[Misc]: Curious why this is happening: Running phi-3-vision on a RTX 3070 (8GB VRAM) works with transformer but not with vllm (goes out of memory)

[chandeldivyam]

Anything you want to discuss about vllm.

I was wondering why does this happen? I am new to this space and was playing around with different machines, models and frameworks.

I am able to inference single image (on RTX3070) in around 70s using huggingface transformer. Tried similar thing using vllm (current main branch), it got out of memory which got me curious.

from transformers import AutoModelForCausalLM, AutoProcessor
from PIL import Image
import torch

model_id = "microsoft/Phi-3-vision-128k-instruct"
device = "cuda:0"

model = AutoModelForCausalLM.from_pretrained(model_id, cache_dir="/content/my_models/phi_3_vision",
                                             device_map="cuda",
                                             trust_remote_code=True,
                                             torch_dtype="auto",
                                             _attn_implementation="eager")

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

def process_image(image_path):
    """Processes a single image and returns the model's response."""
    messages = [
        {
            "role": "user",
            "content": "<|image_1|>\nWhat is the destination address?",
        }
    ]

    prompt = processor.tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    image = Image.open(image_path)

    inputs = processor(prompt, [image], return_tensors="pt").to("cuda:0")

    generation_args = {
        "max_new_tokens": 500,
        "temperature": 0.0,
        "do_sample": False,
    }

    generate_ids = model.generate(
        **inputs, eos_token_id=processor.tokenizer.eos_token_id, **generation_args
    )

    generate_ids = generate_ids[:, inputs["input_ids"].shape[1] :]
    response = processor.batch_decode(
        generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )[0]
    return response

Vllm

import os
import subprocess

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizer

from PIL import Image

from vllm import LLM, SamplingParams
from vllm.multimodal.image import ImagePixelData

def run_phi3v():
    model_path = "microsoft/Phi-3-vision-128k-instruct"
    llm = LLM(
        model=model_path,
        trust_remote_code=True,
        image_input_type="pixel_values",
        image_token_id=32044,
        image_input_shape="1,3,1008,1344",
        image_feature_size=1921,
        disable_image_processor=False,
    gpu_memory_utilization=0.7,
    )

    image = Image.open("images/iamge2.png")

    # single-image prompt
    prompt = "<|user|>\n<|image_1|>\nWhat is the destination address?<|end|>\n<|assistant|>\n"  # noqa: E501
    prompt = prompt.replace("<|image_1|>", "<|image|>" * 1921 + "<s>")

    sampling_params = SamplingParams(temperature=0, max_tokens=64)

    outputs = llm.generate(
        {
            "prompt": prompt,
            "multi_modal_data": ImagePixelData(image),
        },
        sampling_params=sampling_params)
    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)

if __name__ == "__main__":
    local_directory = "images"

    # Make sure the local directory exists or create it
    os.makedirs(local_directory, exist_ok=True)

    run_phi3v()

[DarkLight1337]

Can you try setting gpu_memory_utilization to a higher value? It controls the proportion of GPU memory vLLM is allowed to use (the default is 0.9).

[chandeldivyam]

Thanks for the response, tried 3 values = [1, 0.9, 0.7] but neither worked. @DarkLight1337

[DarkLight1337]

Thanks to @ywang96 we have figured out the reason. The model has 128k context length by default so it might not fit in your GPU. Try passing max_model_len=8192 (or some other value that lets it fit in your GPU) to LLM in the example.

[chandeldivyam]

Thanks @DarkLight1337 and @ywang96

It worked like charm