Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -3,9 +3,6 @@ import torch
|
|
| 3 |
import spaces
|
| 4 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 5 |
|
| 6 |
-
# ----------------------------------------------------
|
| 7 |
-
# Globals
|
| 8 |
-
# ----------------------------------------------------
|
| 9 |
model = None
|
| 10 |
tokenizer = None
|
| 11 |
|
|
@@ -20,20 +17,14 @@ alpaca_prompt = """පහත දැක්වෙන්නේ යම් කාර
|
|
| 20 |
### ප්රතිචාරය:
|
| 21 |
{}"""
|
| 22 |
|
| 23 |
-
# ----------------------------------------------------
|
| 24 |
-
# GPU inference — official ZeroGPU style
|
| 25 |
-
# ----------------------------------------------------
|
| 26 |
@spaces.GPU
|
| 27 |
def infer(instruction, input_text=""):
|
| 28 |
global model, tokenizer
|
| 29 |
-
|
| 30 |
if model is None:
|
| 31 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
| 32 |
-
"manthilaffs/Gamunu-4B-Instruct-Alpha"
|
| 33 |
-
)
|
| 34 |
model = AutoModelForCausalLM.from_pretrained(
|
| 35 |
"manthilaffs/Gamunu-4B-Instruct-Alpha",
|
| 36 |
-
|
| 37 |
device_map="auto",
|
| 38 |
)
|
| 39 |
model.eval()
|
|
@@ -46,72 +37,86 @@ def infer(instruction, input_text=""):
|
|
| 46 |
)
|
| 47 |
|
| 48 |
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
| 49 |
-
|
| 50 |
with torch.inference_mode():
|
| 51 |
outputs = model.generate(
|
| 52 |
**inputs,
|
| 53 |
max_new_tokens=256,
|
| 54 |
-
temperature=0.4,
|
| 55 |
-
top_k=64,
|
| 56 |
-
top_p=0.95,
|
| 57 |
repetition_penalty=1.05,
|
| 58 |
)
|
| 59 |
-
|
| 60 |
text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 61 |
if "### ප්රතිචාරය:" in text:
|
| 62 |
text = text.split("### ප්රතිචාරය:")[-1].strip()
|
| 63 |
return text
|
| 64 |
|
| 65 |
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
|
| 75 |
-
⚙️ Pure Transformers Inference | 💠 ZeroGPU GPU Burst
|
| 76 |
-
"""
|
| 77 |
-
)
|
| 78 |
-
|
| 79 |
-
with gr.Row():
|
| 80 |
-
instruction = gr.Textbox(
|
| 81 |
-
label="🧾 Instruction / Question",
|
| 82 |
-
placeholder="උදා: ඊයේ රු. 30 ක ඇපල් ගෙඩියක් අද රු. 60 නම් මිල වෙනස කීයද?",
|
| 83 |
-
lines=2,
|
| 84 |
-
)
|
| 85 |
-
|
| 86 |
-
with gr.Row():
|
| 87 |
-
input_text = gr.Textbox(
|
| 88 |
-
label="📥 Additional Context (Optional)",
|
| 89 |
-
placeholder="අමතර තොරතුරු (ඇත්නම්) එහි සටහන් කරන්න",
|
| 90 |
-
lines=2,
|
| 91 |
-
)
|
| 92 |
-
|
| 93 |
-
output = gr.Markdown(label="🧩 Gamunu Response")
|
| 94 |
-
run_btn = gr.Button("🔮 Generate Response")
|
| 95 |
run_btn.click(infer, inputs=[instruction, input_text], outputs=output)
|
| 96 |
|
| 97 |
-
gr.
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
inputs=[instruction],
|
| 104 |
-
label="🪄 Example Questions (Click to try)",
|
| 105 |
-
)
|
| 106 |
-
|
| 107 |
-
gr.Markdown(
|
| 108 |
-
"""
|
| 109 |
-
---
|
| 110 |
-
🪶 **Model:** `manthilaffs/Gamunu-4B-Instruct-Alpha`
|
| 111 |
-
🧰 **Stack:** Transformers + Torch + Gradio + ZeroGPU
|
| 112 |
-
© 2025 Gamunu Project | Experimental Release
|
| 113 |
-
"""
|
| 114 |
-
)
|
| 115 |
|
| 116 |
if __name__ == "__main__":
|
| 117 |
demo.launch()
|
|
|
|
| 3 |
import spaces
|
| 4 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 5 |
|
|
|
|
|
|
|
|
|
|
| 6 |
model = None
|
| 7 |
tokenizer = None
|
| 8 |
|
|
|
|
| 17 |
### ප්රතිචාරය:
|
| 18 |
{}"""
|
| 19 |
|
|
|
|
|
|
|
|
|
|
| 20 |
@spaces.GPU
|
| 21 |
def infer(instruction, input_text=""):
|
| 22 |
global model, tokenizer
|
|
|
|
| 23 |
if model is None:
|
| 24 |
+
tokenizer = AutoTokenizer.from_pretrained("manthilaffs/Gamunu-4B-Instruct-Alpha")
|
|
|
|
|
|
|
| 25 |
model = AutoModelForCausalLM.from_pretrained(
|
| 26 |
"manthilaffs/Gamunu-4B-Instruct-Alpha",
|
| 27 |
+
dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
|
| 28 |
device_map="auto",
|
| 29 |
)
|
| 30 |
model.eval()
|
|
|
|
| 37 |
)
|
| 38 |
|
| 39 |
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
|
|
|
| 40 |
with torch.inference_mode():
|
| 41 |
outputs = model.generate(
|
| 42 |
**inputs,
|
| 43 |
max_new_tokens=256,
|
|
|
|
|
|
|
|
|
|
| 44 |
repetition_penalty=1.05,
|
| 45 |
)
|
|
|
|
| 46 |
text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 47 |
if "### ප්රතිචාරය:" in text:
|
| 48 |
text = text.split("### ප්රතිචාරය:")[-1].strip()
|
| 49 |
return text
|
| 50 |
|
| 51 |
|
| 52 |
+
with gr.Blocks(css="""
|
| 53 |
+
.gradio-container {max-width: 1100px !important; margin:auto;}
|
| 54 |
+
h1, h2, h3, h4, h5 {text-align:center;}
|
| 55 |
+
#title-bar {background:linear-gradient(90deg,#804dee,#ee82ee);color:white;padding:0.6rem;border-radius:1rem;margin-bottom:0.6rem;}
|
| 56 |
+
textarea, input {font-family:'Noto Sans Sinhala',sans-serif;}
|
| 57 |
+
""", theme=gr.themes.Soft()) as demo:
|
| 58 |
+
gr.HTML("<div id='title-bar'><h1>🧠 Gamunu 4B Instruct Alpha</h1><h4>Sinhala Instruct LLM — ZeroGPU Demo</h4></div>")
|
| 59 |
+
gr.Markdown("⚙️ **Pure Transformers Inference** | 💠 ZeroGPU GPU Burst")
|
| 60 |
+
|
| 61 |
+
with gr.Row(equal_height=True):
|
| 62 |
+
with gr.Column(scale=1, min_width=350):
|
| 63 |
+
instruction = gr.Textbox(
|
| 64 |
+
label="🧾 Instruction / Question",
|
| 65 |
+
placeholder="උදා: ඊයේ රු. 30 ක ඇපල් ගෙඩියක් අද රු. 60 නම් මිල වෙනස කීයද?",
|
| 66 |
+
lines=4,
|
| 67 |
+
)
|
| 68 |
+
input_text = gr.Textbox(
|
| 69 |
+
label="📥 Additional Context (Optional)",
|
| 70 |
+
placeholder="අමතර තොරතුරු ඇතුළත් කරන්න (ඇත්නම්)",
|
| 71 |
+
lines=3,
|
| 72 |
+
)
|
| 73 |
+
run_btn = gr.Button("🔮 Generate Response", variant="primary", scale=1)
|
| 74 |
+
with gr.Column(scale=1, min_width=350):
|
| 75 |
+
output = gr.Markdown(label="🧩 Gamunu Response", elem_id="output-box")
|
| 76 |
+
|
| 77 |
+
# Example categories
|
| 78 |
+
with gr.Accordion("🧮 Example Prompts by Category", open=False):
|
| 79 |
+
with gr.Tab("Maths"):
|
| 80 |
+
gr.Examples(
|
| 81 |
+
examples=[
|
| 82 |
+
["රු. 30 කින් මිලදී ගත් දේ රු. 60 නම් මිල වෙනස ප්රතිශතයකින් කීයද?", ""],
|
| 83 |
+
["සියයට 10 ක වර්ධනයක් තිබේ නම් අලුත් අගය කීයද?", ""],
|
| 84 |
+
],
|
| 85 |
+
inputs=[instruction],
|
| 86 |
+
)
|
| 87 |
+
with gr.Tab("Roleplay"):
|
| 88 |
+
gr.Examples(
|
| 89 |
+
examples=[
|
| 90 |
+
["ඔබ ගුරුවරයෙකු ලෙස ක්රියාකරන්න. ශිෂ්යයාට ගණිතය උගන්වන්න.", ""],
|
| 91 |
+
["ඔබ පරිසර විද්යාඥයෙකු ලෙස වායු මණ්ඩලය පැහැදිලි කරන්න.", ""],
|
| 92 |
+
],
|
| 93 |
+
inputs=[instruction],
|
| 94 |
+
)
|
| 95 |
+
with gr.Tab("QA"):
|
| 96 |
+
gr.Examples(
|
| 97 |
+
examples=[
|
| 98 |
+
["ශ්රී ලංකාවේ මුල්ම අගමැති කවුද?", ""],
|
| 99 |
+
["ඉන්දියානු මහා සමුද්රය යනු කොහෙද?", ""],
|
| 100 |
+
],
|
| 101 |
+
inputs=[instruction],
|
| 102 |
+
)
|
| 103 |
+
with gr.Tab("NLP"):
|
| 104 |
+
gr.Examples(
|
| 105 |
+
examples=[
|
| 106 |
+
["මෙම වාක්යය සිංහලයට පරිවර්තනය කරන්න: 'The sun rises in the east.'", ""],
|
| 107 |
+
["වචන 'ආදරය' සඳහා පරිවර්තන 3ක් දෙන්න.", ""],
|
| 108 |
+
],
|
| 109 |
+
inputs=[instruction],
|
| 110 |
+
)
|
| 111 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
run_btn.click(infer, inputs=[instruction, input_text], outputs=output)
|
| 113 |
|
| 114 |
+
gr.Markdown("""
|
| 115 |
+
---
|
| 116 |
+
🪶 **Model:** `manthilaffs/Gamunu-4B-Instruct-Alpha`
|
| 117 |
+
🧰 **Stack:** Transformers + Torch + Gradio + ZeroGPU
|
| 118 |
+
© 2025 Gamunu Project | Experimental Release
|
| 119 |
+
""")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 120 |
|
| 121 |
if __name__ == "__main__":
|
| 122 |
demo.launch()
|