Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -7,6 +7,7 @@ import time
|
|
7 |
import uuid
|
8 |
import json
|
9 |
from typing import Optional, List, Union, Dict, Any
|
|
|
10 |
|
11 |
# --- Configuration ---
|
12 |
MODEL_ID = "deepseek-ai/deepseek-coder-1.3b-instruct"
|
@@ -20,12 +21,16 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
20 |
device_map=DEVICE
|
21 |
)
|
22 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
|
|
|
|
|
|
|
|
23 |
print("Modèle et tokenizer chargés avec succès sur le CPU.")
|
24 |
|
25 |
# --- Création de l'application API ---
|
26 |
app = FastAPI()
|
27 |
|
28 |
-
# --- Modèles de données
|
29 |
class ContentPart(BaseModel):
|
30 |
type: str
|
31 |
text: str
|
@@ -38,10 +43,30 @@ class ChatCompletionRequest(BaseModel):
|
|
38 |
model: Optional[str] = None
|
39 |
messages: List[ChatMessage]
|
40 |
stream: Optional[bool] = False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
|
42 |
class Config:
|
43 |
extra = Extra.ignore
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
class ModelData(BaseModel):
|
46 |
id: str
|
47 |
object: str = "model"
|
@@ -50,19 +75,15 @@ class ModelData(BaseModel):
|
|
50 |
class ModelList(BaseModel):
|
51 |
object: str = "list"
|
52 |
data: List[ModelData]
|
53 |
-
|
54 |
# --- Définition des API ---
|
55 |
|
56 |
@app.get("/models", response_model=ModelList)
|
57 |
async def list_models():
|
58 |
-
"""Répond à la requête GET /models pour satisfaire l'extension."""
|
59 |
return ModelList(data=[ModelData(id=MODEL_ID)])
|
60 |
|
61 |
@app.post("/chat/completions")
|
62 |
async def create_chat_completion(request: ChatCompletionRequest):
|
63 |
-
"""Endpoint principal qui gère la génération de texte en streaming."""
|
64 |
-
|
65 |
-
# On extrait le prompt de l'utilisateur de la structure complexe
|
66 |
user_prompt = ""
|
67 |
last_message = request.messages[-1]
|
68 |
if isinstance(last_message.content, list):
|
@@ -75,61 +96,41 @@ async def create_chat_completion(request: ChatCompletionRequest):
|
|
75 |
if not user_prompt:
|
76 |
return {"error": "Prompt non trouvé."}
|
77 |
|
78 |
-
# Préparation pour le modèle DeepSeek
|
79 |
messages_for_model = [{'role': 'user', 'content': user_prompt}]
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
-
|
83 |
-
outputs = model.generate(inputs, max_new_tokens=250, do_sample=True, temperature=0.2, top_k=50, top_p=0.95, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id)
|
84 |
-
response_text = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)
|
85 |
|
86 |
-
# Fonction génératrice pour le streaming
|
87 |
async def stream_generator():
|
88 |
response_id = f"chatcmpl-{uuid.uuid4()}"
|
89 |
|
90 |
-
# On envoie la réponse caractère par caractère, au format attendu
|
91 |
for char in response_text:
|
92 |
-
chunk = {
|
93 |
-
"id": response_id,
|
94 |
-
"object": "chat.completion.chunk",
|
95 |
-
"created": int(time.time()),
|
96 |
-
"model": MODEL_ID,
|
97 |
-
"choices": [{
|
98 |
-
"index": 0,
|
99 |
-
"delta": {"content": char},
|
100 |
-
"finish_reason": None
|
101 |
-
}]
|
102 |
-
}
|
103 |
yield f"data: {json.dumps(chunk)}\n\n"
|
104 |
-
await asyncio.sleep(0.01)
|
105 |
|
106 |
-
|
107 |
-
final_chunk = {
|
108 |
-
"id": response_id,
|
109 |
-
"object": "chat.completion.chunk",
|
110 |
-
"created": int(time.time()),
|
111 |
-
"model": MODEL_ID,
|
112 |
-
"choices": [{
|
113 |
-
"index": 0,
|
114 |
-
"delta": {},
|
115 |
-
"finish_reason": "stop"
|
116 |
-
}]
|
117 |
-
}
|
118 |
yield f"data: {json.dumps(final_chunk)}\n\n"
|
119 |
-
|
120 |
-
# On envoie le signal [DONE]
|
121 |
yield "data: [DONE]\n\n"
|
122 |
|
123 |
-
# Si l'extension demande un stream, on renvoie le générateur
|
124 |
if request.stream:
|
125 |
return StreamingResponse(stream_generator(), media_type="text/event-stream")
|
126 |
else:
|
127 |
-
# Code de secours si le stream n'est pas demandé (peu probable)
|
128 |
return {"choices": [{"message": {"role": "assistant", "content": response_text}}]}
|
129 |
|
130 |
@app.get("/")
|
131 |
def root():
|
132 |
-
return {"status": "API compatible OpenAI en ligne (avec streaming)", "model_id": MODEL_ID}
|
133 |
-
|
134 |
-
# On a besoin de asyncio pour la pause dans le stream
|
135 |
-
import asyncio
|
|
|
7 |
import uuid
|
8 |
import json
|
9 |
from typing import Optional, List, Union, Dict, Any
|
10 |
+
import asyncio
|
11 |
|
12 |
# --- Configuration ---
|
13 |
MODEL_ID = "deepseek-ai/deepseek-coder-1.3b-instruct"
|
|
|
21 |
device_map=DEVICE
|
22 |
)
|
23 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
24 |
+
if tokenizer.pad_token is None:
|
25 |
+
tokenizer.pad_token = tokenizer.eos_token
|
26 |
+
print("Le pad_token a été défini sur eos_token.")
|
27 |
+
|
28 |
print("Modèle et tokenizer chargés avec succès sur le CPU.")
|
29 |
|
30 |
# --- Création de l'application API ---
|
31 |
app = FastAPI()
|
32 |
|
33 |
+
# --- Modèles de données ---
|
34 |
class ContentPart(BaseModel):
|
35 |
type: str
|
36 |
text: str
|
|
|
43 |
model: Optional[str] = None
|
44 |
messages: List[ChatMessage]
|
45 |
stream: Optional[bool] = False
|
46 |
+
max_tokens: Optional[int] = 512 # Augmenté pour des réponses plus longues
|
47 |
+
|
48 |
+
# --- LES NOUVEAUX CHAMPS SONT ICI ---
|
49 |
+
# Ajout des paramètres de génération avec des valeurs par défaut.
|
50 |
+
temperature: Optional[float] = 0.4
|
51 |
+
top_p: Optional[float] = 0.95
|
52 |
+
top_k: Optional[int] = 50
|
53 |
|
54 |
class Config:
|
55 |
extra = Extra.ignore
|
56 |
|
57 |
+
# ... (le reste des modèles de données est inchangé) ...
|
58 |
+
class ChatCompletionResponseChoice(BaseModel):
|
59 |
+
index: int = 0
|
60 |
+
message: ChatMessage
|
61 |
+
finish_reason: str = "stop"
|
62 |
+
|
63 |
+
class ChatCompletionResponse(BaseModel):
|
64 |
+
id: str
|
65 |
+
object: str = "chat.completion"
|
66 |
+
created: int
|
67 |
+
model: str
|
68 |
+
choices: List[ChatCompletionResponseChoice]
|
69 |
+
|
70 |
class ModelData(BaseModel):
|
71 |
id: str
|
72 |
object: str = "model"
|
|
|
75 |
class ModelList(BaseModel):
|
76 |
object: str = "list"
|
77 |
data: List[ModelData]
|
78 |
+
|
79 |
# --- Définition des API ---
|
80 |
|
81 |
@app.get("/models", response_model=ModelList)
|
82 |
async def list_models():
|
|
|
83 |
return ModelList(data=[ModelData(id=MODEL_ID)])
|
84 |
|
85 |
@app.post("/chat/completions")
|
86 |
async def create_chat_completion(request: ChatCompletionRequest):
|
|
|
|
|
|
|
87 |
user_prompt = ""
|
88 |
last_message = request.messages[-1]
|
89 |
if isinstance(last_message.content, list):
|
|
|
96 |
if not user_prompt:
|
97 |
return {"error": "Prompt non trouvé."}
|
98 |
|
|
|
99 |
messages_for_model = [{'role': 'user', 'content': user_prompt}]
|
100 |
+
text_prompt = tokenizer.apply_chat_template(messages_for_model, tokenize=False, add_generation_prompt=True)
|
101 |
+
inputs = tokenizer(text_prompt, return_tensors="pt", padding=True).to(DEVICE)
|
102 |
+
|
103 |
+
# --- LA MISE À JOUR EST ICI ---
|
104 |
+
# On utilise maintenant les paramètres de la requête pour la génération.
|
105 |
+
outputs = model.generate(
|
106 |
+
**inputs,
|
107 |
+
max_new_tokens=request.max_tokens,
|
108 |
+
do_sample=True, # do_sample doit être True pour que temp, top_p et top_k aient un effet
|
109 |
+
temperature=request.temperature,
|
110 |
+
top_p=request.top_p,
|
111 |
+
top_k=request.top_k,
|
112 |
+
eos_token_id=tokenizer.eos_token_id
|
113 |
+
)
|
114 |
|
115 |
+
response_text = tokenizer.decode(outputs[0, inputs['input_ids'].shape[1]:], skip_special_tokens=True)
|
|
|
|
|
116 |
|
|
|
117 |
async def stream_generator():
|
118 |
response_id = f"chatcmpl-{uuid.uuid4()}"
|
119 |
|
|
|
120 |
for char in response_text:
|
121 |
+
chunk = { "id": response_id, "object": "chat.completion.chunk", "created": int(time.time()), "model": MODEL_ID, "choices": [{"index": 0, "delta": {"content": char}, "finish_reason": None }] }
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
yield f"data: {json.dumps(chunk)}\n\n"
|
123 |
+
await asyncio.sleep(0.01)
|
124 |
|
125 |
+
final_chunk = { "id": response_id, "object": "chat.completion.chunk", "created": int(time.time()), "model": MODEL_ID, "choices": [{"index": 0, "delta": {}, "finish_reason": "stop" }] }
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
yield f"data: {json.dumps(final_chunk)}\n\n"
|
|
|
|
|
127 |
yield "data: [DONE]\n\n"
|
128 |
|
|
|
129 |
if request.stream:
|
130 |
return StreamingResponse(stream_generator(), media_type="text/event-stream")
|
131 |
else:
|
|
|
132 |
return {"choices": [{"message": {"role": "assistant", "content": response_text}}]}
|
133 |
|
134 |
@app.get("/")
|
135 |
def root():
|
136 |
+
return {"status": "API compatible OpenAI en ligne (avec streaming et paramètres dynamiques)", "model_id": MODEL_ID}
|
|
|
|
|
|