Spaces:
Sleeping
Sleeping
Update main.py
Browse files
main.py
CHANGED
@@ -5,23 +5,25 @@ from fastapi.responses import StreamingResponse
|
|
5 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
6 |
import torch
|
7 |
import os
|
8 |
-
|
9 |
|
|
|
|
|
|
|
10 |
|
11 |
-
|
12 |
-
# Load Qwen model and tokenizer (once)
|
13 |
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
|
14 |
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
15 |
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
16 |
|
17 |
-
# Set device
|
18 |
-
device = torch.device("
|
19 |
model.to(device)
|
20 |
|
21 |
-
# FastAPI app
|
22 |
app = FastAPI()
|
23 |
|
24 |
-
# CORS settings
|
25 |
app.add_middleware(
|
26 |
CORSMiddleware,
|
27 |
allow_origins=["*"],
|
@@ -30,26 +32,25 @@ app.add_middleware(
|
|
30 |
allow_headers=["*"],
|
31 |
)
|
32 |
|
33 |
-
# Request
|
34 |
class Question(BaseModel):
|
35 |
question: str
|
36 |
|
37 |
-
# System prompt
|
38 |
SYSTEM_PROMPT = "You are Orion, an intelligent AI assistant created by Abdullah Ali, a 13-year-old from Lahore. Respond kindly and wisely."
|
39 |
|
40 |
-
#
|
41 |
async def generate_response_chunks(prompt: str):
|
42 |
-
# Build prompt using Qwen's expected format
|
43 |
qwen_prompt = (
|
44 |
f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
|
45 |
f"<|im_start|>user\n{prompt}<|im_end|>\n"
|
46 |
f"<|im_start|>assistant\n"
|
47 |
)
|
48 |
|
49 |
-
# Tokenize
|
50 |
inputs = tokenizer(qwen_prompt, return_tensors="pt").to(device)
|
51 |
|
52 |
-
# Generate
|
53 |
outputs = model.generate(
|
54 |
**inputs,
|
55 |
max_new_tokens=256,
|
@@ -59,16 +60,16 @@ async def generate_response_chunks(prompt: str):
|
|
59 |
pad_token_id=tokenizer.eos_token_id
|
60 |
)
|
61 |
|
62 |
-
# Decode
|
63 |
full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
64 |
reply = full_output.split("<|im_start|>assistant\n")[-1].strip()
|
65 |
|
66 |
-
|
67 |
-
|
|
|
|
|
68 |
|
|
|
69 |
@app.post("/ask")
|
70 |
async def ask(question: Question):
|
71 |
-
return StreamingResponse(
|
72 |
-
generate_response_chunks(question.question),
|
73 |
-
media_type="text/plain"
|
74 |
-
)
|
|
|
5 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
6 |
import torch
|
7 |
import os
|
8 |
+
import asyncio
|
9 |
|
10 |
+
# β
Set a safe and writable HF cache directory
|
11 |
+
os.environ["HF_HOME"] = "./hf_home"
|
12 |
+
os.makedirs(os.environ["HF_HOME"], exist_ok=True)
|
13 |
|
14 |
+
# β
Model and tokenizer (only loaded once)
|
|
|
15 |
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
|
16 |
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
17 |
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
18 |
|
19 |
+
# β
Set device (use GPU if available)
|
20 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
21 |
model.to(device)
|
22 |
|
23 |
+
# β
FastAPI app
|
24 |
app = FastAPI()
|
25 |
|
26 |
+
# β
CORS settings
|
27 |
app.add_middleware(
|
28 |
CORSMiddleware,
|
29 |
allow_origins=["*"],
|
|
|
32 |
allow_headers=["*"],
|
33 |
)
|
34 |
|
35 |
+
# β
Request schema
|
36 |
class Question(BaseModel):
|
37 |
question: str
|
38 |
|
39 |
+
# β
System prompt
|
40 |
SYSTEM_PROMPT = "You are Orion, an intelligent AI assistant created by Abdullah Ali, a 13-year-old from Lahore. Respond kindly and wisely."
|
41 |
|
42 |
+
# β
Streaming generator
|
43 |
async def generate_response_chunks(prompt: str):
|
|
|
44 |
qwen_prompt = (
|
45 |
f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
|
46 |
f"<|im_start|>user\n{prompt}<|im_end|>\n"
|
47 |
f"<|im_start|>assistant\n"
|
48 |
)
|
49 |
|
50 |
+
# Tokenize prompt
|
51 |
inputs = tokenizer(qwen_prompt, return_tensors="pt").to(device)
|
52 |
|
53 |
+
# Generate output
|
54 |
outputs = model.generate(
|
55 |
**inputs,
|
56 |
max_new_tokens=256,
|
|
|
60 |
pad_token_id=tokenizer.eos_token_id
|
61 |
)
|
62 |
|
63 |
+
# Decode output
|
64 |
full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
65 |
reply = full_output.split("<|im_start|>assistant\n")[-1].strip()
|
66 |
|
67 |
+
# Yield chunks word by word (simulating stream)
|
68 |
+
for word in reply.split():
|
69 |
+
yield word + " "
|
70 |
+
await asyncio.sleep(0.01) # slight delay for streaming effect
|
71 |
|
72 |
+
# β
POST endpoint
|
73 |
@app.post("/ask")
|
74 |
async def ask(question: Question):
|
75 |
+
return StreamingResponse(generate_response_chunks(question.question), media_type="text/plain")
|
|
|
|
|
|