Instructions to use hiiamsid/llama-3.2-vision-11B-ROCO with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use hiiamsid/llama-3.2-vision-11B-ROCO with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="hiiamsid/llama-3.2-vision-11B-ROCO")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("hiiamsid/llama-3.2-vision-11B-ROCO")
model = AutoModelForImageTextToText.from_pretrained("hiiamsid/llama-3.2-vision-11B-ROCO")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use hiiamsid/llama-3.2-vision-11B-ROCO with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "hiiamsid/llama-3.2-vision-11B-ROCO"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hiiamsid/llama-3.2-vision-11B-ROCO",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/hiiamsid/llama-3.2-vision-11B-ROCO

SGLang

How to use hiiamsid/llama-3.2-vision-11B-ROCO with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "hiiamsid/llama-3.2-vision-11B-ROCO" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hiiamsid/llama-3.2-vision-11B-ROCO",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "hiiamsid/llama-3.2-vision-11B-ROCO" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hiiamsid/llama-3.2-vision-11B-ROCO",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use hiiamsid/llama-3.2-vision-11B-ROCO with Docker Model Runner:
```
docker model run hf.co/hiiamsid/llama-3.2-vision-11B-ROCO
```

Model Card for hiiamsid/llama-3.2-vision-11B-ROCO

This is the finetuned version of meta-llama/Llama-3.2-11B-Vision-Instruct trained on MedIR/roco dataset using FSDP on 2 A100s.

Model Details

Model Description

Developed by: hiiamsid
Model type: multimodal (Image/Text to Text)
Language(s) (NLP): multilingual
License: Apache License 2.0
Finetuned from model [optional]: meta-llama/Llama-3.2-11B-Vision-Instruct

How to Get Started with the Model


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]))

Training Details

Training Data

MedIR/roco: https://huggingface.co/datasets/MedIR/roco (only 1000 samples where used for training)

Training Procedure

-Trained using FSDP activating wraping policy, MixedPrecision Policy (on bfloat16), activationcheckpointing etc and saved using Type FULL_STATE_DICT

Training Hyperparameters

  @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"

Model Architecture and Objective

This was just trained to see how much improvement can be seen when finetuned llama 3.2 vision.

Compute Infrastructure

Trained on 2 A100 (80GB) from runpods.

Citation

https://github.com/meta-llama/llama-recipes [More Information Needed]

Downloads last month: 1

Safetensors

Model size

11B params

Tensor type

BF16

Model tree for hiiamsid/llama-3.2-vision-11B-ROCO

Base model

meta-llama/Llama-3.2-11B-Vision-Instruct

Finetuned

(163)

this model

hiiamsid
/

llama-3.2-vision-11B-ROCO