README.md

metadata

license: apache-2.0

This repo deploys a pix2struct model to Inference Endpoints to a single GPU. This is not meant for multiple GPUs (single gpu replicas are ok; don't use the 4xT4 option, though) or CPU.

Options:

• modify the model_name with whichever model you want to use
  • if it is a private model, just add handler.py to that model repo and changed model_name to "./"
• modify the dtype to whichever one you want
  • see notes here for dtype tradeoffs

After deploying the model, inference can be done as follows:

import base64

with open("path/to/image", "rb") as f:
    b64 = base64.b64encode(f.read())

question = "question to model"

payload = {
    "inputs": {
        "image": [b64.decode("utf-8")], # for batched inference, send list of images/questions
        "question": [question] 
        }, 
    "parameters":{
        "max_new_tokens": 10, # can use any generation parameters
    }}

import requests

API_URL = "url_to_endpoint"
headers = {
    "Authorization": "Bearer HF_TOKEN",
    "Content-Type": "application/json"
}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query(payload)

# {'output': ['55%']}