feat(grounding): pipeline centralisé + serveur UI-TARS transformers + nettoyage code mort
Architecture grounding complète :
- core/grounding/server.py : serveur FastAPI (port 8200) avec UI-TARS-1.5-7B en 4-bit NF4
Process séparé avec son propre contexte CUDA (résout le crash Flask/CUDA)
- core/grounding/pipeline.py : orchestrateur cascade template→OCR→UI-TARS→static
- core/grounding/template_matcher.py : TemplateMatcher centralisé (remplace 5 copies)
- core/grounding/ui_tars_grounder.py : client HTTP vers le serveur de grounding
- core/grounding/target.py : GroundingTarget + GroundingResult
ORA modifié :
- _act_click() : capture unique de l'écran envoyée au serveur de grounding
- Pre-check VLM skippé pour ui_tars (redondant, et Ollama n'a plus de VRAM)
- verify_level='none' par défaut (vérification titre OCR prévue en Phase 2)
- Détection réponses négatives UI-TARS ("I don't see it" → fallback OCR)
Nettoyage :
- 9 fichiers morts archivés dans _archive/ (~6300 lignes supprimées)
- 21 tests ajoutés pour TemplateMatcher
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Dom
433
core/grounding/server.py
Normal file
433
core/grounding/server.py
Normal file
@@ -0,0 +1,433 @@
|
||||
"""
|
||||
core/grounding/server.py — Serveur FastAPI de grounding visuel (port 8200)
|
||||
|
||||
Charge UI-TARS-1.5-7B en 4-bit NF4 dans son propre process Python avec son
|
||||
propre contexte CUDA. Le backend Flask VWB (port 5002) et la boucle ORA
|
||||
appellent ce serveur en HTTP au lieu de charger le modele in-process.
|
||||
|
||||
Lancement :
|
||||
.venv/bin/python3 -m core.grounding.server
|
||||
|
||||
Endpoints :
|
||||
GET /health — verifie que le modele est charge
|
||||
POST /ground — localise un element UI sur un screenshot
|
||||
"""
|
||||
|
||||
import base64
|
||||
import gc
|
||||
import io
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
import time
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
from fastapi import FastAPI, HTTPException
|
||||
from pydantic import BaseModel
|
||||
import uvicorn
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Configuration
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
PORT = int(os.environ.get("GROUNDING_PORT", 8200))
|
||||
MODEL_ID = "ByteDance-Seed/UI-TARS-1.5-7B"
|
||||
MIN_PIXELS = 100 * 28 * 28
|
||||
MAX_PIXELS = 16384 * 28 * 28
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Smart resize — identique a /tmp/test_uitars.py
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _smart_resize(height: int, width: int, factor: int = 28,
|
||||
min_pixels: int = MIN_PIXELS, max_pixels: int = MAX_PIXELS):
|
||||
"""UI-TARS smart resize (memes defaults que le test valide)."""
|
||||
h_bar = max(factor, round(height / factor) * factor)
|
||||
w_bar = max(factor, round(width / factor) * factor)
|
||||
if h_bar * w_bar > max_pixels:
|
||||
beta = math.sqrt((height * width) / max_pixels)
|
||||
h_bar = math.floor(height / beta / factor) * factor
|
||||
w_bar = math.floor(width / beta / factor) * factor
|
||||
elif h_bar * w_bar < min_pixels:
|
||||
beta = math.sqrt(min_pixels / (height * width))
|
||||
h_bar = math.ceil(height * beta / factor) * factor
|
||||
w_bar = math.ceil(width * beta / factor) * factor
|
||||
return h_bar, w_bar
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Prompt officiel UI-TARS — identique a /tmp/test_uitars.py
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
_GROUNDING_PROMPT = """You are a GUI agent. You are given a task and your action history, with screenshots. You need to perform the next action to complete the task.
|
||||
|
||||
## Output Format
|
||||
|
||||
Thought: ...
|
||||
Action: ...
|
||||
|
||||
|
||||
## Action Space
|
||||
click(start_box='(x1, y1)')
|
||||
|
||||
|
||||
## User Instruction
|
||||
{instruction}"""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Modele singleton
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
_model = None
|
||||
_processor = None
|
||||
_model_loaded = False
|
||||
|
||||
|
||||
def _evict_ollama_models():
|
||||
"""Libere les modeles Ollama de la VRAM avant de charger UI-TARS."""
|
||||
try:
|
||||
import requests
|
||||
try:
|
||||
ps_resp = requests.get('http://localhost:11434/api/ps', timeout=3)
|
||||
if ps_resp.status_code == 200:
|
||||
loaded = ps_resp.json().get('models', [])
|
||||
model_names = [m.get('name', '') for m in loaded if m.get('name')]
|
||||
else:
|
||||
model_names = []
|
||||
except Exception:
|
||||
model_names = []
|
||||
|
||||
if not model_names:
|
||||
print("[grounding-server] Aucun modele Ollama en VRAM")
|
||||
return
|
||||
|
||||
for model_name in model_names:
|
||||
try:
|
||||
requests.post(
|
||||
'http://localhost:11434/api/generate',
|
||||
json={'model': model_name, 'keep_alive': '0'},
|
||||
timeout=5,
|
||||
)
|
||||
print(f"[grounding-server] Ollama: eviction de '{model_name}'")
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
time.sleep(1.0)
|
||||
print("[grounding-server] Modeles Ollama liberes")
|
||||
except ImportError:
|
||||
print("[grounding-server] requests non dispo, skip eviction Ollama")
|
||||
|
||||
|
||||
def _load_model():
|
||||
"""Charge UI-TARS-1.5-7B en 4-bit NF4 — code identique a /tmp/test_uitars.py."""
|
||||
global _model, _processor, _model_loaded
|
||||
|
||||
if _model_loaded:
|
||||
return
|
||||
|
||||
print("=" * 60)
|
||||
print(f"[grounding-server] Chargement de {MODEL_ID}")
|
||||
print("=" * 60)
|
||||
|
||||
if not torch.cuda.is_available():
|
||||
raise RuntimeError("CUDA non disponible — le serveur de grounding necessite un GPU")
|
||||
|
||||
# Liberer la VRAM Ollama
|
||||
_evict_ollama_models()
|
||||
|
||||
torch.cuda.empty_cache()
|
||||
gc.collect()
|
||||
|
||||
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor, BitsAndBytesConfig
|
||||
|
||||
bnb_config = BitsAndBytesConfig(
|
||||
load_in_4bit=True,
|
||||
bnb_4bit_quant_type="nf4",
|
||||
bnb_4bit_compute_dtype=torch.bfloat16,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
)
|
||||
|
||||
t0 = time.time()
|
||||
_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
||||
MODEL_ID,
|
||||
quantization_config=bnb_config,
|
||||
device_map="auto",
|
||||
)
|
||||
_model.eval()
|
||||
|
||||
_processor = AutoProcessor.from_pretrained(
|
||||
MODEL_ID,
|
||||
min_pixels=MIN_PIXELS,
|
||||
max_pixels=MAX_PIXELS,
|
||||
)
|
||||
|
||||
_model_loaded = True
|
||||
load_time = time.time() - t0
|
||||
alloc = torch.cuda.memory_allocated() / 1024**3
|
||||
peak = torch.cuda.max_memory_allocated() / 1024**3
|
||||
print(f"[grounding-server] Modele charge en {load_time:.1f}s | "
|
||||
f"VRAM: {alloc:.2f} GB (peak: {peak:.2f} GB)")
|
||||
|
||||
|
||||
def _capture_screen():
|
||||
"""Capture l'ecran complet via mss. Retourne PIL Image ou None."""
|
||||
try:
|
||||
import mss as mss_lib
|
||||
from PIL import Image
|
||||
with mss_lib.mss() as sct:
|
||||
mon = sct.monitors[0]
|
||||
grab = sct.grab(mon)
|
||||
return Image.frombytes('RGB', grab.size, grab.bgra, 'raw', 'BGRX')
|
||||
except Exception as e:
|
||||
print(f"[grounding-server] Erreur capture ecran: {e}")
|
||||
return None
|
||||
|
||||
|
||||
def _parse_coordinates(raw: str, orig_w: int, orig_h: int,
|
||||
resized_w: int, resized_h: int):
|
||||
"""Parse les coordonnees du modele — identique a /tmp/test_uitars.py.
|
||||
|
||||
Retourne (px, py, method_detail, confidence) ou None.
|
||||
"""
|
||||
cx, cy = None, None
|
||||
|
||||
# Format 1: <point>x y</point>
|
||||
pm = re.search(r'<point>\s*(\d+)\s+(\d+)\s*</point>', raw)
|
||||
if pm:
|
||||
cx, cy = int(pm.group(1)), int(pm.group(2))
|
||||
|
||||
# Format 2: start_box='(x, y)'
|
||||
if cx is None:
|
||||
bm = re.search(r"start_box=\s*['\"]?\((\d+)\s*,\s*(\d+)\)", raw)
|
||||
if bm:
|
||||
cx, cy = int(bm.group(1)), int(bm.group(2))
|
||||
|
||||
# Format 3: fallback x, y
|
||||
if cx is None:
|
||||
fm = re.search(r'(\d+)\s*,\s*(\d+)', raw)
|
||||
if fm:
|
||||
cx, cy = int(fm.group(1)), int(fm.group(2))
|
||||
|
||||
if cx is None or cy is None:
|
||||
return None
|
||||
|
||||
# Conversion : tester les 2 interpretations, garder la meilleure
|
||||
# Methode A : coordonnees dans l'espace de l'image resizee
|
||||
px_r = int(cx / resized_w * orig_w)
|
||||
py_r = int(cy / resized_h * orig_h)
|
||||
delta_r = ((px_r - orig_w / 2) ** 2 + (py_r - orig_h / 2) ** 2) ** 0.5
|
||||
|
||||
# Methode B : coordonnees 0-1000
|
||||
px_1k = int(cx / 1000 * orig_w)
|
||||
py_1k = int(cy / 1000 * orig_h)
|
||||
delta_1k = ((px_1k - orig_w / 2) ** 2 + (py_1k - orig_h / 2) ** 2) ** 0.5
|
||||
|
||||
# Heuristique du script valide : si coords dans les limites du resize,
|
||||
# les deux sont possibles. UI-TARS utilise l'espace resize en natif.
|
||||
if cx <= resized_w and cy <= resized_h:
|
||||
in_screen_r = (0 <= px_r <= orig_w and 0 <= py_r <= orig_h)
|
||||
in_screen_1k = (0 <= px_1k <= orig_w and 0 <= py_1k <= orig_h)
|
||||
|
||||
if in_screen_r and in_screen_1k:
|
||||
px, py = px_r, py_r
|
||||
method_detail = "resized"
|
||||
elif in_screen_r:
|
||||
px, py = px_r, py_r
|
||||
method_detail = "resized"
|
||||
else:
|
||||
px, py = px_1k, py_1k
|
||||
method_detail = "0-1000"
|
||||
else:
|
||||
px, py = px_1k, py_1k
|
||||
method_detail = "0-1000"
|
||||
|
||||
confidence = 0.85 if ("start_box" in raw or "<point>" in raw) else 0.70
|
||||
|
||||
print(f"[grounding-server] model=({cx},{cy}) -> pixel=({px},{py}) "
|
||||
f"[{method_detail}] resized={resized_w}x{resized_h} orig={orig_w}x{orig_h}")
|
||||
|
||||
return px, py, method_detail, confidence
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# FastAPI app
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
app = FastAPI(title="RPA Vision Grounding Server", version="1.0.0")
|
||||
|
||||
|
||||
class GroundRequest(BaseModel):
|
||||
target_text: str = ""
|
||||
target_description: str = ""
|
||||
image_b64: str = ""
|
||||
|
||||
|
||||
class GroundResponse(BaseModel):
|
||||
x: Optional[int] = None
|
||||
y: Optional[int] = None
|
||||
method: str = "ui_tars"
|
||||
confidence: float = 0.85
|
||||
time_ms: float = 0.0
|
||||
raw_output: str = ""
|
||||
|
||||
|
||||
@app.get("/health")
|
||||
def health():
|
||||
return {
|
||||
"status": "ok" if _model_loaded else "loading",
|
||||
"model": MODEL_ID,
|
||||
"model_loaded": _model_loaded,
|
||||
"cuda_available": torch.cuda.is_available(),
|
||||
"vram_allocated_gb": round(torch.cuda.memory_allocated() / 1024**3, 2) if torch.cuda.is_available() else 0,
|
||||
}
|
||||
|
||||
|
||||
@app.post("/ground", response_model=GroundResponse)
|
||||
def ground(req: GroundRequest):
|
||||
if not _model_loaded:
|
||||
raise HTTPException(status_code=503, detail="Modele pas encore charge")
|
||||
|
||||
from PIL import Image
|
||||
from qwen_vl_utils import process_vision_info
|
||||
|
||||
# Construire l'instruction
|
||||
parts = []
|
||||
if req.target_text:
|
||||
parts.append(req.target_text)
|
||||
if req.target_description:
|
||||
parts.append(req.target_description)
|
||||
if not parts:
|
||||
raise HTTPException(status_code=400, detail="target_text ou target_description requis")
|
||||
|
||||
instruction = f"Click on the {' — '.join(parts)}"
|
||||
|
||||
# Obtenir l'image (fournie en b64 ou capture ecran)
|
||||
if req.image_b64:
|
||||
try:
|
||||
raw_b64 = req.image_b64.split(',')[1] if ',' in req.image_b64 else req.image_b64
|
||||
img_data = base64.b64decode(raw_b64)
|
||||
screen_pil = Image.open(io.BytesIO(img_data)).convert('RGB')
|
||||
except Exception as e:
|
||||
raise HTTPException(status_code=400, detail=f"Erreur decodage image: {e}")
|
||||
else:
|
||||
screen_pil = _capture_screen()
|
||||
if screen_pil is None:
|
||||
raise HTTPException(status_code=500, detail="Capture ecran echouee")
|
||||
|
||||
W, H = screen_pil.size
|
||||
rH, rW = _smart_resize(H, W, min_pixels=MIN_PIXELS, max_pixels=MAX_PIXELS)
|
||||
|
||||
# Sauver temporairement l'image pour qwen_vl_utils
|
||||
import tempfile
|
||||
tmp_path = os.path.join(tempfile.gettempdir(), f"grounding_screen_{os.getpid()}.png")
|
||||
screen_pil.save(tmp_path)
|
||||
|
||||
try:
|
||||
system_prompt = _GROUNDING_PROMPT.format(instruction=instruction)
|
||||
|
||||
messages = [
|
||||
{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{
|
||||
"type": "image",
|
||||
"image": f"file://{tmp_path}",
|
||||
"min_pixels": MIN_PIXELS,
|
||||
"max_pixels": MAX_PIXELS,
|
||||
},
|
||||
{
|
||||
"type": "text",
|
||||
"text": system_prompt,
|
||||
},
|
||||
],
|
||||
}
|
||||
]
|
||||
|
||||
text = _processor.apply_chat_template(
|
||||
messages, tokenize=False, add_generation_prompt=True
|
||||
)
|
||||
image_inputs, video_inputs = process_vision_info(messages)
|
||||
inputs = _processor(
|
||||
text=[text],
|
||||
images=image_inputs,
|
||||
videos=video_inputs,
|
||||
padding=True,
|
||||
return_tensors="pt",
|
||||
).to(_model.device)
|
||||
|
||||
# Inference
|
||||
t0 = time.time()
|
||||
with torch.no_grad():
|
||||
gen = _model.generate(**inputs, max_new_tokens=256)
|
||||
infer_ms = (time.time() - t0) * 1000
|
||||
|
||||
# Decoder
|
||||
trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, gen)]
|
||||
raw = _processor.batch_decode(
|
||||
trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
||||
)[0].strip()
|
||||
|
||||
print(f"[grounding-server] '{instruction}' -> raw='{raw[:150]}' ({infer_ms:.0f}ms)")
|
||||
|
||||
# Détecter les réponses négatives (le modèle dit qu'il ne voit pas l'élément)
|
||||
_raw_lower = raw.lower()
|
||||
_negative_markers = ["don't see", "do not see", "cannot find", "can't find",
|
||||
"not visible", "not found", "doesn't appear", "does not appear",
|
||||
"i don't", "unable to find", "unable to locate"]
|
||||
for _neg in _negative_markers:
|
||||
if _neg in _raw_lower:
|
||||
print(f"[grounding-server] NÉGATIF détecté: '{_neg}' → élément non trouvé")
|
||||
return GroundResponse(x=None, y=None, method="ui_tars", confidence=0.0,
|
||||
time_ms=round(infer_ms, 1), raw_output=raw[:300])
|
||||
|
||||
# Parser les coordonnees
|
||||
parsed = _parse_coordinates(raw, W, H, rW, rH)
|
||||
if parsed is None:
|
||||
raise HTTPException(
|
||||
status_code=422,
|
||||
detail=f"Coordonnees non parsees dans la reponse: {raw[:200]}"
|
||||
)
|
||||
|
||||
px, py, method_detail, confidence = parsed
|
||||
|
||||
print(f"[grounding-server] Resultat: ({px}, {py}) conf={confidence:.2f} "
|
||||
f"[{method_detail}] ({infer_ms:.0f}ms)")
|
||||
|
||||
return GroundResponse(
|
||||
x=px,
|
||||
y=py,
|
||||
method="ui_tars",
|
||||
confidence=confidence,
|
||||
time_ms=round(infer_ms, 1),
|
||||
raw_output=raw[:300],
|
||||
)
|
||||
|
||||
finally:
|
||||
try:
|
||||
os.unlink(tmp_path)
|
||||
except OSError:
|
||||
pass
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Entrypoint
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@app.on_event("startup")
|
||||
async def startup_event():
|
||||
"""Charge le modele au demarrage du serveur."""
|
||||
print(f"[grounding-server] Demarrage sur port {PORT}...")
|
||||
_load_model()
|
||||
print(f"[grounding-server] Pret a recevoir des requetes sur http://localhost:{PORT}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
uvicorn.run(
|
||||
"core.grounding.server:app",
|
||||
host="0.0.0.0",
|
||||
port=PORT,
|
||||
log_level="info",
|
||||
workers=1, # 1 seul worker (1 seul GPU)
|
||||
)
|
||||
Reference in New Issue
Block a user