# from transformers import AutoTokenizer, AutoModel

# torch.set_float32_matmul_precision("high")

# model_id = "FacebookAI/roberta-large"
# tokenizer = AutoTokenizer.from_pretrained(model_id)
# model = AutoModel.from_pretrained(model_id).to("cuda")

# text = "The capital of France is [MASK]."
# inputs = tokenizer(text, return_tensors="pt").to("cuda")
# outputs = model(**inputs)

# # To get predictions for the mask:
# masked_index = inputs["input_ids"][0].tolist().index(tokenizer.mask_token_id)
# predicted_token_id = outputs.logits[0, masked_index].argmax(axis=-1)
# predicted_token = tokenizer.decode(predicted_token_id)
# print("Predicted token:", predicted_token)

# from transformers import AutoModel, AutoTokenizer

# model_name = "chandar-lab/NeoBERT"
# tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
 
# text = "NeoBERT is the most efficient model of its kind!"
# inputs = tokenizer(text, return_tensors="pt")

# # Generate embeddings
# outputs = model(**inputs)
# embedding = outputs.last_hidden_state[:, 0, :]
# print(embedding.shape)


# import sys

# # Add the folder to sys.path
# sys.path.append("/gpfs/home/ym621/UniPointMap")

# import torch
# from PIL import Image
# sys.path.append("/home/m50048399/transfered/ye_project/UniPointMap")
# import open_clip

# # Create model & transforms
# model, _, preprocess = open_clip.create_model_and_transforms('ViT-B-16-quickgelu', pretrained='dfn2b')
# model.eval()  # Set model to eval mode

# # Get tokenizer
# tokenizer = open_clip.get_tokenizer('ViT-B-16-quickgelu')


# image = preprocess(Image.open("docs/CLIP.png")).unsqueeze(0)
# text = tokenizer(["a diagram", "a dog", "a cat"])

# with torch.no_grad(), torch.autocast("cuda"):
#     image_features = model.encode_image(image)
#     text_features = model.encode_text(text)
#     image_features /= image_features.norm(dim=-1, keepdim=True)
#     text_features /= text_features.norm(dim=-1, keepdim=True)

#     text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)

# print("Label probs:", text_probs)  # prints: [[1., 0., 0.]]

# import torch
# from PIL import Image
# from transformers import (
#     AutoImageProcessor,
#     AutoTokenizer,
#     AutoModelForCausalLM,
# )


# model_root = "jina"
# image_size=224
# model = AutoModelForCausalLM.from_pretrained(model_root,trust_remote_code=True).cuda()

# device = model.device

# from transformers import AutoModel, AutoTokenizer, AutoImageProcessor
# model_root = 'fg-clip-base'
# tokenizer = AutoTokenizer.from_pretrained(model_root)
# image_processor = AutoImageProcessor.from_pretrained(model_root)

# text_encoder = AutoModel.from_pretrained('jinaai/jina-clip-v2', trust_remote_code=True)
# tokenizer =  AutoTokenizer.from_pretrained('jinaai/jina-clip-v2', trust_remote_code=True)

# pip install -U huggingface_hub
from huggingface_hub import snapshot_download

# Download ONLY the light_cc3m subfolder into HF cache
snapshot_download(
    repo_id="MatchLab/ScenePoint",
    repo_type="dataset",
    allow_patterns=["light_3rscan/**", "light_arkitscenes/**"],  # only this subfolder
    resume_download=True,              # safe to re-run
    max_workers=8                      # parallel downloads
)

print("Downloaded to Hugging Face cache only.")