import requests
from PIL import Image

import torch
from transformers import AutoModelForVision2Seq, LlavaConfig
from openmind import AutoProcessor
import transformers
from openmind_hub import snapshot_download
from torch import nn

class LlavaMultiModalProjectorYiVL(nn.Module):
    def __init__(self, config: "LlavaConfig"):
        super().__init__()
        self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.linear_2 = nn.LayerNorm(config.text_config.hidden_size, bias=True)
        self.linear_3 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.linear_4 = nn.LayerNorm(config.text_config.hidden_size, bias=True)
        self.act = nn.GELU()

    def forward(self, image_features):
        hidden_states = self.linear_1(image_features)
        hidden_states = self.linear_2(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_3(hidden_states)
        hidden_states = self.linear_4(hidden_states)
        return hidden_states
# Monkey patch of LlavaMultiModalProjector is mandatory
transformers.models.llava.modeling_llava.LlavaMultiModalProjector = LlavaMultiModalProjectorYiVL

model_id = snapshot_download("AI-Research/Yi-VL-6B-hf", revision="main", resume_download=True,
                                    ignore_patterns=["*.h5", "*.ot", "*.mspack"])

messages = [
  { "role": "user", "content": "<image>What's in the picture?" }
]
image_file = "http://images.cocodataset.org/val2017/000000039769.jpg"

model = AutoModelForVision2Seq.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True, 
).to("npu")
processor = AutoProcessor.from_pretrained(model_id)

text = [processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)]
images = [Image.open(requests.get(image_file, stream=True).raw)]
inputs = processor(text=text, images=images, return_tensors='pt').to("npu", torch.float16)

output = model.generate(**inputs, max_new_tokens=200)
output = processor.batch_decode(output, skip_special_tokens=True)
print(output[0].split("Assistant:")[-1].strip())

llamafactory-cli webchat \
--model_name_or_path BUAADreamer/Yi-VL-6B-hf \
--template yivl \
--visual_inputs