Upload app.py
Browse files
app.py
CHANGED
|
@@ -1,10 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
gr.Markdown("# Inference Provider")
|
| 6 |
-
gr.Markdown("This Space showcases the stabilityai/stable-diffusion-3.5-large model, served by the fal-ai API. Sign in with your Hugging Face account to use this API.")
|
| 7 |
-
button = gr.LoginButton("Sign in")
|
| 8 |
-
gr.load("models/stabilityai/stable-diffusion-3.5-large", accept_token=button, provider="fal-ai")
|
| 9 |
-
|
| 10 |
-
demo.launch()
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Gradio App for EeshaAI/Zeeb Training Space
|
| 4 |
+
==========================================
|
| 5 |
+
Provides a web UI to trigger LoRA fine-tuning of OLMo 2 1B Instruct
|
| 6 |
+
on tokenized video data, then push the trained model to EeshaAI/zeeb.
|
| 7 |
+
"""
|
| 8 |
+
|
| 9 |
import gradio as gr
|
| 10 |
+
from train_on_hf_spaces import train
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
def run_training():
|
| 14 |
+
"""Run the training pipeline and stream logs to the UI."""
|
| 15 |
+
for log_msg in train("tokenized_dataset.json"):
|
| 16 |
+
yield log_msg
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
with gr.Blocks(
|
| 20 |
+
title="Zeeb β Video-LLM Trainer",
|
| 21 |
+
theme=gr.themes.Soft(),
|
| 22 |
+
css="""
|
| 23 |
+
#logbox { font-family: 'Courier New', monospace; font-size: 13px; line-height: 1.5; }
|
| 24 |
+
#title { text-align: center; }
|
| 25 |
+
.contain { max-width: 800px; margin: auto; }
|
| 26 |
+
"""
|
| 27 |
+
) as demo:
|
| 28 |
+
|
| 29 |
+
gr.Markdown(
|
| 30 |
+
"""
|
| 31 |
+
# π¬ Zeeb β Video-LLM Trainer
|
| 32 |
+
Fine-tune **OLMo 2 1B Instruct** with **LoRA (r=4)** to generate video tokens.
|
| 33 |
+
Trained model is automatically pushed to [EeshaAI/zeeb](https://huggingface.co/EeshaAI/zeeb).
|
| 34 |
+
""",
|
| 35 |
+
elem_id="title",
|
| 36 |
+
)
|
| 37 |
+
|
| 38 |
+
with gr.Row():
|
| 39 |
+
train_btn = gr.Button("π Start Training", variant="primary", size="lg")
|
| 40 |
+
|
| 41 |
+
logbox = gr.Textbox(
|
| 42 |
+
label="Training Log",
|
| 43 |
+
elem_id="logbox",
|
| 44 |
+
lines=25,
|
| 45 |
+
max_lines=100,
|
| 46 |
+
interactive=False,
|
| 47 |
+
show_copy_button=True,
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
gr.Markdown(
|
| 51 |
+
"""
|
| 52 |
+
### What happens when you click "Start Training"?
|
| 53 |
+
1. π¦ Downloads **OLMo 2 1B Instruct** from HuggingFace
|
| 54 |
+
2. π€ Expands vocabulary with **1,024 visual tokens** (`<v_0>` ... `<v_1023>`)
|
| 55 |
+
3. π§ Applies **LoRA r=4** to q_proj & v_proj (minimal memory)
|
| 56 |
+
4. π₯ Trains for **3 epochs** on the tokenized video dataset
|
| 57 |
+
5. π Merges LoRA weights back into the base model
|
| 58 |
+
6. π Pushes the merged model to **EeshaAI/zeeb**
|
| 59 |
+
|
| 60 |
+
β οΈ Training on CPU takes time (~10-30 min depending on dataset size).
|
| 61 |
+
"""
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
train_btn.click(fn=run_training, outputs=logbox)
|
| 65 |
+
|
| 66 |
|
| 67 |
+
if __name__ == "__main__":
|
| 68 |
+
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|