InternVL3-1B model app

Dockerfile

@@ -0,0 +1,13 @@
+FROM python:3.10
+
+RUN useradd -m -u 1000 user
+USER user
+ENV PATH="/home/user/.local/bin:$PATH"
+
+WORKDIR /app
+
+COPY --chown=user ./requirements.txt requirements.txt
+RUN pip install --no-cache-dir --upgrade -r requirements.txt
+
+COPY --chown=user . /app
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

app.py

@@ -0,0 +1,25 @@
+import torch
+from fastapi import FastAPI, Depends
+from fastapi.responses import JSONResponse
+from authenticator import authenticate_token
+from payload_model import PayloadModel
+from models import InternVL3
+from internvl_utils import internvl_inference
+
+
+app = FastAPI()
+
+model = InternVL3("OpenGVLab/InternVL3-1B-Instruct")
+
+@app.get("/healthcheck")
+def healthcheck():
+    return JSONResponse(status_code=200, content={"status": "ok"})
+
+@app.post("/internvl_inference")
+async def inference(payload: PayloadModel, token: str = Depends(authenticate_token)):
+    try:
+        model_response = await internvl_inference(model, payload)
+        return JSONResponse(status_code=200, content={"status": "ok", "response": model_response})
+    except Exception as e:
+        print(f"Error: {e}")
+        return JSONResponse(status_code=500, content={"status": "error", "message": str(e)})

authenticator.py

@@ -0,0 +1,21 @@
+from fastapi import Header, HTTPException
+import os
+import jwt
+from dotenv import load_dotenv
+
+load_dotenv()
+
+secret_token = os.getenv("AUTH_TOKEN")
+
+async def authenticate_token(authorization: str = Header(...)):
+    token_type, token = authorization.split()
+    if token_type != "Bearer":
+        raise HTTPException(status_code=401, detail="Unauthorized")
+    try:
+        return jwt.decode(token, secret_token, algorithms=["HS256"])
+    except jwt.ExpiredSignatureError:
+        raise HTTPException(status_code=401, detail="Token has expired") from None
+    except (jwt.InvalidTokenError, IndexError):
+        raise HTTPException(status_code=401, detail="Invalid token") from None
+
+

internvl_utils.py

@@ -0,0 +1,115 @@
+import math
+import numpy as np
+import torch
+import torchvision.transforms as T
+from torchvision.transforms.functional import InterpolationMode
+from transformers import AutoConfig
+from models import InternVL3
+from payload_model import PayloadModel
+from models.misc_utils import convert_base64_to_pil
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+
+
+def build_transform(input_size):
+    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
+    transform = T.Compose([
+        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+        T.ToTensor(),
+        T.Normalize(mean=MEAN, std=STD)
+    ])
+    return transform
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    return best_ratio
+
+def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images
+
+def load_image(image, input_size=448, max_num=12):
+    # image = Image.open(image_file).convert('RGB')
+    pil_image = convert_base64_to_pil(image)
+    transform = build_transform(input_size=input_size)
+    images = dynamic_preprocess(pil_image, image_size=input_size, use_thumbnail=True, max_num=max_num)
+    pixel_values = [transform(image) for image in images]
+    pixel_values = torch.stack(pixel_values)
+    pixel_values = pixel_values.to(torch.bfloat16).to("cuda")
+    return pixel_values
+
+def split_model(model_name):
+    device_map = {}
+    world_size = torch.cuda.device_count()
+    config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+    num_layers = config.llm_config.num_hidden_layers
+    # Since the first GPU will be used for ViT, treat it as half a GPU.
+    num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
+    num_layers_per_gpu = [num_layers_per_gpu] * world_size
+    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
+    layer_cnt = 0
+    for i, num_layer in enumerate(num_layers_per_gpu):
+        for j in range(num_layer):
+            device_map[f'language_model.model.layers.{layer_cnt}'] = i
+            layer_cnt += 1
+    device_map['vision_model'] = 0
+    device_map['mlp1'] = 0
+    device_map['language_model.model.tok_embeddings'] = 0
+    device_map['language_model.model.embed_tokens'] = 0
+    device_map['language_model.output'] = 0
+    device_map['language_model.model.norm'] = 0
+    device_map['language_model.model.rotary_emb'] = 0
+    device_map['language_model.lm_head'] = 0
+    device_map[f'language_model.model.layers.{num_layers - 1}'] = 0
+
+    return device_map
+
+async def internvl_inference(model: InternVL3, payload: PayloadModel):
+    return await model(payload)

models/InternVL3/intervl3.py

@@ -0,0 +1,68 @@
+from transformers import AutoModel, AutoTokenizer
+import torch
+from payload_model import PayloadModel
+from internvl_utils import load_image
+from pydantic import BaseModel, Field
+from typing import Optional
+
+class InternVL3(BaseModel):
+    model_name: str
+    model: Optional[AutoModel] = None
+    tokenizer: Optional[AutoTokenizer] = None
+    generation_config: dict = Field(default_factory=lambda: {"max_new_tokens": 1024, "do_sample": True})
+
+    model_config = {
+        "arbitrary_types_allowed": True,
+        "from_attributes": True
+    }
+
+    def __init__(self, model_name: str, **kwargs):
+        super().__init__(model_name=model_name, **kwargs)
+        self.model = AutoModel.from_pretrained(
+            self.model_name,
+            torch_dtype=torch.bfloat16,
+            load_in_8bit=False,
+            low_cpu_mem_usage=True,
+            use_flash_attn=True,
+            trust_remote_code=True,
+            device_map="auto",
+        ).eval()
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            self.model_name,
+            trust_remote_code=True,
+            use_fast=False,
+        )
+
+    def get_query_prompt(self, prompt_keyword: str):
+        if prompt_keyword == "person_running":
+            query_prompt = """
+<image>\nCheck if person is running or not? If they are running
+respond with "Yes" else respond with "No". Limit your response to either "Yes" or "No"
+"""
+        else:
+            query_prompt = None
+        return query_prompt
+
+    def predict(self, payload: PayloadModel):
+        pixel_values = load_image(payload.image)
+        query_prompt = self.get_query_prompt(payload.prompt_keyword)
+        if query_prompt is None:
+            model_response = f"Invalid prompt keyword: {payload.prompt_keyword}"
+        else:
+            model_response = self.model.chat(
+                self.tokenizer,
+                pixel_values,
+                query_prompt,
+                generation_config=self.generation_config,
+            )
+
+        return model_response
+    
+    def extract_model_response(self, model_response: str):
+        return "Yes" in model_response
+
+    async def __call__(self, payload: PayloadModel):
+        model_response = self.predict(payload)
+        extracted_response = self.extract_model_response(model_response)
+        return extracted_response
+

models/__init__.py

@@ -0,0 +1,2 @@
+from .InternVL3.intervl3 import InternVL3
+from .misc_utils import *

models/misc_utils.py

@@ -0,0 +1,15 @@
+import cv2
+import numpy as np
+import base64
+from PIL import Image
+
+
+def convert_base64_to_cv2(base64_string: str):
+    return cv2.imdecode(np.frombuffer(base64.b64decode(base64_string), np.uint8), cv2.IMREAD_COLOR)
+
+def convert_cv2_to_pil(image: np.ndarray):
+    return Image.fromarray(image).convert('RGB')
+
+def convert_base64_to_pil(base64_string: str):
+    return convert_cv2_to_pil(convert_base64_to_cv2(base64_string))
+

payload_model.py

@@ -0,0 +1,6 @@
+from pydantic import BaseModel
+
+class PayloadModel(BaseModel):
+    """Type check for payload parameters"""
+    image: str
+    prompt_keyword: str

requirements.txt

@@ -0,0 +1,13 @@
+fastapi
+uvicorn[standard]
+transformers>=4.37.2
+decord
+numpy~=1.26.4
+torch
+torchvision
+pillow
+PyJWT
+opencv-python==4.9.0.80
+einops
+timm
+accelerate