app.py

import gradio as gr
import onnx
from huggingface_hub import HfApi
import json
import sys
import os
import io
import requests
from urllib.parse import urlparse


HfClient = HfApi();

ONNX_PREFERED = [
    "model.onnx",
    "onnx/model.onnx"    
]

ONNX_CACHE = {
    'models': {
        
    }
}

def is_url(path):
    try:
        result = urlparse(path)
        return all([result.scheme, result.netloc])
    except ValueError:
        return False


def load_model(path):
    if is_url(path):
        print(f"Downloading model from: {path}...", file=sys.stderr)
        try:
            response = requests.get(path)
            response.raise_for_status() # Check for HTTP errors
            # Load from binary stream
            return onnx.load(io.BytesIO(response.content))
        except requests.exceptions.RequestException as e:
            print(f"Error downloading model: {e}", file=sys.stderr)
            sys.exit(1)
    else:
        # Check if local file exists before loading
        if not os.path.exists(path):
            print(f"Error: File not found at {path}", file=sys.stderr)
            sys.exit(1)
        
        return onnx.load(path)

def CheckSqlOnnx(path):
    
    OnnxModel = load_model(path);
    
    initializer_names = {init.name for init in OnnxModel.graph.initializer}
    inputs = [inp.name for inp in OnnxModel.graph.input if inp.name not in initializer_names]
    outputs = [out.name for out in OnnxModel.graph.output]

    required_inputs = {"input_ids", "attention_mask"}
    required_outputs = {"token_embeddings", "sentence_embedding"}

    is_supported = (
        required_inputs.issubset(inputs) and
        required_outputs.issubset(outputs)
    )

    OnnxInouts = {
        "supported": is_supported,
        "inputs": inputs,
        "outputs": outputs
    }
    
    return OnnxInouts;

def CheckModel(repo_id: str, path: str | None = None):
    
    MODELS_CACHE = ONNX_CACHE['models'];
    
    CacheSlot = MODELS_CACHE.get(repo_id);
    
    if CacheSlot:
        return json.dumps(CacheSlot, indent = 2);
    
    
    
    model_info = HfClient.model_info(repo_id=repo_id)

    # Extract filenames from RepoSibling objects
    sibling_files = [s.rfilename for s in model_info.siblings]

    onnx_path = None

    if path:
        if path in sibling_files:
            onnx_path = path
        else:
            return f"Error: ONNX file not found: {path}"
    else:
        for p in ONNX_PREFERED:
            if p in sibling_files:
                onnx_path = p
                break

        if not onnx_path:
            onnx_path = next(
                (f for f in sibling_files if f.lower().endswith(".onnx")),
                None
            )

    if not onnx_path:
        raise "Error: No ONNX model found in repository";

    # Build Hugging Face raw file URL
    file_url = f"https://huggingface.co/{repo_id}/resolve/main/{onnx_path}"

    # Check SQL ONNX compatibility
    OnnxInfo = CheckSqlOnnx(file_url)
    
    CacheSlot = {
         'url': file_url
        ,'onnx': OnnxInfo
    }
    
    MODELS_CACHE[repo_id] = CacheSlot

    return json.dumps({**CacheSlot, 'cached': False}, indent = 2);


with gr.Blocks() as demo:
    
    gr.Markdown("""
        This sample app test if a given model repository can be used with SQL Server ONNX.  
        In some tests, discovered that is not any ONNX model that works with sql server CREATE EXTERNAL MODEL.  
        For works, the input parameters of neural network must contains specific names, and output also.  
        I dont know if this behavior will be ketp in future verisons of SQL 2025...  
        
        But, while we dont have official doc about this, this repo can help discovery if a given model will work with sql server if you plan download to use with ONNX.  
        Just input the model name bellow in format user/model-name (check examples).  
        
        Look at JSON output. If "supported" is True, then you can use with SQL...
        
        Soon bring a default tested list!
        
        **IMPORTANT**: To check, this space will attempt donwload the model onnx file. If is big can take several minutes.
    """)
    
    ModelPath = gr.Textbox(label="Model Repository", submit_btn = True);
    ModelInfoOut = gr.Textbox(label="Model Info", lines = 10)
    
    ModelPath.submit(fn=CheckModel, inputs=ModelPath, outputs=ModelInfoOut)
    
    gr.Examples([
         ["intfloat/multilingual-e5-large"]
        ,["mixedbread-ai/mxbai-embed-xsmall-v1"]
        ,["nsense/all-MiniLM-L6-v2-onnx"]
    
    ], ModelPath)

demo.launch(
    server_name = '0.0.0.0'
)