Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
import gradio as gr
|
2 |
-
from langchain_community.llms import LlamaCpp
|
3 |
import os
|
4 |
import json
|
5 |
import torch
|
@@ -15,6 +15,7 @@ import requests
|
|
15 |
from pathlib import Path
|
16 |
from tqdm import tqdm
|
17 |
from contextlib import asynccontextmanager
|
|
|
18 |
|
19 |
# Configure logging
|
20 |
logging.basicConfig(level=logging.INFO)
|
@@ -27,33 +28,30 @@ class ChatCompletionRequest(BaseModel):
|
|
27 |
max_tokens: Optional[int] = 2048
|
28 |
stream: Optional[bool] = False
|
29 |
|
30 |
-
def
|
31 |
-
"""Download the model file
|
32 |
-
|
33 |
-
logger.info(
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
pbar.update(size)
|
52 |
-
|
53 |
-
return local_path
|
54 |
|
55 |
class QwenModel:
|
56 |
-
def __init__(self
|
57 |
"""Initialize the Qwen model with automatic device detection."""
|
58 |
try:
|
59 |
# Check for GPU availability
|
@@ -61,12 +59,9 @@ class QwenModel:
|
|
61 |
self.device_count = torch.cuda.device_count() if self.has_gpu else 0
|
62 |
logger.info(f"GPU available: {self.has_gpu}, Device count: {self.device_count}")
|
63 |
|
64 |
-
#
|
65 |
-
model_path =
|
66 |
-
|
67 |
-
# If model doesn't exist locally, download it
|
68 |
-
model_url = "https://huggingface.co/G17c21ds/Qwen2.5-14B-Instruct-Uncensored-Q8_0-GGUF/resolve/main/model.gguf"
|
69 |
-
model_path = download_model(model_url, model_path)
|
70 |
|
71 |
# Configure model parameters based on available hardware
|
72 |
n_gpu_layers = 40 if self.has_gpu else 0
|
@@ -76,13 +71,13 @@ class QwenModel:
|
|
76 |
model_path=str(model_path),
|
77 |
n_gpu_layers=n_gpu_layers,
|
78 |
n_ctx=4096,
|
79 |
-
n_batch=512 if self.has_gpu else 128,
|
80 |
verbose=True,
|
81 |
temperature=0.7,
|
82 |
max_tokens=2048,
|
83 |
top_p=0.95,
|
84 |
top_k=50,
|
85 |
-
f16_kv=self.has_gpu,
|
86 |
use_mlock=True,
|
87 |
use_mmap=True,
|
88 |
)
|
@@ -107,8 +102,7 @@ async def lifespan(app: FastAPI):
|
|
107 |
"""Lifespan context manager for FastAPI startup and shutdown events."""
|
108 |
global model
|
109 |
try:
|
110 |
-
|
111 |
-
model = QwenModel(model_path)
|
112 |
logger.info("Model initialized successfully")
|
113 |
yield
|
114 |
finally:
|
@@ -117,18 +111,16 @@ async def lifespan(app: FastAPI):
|
|
117 |
|
118 |
app = FastAPI(lifespan=lifespan)
|
119 |
|
120 |
-
# ... [rest of the FastAPI routes
|
121 |
|
122 |
def main():
|
123 |
"""Main function to initialize and launch the application."""
|
124 |
try:
|
125 |
global model
|
126 |
-
# Model path
|
127 |
-
model_path = Path("models/qwen-2.5-14b-gguf")
|
128 |
|
129 |
# Initialize the model if not already initialized
|
130 |
if model is None:
|
131 |
-
model = QwenModel(
|
132 |
|
133 |
# Create and launch the Gradio interface
|
134 |
interface = create_gradio_interface(model)
|
|
|
1 |
import gradio as gr
|
2 |
+
from langchain_community.llms import LlamaCpp
|
3 |
import os
|
4 |
import json
|
5 |
import torch
|
|
|
15 |
from pathlib import Path
|
16 |
from tqdm import tqdm
|
17 |
from contextlib import asynccontextmanager
|
18 |
+
from huggingface_hub import hf_hub_download
|
19 |
|
20 |
# Configure logging
|
21 |
logging.basicConfig(level=logging.INFO)
|
|
|
28 |
max_tokens: Optional[int] = 2048
|
29 |
stream: Optional[bool] = False
|
30 |
|
31 |
+
def download_model_from_hf():
|
32 |
+
"""Download the model file from Hugging Face."""
|
33 |
+
try:
|
34 |
+
logger.info("Downloading model from Hugging Face Hub...")
|
35 |
+
|
36 |
+
# Create models directory if it doesn't exist
|
37 |
+
model_dir = Path("models")
|
38 |
+
model_dir.mkdir(exist_ok=True)
|
39 |
+
|
40 |
+
# Download the model using huggingface_hub
|
41 |
+
local_path = hf_hub_download(
|
42 |
+
repo_id="G17c21ds/Qwen2.5-14B-Instruct-Uncensored-Q8_0-GGUF",
|
43 |
+
filename="model.gguf",
|
44 |
+
local_dir=model_dir,
|
45 |
+
local_dir_use_symlinks=False
|
46 |
+
)
|
47 |
+
|
48 |
+
return Path(local_path)
|
49 |
+
except Exception as e:
|
50 |
+
logger.error(f"Error downloading model: {str(e)}")
|
51 |
+
raise
|
|
|
|
|
|
|
52 |
|
53 |
class QwenModel:
|
54 |
+
def __init__(self):
|
55 |
"""Initialize the Qwen model with automatic device detection."""
|
56 |
try:
|
57 |
# Check for GPU availability
|
|
|
59 |
self.device_count = torch.cuda.device_count() if self.has_gpu else 0
|
60 |
logger.info(f"GPU available: {self.has_gpu}, Device count: {self.device_count}")
|
61 |
|
62 |
+
# Download or get the model
|
63 |
+
model_path = download_model_from_hf()
|
64 |
+
logger.info(f"Model path: {model_path}")
|
|
|
|
|
|
|
65 |
|
66 |
# Configure model parameters based on available hardware
|
67 |
n_gpu_layers = 40 if self.has_gpu else 0
|
|
|
71 |
model_path=str(model_path),
|
72 |
n_gpu_layers=n_gpu_layers,
|
73 |
n_ctx=4096,
|
74 |
+
n_batch=512 if self.has_gpu else 128,
|
75 |
verbose=True,
|
76 |
temperature=0.7,
|
77 |
max_tokens=2048,
|
78 |
top_p=0.95,
|
79 |
top_k=50,
|
80 |
+
f16_kv=self.has_gpu,
|
81 |
use_mlock=True,
|
82 |
use_mmap=True,
|
83 |
)
|
|
|
102 |
"""Lifespan context manager for FastAPI startup and shutdown events."""
|
103 |
global model
|
104 |
try:
|
105 |
+
model = QwenModel()
|
|
|
106 |
logger.info("Model initialized successfully")
|
107 |
yield
|
108 |
finally:
|
|
|
111 |
|
112 |
app = FastAPI(lifespan=lifespan)
|
113 |
|
114 |
+
# ... [rest of the FastAPI routes remain the same] ...
|
115 |
|
116 |
def main():
|
117 |
"""Main function to initialize and launch the application."""
|
118 |
try:
|
119 |
global model
|
|
|
|
|
120 |
|
121 |
# Initialize the model if not already initialized
|
122 |
if model is None:
|
123 |
+
model = QwenModel()
|
124 |
|
125 |
# Create and launch the Gradio interface
|
126 |
interface = create_gradio_interface(model)
|