meta-llama

/

Meta-Llama-3-8B-Instruct

like 3.48k

when calling model.generate, setting pad_token_id=tokenizer.eos_token_id seems to remove the warning.

https://stackoverflow.com/questions/69609401/suppress-huggingface-logging-warning-setting-pad-token-id-to-eos-token-id

it seems to work fine still with that, although some explanation and reasoning wouldn't hurt. maybe i saw one when searching for the solution but i forget.. :)

model.generate(**encoded_input, pad_token_id=tokenizer.eos_token_id).just change this will be ok.

Hey there, i was also facing the same error and after tinkering with the tokenizer for a bit i found out that inputs = tokenizer.encode(prompt) returns just the input_ids but not the attention mask, Whereas inputs = tokenizer(prompt) returns both the input_ids and the attention mask.

So if you replace this code

inputs = tokenizer.encode(prompt)
output = model.generate(inputs)

With this

inputs = tokenizer(prompt)
output = model.generate(**inputs)

The error warning goes away. Hope that helps 😊.

I had the same warning as well, and it took me looking at the huggingface transformers code quite a bit but was able to come to a solution:
original warning: The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's attention_mask to obtain reliable results.

If you choose to use apply_chat_template() with your tokenizer, an instance of PreTrainedTokenizer for instance, then set return_dict=True, e.g.
return_output = tokenizer.apply_chat_template(
conversation,
add_generation_prompt=True,
return_tensors="pt",
return_dict=True).to(model.device)

I do the to(model.device) at the end to move those (PyTorch?) tensors into the model's device (e.g. "cuda").

return_output is a dict of 2 keys: "input_ids" and "attention_mask". Use both when you run generate(..) for example:

        return model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=generation_configuration.max_new_tokens,
            do_sample=do_sample,
            top_k=generation_configuration.top_k,
            top_p=generation_configuration.top_p,
            temperature=generation_configuration.temperature,
            eos_token_id=eos_token_id,
            streamer=streamer)

and that gets called in my function:

        with torch.no_grad():
            generate_output = run_model_generate(
                input_ids=return_output["input_ids"],
                model=model,
                streamer=streamer,
                eos_token_id=generation_configuration.eos_token_id,
                generation_configuration=generation_configuration,
                attention_mask=return_output["attention_mask"])

My code here: https://github.com/InServiceOfX/InServiceOfX/blob/master/PythonLibraries/HuggingFace/MoreTransformers/executable_scripts/terminal_only_infinite_loop_instruct.py (I'm trying to build out my own library so I'm calling a number of wrappers I made)