Hey there, I have a fine-tuned PEFT / Alpaca-LoRA model that I want to use for model inference in an API for an app that we’re developing.
Right now when I go to the model hub, I see the error message in the subject line. From my reading of the linked documentation, you need to specify the model architecture in the config.json
file, and I did that with the following lines:
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM
config = PeftConfig.from_pretrained("dsml-ngpt/train_0_1_0")
model = AutoModelForCausalLM.from_pretrained("yahma/llama-7b-hf")
model = PeftModel.from_pretrained(model, "dsml-ngpt/train_0_1_0")
model.config.to_json_file('config.json')
My config.json
file looks like this:
{
"_name_or_path": "yahma/llama-7b-hf",
"architectures": [
"LlamaForCausalLM"
],
"bos_token_id": 1,
"eos_token_id": 2,
"hidden_act": "silu",
"hidden_size": 4096,
"initializer_range": 0.02,
"intermediate_size": 11008,
"max_position_embeddings": 2048,
"model_type": "llama",
"num_attention_heads": 32,
"num_hidden_layers": 32,
"num_key_value_heads": 32,
"pad_token_id": 0,
"pretraining_tp": 1,
"rms_norm_eps": 1e-06,
"rope_scaling": null,
"tie_word_embeddings": false,
"torch_dtype": "float16",
"transformers_version": "4.32.1",
"use_cache": true,
"vocab_size": 32000
}
However, when I added that to the huggingface repo, the problem has not gone away.
Any suggestions?
Thank you.
Model repo can be found here: dsml-ngpt/train_0_1_0 at main