feat: support LongCat-Image-Edit on cuda device.

[Dragonliu2018]

This PR supports LongCat-Image-Edit on cuda device.

The test program and the generated image are as follows.

• OS: Ubuntu22.04
• device: NVIDIA A800-SXM4-40GB
• model: LongCat-Image-Edit

import requests
import json
import base64
from PIL import Image
from io import BytesIO

# Test prompt for image generation
url = "http://localhost:9977/v1/image/generation"

img = Image.open("cat.png").convert("RGB")
buf = BytesIO()
img.save(buf, format="PNG")        # 和服务端 OpenCV 解码兼容
img_bytes = buf.getvalue()
image_base64 = base64.b64encode(img_bytes).decode("utf-8")

prompt = "将猫变成狗"
request_data = {
    "model": "LongCat-Image-Edit",
    "input": {
        "prompt": prompt,
        "negative_prompt": "",
        "image": image_base64
    },
    "parameters": {
        "guidance_scale": 1,
        "num_inference_steps": 8,
        "num_images_per_prompt": 1,
        "seed":43
    }
}

print("Testing LongCat-Image-Edit model...")
print(f"Request URL: {url}")
print(f"Request data: {json.dumps(request_data, indent=2, ensure_ascii=False)}")

response = requests.post(url, json=request_data)
if response.status_code != 200:
    print(f"Error: {response.status_code}")
    print(f"Response: {response.text}")
else:
    try:
        result = json.loads(response.text)
        print("Success! Response:")
        print(json.dumps(result, indent=2, ensure_ascii=False))
        
        # Handle image response
        if "output" in result and "results" in result["output"]:
            for i, image_data in enumerate(result["output"]["results"]):
                if "image" in image_data:
                    # Decode base64 image
                    image_bytes = base64.b64decode(image_data["image"])
                    image = Image.open(BytesIO(image_bytes))
                    
                    # Save image
                    filename = f"edited_image_{i+1}.png"
                    image.save(filename)
                    print(f"\nGenerated image saved as: {filename}")
                    print(f"Image size: {image_data.get('width', 'unknown')}x{image_data.get('height', 'unknown')}")
                    print(f"Seed: {image_data.get('seed', 'unknown')}")
    except json.JSONDecodeError as e:
        print(f"Failed to parse JSON response: {e}")
        print(f"Raw response: {response.text}")

---

Input image:

Output image:

[gemini-code-assist]

Code Review

This pull request adds support for the LongCat-Image-Edit model on CUDA devices. The changes include a new pipeline implementation, modifications to the model loader to handle preprocessor configurations, and several improvements and bug fixes in the attention and rotary embedding kernels. The implementation of the new pipeline is comprehensive. I've identified one issue in the model loader's error handling for the preprocessor configuration that could lead to silent failures.