EAGLE-2 is slower than EAGlE-1

[yjdy]

Thanks for this great repo. I have test EAGLE-1 and EAGLE-2 on vicuna-7b.
But I found that EAGLE-2 is slower than EAGLE-1, 69 tokens/s and 66 tokens/s respectively.
The inference test is on MT-bentch, using V100 with 32G memory, and batch size is 1.

Is it normal?

Best regards

[yjdy]

I make a mistake above, the inference speed of EAGLE-2 is 66 and EAGLE-1 is 69. Besides, temperature is 0

[hongyanz]

It is not normal. Can you provide more details (e.g., if you are running something else on your machine, what is your environment, which codes are your running)? Without them, it is hard to debug your code.

[yjdy]

Some details of my environment are listed as follow:
1 GPU V100 32G memory
python 3.10.14
CUDA 11.7
Driver Version: 515.65.01
torch 2.1.0
triton 2.1.0
transformers 4.36.2

I just run the evaluation script gen_ea_answer_vicuna.py as suggested in Readme
batch size= 1
temperature=0

[Liyuhui-12]

The possible reason is that total_token was not set correctly.

[yjdy]

Thanks for the response. Can you provide me some advise to set total_token? Should I set larger or smaller?

[Lucas-TY]

HI, the benchmark can't record new token correctly, I don't know if that's normal.

python -m eagle.evaluation.gen_ea_answer_vicuna\
        --ea-model-path yuhuili/EAGLE-Vicuna-7B-v1.3\
        --base-model-path lmsys/vicuna-7b-v1.3

python -m eagle.evaluation.gen_baseline_answer_vicuna\
		 --ea-model-path yuhuili/EAGLE-Vicuna-7B-v1.3\
		 --base-model-path lmsys/vicuna-7b-v1.3

{"question_id": 81, "answer_id": "TP4CRrbLYBqFHdQqoeb7ug", "model_id": "ess-vicuna-70b-fp16-baseline-temperature-1.0", "choices": [{"index": 0, "turns": ["....... "idxs": [603, 603], "new_tokens": [0, 0], "wall_time": [8.09636378288269, 7.946403741836548]}], "tstamp": 1720166094.253764}

[Liyuhui-12]

Thanks for the response. Can you provide me some advise to set total_token? Should I set larger or smaller?

Overall, the smaller the model and the more powerful the computational capacity, the larger this value should be.

[Liyuhui-12]

HI, the benchmark can't record new token correctly, I don't know if that's normal.

It is normal for the baseline not to return new tokens.