Spaces:
Paused
Paused
encryptd commited on
Commit ·
903e822
1
Parent(s): 9139a1b
prog update
Browse files
app.py
CHANGED
|
@@ -1,128 +1,128 @@
|
|
| 1 |
import os
|
| 2 |
import subprocess
|
| 3 |
-
import time
|
| 4 |
import sys
|
| 5 |
import httpx
|
|
|
|
|
|
|
|
|
|
| 6 |
from fastapi import FastAPI, Request
|
| 7 |
-
from fastapi.responses import StreamingResponse
|
| 8 |
import uvicorn
|
| 9 |
import gradio as gr
|
| 10 |
-
from openai import OpenAI
|
| 11 |
-
import base64
|
| 12 |
-
from io import BytesIO
|
| 13 |
|
| 14 |
# --- CONFIGURATION ---
|
| 15 |
MODEL_ID = "numind/NuMarkdown-8B-Thinking"
|
| 16 |
GPU_UTILIZATION = 0.95
|
| 17 |
MAX_MODEL_LEN = 32768
|
| 18 |
VLLM_PORT = 8000
|
| 19 |
-
|
| 20 |
|
| 21 |
-
# --- STEP 1:
|
| 22 |
def start_vllm():
|
| 23 |
if "VLLM_PID" in os.environ:
|
| 24 |
return
|
| 25 |
-
|
| 26 |
-
print(f"Starting vLLM server on port {VLLM_PORT}...")
|
| 27 |
-
|
| 28 |
-
# JSON formatted limit string to fix parsing error
|
| 29 |
-
limit_mm_config = '{"image": 1}'
|
| 30 |
-
|
| 31 |
command = [
|
| 32 |
"vllm", "serve", MODEL_ID,
|
| 33 |
-
"--host", "
|
| 34 |
"--port", str(VLLM_PORT),
|
| 35 |
"--trust-remote-code",
|
| 36 |
"--gpu-memory-utilization", str(GPU_UTILIZATION),
|
| 37 |
"--max-model-len", str(MAX_MODEL_LEN),
|
| 38 |
"--dtype", "bfloat16",
|
| 39 |
-
"--limit-mm-per-prompt",
|
| 40 |
]
|
| 41 |
-
|
| 42 |
-
# Redirect stdout/stderr to see download progress
|
| 43 |
proc = subprocess.Popen(command, stdout=sys.stdout, stderr=sys.stderr)
|
| 44 |
os.environ["VLLM_PID"] = str(proc.pid)
|
| 45 |
-
|
| 46 |
-
# We do NOT block here anymore. We let vLLM load in the background
|
| 47 |
-
# while the UI starts. This allows you to see the UI immediately.
|
| 48 |
-
print("vLLM started in background. Please wait for model download...")
|
| 49 |
|
| 50 |
start_vllm()
|
| 51 |
|
| 52 |
-
# --- STEP 2: FASTAPI
|
| 53 |
app = FastAPI()
|
| 54 |
|
| 55 |
@app.api_route("/v1/{path:path}", methods=["GET", "POST", "PUT", "DELETE"])
|
| 56 |
-
async def
|
| 57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
async with httpx.AsyncClient() as client:
|
| 59 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
proxy_req = client.build_request(
|
| 61 |
-
request.method,
|
| 62 |
-
target_url,
|
| 63 |
-
headers=request.headers.raw,
|
| 64 |
-
content=await request.body(),
|
| 65 |
-
timeout=300.0
|
| 66 |
)
|
| 67 |
r = await client.send(proxy_req, stream=True)
|
| 68 |
-
return StreamingResponse(
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
background=None
|
| 73 |
-
)
|
| 74 |
-
except httpx.ConnectError:
|
| 75 |
-
return JSONResponse(status_code=503, content={"error": "Model is still loading. Please wait."})
|
| 76 |
|
| 77 |
-
# --- STEP 3: GRADIO UI ---
|
| 78 |
def run_ui_test(image, prompt):
|
| 79 |
-
if image is None:
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
client = OpenAI(base_url=f"http://localhost:{VLLM_PORT}/v1", api_key="EMPTY")
|
| 83 |
-
|
| 84 |
-
# Encode Image
|
| 85 |
try:
|
| 86 |
buffered = BytesIO()
|
| 87 |
image.save(buffered, format="JPEG")
|
| 88 |
b64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
|
| 89 |
-
|
| 90 |
-
return f"Error processing image: {e}"
|
| 91 |
-
|
| 92 |
-
if not prompt: prompt = "Convert to markdown."
|
| 93 |
-
|
| 94 |
-
try:
|
| 95 |
completion = client.chat.completions.create(
|
| 96 |
model=MODEL_ID,
|
| 97 |
messages=[{"role": "user", "content": [
|
| 98 |
-
{"type": "text", "text": prompt},
|
| 99 |
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}}
|
| 100 |
-
]}]
|
| 101 |
-
max_tokens=4096
|
| 102 |
)
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
return "⏳ Model is still downloading/loading... Check the 'Logs' tab. This takes 2-3 minutes on a fresh GPU."
|
| 106 |
except Exception as e:
|
| 107 |
-
return f"Error: {
|
| 108 |
|
| 109 |
-
with gr.Blocks() as demo:
|
| 110 |
-
gr.Markdown("# NuMarkdown L40S
|
| 111 |
-
gr.Markdown("
|
| 112 |
-
|
| 113 |
with gr.Row():
|
| 114 |
with gr.Column():
|
| 115 |
-
img_input = gr.Image(type="pil"
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
btn = gr.Button("Test Inference")
|
| 119 |
with gr.Column():
|
| 120 |
-
out = gr.
|
| 121 |
-
|
| 122 |
-
# FIXED: Passed both inputs [img_input, txt_input]
|
| 123 |
btn.click(run_ui_test, inputs=[img_input, txt_input], outputs=[out])
|
| 124 |
|
| 125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
|
| 127 |
if __name__ == "__main__":
|
| 128 |
-
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
import subprocess
|
|
|
|
| 3 |
import sys
|
| 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
|
| 10 |
import uvicorn
|
| 11 |
import gradio as gr
|
| 12 |
+
from openai import OpenAI
|
|
|
|
|
|
|
| 13 |
|
| 14 |
# --- CONFIGURATION ---
|
| 15 |
MODEL_ID = "numind/NuMarkdown-8B-Thinking"
|
| 16 |
GPU_UTILIZATION = 0.95
|
| 17 |
MAX_MODEL_LEN = 32768
|
| 18 |
VLLM_PORT = 8000
|
| 19 |
+
HF_PORT = int(os.getenv("PORT", 7860))
|
| 20 |
|
| 21 |
+
# --- STEP 1: START vLLM ---
|
| 22 |
def start_vllm():
|
| 23 |
if "VLLM_PID" in os.environ:
|
| 24 |
return
|
| 25 |
+
print("Launching vLLM...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
command = [
|
| 27 |
"vllm", "serve", MODEL_ID,
|
| 28 |
+
"--host", "127.0.0.1", # Local only
|
| 29 |
"--port", str(VLLM_PORT),
|
| 30 |
"--trust-remote-code",
|
| 31 |
"--gpu-memory-utilization", str(GPU_UTILIZATION),
|
| 32 |
"--max-model-len", str(MAX_MODEL_LEN),
|
| 33 |
"--dtype", "bfloat16",
|
| 34 |
+
"--limit-mm-per-prompt", '{"image": 1}'
|
| 35 |
]
|
|
|
|
|
|
|
| 36 |
proc = subprocess.Popen(command, stdout=sys.stdout, stderr=sys.stderr)
|
| 37 |
os.environ["VLLM_PID"] = str(proc.pid)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
start_vllm()
|
| 40 |
|
| 41 |
+
# --- STEP 2: FASTAPI GATEKEEPER ---
|
| 42 |
app = FastAPI()
|
| 43 |
|
| 44 |
@app.api_route("/v1/{path:path}", methods=["GET", "POST", "PUT", "DELETE"])
|
| 45 |
+
async def gatekeeper_proxy(path: str, request: Request):
|
| 46 |
+
"""
|
| 47 |
+
Proxies external API calls to internal vLLM and cleans the output.
|
| 48 |
+
"""
|
| 49 |
+
target_url = f"http://127.0.0.1:{VLLM_PORT}/v1/{path}"
|
| 50 |
+
|
| 51 |
+
# Strip host header to prevent Hugging Face routing loops
|
| 52 |
+
headers = {k: v for k, v in request.headers.items() if k.lower() not in ["host", "content-length"]}
|
| 53 |
+
|
| 54 |
async with httpx.AsyncClient() as client:
|
| 55 |
try:
|
| 56 |
+
# Handle API calls from Docling (Non-streaming)
|
| 57 |
+
if path == "chat/completions" and request.method == "POST":
|
| 58 |
+
body = await request.json()
|
| 59 |
+
|
| 60 |
+
# If non-streaming, we intercept and clean reasoning
|
| 61 |
+
if not body.get("stream", False):
|
| 62 |
+
resp = await client.post(target_url, headers=headers, json=body, timeout=300.0)
|
| 63 |
+
if resp.status_code == 200:
|
| 64 |
+
data = resp.json()
|
| 65 |
+
content = data["choices"][0]["message"].get("content", "")
|
| 66 |
+
# SUPPRESS REASONING
|
| 67 |
+
if "</think>" in content:
|
| 68 |
+
data["choices"][0]["message"]["content"] = content.split("</think>")[-1].strip()
|
| 69 |
+
return JSONResponse(content=data)
|
| 70 |
+
return JSONResponse(status_code=resp.status_code, content=resp.json())
|
| 71 |
+
|
| 72 |
+
# Fallback for Models list or Streaming requests
|
| 73 |
proxy_req = client.build_request(
|
| 74 |
+
request.method, target_url, headers=headers, content=await request.body()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
)
|
| 76 |
r = await client.send(proxy_req, stream=True)
|
| 77 |
+
return StreamingResponse(r.aiter_raw(), status_code=r.status_code, headers=dict(r.headers))
|
| 78 |
+
|
| 79 |
+
except Exception as e:
|
| 80 |
+
return JSONResponse(status_code=503, content={"error": f"Internal Proxy Error: {str(e)}"})
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
|
| 82 |
+
# --- STEP 3: GRADIO UI (INTERNAL TESTING) ---
|
| 83 |
def run_ui_test(image, prompt):
|
| 84 |
+
if image is None: return "Please upload an image."
|
| 85 |
+
client = OpenAI(base_url=f"http://127.0.0.1:{VLLM_PORT}/v1", api_key="EMPTY")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
try:
|
| 87 |
buffered = BytesIO()
|
| 88 |
image.save(buffered, format="JPEG")
|
| 89 |
b64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
|
| 90 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
completion = client.chat.completions.create(
|
| 92 |
model=MODEL_ID,
|
| 93 |
messages=[{"role": "user", "content": [
|
| 94 |
+
{"type": "text", "text": prompt or "Convert to markdown."},
|
| 95 |
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}}
|
| 96 |
+
]}]
|
|
|
|
| 97 |
)
|
| 98 |
+
content = completion.choices[0].message.content
|
| 99 |
+
return content.split("</think>")[-1].strip() if "</think>" in content else content
|
|
|
|
| 100 |
except Exception as e:
|
| 101 |
+
return f"Error: {e}"
|
| 102 |
|
| 103 |
+
with gr.Blocks(title="NuMarkdown API Server") as demo:
|
| 104 |
+
gr.Markdown("# NuMarkdown L40S API Server")
|
| 105 |
+
gr.Markdown("The API is available externally at `/v1/chat/completions` (OpenAI compatible)")
|
|
|
|
| 106 |
with gr.Row():
|
| 107 |
with gr.Column():
|
| 108 |
+
img_input = gr.Image(type="pil")
|
| 109 |
+
txt_input = gr.Textbox(value="Convert to markdown.")
|
| 110 |
+
btn = gr.Button("Run Local Test")
|
|
|
|
| 111 |
with gr.Column():
|
| 112 |
+
out = gr.Markdown()
|
|
|
|
|
|
|
| 113 |
btn.click(run_ui_test, inputs=[img_input, txt_input], outputs=[out])
|
| 114 |
|
| 115 |
+
# --- STEP 4: CORRECT MOUNTING ---
|
| 116 |
+
# We mount Gradio to a subpath to prevent it from hijacking asset routes
|
| 117 |
+
# Hugging Face will still show this UI at the root because of how it handles FastAPI mounts
|
| 118 |
+
app = gr.mount_gradio_app(app, demo, path="/ui")
|
| 119 |
+
|
| 120 |
+
# Redirect root to /ui for browser users
|
| 121 |
+
@app.get("/")
|
| 122 |
+
async def redirect_to_ui():
|
| 123 |
+
from fastapi.responses import RedirectResponse
|
| 124 |
+
return RedirectResponse(url="/ui")
|
| 125 |
|
| 126 |
if __name__ == "__main__":
|
| 127 |
+
# Use 1 worker only to prevent port conflicts on HF L40S
|
| 128 |
+
uvicorn.run(app, host="0.0.0.0", port=HF_PORT, workers=1)
|