EAGLE
EAGLE copied to clipboard
SafeAILab
Official Implementation of EAGLE-1 (ICML'24) and EAGLE-2 (EMNLP'24)
I see the corresponding code for training data generation is removed. Any plan to release that for EAGLE-3?
Hi @hongyanz 🤗 I'm Niels and work as part of the open-source team at Hugging Face. I discovered your work through Hugging Face's daily papers as yours got featured: https://huggingface.co/papers/2503.01840....
Hello, our data is for an agent scenario, with 100 tokens as input and 20 tokens as output. The input is relatively fixed. The acceleration effect of using Eagle's code...
After 3 epochs training，there are there dirs：state_0, state_1, state_2，and the model weights files are model.safetensors(2.9G) and model_1.safetensors(2G)，are they right?
when I run "python -m eagle.ge_data.allocation --outdir [path of data]" It seems that: Token indices sequence length is longer than the specified maximum sequence length for this model (2319 >...
In [training script](https://github.com/SafeAILab/EAGLE/blob/a929ad7bb036b29d95246ba534776e5fa88a804a/eagle/train/main.py#L128C9-L128C83), the uniform random noise (with std=0.1 as written on paper) is additionally scaled by 512/seq_len: ```py noise = (torch.rand_like(tensor) - 0.5) * self.std * 512 / tensor.shape[1]...
The paper mentions using several datasets, including MT-bench, HumanEval, GSM8K, and Alpaca, to test Eagle’s speedup effect. Would it be possible to provide the specific paths to the complete datasets?...
I am curious how did you guys implemented this experiments. I mean given Figure 6 in eagle 1 as an example, Feautre & unshifted token can be concatenated for tokens...
Hi, thanks for you great work! When I used the EAGLE-llama2-chat-7B you provided for testing, the average acceptance length I measured was lower than the value in the paper. The...
