I find I cannot load from fined LoRA checkpoint

[YerongLi]

import torch
from transformers import AutoModel, AutoTokenizer, AutoConfig

# Load the base model and tokenizer
model_path = "/home/yerong2/models/internlm-xcomposer2d5-7b"
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Load the adapter/finetuned weights
checkpoint_path = "lorra_finetune/src/finetune/output/finetune_lora/checkpoint-2/"
adapter_weights = torch.load(f"{checkpoint_path}/adapter_model.bin")

# Merge the adapter weights with the base model
model.load_state_dict(adapter_weights, strict=True)

# Verify that the model has loaded the weights
print("Model successfully loaded with finetuned weights.")

model.tokenizer = tokenizer

query = 'Image1 <ImageHere>; Image2 <ImageHere>; Image3 <ImageHere>; I want to buy a car from the three given cars, analyze their advantages and weaknesses one by one'
image = ['../examples/cars1.jpg',
        '../examples/cars2.jpg',
        '../examples/cars3.jpg',]
with torch.autocast(device_type='cuda', dtype=torch.float16):
    response, his = model.chat(tokenizer, query, image, do_sample=False, num_beams=3, use_meta=True)
print(response)

This reports error:

[rank1]: Traceback (most recent call last):
[rank1]:   File "/home/yerong2/representation-engineering/lorra_finetune/src/mllm_lorra.py", line 580, in <module>
[rank1]:     train()
[rank1]:   File "/home/yerong2/representation-engineering/lorra_finetune/src/mllm_lorra.py", line 491, in train
[rank1]:     model.load_state_dict(adapter_weights, strict=True)
[rank1]:   File "/home/yerong2/local/miniconda3/envs/mllm/lib/python3.11/site-packages/torch/nn/modules/module.py", line 2189, in load_state_dict
[rank1]:     raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
[rank1]: RuntimeError: Error(s) in loading state_dict for InternLMXComposer2ForCausalLM:
[rank1]:        Missing key(s) in state_dict: "plora_glb_GN", "plora_sub_GN", "model.tok_embeddings.weight", "model.layers.0.attention.wqkv.weight", "model.layers.0.attention.wqkv.Plora_A.weight", "model.layers.0.attention.wqkv.Plora_B.weight", "model.layers.0.attention.wqkv.lora_sft_A.weight", "model.layers.0.attention.wqkv.lora_sft_B.weight", "model.layers.0.attention.wqkv.lora_dpo_A.weight", "model.layers.0.attention.wqkv.lora_dpo_B.weight", "model.layers.0.attention.wqkv.lora_web_A.weight", "model.layers.0.attention.wqkv.lora_web_B.weight", "model.layers.0.attention.wo.weight", "model.layers.0.attention.wo.Plora_A.weight", "model.layers.0.attention.wo.Plora_B.weight", "model.layers.0.attention.wo.lora_sft_A.weight", "model.layers.0.attention.wo.lora_sft_B.weight", "model.layers.0.attention.wo.lora_dpo_A.weight", "model.layers.0.attention.wo.lora_dpo_B.weight", "model.layers.0.attention.wo.lora_web_A.weight", "model.layers.0.attention.wo.lora_web_B.weight", "model.layers.0.feed_forward.w1.weight", "model.layers.0.feed_forward.w1.Plora_A.weight", "model.layers.0.feed_forward.w1.Plora_B.weight", "model.layers.0.feed_forward.w1.lora_sft_A.weight", "model.layers.0.feed_forward.w1.lora_sft_B.weight", "model.layers.0.feed_forward.w1.lora_dpo_A.weight", "model.layers.0.feed_forward.w1.lora_dpo_B.weight", "model.layers.0.feed_forward.w1.lora_web_A.weight", "model.layers.0.feed_forward.w1.lora_web_B.weight", "model.layers.0.feed_forward.w3.weight", "model.layers.0.feed_forward.w3.Plora_A.weight", "model.layers.0.feed_forward.w3.Plora_B.weight", "model.layers.0.feed_forward.w3.lora_sft_A.weight", "model.layers.0.feed_forward.w3.lora_sft_B.weight", "model.layers.0.feed_forward.w3.lora_dpo_A.weight", "model.layers.0.feed_forward.w3.lora_dpo_B.weight", "model.layers.0.feed_forward.w3.lora_web_A.weight", "model.layers.0.feed_forward.w3.lora_web_B.weight", "model.layers.0.feed_forward.w2.weight", "model.layers.0.feed_forward.w2.Plora_A.weight", "model.layers.0.feed_forward.w2.Plora_B.weight", "model.layers.0.feed_forward.w2.lora_sft_A.weight", "model.layers.0.feed_forward.w2.lora_sft_B.weight", "model.layers.0.feed_forward.w2.lora_dpo_A.weight", "model.layers.0.feed_forward.w2.lora_dpo_B.weight", "m

[YerongLi]

Use PeftModel from pretrained seems to work

import torch
from transformers import AutoModel, AutoTokenizer, AutoConfig
from peft import PeftModel    

# Load the base model and tokenizer
model_path = "/home/yerong2/models/internlm-xcomposer2d5-7b"
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Load the adapter/finetuned weights
checkpoint_path = "lorra_finetune/math/best"
# adapter_weights = torch.load(f"{checkpoint_path}/adapter_model.bin")
model = PeftModel.from_pretrained(model, checkpoint_path)
model = model.merge_and_unload()
# Verify that the model has loaded the weights
print("Model successfully loaded with finetuned weights.")

model.tokenizer = tokenizer

query = 'Image1 <ImageHere>; Image2 <ImageHere>; Image3 <ImageHere>; I want to buy a car from the three given cars, analyze their advantages and weaknesses one by one'
image = ['./examples/cars1.jpg',
        './examples/cars2.jpg',
        './examples/cars3.jpg',]
with torch.autocast(device_type='cuda', dtype=torch.float16):
    response, his = model.chat(tokenizer, query, image, do_sample=False, num_beams=3, use_meta=True)
print(response)

[YerongLi]

Use PeftModel from pretrained seems to work

import torch
from transformers import AutoModel, AutoTokenizer, AutoConfig
from peft import PeftModel    

# Load the base model and tokenizer
model_path = "/home/yerong2/models/internlm-xcomposer2d5-7b"
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Load the adapter/finetuned weights
checkpoint_path = "lorra_finetune/math/best"
# adapter_weights = torch.load(f"{checkpoint_path}/adapter_model.bin")
model = PeftModel.from_pretrained(model, checkpoint_path)
model = model.merge_and_unload()
# Verify that the model has loaded the weights
print("Model successfully loaded with finetuned weights.")

model.tokenizer = tokenizer

query = 'Image1 <ImageHere>; Image2 <ImageHere>; Image3 <ImageHere>; I want to buy a car from the three given cars, analyze their advantages and weaknesses one by one'
image = ['./examples/cars1.jpg',
        './examples/cars2.jpg',
        './examples/cars3.jpg',]
with torch.autocast(device_type='cuda', dtype=torch.float16):
    response, his = model.chat(tokenizer, query, image, do_sample=False, num_beams=3, use_meta=True)
print(response)

Actually this merge is not effective, I think it just loaded the original model from HF

[yuhangzang]

You may read the document and use the merge_peft_adapter.py to merge the LoRA weight.

[YerongLi]

Here is a document for 2.0 https://github.com/InternLM/InternLM-XComposer/blob/0a33f58f3cbc7d20be4bae4a3d4a812289c63406/InternLM-XComposer-2.0/finetune/README.md

[YerongLi]

#423 I found with multiple GPU, deepspeed+LoRA saves an invalid checkpoint for some reason.