--- language: - multilingual - en license: apache-2.0 library_name: transformers tags: - nlp - code - vision - chemistry - engineering - biology - bio-inspired - text-generation-inference - materials science - mixture-of-experts - science - latex datasets: - lamm-mit/Cephalo-Bioinspired-Mechanics-Materials - lamm-mit/Cephalo-Wikipedia-Materials pipeline_tag: image-text-to-text inference: parameters: temperature: 0.3 widget: - messages: - role: user content: <|image_1|>Can you describe what you see in the image? --- ## Model Summary Cephalo is a series of multimodal materials science focused vision large language models (V-LLMs) designed to integrate visual and linguistic data for advanced understanding and interaction in human-AI or multi-agent AI frameworks. The model is developed to process diverse inputs, including images and text, facilitating a broad range of applications such as image captioning, visual question answering, and multimodal content generation. The architecture combines a vision encoder model and an autoregressive transformer to process complex natural language understanding. ![image/png](https://cdn-uploads.huggingface.co/production/uploads/623ce1c6b66fedf374859fe7/kl5GWBP9WS0D4uwd1t3S7.png) Cephalo provides a robust framework for multimodal interaction and understanding, including the development of complex generative pipelines to create 2D and 3D renderings of material microstructures as input for additive manufacturing methods. This version of Cephalo, lamm-mit/Cephalo-Idefics2-3x8b-beta, is a Mixture-of-Expert model based on variants and fine-tuned versions of the Idefics-2 model. The basic model architecture is as follows: ![image/png](https://cdn-uploads.huggingface.co/production/uploads/623ce1c6b66fedf374859fe7/b7BK8ZtDzTMsyFDi0wP3w.png) The model has 20b parameters (3 experts, each 8b each, 8b active parameters during inference). ### Download Idefics-2 MoE Model and Sample inference code ```python pip install transformers -U ``` ```python import torch from transformers import AutoModelForCausalLM, AutoProcessor, AutoConfig def count_parameters(model): total_params = sum(p.numel() for p in model.parameters()) trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad) #number of parameters in b return total_params/1e9, trainable_params/1e9 device = torch.device("cuda" if torch.cuda.is_available() else "cpu") model_name_moe = f"lamm-mit/Cephalo-Idefics2-3x8b-beta" config = AutoConfig.from_pretrained(model_name_moe, trust_remote_code=True) processor = AutoProcessor.from_pretrained(model_name_moe, trust_remote_code=True) moe_model = AutoModelForCausalLM.from_pretrained( model_name_moe,config=config, trust_remote_code=True, torch_dtype=torch.bfloat16, # quantization_config=quantization_config, ).to(device) count_parameters(moe_model) ``` Now use downloaded model for inference: ```python from transformers.image_utils import load_image DEVICE='cuda' image = load_image("https://d2r55xnwy6nx47.cloudfront.net/uploads/2018/02/Ants_Lede1300.jpg") # Create inputs messages = [ { "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI."}, ] }, ] prompt = processor.apply_chat_template(messages, add_generation_prompt=True) # Get inputs using the processor inputs = processor(text=prompt, images=[image], return_tensors="pt") inputs = {k: v.to(DEVICE) for k, v in inputs.items()} # Generate generated_ids = moe_model.generate(**inputs, max_new_tokens=500) generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True) print(generated_texts) ``` ## Make a Idefics-2-MoE model from scratch using several pre-trained models Download .py files that implement the Phi-3-V and the Mixture-of-Expert Vision model ```python pip install huggingface_hub ``` ```python from huggingface_hub import HfApi, hf_hub_download from tqdm.notebook import tqdm import os import shutil # Repository details repo_id = "lamm-mit/Cephalo-Idefics2-3x8b-beta" api = HfApi() # List all files in the repository files_in_repo = api.list_repo_files(repo_id) # Filter for .py files py_files = [file for file in files_in_repo if file.endswith('.py')] # Directory to save the downloaded files save_dir = "./Idefics2_MoE/" os.makedirs(save_dir, exist_ok=True) # Download each .py file for file_name in tqdm(py_files): file_path = hf_hub_download(repo_id=repo_id, filename=file_name) new_path = os.path.join(save_dir, file_name) shutil.move(file_path, new_path) print(f"Downloaded: {file_name}") print("Download completed.") ``` Download models that will form the experts, as well as the base model. As a simple example, we use 1) Materials-science fine-tuned model: lamm-mit/Cephalo-Idefics-2-vision-8b-beta (model_1) 2) A chatty version: HuggingFaceM4/idefics2-8b-chatty (model_1) (model_2) 3) A basic variant: HuggingFaceM4/idefics2-8b (model_3) ```python from transformers import AutoProcessor, Idefics2ForConditionalGeneration , AutoTokenizer from transformers import BitsAndBytesConfig DEVICE='cuda' quantization_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.bfloat16 ) model_id_1='lamm-mit/Cephalo-Idefics-2-vision-8b-beta' model_1 = Idefics2ForConditionalGeneration.from_pretrained( model_id_1, torch_dtype=torch.bfloat16, #if your GPU allows _attn_implementation="flash_attention_2", #make sure Flash Attention 2 is installed trust_remote_code=True, #quantization_config=quantization_config, ) processor = AutoProcessor.from_pretrained( f"{model_id_1}", do_image_splitting=True ) config = AutoConfig.from_pretrained(model_id_1, trust_remote_code=True) IDEFICS2_CHAT_TEMPLATE = "{% for message in messages %}{{message['role'].capitalize()}}{% if message['content'][0]['type'] == 'image' %}{{':'}}{% else %}{{': '}}{% endif %}{% for line in message['content'] %}{% if line['type'] == 'text' %}{{line['text']}}{% elif line['type'] == 'image' %}{{ '' }}{% endif %}{% endfor %}\n{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}" processor.chat_template = IDEFICS2_CHAT_TEMPLATE ``` Now, load the rest of the models: ```python model_id_2='HuggingFaceM4/idefics2-8b-chatty' model_2 = Idefics2ForConditionalGeneration.from_pretrained( model_id_2, torch_dtype=torch.bfloat16, #if your GPU allows _attn_implementation="flash_attention_2", #make sure Flash Attention 2 is installed trust_remote_code=True, #quantization_config=quantization_config, ) model_id_3='HuggingFaceM4/idefics2-8b' model_3 = Idefics2ForConditionalGeneration.from_pretrained( model_id_3, torch_dtype=torch.bfloat16, #if your GPU allows _attn_implementation="flash_attention_2", #make sure Flash Attention 2 is installed trust_remote_code=True, #quantization_config=quantization_config, ) ``` Put on device: ```python model_1.to(DEVICE) model_2.to(DEVICE) model_3.to(DEVICE) ``` ### Construct MoE Here we show how a MoE is constructed from the set of expert models loaded earlier. We consider three models, model_1, model_2 and model_3. ```python dtype = torch.bfloat16 # Desired dtype for new layers base_model = copy.deepcopy(model_1) # Your base model expert_models = [ model_1, model_2, model_3 ] # List of expert models moe_config = Idefics2ForCausalLMMoEConfig(config=config, k=1, num_expert_models=len (expert_models)) moe_model = Idefics2ForCausalLMMoE(moe_config, base_model, expert_models, layer_dtype = dtype) count_parameters(expert_models[0]),count_parameters(moe_model) ``` Delete models no longer needed: ```python del model_1 del model_2 del model_3 ``` Put MoE model on device: ```python moe_model.to(DEVICE) ``` Test if it works (untrained, may not produce desirable putput since gating layers have not been trained): ```python from transformers.image_utils import load_image image = load_image("https://d2r55xnwy6nx47.cloudfront.net/uploads/2018/02/Ants_Lede1300.jpg") # Create inputs messages = [ { "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI."}, ] }, ] prompt = processor.apply_chat_template(messages, add_generation_prompt=True) # Get inputs using the processor inputs = processor(text=prompt, images=[image], return_tensors="pt") inputs = {k: v.to(DEVICE) for k, v in inputs.items()} # Generate generated_ids = moe_model.generate(**inputs, max_new_tokens=500) generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True) print(generated_texts) ``` ### Now train MoE gating function We train the gating layers by providing sample images/prompts for each of the three experts. Here is a simple example training set: ```python image_1 = Image.open("./VALIDATION/Q15.jpg") image_1a = Image.open("./VALIDATION/Q31.jpg") image_2 = Image.open(requests.get("https://media.wired.com/photos/5aa32b912ba43111d1213e0c/master/w_2240,c_limit/akhacouple.jpg", stream=True).raw) image_2a = Image.open(requests.get("https://media.wired.com/photos/5aa32b912ba43111d1213e0c/master/w_2240,c_limit/akhacouple.jpg", stream=True).raw) image_3 = Image.open(requests.get("https://i5.walmartimages.com/seo/Amazing-Andrea-Apple-Tree-Seeds-20-Seeds-Grow-Fresh-Apples_ff218043-bcd4-4437-8418-6631d8e97bb3.638ac0120ff05c8913e85ebb74f45f6c.jpeg?odnHeight=640&odnWidth=640&odnBg=FFFFFF", stream=True).raw) image_3a = Image.open(requests.get("https://i5.walmartimages.com/seo/Amazing-Andrea-Apple-Tree-Seeds-20-Seeds-Grow-Fresh-Apples_ff218043-bcd4-4437-8418-6631d8e97bb3.638ac0120ff05c8913e85ebb74f45f6c.jpeg?odnHeight=640&odnWidth=640&odnBg=FFFFFF", stream=True).raw) prompts_per_expert = [ [{"text": "User:What is shown in this image. Explain the importance for materials design.Assistant: The image shows", "image": [image_1]}, {"text": "User:What is shown in this image. Explain the importance for materials design.Assistant: The image shows", "image": [image_1a]}, ], [{"text": "User:What is shown in this image. Assistant: The image shows a human.", "image": [image_2]}, {"text": "User:What is shown in this image, and what does it mean in terms of human history? Assistant: The image shows a historical image of human development.", "image": [image_2a]}, ], [{"text": "User:What is shown in this image. Provide a brief answer. Assistant: This is an apple.", "image": [image_3]}, {"text": "User:What is shown in this image. Brief and concise answer. Assistant: The image shows an apple.", "image": [image_3a]}, ], ] gating_layer_params = moe_model.train_gating_layer_params_from_hidden_states(processor, prompts_per_expert, epochs=1000, loss_steps=100, lr=5e-5, layer_offset=0) # Set parameters for a specific layer moe_model.set_gating_layer_params(gating_layer_params) ``` ![image/png](https://cdn-uploads.huggingface.co/production/uploads/623ce1c6b66fedf374859fe7/mh4eFDuFsTBOYbjc38PYz.png) Now that the MoE model has been trained, we can try inference. Inference after MoE gating layers are trained: ```python from transformers.image_utils import load_image image = load_image("https://d2r55xnwy6nx47.cloudfront.net/uploads/2018/02/Ants_Lede1300.jpg") # Create inputs messages = [ { "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI."}, ] }, ] prompt = processor.apply_chat_template(messages, add_generation_prompt=True) # Get inputs using the processor inputs = processor(text=prompt, images=[image], return_tensors="pt") inputs = {k: v.to(DEVICE) for k, v in inputs.items()} # Generate generated_ids = moe_model.generate(**inputs, max_new_tokens=500) generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True) print(generated_texts) ``` ### Push to hub and save locally We can save the MoE model either in Hugging Face Hub or locally: ```python repo_id='...' moe_name='Cephalo-Idefics2-3x8b-beta' processor.push_to_hub (f'{repo_id}/'+moe_name, ) moe_model.push_to_hub (f'{repo_id}/'+merged_name, ) ``` Save locally: ```python processor.save_pretrained(moe_name, ) moe_model.save_pretrained(moe_name, ) ``` Loading the model works as done above. Here included again for completeness: ```python model_name_moe = f'{repo_id}/'+moe_name config = AutoConfig.from_pretrained(model_name_moe, trust_remote_code=True) processor = AutoProcessor.from_pretrained(model_name_moe, trust_remote_code=True) moe_model = AutoModelForCausalLM.from_pretrained( model_name_moe,config=config, trust_remote_code=True, torch_dtype=torch.bfloat16, # quantization_config=quantization_config, ).to(device) count_parameters(moe_model) ```