Nafiul Huda commited on
Update app.py
Browse files
app.py
CHANGED
|
@@ -2,27 +2,33 @@ import os
|
|
| 2 |
import gradio as gr
|
| 3 |
import spaces
|
| 4 |
from huggingface_hub import hf_hub_download
|
| 5 |
-
from llama_cpp import Llama
|
| 6 |
|
| 7 |
-
#
|
|
|
|
|
|
|
|
|
|
| 8 |
print("Downloading model...")
|
| 9 |
model_path = hf_hub_download(
|
| 10 |
repo_id="Abiray/MiniCPM5-1B-GGUF",
|
| 11 |
filename="minicpm5-1b-Q6_K.gguf"
|
| 12 |
)
|
| 13 |
|
| 14 |
-
# 2.
|
| 15 |
-
print("Initializing model on GPU paths...")
|
| 16 |
-
llm = Llama(
|
| 17 |
-
model_path=model_path,
|
| 18 |
-
n_ctx=2048,
|
| 19 |
-
n_gpu_layers=-1
|
| 20 |
-
)
|
| 21 |
-
|
| 22 |
-
# 3. Use the spaces decorator to temporarily claim the ZeroGPU instance for inference
|
| 23 |
@spaces.GPU(duration=60)
|
| 24 |
def generate_response(message, history):
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
prompt = f"<|user|>\n{message}\n<|assistant|>\n"
|
| 27 |
|
| 28 |
# Run high-speed GPU accelerated inference
|
|
@@ -33,11 +39,10 @@ def generate_response(message, history):
|
|
| 33 |
echo=False
|
| 34 |
)
|
| 35 |
|
| 36 |
-
|
| 37 |
-
return response_text
|
| 38 |
|
| 39 |
-
#
|
| 40 |
-
with gr.Blocks(
|
| 41 |
gr.Markdown("### MiniCPM-1B GGUF ZeroGPU Pipeline")
|
| 42 |
chatbot = gr.ChatInterface(
|
| 43 |
fn=generate_response,
|
|
|
|
| 2 |
import gradio as gr
|
| 3 |
import spaces
|
| 4 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 5 |
|
| 6 |
+
# Global variable to hold our model instance once loaded
|
| 7 |
+
llm = None
|
| 8 |
+
|
| 9 |
+
# 1. Download model at root (this is safe, it's just fetching a file)
|
| 10 |
print("Downloading model...")
|
| 11 |
model_path = hf_hub_download(
|
| 12 |
repo_id="Abiray/MiniCPM5-1B-GGUF",
|
| 13 |
filename="minicpm5-1b-Q6_K.gguf"
|
| 14 |
)
|
| 15 |
|
| 16 |
+
# 2. Wrap initialization inside the GPU context
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
@spaces.GPU(duration=60)
|
| 18 |
def generate_response(message, history):
|
| 19 |
+
global llm
|
| 20 |
+
|
| 21 |
+
# Lazy import/load only when the GPU is actually active and attached
|
| 22 |
+
if llm is None:
|
| 23 |
+
print("GPU attached! Loading llama_cpp and initializing model...")
|
| 24 |
+
from llama_cpp import Llama
|
| 25 |
+
llm = Llama(
|
| 26 |
+
model_path=model_path,
|
| 27 |
+
n_ctx=2048,
|
| 28 |
+
n_gpu_layers=-1 # Safely sends layers to the active GPU
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
# Construct prompt format
|
| 32 |
prompt = f"<|user|>\n{message}\n<|assistant|>\n"
|
| 33 |
|
| 34 |
# Run high-speed GPU accelerated inference
|
|
|
|
| 39 |
echo=False
|
| 40 |
)
|
| 41 |
|
| 42 |
+
return output["choices"][0]["text"].strip()
|
|
|
|
| 43 |
|
| 44 |
+
# 3. Gradio Interface
|
| 45 |
+
with gr.Blocks() as demo:
|
| 46 |
gr.Markdown("### MiniCPM-1B GGUF ZeroGPU Pipeline")
|
| 47 |
chatbot = gr.ChatInterface(
|
| 48 |
fn=generate_response,
|