import requests
from PIL import Image
import torch
from transformers import MllamaForConditionalGeneration, AutoProcessor

base_model = "hiiamsid/llama-3.2-vision-11B-ROCO"

processor = AutoProcessor.from_pretrained(base_model)

model = MllamaForConditionalGeneration.from_pretrained(
    base_model,
    low_cpu_mem_usage=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

url = "https://lh7-rt.googleusercontent.com/docsz/AD_4nXcz-J3iR2bEGcCSLzay07Rqfj5tTakp2EMTTN0x6nKYGLS5yWl0unoSpj2S0-mrWpDtMqjl1fAgH6pVkKJekQEY_kwzL6QNOdf143Yt66znQ0EpfLvx6CLFOqw41oeOYmhPZ6Qrlb5AjEr4AenIOgBMTWTD?key=vhLUYntaS9QOx531XpJH3g"
image = Image.open(requests.get(url, stream=True).raw)

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Describe the tutorial feature image."}
    ]}
]
input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(image, input_text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=120)
print(processor.decode(output[0]))

  @dataclass
  class train_config:
    model_name: str="meta-llama/Llama-3.2-11B-Vision-Instruct"
    batch_size_training: int=8
    batching_strategy: str="padding" #alternative is packing but vision model doesn't work with packing.
    context_length: int =4096
    gradient_accumulation_steps: int=1
    num_epochs: int=3
    lr: float=1e-5
    weight_decay: float=0.0
    gamma: float= 0.85 # multiplicatively decay the learning rate by gamma after each epoch
    seed: int=42
    use_fp16: bool=False
    mixed_precision: bool=True
    val_batch_size:int = 1
    use_peft: bool = False
    output_dir: str = "workspace/models"
    enable_fsdp: bool = True
    dist_checkpoint_root_folder: str="workspace/FSDP/model" # will be used if using FSDP
    dist_checkpoint_folder: str="fine-tuned" # will be used if using FSDP
    save_optimizer: bool=False # will be used if using FSDP
    
  @dataclass
  class fsdp_config:
      mixed_precision: bool = True
      use_fp16: bool=False
      sharding_strategy: ShardingStrategy = ShardingStrategy.FULL_SHARD # HYBRID_SHARD "Full Shard within a node DDP cross Nodes", SHARD_GRAD_OP "Shard only Gradients and Optimizer States", NO_SHARD "Similar to DDP".
      hsdp : bool =False # Require HYBRID_SHARD to be set. This flag can extend the HYBRID_SHARD by allowing sharding a model on customized number of GPUs (Sharding_group) and Replicas over Sharding_group.
      sharding_group_size: int=0 # requires hsdp to be set. This specifies the sharding group size, number of GPUs that you model can fit into to form a replica of a model.
      replica_group_size: int=0 #requires hsdp to be set. This specifies the replica group size, which is world_size/sharding_group_size.
      checkpoint_type: StateDictType = StateDictType.FULL_STATE_DICT  # alternatively FULL_STATE_DICT can be used. SHARDED_STATE_DICT saves one file with sharded weights per rank while FULL_STATE_DICT will collect all weights on rank 0 and save them in a single file.
      fsdp_activation_checkpointing: bool=True
      fsdp_cpu_offload: bool=False
      pure_bf16: bool = True
      optimizer: str= "AdamW"