Spaces:
Paused
Paused
encryptd commited on
Commit ·
cd46a2d
1
Parent(s): 5c7ad0b
prog update
Browse files
app.py
CHANGED
|
@@ -4,10 +4,9 @@ import sys
|
|
| 4 |
import httpx
|
| 5 |
import json
|
| 6 |
import base64
|
| 7 |
-
import time
|
| 8 |
from io import BytesIO
|
| 9 |
-
from fastapi import Request
|
| 10 |
-
from fastapi.responses import StreamingResponse, JSONResponse
|
| 11 |
import uvicorn
|
| 12 |
import gradio as gr
|
| 13 |
from openai import OpenAI
|
|
@@ -19,12 +18,12 @@ MAX_MODEL_LEN = 16384
|
|
| 19 |
VLLM_PORT = 8000
|
| 20 |
HF_PORT = 7860
|
| 21 |
|
| 22 |
-
# --- STEP 1: START vLLM
|
| 23 |
def start_vllm():
|
| 24 |
if "VLLM_PID" in os.environ:
|
| 25 |
return
|
| 26 |
print("--- STARTING VLLM ---")
|
| 27 |
-
command =
|
| 28 |
"python3", "-m", "vllm.entrypoints.openai.api_server",
|
| 29 |
"--model", MODEL_ID,
|
| 30 |
"--host", "127.0.0.1",
|
|
@@ -35,66 +34,79 @@ def start_vllm():
|
|
| 35 |
"--dtype", "bfloat16",
|
| 36 |
"--limit-mm-per-prompt", '{"image": 1}'
|
| 37 |
]
|
| 38 |
-
# Keep stdout/stderr connected to sys so you can see vLLM errors in HF logs
|
| 39 |
proc = subprocess.Popen(command, stdout=sys.stdout, stderr=sys.stderr)
|
| 40 |
os.environ["VLLM_PID"] = str(proc.pid)
|
| 41 |
-
print(f"vLLM process started with PID: {proc.pid}")
|
| 42 |
|
| 43 |
start_vllm()
|
| 44 |
|
| 45 |
-
# --- STEP 2:
|
| 46 |
-
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
-
|
| 50 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
try:
|
| 54 |
with httpx.Client() as check:
|
| 55 |
check.get(f"http://127.0.0.1:{VLLM_PORT}/v1/models", timeout=2.0)
|
| 56 |
except Exception:
|
| 57 |
-
|
| 58 |
-
return "⏳ Model is still loading (takes 3-5 mins). Watch the logs for 'Uvicorn running on http://127.0.0.1:8000'."
|
| 59 |
|
| 60 |
client = OpenAI(base_url=f"http://127.0.0.1:{VLLM_PORT}/v1", api_key="EMPTY")
|
| 61 |
-
|
| 62 |
try:
|
| 63 |
-
# Pre-process image
|
| 64 |
-
print(">>> Converting image to RGB...")
|
| 65 |
image = image.convert("RGB")
|
| 66 |
buffered = BytesIO()
|
| 67 |
image.save(buffered, format="JPEG")
|
| 68 |
b64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
|
| 69 |
|
| 70 |
-
print(">>> Sending to vLLM...")
|
| 71 |
completion = client.chat.completions.create(
|
| 72 |
model=MODEL_ID,
|
| 73 |
-
messages=[{"role": "user", "content":
|
| 74 |
{"type": "text", "text": prompt or "Convert to markdown."},
|
| 75 |
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}}
|
| 76 |
]}],
|
| 77 |
max_tokens=4096,
|
| 78 |
timeout=300.0
|
| 79 |
)
|
| 80 |
-
|
| 81 |
content = completion.choices[0].message.content
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
if "</think>" in content:
|
| 85 |
-
return content.split("</think>")[-1].strip()
|
| 86 |
-
return content.strip()
|
| 87 |
-
|
| 88 |
except Exception as e:
|
| 89 |
-
print(f">>> ERROR: {str(e)}")
|
| 90 |
return f"❌ Error: {str(e)}"
|
| 91 |
|
| 92 |
-
|
| 93 |
-
with gr.Blocks() as demo:
|
| 94 |
gr.Markdown("# NuMarkdown L40S API Server")
|
|
|
|
| 95 |
with gr.Row():
|
| 96 |
with gr.Column():
|
| 97 |
-
img_input = gr.Image(type="pil", label="Input")
|
| 98 |
txt_input = gr.Textbox(value="Convert to markdown.", label="Prompt")
|
| 99 |
btn = gr.Button("Extract Markdown", variant="primary")
|
| 100 |
with gr.Column():
|
|
@@ -102,43 +114,14 @@ with gr.Blocks() as demo:
|
|
| 102 |
|
| 103 |
btn.click(run_ui_test, inputs=[img_input, txt_input], outputs=[out])
|
| 104 |
|
| 105 |
-
# --- STEP 4: MOUNTING
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
# demo.queue() is CRITICAL for the button to work on HF Spaces
|
| 109 |
-
demo.queue()
|
| 110 |
-
demo.render()
|
| 111 |
-
# 2. Patch the bug in Gradio 5 that causes the AttributeError
|
| 112 |
-
demo.max_file_size = 100 * 1024 * 1024
|
| 113 |
-
# 2. Grab the FastAPI app instance from Gradio
|
| 114 |
-
app = demo.app
|
| 115 |
-
|
| 116 |
-
# 3. Add the Proxy Route for your external Docling API
|
| 117 |
-
@app.api_route("/v1/{path:path}", methods=["GET", "POST", "PUT", "DELETE"])
|
| 118 |
-
async def gatekeeper_proxy(path: str, request: Request):
|
| 119 |
-
target_url = f"http://127.0.0.1:{VLLM_PORT}/v1/{path}"
|
| 120 |
-
headers = {k: v for k, v in request.headers.items() if k.lower() not in ["host", "content-length"]}
|
| 121 |
-
|
| 122 |
-
async with httpx.AsyncClient(timeout=300.0) as client:
|
| 123 |
-
try:
|
| 124 |
-
if path == "chat/completions" and request.method == "POST":
|
| 125 |
-
body = await request.json()
|
| 126 |
-
if not body.get("stream", False):
|
| 127 |
-
resp = await client.post(target_url, headers=headers, json=body)
|
| 128 |
-
if resp.status_code == 200:
|
| 129 |
-
data = resp.json()
|
| 130 |
-
content = data["choices"][0]["message"].get("content", "")
|
| 131 |
-
if "</think>" in content:
|
| 132 |
-
data["choices"][0]["message"]["content"] = content.split("</think>")[-1].strip()
|
| 133 |
-
return JSONResponse(content=data)
|
| 134 |
-
return JSONResponse(status_code=resp.status_code, content=resp.json())
|
| 135 |
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
return JSONResponse(status_code=503, content={"error": str(e)})
|
| 141 |
|
| 142 |
if __name__ == "__main__":
|
| 143 |
-
# Standard HF uvicorn launch
|
| 144 |
uvicorn.run(app, host="0.0.0.0", port=HF_PORT, workers=1)
|
|
|
|
| 4 |
import httpx
|
| 5 |
import json
|
| 6 |
import base64
|
|
|
|
| 7 |
from io import BytesIO
|
| 8 |
+
from fastapi import FastAPI, Request
|
| 9 |
+
from fastapi.responses import StreamingResponse, JSONResponse, RedirectResponse
|
| 10 |
import uvicorn
|
| 11 |
import gradio as gr
|
| 12 |
from openai import OpenAI
|
|
|
|
| 18 |
VLLM_PORT = 8000
|
| 19 |
HF_PORT = 7860
|
| 20 |
|
| 21 |
+
# --- STEP 1: START vLLM ---
|
| 22 |
def start_vllm():
|
| 23 |
if "VLLM_PID" in os.environ:
|
| 24 |
return
|
| 25 |
print("--- STARTING VLLM ---")
|
| 26 |
+
command =[
|
| 27 |
"python3", "-m", "vllm.entrypoints.openai.api_server",
|
| 28 |
"--model", MODEL_ID,
|
| 29 |
"--host", "127.0.0.1",
|
|
|
|
| 34 |
"--dtype", "bfloat16",
|
| 35 |
"--limit-mm-per-prompt", '{"image": 1}'
|
| 36 |
]
|
|
|
|
| 37 |
proc = subprocess.Popen(command, stdout=sys.stdout, stderr=sys.stderr)
|
| 38 |
os.environ["VLLM_PID"] = str(proc.pid)
|
|
|
|
| 39 |
|
| 40 |
start_vllm()
|
| 41 |
|
| 42 |
+
# --- STEP 2: FASTAPI GATEKEEPER (EXTERNAL API) ---
|
| 43 |
+
app = FastAPI()
|
| 44 |
+
|
| 45 |
+
@app.api_route("/v1/{path:path}", methods=["GET", "POST", "PUT", "DELETE"])
|
| 46 |
+
async def gatekeeper_proxy(path: str, request: Request):
|
| 47 |
+
"""Intercepts external Docling API calls and strips reasoning tags."""
|
| 48 |
+
target_url = f"http://127.0.0.1:{VLLM_PORT}/v1/{path}"
|
| 49 |
+
headers = {k: v for k, v in request.headers.items() if k.lower() not in ["host", "content-length"]}
|
| 50 |
|
| 51 |
+
async with httpx.AsyncClient(timeout=300.0) as client:
|
| 52 |
+
try:
|
| 53 |
+
if path == "chat/completions" and request.method == "POST":
|
| 54 |
+
body = await request.json()
|
| 55 |
+
if not body.get("stream", False):
|
| 56 |
+
resp = await client.post(target_url, headers=headers, json=body)
|
| 57 |
+
if resp.status_code == 200:
|
| 58 |
+
data = resp.json()
|
| 59 |
+
content = data["choices"][0]["message"].get("content", "")
|
| 60 |
+
# SUPPRESS REASONING TAGS FOR EXTERNAL API
|
| 61 |
+
if "</think>" in content:
|
| 62 |
+
data["choices"][0]["message"]["content"] = content.split("</think>")[-1].strip()
|
| 63 |
+
return JSONResponse(content=data)
|
| 64 |
+
return JSONResponse(status_code=resp.status_code, content=resp.json())
|
| 65 |
+
|
| 66 |
+
proxy_req = client.build_request(request.method, target_url, headers=headers, content=await request.body())
|
| 67 |
+
r = await client.send(proxy_req, stream=True)
|
| 68 |
+
return StreamingResponse(r.aiter_raw(), status_code=r.status_code, headers=dict(r.headers))
|
| 69 |
+
except Exception as e:
|
| 70 |
+
return JSONResponse(status_code=503, content={"error": str(e)})
|
| 71 |
|
| 72 |
+
# --- STEP 3: GRADIO UI ---
|
| 73 |
+
def run_ui_test(image, prompt):
|
| 74 |
+
print(">>> UI Request Received")
|
| 75 |
+
if image is None: return "⚠️ Please upload an image."
|
| 76 |
+
|
| 77 |
try:
|
| 78 |
with httpx.Client() as check:
|
| 79 |
check.get(f"http://127.0.0.1:{VLLM_PORT}/v1/models", timeout=2.0)
|
| 80 |
except Exception:
|
| 81 |
+
return "⏳ Model is still loading (takes 3-5 mins). Please wait."
|
|
|
|
| 82 |
|
| 83 |
client = OpenAI(base_url=f"http://127.0.0.1:{VLLM_PORT}/v1", api_key="EMPTY")
|
|
|
|
| 84 |
try:
|
|
|
|
|
|
|
| 85 |
image = image.convert("RGB")
|
| 86 |
buffered = BytesIO()
|
| 87 |
image.save(buffered, format="JPEG")
|
| 88 |
b64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
|
| 89 |
|
|
|
|
| 90 |
completion = client.chat.completions.create(
|
| 91 |
model=MODEL_ID,
|
| 92 |
+
messages=[{"role": "user", "content":[
|
| 93 |
{"type": "text", "text": prompt or "Convert to markdown."},
|
| 94 |
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}}
|
| 95 |
]}],
|
| 96 |
max_tokens=4096,
|
| 97 |
timeout=300.0
|
| 98 |
)
|
|
|
|
| 99 |
content = completion.choices[0].message.content
|
| 100 |
+
return content.split("</think>")[-1].strip() if "</think>" in content else content
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
except Exception as e:
|
|
|
|
| 102 |
return f"❌ Error: {str(e)}"
|
| 103 |
|
| 104 |
+
with gr.Blocks(title="NuMarkdown API Server") as demo:
|
|
|
|
| 105 |
gr.Markdown("# NuMarkdown L40S API Server")
|
| 106 |
+
gr.Markdown("The Markdown API endpoint is live externally at `/v1/chat/completions`.")
|
| 107 |
with gr.Row():
|
| 108 |
with gr.Column():
|
| 109 |
+
img_input = gr.Image(type="pil", label="Input Document")
|
| 110 |
txt_input = gr.Textbox(value="Convert to markdown.", label="Prompt")
|
| 111 |
btn = gr.Button("Extract Markdown", variant="primary")
|
| 112 |
with gr.Column():
|
|
|
|
| 114 |
|
| 115 |
btn.click(run_ui_test, inputs=[img_input, txt_input], outputs=[out])
|
| 116 |
|
| 117 |
+
# --- STEP 4: SAFE MOUNTING ---
|
| 118 |
+
# We mount Gradio to a subpath to prevent it from fighting with the Root path and breaking CSS/JS.
|
| 119 |
+
app = gr.mount_gradio_app(app, demo, path="/ui")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 120 |
|
| 121 |
+
# Redirect anyone visiting the root URL to the UI.
|
| 122 |
+
@app.get("/")
|
| 123 |
+
def redirect_to_ui():
|
| 124 |
+
return RedirectResponse(url="/ui")
|
|
|
|
| 125 |
|
| 126 |
if __name__ == "__main__":
|
|
|
|
| 127 |
uvicorn.run(app, host="0.0.0.0", port=HF_PORT, workers=1)
|