Optimisation performance (30t/s --> 50t/s) & adaptation parsing json --> llama.cpp & Fix système mémoire

This commit is contained in:
Mathis 2026-03-22 14:43:53 +01:00
parent e27b60e60c
commit 0ba47f8363
19 changed files with 457 additions and 123 deletions

View file

@ -40,34 +40,60 @@ class LLMManager:
self,
prompt: str,
system_prompt: Optional[str] = None,
structured_history: Optional[list] = None,
vision_model: bool = False,
attachments: Optional[list] = None
) -> Dict[str, Any]:
"""Build the API payload for Ollama."""
payload = {
"model": self.ollama_model,
"prompt": prompt,
"stream": True,
"options": {"temperature": self.ollama_temperature}
}
"""Build the API payload for Ollama or OpenAI Chat API (llama.cpp)."""
is_openai = "/v1/chat/completions" in self.ollama_api_url
if system_prompt:
payload["system"] = system_prompt
if is_openai:
# OpenAI Chat API format
messages = []
if system_prompt:
messages.append({"role": "system", "content": system_prompt})
if structured_history:
messages.extend(structured_history)
messages.append({"role": "user", "content": prompt})
payload = {
"model": self.ollama_model,
"messages": messages,
"stream": True,
"temperature": self.ollama_temperature,
"max_tokens": 2048
}
else:
# Ollama /api/generate payload (Fallback)
full_prompt = prompt
if structured_history:
hist_str = "\n".join([f"{m['role'].capitalize()}: {m['content']}" for m in structured_history])
full_prompt = f"Historique:\n{hist_str}\n\nUser: {prompt}"
payload = {
"model": self.ollama_model,
"prompt": full_prompt,
"stream": True,
"options": {"temperature": self.ollama_temperature}
}
if system_prompt:
payload["system"] = system_prompt
# Vision support (simplified)
if vision_model and attachments:
for attachment in attachments:
if attachment.content_type and attachment.content_type.startswith('image/'):
if any(v in self.ollama_model.lower() for v in
['llava', 'bakllava', 'moondream', 'llama3.2-vision', 'minicpm-v']):
# Note: actual image handling would require download
pass
# actual image handling would require download and base64 for OpenAI
pass
return payload
async def call_ollama(self, payload: Dict[str, Any]) -> str:
"""Call Ollama API and accumulate response."""
"""Call LLM API and accumulate response."""
session = await self.get_session()
is_openai = "/v1/chat/completions" in self.ollama_api_url
async with session.post(
self.ollama_api_url,
@ -76,86 +102,132 @@ class LLMManager:
) as response:
response.raise_for_status()
# Si on ne streame pas, on récupère tout d'un coup
if not payload.get("stream", True):
data = await response.json()
return data.get("response", "")
accumulated_reply = ""
async for line in response.content:
line = line.decode('utf-8').strip()
accumulated_thoughts = ""
async for line_bytes in response.content:
line = line_bytes.decode('utf-8').strip()
if not line:
continue
# Strip SSE prefix if present
if line.startswith("data: "):
line = line[6:]
elif line.startswith("event:"):
continue
if line == "[DONE]":
break
try:
data = json.loads(line)
if 'thinking' in data and data['thinking']:
print(f"🤔 Le superviseur réfléchit : {data['thinking']}", flush=True)
continue
# 1. Handle Thinking (OpenAI style reasoning or special field)
thought = None
if 'thinking' in data: # Custom llama.cpp field
thought = data['thinking']
elif 'choices' in data and len(data['choices']) > 0:
delta = data['choices'][0].get('delta', {})
thought = delta.get('reasoning_content')
if thought:
accumulated_thoughts += thought
print(f"🤔 Le superviseur réfléchit : {thought}", flush=True)
if 'response' in data:
# 2. Handle Content
chunk = ""
if 'response' in data: # Ollama
chunk = data['response']
elif 'choices' in data and len(data['choices']) > 0: # OpenAI
delta = data['choices'][0].get('delta', {})
if delta.get('content') is not None:
chunk = delta.get('content', "")
elif 'content' in data: # llama.cpp legacy
chunk = data['content']
if chunk:
accumulated_reply += chunk
if chunk.strip():
print(f"🤖 Réponse du Superviseur (fragment) : {chunk}", flush=True)
elif 'action' in data:
elif 'action' in data: # Action already in JSON (unlikely for chat API)
accumulated_reply = line
break
except json.JSONDecodeError:
accumulated_reply += line
# In case of malformed line or raw text
if not line.startswith("{"):
accumulated_reply += line
# Fallback for models that put the final JSON response inside the thought block
if not accumulated_reply.strip() and accumulated_thoughts.strip():
# Test if the thought block contains our strict format using robust search
if re.search(r'\{[\s\n]*[\"\']action[\"\']', accumulated_thoughts, re.IGNORECASE):
print("⚠️ Aucun texte de contenu final mais du JSON détecté dans les pensées. Récupération.", flush=True)
accumulated_reply = accumulated_thoughts
return accumulated_reply
def extract_json_actions(self, text: str) -> Optional[str]:
"""Extract JSON action objects from text more robustly."""
"""Extract JSON action objects from text more robustly using an O(N) brace matcher."""
# 1. First, try to extract from markdown blocks if present
markdown_blocks = re.findall(r'```(?:json)?\s*(.*?)\s*```', text, re.DOTALL)
candidates = markdown_blocks if markdown_blocks else [text]
actions = []
# Regex plus laxiste pour trouver des blocs JSON potentiels
# On cherche des blocs commençant par {"action" ou {'action'
matches = re.finditer(r'\{[\s\n]*[\"\']action[\"\'].*?\}', text, re.DOTALL)
for match in matches:
json_str = match.group(0)
for candidate_text in candidates:
# Try to parse the whole text as a list or dict first
try:
# Tentative de parsing standard
parsed = json.loads(json_str)
if isinstance(parsed, dict) and 'action' in parsed:
parsed = json.loads(candidate_text.strip())
if isinstance(parsed, list):
for item in parsed:
if isinstance(item, dict) and 'action' in item:
actions.append(item)
if actions: return json.dumps(actions)
elif isinstance(parsed, dict) and 'action' in parsed:
actions.append(parsed)
except json.JSONDecodeError:
# Si erreur, on tente de remplacer les simples quotes par des doubles
# (Cas fréquent si le LLM mélange les styles)
try:
# Remplacement très basique (peut échouer si contenu complexe)
json_corrected = json_str.replace("'", '"')
parsed = json.loads(json_corrected)
if isinstance(parsed, dict) and 'action' in parsed:
actions.append(parsed)
continue
except: pass
return json.dumps(actions)
except:
pass
# Méthode récursive des accolades si le bloc est tronqué ou complexe
start = match.start()
brace_count = 0
for i in range(start, len(text)):
if text[i] == '{': brace_count += 1
elif text[i] == '}':
# Fallback: O(N) scan for top-level { ... } blocks
brace_count = 0
current_block = []
for char in candidate_text:
if char == '{':
if brace_count < 0:
brace_count = 0
current_block = []
brace_count += 1
if brace_count > 0:
current_block.append(char)
if char == '}':
if brace_count > 0:
brace_count -= 1
if brace_count == 0:
if brace_count == 0 and current_block:
json_str = "".join(current_block)
current_block = []
try:
candidate = text[start:i+1]
try:
parsed = json.loads(candidate)
except:
parsed = json.loads(candidate.replace("'", '"'))
parsed = json.loads(json_str)
if isinstance(parsed, dict) and 'action' in parsed:
actions.append(parsed)
break
except: pass
if parsed not in actions: # avoid duplicates
actions.append(parsed)
except:
# Fallback with single quote replace
try:
parsed = json.loads(json_str.replace("'", '"'))
if isinstance(parsed, dict) and 'action' in parsed:
if parsed not in actions:
actions.append(parsed)
except:
pass
else:
brace_count = 0
current_block = []
if actions:
return json.dumps(actions)
return None