |
- **Standard Speculative Decoding**: Falls back to standard speculative decoding acceptance when FSD threshold is not met
|
| 31 |
-
- **Raw Logits Support**: Returns both processed and raw logits for advanced use cases
|
| 32 |
-
|
| 33 |
-
## Installation
|
| 34 |
-
|
| 35 |
-
```bash
|
| 36 |
-
pip install -r custom_generate/requirements.txt
|
| 37 |
-
```
|
| 38 |
|
| 39 |
## Usage
|
| 40 |
|
|
@@ -64,7 +57,6 @@ outputs = target_model.generate(
|
|
| 64 |
do_sample=True,
|
| 65 |
temperature=0.7,
|
| 66 |
max_new_tokens=100,
|
| 67 |
-
output_logits=True, # Enable raw logits output
|
| 68 |
)
|
| 69 |
|
| 70 |
# Decode result
|
|
@@ -89,12 +81,6 @@ print(generated_text)
|
|
| 89 |
- Otherwise: standard speculative decoding acceptance is applied
|
| 90 |
4. Accepted tokens are kept, rejected tokens trigger resampling from the target model
|
| 91 |
|
| 92 |
-
## Requirements
|
| 93 |
-
|
| 94 |
-
- `torch>=2.0.0`
|
| 95 |
-
- `transformers>=4.40.0`
|
| 96 |
-
- `scikit-learn` (optional, for confidence threshold features)
|
| 97 |
-
|
| 98 |
## Citation
|
| 99 |
|
| 100 |
If you use this code in your research, please cite the original paper:
|
|
|
|
| 28 |
- `js`: Jensen-Shannon divergence
|
| 29 |
- `draft_tokens`: Absolute difference in draft token probabilities
|
| 30 |
- **Standard Speculative Decoding**: Falls back to standard speculative decoding acceptance when FSD threshold is not met
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
|
| 32 |
## Usage
|
| 33 |
|
|
|
|
| 57 |
do_sample=True,
|
| 58 |
temperature=0.7,
|
| 59 |
max_new_tokens=100,
|
|
|
|
| 60 |
)
|
| 61 |
|
| 62 |
# Decode result
|
|
|
|
| 81 |
- Otherwise: standard speculative decoding acceptance is applied
|
| 82 |
4. Accepted tokens are kept, rejected tokens trigger resampling from the target model
|
| 83 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
## Citation
|
| 85 |
|
| 86 |
If you use this code in your research, please cite the original paper:
|