Upload folder using huggingface_hub

app.py

@@ -1,23 +1,41 @@
 import os
 import gradio as gr
-import torch
 import subprocess
 from huggingface_hub import login
+import sys
 
 def run_finetuning(hf_token):
     if not hf_token:
-        return "ERROR: HF Token is required!"
+        yield "ERROR: HF Token is required!"
+        return
     
     yield "Logging in to Hugging Face..."
     try:
         login(token=hf_token)
-        yield "Login successful! Installing dependencies (unsloth etc)..."
+        yield "Login successful!"
     except Exception as e:
-        return f"Login failed: {e}"
+        yield f"Login failed: {e}"
+        return
 
-    # To avoid blocking the Gradio UI and memory issues, we can run the actual training as a subprocess.
-    # But since we are generating the logic directly, let's write the training script and run it.
+    yield "Installing GPU dependencies dynamically (Unsloth, xformers, trl). Please wait ~3 minutes..."
+    
+    # Run pip install dynamically at runtime (when GPU is attached)
+    pip_cmd = [
+        sys.executable, "-m", "pip", "install", 
+        "unsloth", "xformers", "trl", "peft", "accelerate", "bitsandbytes"
+    ]
     
+    p_pip = subprocess.Popen(pip_cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
+    for line in p_pip.stdout:
+        yield f"[PIP] {line.strip()}"
+    p_pip.wait()
+    
+    if p_pip.returncode != 0:
+        yield "ERROR: Failed to install dependencies."
+        return
+        
+    yield "Dependencies installed successfully! Starting Training Script..."
+
     train_script = """
 import os
 import torch
@@ -98,7 +116,7 @@ for row in ds_njirlah:
             {'role': 'assistant', 'content': str(out)},
         ]
         all_texts.append(tokenizer.apply_chat_template(convo, tokenize=False, add_generation_prompt=False))
-    elif 'text' in row and '<|im_start|>' in row['text']: # If already formatted in chatml
+    elif 'text' in row and '<|im_start|>' in row['text']: 
         all_texts.append(row['text'])
 
 merged_dataset = Dataset.from_dict({'text': all_texts})
@@ -146,11 +164,10 @@ print('ALL DONE!')
     with open("train.py", "w") as f:
         f.write(train_script)
     
-    yield "Training script generated! Starting subprocess..."
-    process = subprocess.Popen(["python", "train.py"], stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
+    process = subprocess.Popen([sys.executable, "train.py"], stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
     
     for line in process.stdout:
-        yield f"Logs: {line.strip()}"
+        yield f"[TRAIN] {line.strip()}"
         
     process.wait()
     if process.returncode == 0:
@@ -160,13 +177,13 @@ print('ALL DONE!')
 
 with gr.Blocks(title="NJIRLAH-OSS-1 Mega Finetune", theme=gr.themes.Monochrome()) as app:
     gr.Markdown("# 🚀 NJIRLAH-OSS-1 Mega Finetune Engine (Hugging Face Spaces)")
-    gr.Markdown("Tool ini akan menjalankan seluruh logic training dari Kaggle sebelumnya secara otomatis menggunakan GPU di Hugging Face Spaces.")
+    gr.Markdown("Tool ini akan menjalankan seluruh logic training dari Kaggle sebelumnya secara otomatis menggunakan GPU di Hugging Face Spaces. **Pastikan Anda sudah mengaktifkan GPU A10G atau L4 di Settings Space ini!**")
     
     with gr.Row():
         hf_token_input = gr.Textbox(label="Hugging Face Token", type="password", placeholder="hf_...")
         start_btn = gr.Button("Mulai Finetune & Push ke Hub!", variant="primary")
         
-    output_logs = gr.Textbox(label="Training Logs", lines=20, max_lines=30)
+    output_logs = gr.Textbox(label="Live Logs", lines=20, max_lines=30)
     
     start_btn.click(fn=run_finetuning, inputs=hf_token_input, outputs=output_logs)
 

requirements.txt

@@ -1,9 +1,3 @@
 gradio
-torch
-transformers
+huggingface_hub
 datasets
-trl
-peft
-accelerate
-bitsandbytes
-unsloth[cu121-ampere] @ git+https://github.com/unslothai/unsloth.git