AttributeError: 'NoneType' object has no attribute 'device'

[imrankh46]

why this happening?

batch = tokenizer("Two things are infinite: ", return_tensors="pt")

with torch.cuda.amp.autocast():
    output_tokens = model.generate(**batch, max_new_tokens=50)

print("\n\n", tokenizer.decode(output_tokens[0], skip_special_tokens=True))

its give the following error

AttributeError: 'NoneType' object has no attribute 'device'

[younesbelkada]

Hi @imrankh46 Thanks for the issue! We are aware of the issue, for now the solution is to pass device_map={"":0} when calling PeftModel.from_pretrained, we will work on a proper fix soon The issue is due to a call of dispatch_model inside from_pretrained of PeftModel, that breaks few things with Linear8bitLt layers

[imrankh46]

Hi @imrankh46 Thanks for the issue! We are aware of the issue, for now the solution is to pass device_map={"":0} when calling PeftModel.from_pretrained, we will work on a proper fix soon The issue is due to a call of dispatch_model inside from_pretrained of PeftModel, that breaks few things with Linear8bitLt layers

Thanks for response.

[younesbelkada]

@imrankh46 I believe https://github.com/huggingface/accelerate/pull/1237 should have fixed your issue, can you try to download accelerate from source and let us know if you still face the issue?

pip install git+https://github.com/huggingface/accelerate

[HZQ950419]

Hi @younesbelkada , I have tried to install accelerate from source, but I got another Error:

NotImplementedError: Cannot copy out of meta tensor; no data!

Do you know the possible reason for this? Thanks!

[YSLLYW]

Hi @imrankh46 Thanks for the issue! We are aware of the issue, for now the solution is to pass device_map={"":0} when calling PeftModel.from_pretrained, we will work on a proper fix soon The issue is due to a call of dispatch_model inside from_pretrained of PeftModel, that breaks few things with Linear8bitLt layers

I have encountered the same problem, my version is peft=0.2.0. I wonder if you have resolved this issue?

[younesbelkada]

@YSLLYW Can you try by installing accelerate from source?

pip install git+https://github.com/huggingface/accelerate

[imrankh46]

Hi @imrankh46 Thanks for the issue! We are aware of the issue, for now the solution is to pass device_map={"":0} when calling PeftModel.from_pretrained, we will work on a proper fix soon The issue is due to a call of dispatch_model inside from_pretrained of PeftModel, that breaks few things with Linear8bitLt layers

I have encountered the same problem, my version is peft=0.2.0. I wonder if you have resolved this issue?

You need to pass the device_map parameters like this device_map={"":0}

The complete code are here. You just need to pass your model

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import LlamaTokenizer
# pass your model name 
peft_model_id = 'your_model_name'
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=True, device_map='auto')
tokenizer = LlamaTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Lora model
model = PeftModel.from_pretrained(model, peft_model_id, torch_dtype=torch.float16,device_map={"":0})

[imrankh46]

@YSLLYW Can you try by installing accelerate from source?

pip install git+https://github.com/huggingface/accelerate

I already solved the issues. Thanks

[YSLLYW]

@YSLLYW Can you try by installing accelerate from source?

pip install git+https://github.com/huggingface/accelerate

Yes, I just updated the PEFT version to 0.3.0 and resolved this issue. Thank you for your reply

[YSLLYW]

@YSLLYW Can you try by installing accelerate from source?

pip install git+https://github.com/huggingface/accelerate

I already solved the issues. Thanks

Yes, I just updated the PEFT version to 0.3.0 and resolved this issue. Thank you for your reply

[imrankh46]

@YSLLYW Can you try by installing accelerate from source?

pip install git+https://github.com/huggingface/accelerate

I already solved the issues. Thanks

Yes, I just updated the PEFT version to 0.3.0 and resolved this issue. Thank you for your reply

Please also add Contrastive Search for text generation. When I pass penalty_alpha=0.6 they give me error cuda out memory.

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.