A100 can process only 4k tokens

#27
by KubilayCan - opened

I'm using a A100 80Gb GPU, transformers==4.42.3, torch==2.3.1 and bfloat16 precision.

The code is as in the template:

model = AutoModelForCausalLM.from_pretrained(
        model_id,
        device_map="auto",
        torch_dtype=torch.bfloat16,
        attn_implementation='eager'
) 

Whenever the prompt + max_new_tokens exceed 4k tokens, i get CUDA error. I have tried using two GPUs but still have same problem.

What else can I try?

I can generate an ouput with the same prompt in Google studio.

KubilayCan changed discussion title from large value of max_new_tokens crashes the model to A100 can process only 4k tokens
Google org

Hi @KubilayCan , Could you please provide the stack trace error in details to better understand the issue. Thank you.

Hi @Renu11 , the problem was related to the sliding windows. The update function in transformers was using a wrong variable. It is reported here: https://github.com/huggingface/transformers/issues/31781 and here: https://github.com/huggingface/transformers/issues/31848. I updated the function as suggested in the later one and it works now. I guess last transformers update solved this bug according to the last comment in that page.

KubilayCan changed discussion status to closed

Sign up or log in to comment