Upload gemma4_heretic_vlm.ipynb

gemma4_heretic_vlm.ipynb

@@ -1 +1 @@
-{"cells":[{"cell_type":"code","execution_count":null,"metadata":{"id":"pb9bxY0OE81H"},"outputs":[],"source":["from google.colab import drive\n","drive.mount('/content/drive')"]},{"cell_type":"code","source":["#@markdown # Cell 1a: REVISED – Install + CORRECT Direct GPU Load (No Quantization, Exact Model Card)\n","\n","!pip install -q --upgrade --force-reinstall \"pillow<12.0\"\n","!pip install -q --upgrade transformers accelerate bitsandbytes\n","\n","import torch\n","from transformers import AutoProcessor, AutoModelForMultimodalLM\n","from google.colab import userdata\n","import os\n","\n","# Get HF token from secrets\n","hf_token = userdata.get('HF_TOKEN')\n","if not hf_token:\n","    print(\"⚠️ Please add HF_TOKEN to Colab secrets (key icon on the left sidebar)\")\n","\n","# Model ID\n","model_id = \"coder3101/gemma-4-E2B-it-heretic\"\n","\n","print(f\"🔽 Loading Gemma-4-E2B-Heretic **EXACTLY** as recommended on model card...\")\n","\n","processor = AutoProcessor.from_pretrained(model_id, token=hf_token)\n","\n","# ←←← THIS IS THE OFFICIAL RECOMMENDED LOADING (fixes vision hangs + gray-image bug)\n","model = AutoModelForMultimodalLM.from_pretrained(\n","    model_id,\n","    token=hf_token,\n","    dtype=\"auto\",            # ← model card uses \"dtype\" (not torch_dtype)\n","    device_map=\"auto\",       # ← automatically puts vision + text on GPU\n","    low_cpu_mem_usage=True,\n",")\n","\n","print(\"✅ Model loaded DIRECTLY to GPU (vision encoder fully initialized)\")\n","print(f\"   Device: {model.device}\")\n","print(f\"   Dtype:  {model.dtype}\")\n","print(f\"   VRAM used: {torch.cuda.memory_allocated() / 1024**3:.2f} GB\")\n","print(\"   Ready for Cell 1b (clear) and Cell 2 (captions). No more hangs expected.\")"],"metadata":{"cellView":"form","id":"c7TmhIAuAE6M"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["#@markdown # Cell 1b: REVISED – Clean VRAM & Confirm GPU Load\n","\n","import torch\n","import gc\n","\n","# Clear any leftover from previous broken loads\n","torch.cuda.empty_cache()\n","gc.collect()\n","\n","print(\"🧹 VRAM fully cleared.\")\n","\n","# Re-confirm the model is still on GPU (safety check)\n","print(f\"✅ Current model device: {model.device}\")\n","print(f\"   VRAM allocated: {torch.cuda.memory_allocated() / 1024**3:.2f} GB\")\n","print(f\"   GPU: {torch.cuda.get_device_name(0)}\")\n","print(\"\\n✅ Model is ready for Cell 3 debug / Cell 2 captions.\")\n","print(\"   Vision tower is now correctly loaded – the previous hanging should be fixed.\")"],"metadata":{"cellView":"form","id":"pUrpfegeAG8v"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["#@markdown # Cell 2: Upload Images → Generate Accurate Captions → Download ZIP\n","\n","from google.colab import files\n","import zipfile\n","import os\n","from PIL import Image\n","\n","output_dir = \"/content/image_text_pairs\"\n","os.makedirs(output_dir, exist_ok=True)\n","\n","print(\"📤 Upload your image(s) (multiple files allowed)\")\n","uploaded = files.upload()\n","\n","prompt_style = \"Describe this image in detail, including key objects, scene, colors, mood, lighting, and any visible text. Be accurate and comprehensive. you may not use newline or **bold font**.  You may not itemize. I want a single block of text , 400 words long. \"  #@param {type:\"string\"}\n","\n","generated_count = 0\n","\n","for filename in list(uploaded.keys()):\n","    if filename.lower().endswith(('.png', '.jpg', '.jpeg', '.webp', '.bmp')):\n","        image_path = f\"/content/{filename}\"\n","        image = Image.open(image_path).convert(\"RGB\")\n","\n","        # Use local file path (fastest + most reliable for this processor)\n","        messages = [\n","            {\n","                \"role\": \"user\",\n","                \"content\": [\n","                    {\"type\": \"image\", \"url\": image_path},\n","                    {\"type\": \"text\", \"text\": prompt_style}\n","                ]\n","            }\n","        ]\n","\n","        inputs = processor.apply_chat_template(\n","            messages,\n","            add_generation_prompt=True,\n","            tokenize=True,\n","            return_tensors=\"pt\",\n","            return_dict=True\n","        ).to(model.device)\n","\n","        print(f\"📸 Processing {filename} → generating caption...\")\n","\n","        with torch.no_grad():\n","            outputs = model.generate(\n","                **inputs,\n","                max_new_tokens=512,\n","                do_sample=True,\n","                temperature=0.7,\n","                top_p=0.9,\n","            )\n","\n","        # Decode and clean output (Gemma 4 returns structured dict)\n","        input_len = inputs[\"input_ids\"].shape[-1]\n","        raw = processor.decode(outputs[0][input_len:], skip_special_tokens=False)\n","        parsed = processor.parse_response(raw) if hasattr(processor, \"parse_response\") else raw.strip()\n","\n","        # Extract just the text content\n","        caption = parsed[\"content\"] if isinstance(parsed, dict) and \"content\" in parsed else str(parsed)\n","\n","        print(f\"   ✅ Caption for {filename} ready\\n\")\n","\n","        # Save image + text pair\n","        base_name = os.path.splitext(filename)[0]\n","        with open(os.path.join(output_dir, f\"{base_name}.txt\"), \"w\", encoding=\"utf-8\") as f:\n","            f.write(caption)\n","        image.save(os.path.join(output_dir, filename))\n","        generated_count += 1\n","\n","# Create ZIP and download\n","zip_path = \"/content/image_text_pairs.zip\"\n","with zipfile.ZipFile(zip_path, 'w') as zipf:\n","    for root, _, files_list in os.walk(output_dir):\n","        for file in files_list:\n","            file_path = os.path.join(root, file)\n","            zipf.write(file_path, os.path.relpath(file_path, output_dir))\n","\n","print(f\"\\n🎉 Done! Generated {generated_count} accurate image-text pairs.\")\n","files.download(zip_path)"],"metadata":{"cellView":"form","id":"SnsYbQqkAUiP"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["#@title Cell 3: UPDATED – Talk to the LLM + ask it to caption images (Latest-Reply + Image Display Options)\n","\n","import json\n","import os\n","import torch\n","from google.colab import files\n","from IPython.display import Image, display   # ← Added for small image display\n","\n","#@markdown\n","# ==================== CHAT PARAMETERS ====================\n","user_message = \"Describe this image in detail, including key objects, scene, colors, mood, lighting, and any visible text. Be accurate and comprehensive. you may not use newline or **bold font**.  You may not itemize. I want a single block of text , 400 words long. \" #@param {type:\"string\"}\n","attach_image = True #@param {type:\"boolean\"}\n","# ← Check this ONLY when you want to upload a NEW image\n","reset_chat = False #@param {type:\"boolean\"}\n","only_show_latest_reply = True #@param {type:\"boolean\"}\n","# ← NEW: When checked, the cell will ONLY show the latest LLM reply (instead of full history)\n","show_image_with_reply = True #@param {type:\"boolean\"}\n","# ← NEW: When checked (and only_show_latest_reply is also checked), displays the currently loaded image as a small preview with the latest reply\n","# =======================================================\n","\n","chat_history_path = \"/content/chat_history.json\"\n","chat_log_path = \"/content/chat_log.txt\"\n","latest_image_file = \"/content/latest_image_path.txt\"\n","\n","# Safety check – model must be loaded from Cell 1a/1b\n","if 'model' not in globals() or 'processor' not in globals():\n","    print(\"⚠️ ERROR: Model or processor not found. Please run Cell 1a → Cell 1b first!\")\n","else:\n","    print(\"✅ Model ready for multimodal chat.\")\n","    print(\"📸 Latest-image mode ACTIVE + FULL HISTORY MODE:\")\n","    print(\"   → Every query sends the ENTIRE chatlog (all previous user + LLM replies) + latest image + current text.\")\n","\n","    # Reset logic (includes latest image)\n","    if reset_chat:\n","        if os.path.exists(chat_history_path):\n","            os.remove(chat_history_path)\n","        if os.path.exists(chat_log_path):\n","            os.remove(chat_log_path)\n","        if os.path.exists(latest_image_file):\n","            os.remove(latest_image_file)\n","        print(\"🗑️ Chat history, log, and latest image have been completely reset.\")\n","\n","    # Load previous conversation (if not reset)\n","    if os.path.exists(chat_history_path) and not reset_chat:\n","        with open(chat_history_path, \"r\", encoding=\"utf-8\") as f:\n","            messages = json.load(f)\n","        print(f\"   📜 Loaded full chat history ({len(messages)} messages) – will be sent to the LLM.\")\n","    else:\n","        messages = []\n","        print(\"   📜 Starting fresh conversation.\")\n","\n","    # Load latest image path (persistent across cell runs)\n","    latest_image_path = None\n","    if os.path.exists(latest_image_file):\n","        with open(latest_image_file, \"r\", encoding=\"utf-8\") as f:\n","            latest_image_path = f.read().strip()\n","        if not os.path.exists(latest_image_path):\n","            latest_image_path = None\n","\n","    # Add new user message (if provided)\n","    image_path = None\n","    if user_message.strip():\n","        # === IMAGE HANDLING: upload new OR reuse latest ===\n","        if attach_image:\n","            print(\"📤 Upload the image to set as the NEW latest image...\")\n","            uploaded = files.upload()\n","            if uploaded:\n","                filename = next(iter(uploaded.keys()))\n","                if filename.lower().endswith(('.png', '.jpg', '.jpeg', '.webp', '.bmp')):\n","                    image_path = f\"/content/{filename}\"\n","                    latest_image_path = image_path\n","                    with open(latest_image_file, \"w\", encoding=\"utf-8\") as f:\n","                        f.write(image_path)\n","                    print(f\"   📸 New latest image set: {filename}\")\n","                else:\n","                    print(\"⚠️ Uploaded file is not an image – keeping previous latest image.\")\n","\n","        # Reuse latest image if no new upload\n","        if image_path is None and latest_image_path is not None:\n","            image_path = latest_image_path\n","            print(f\"   📸 Re-using latest image: {os.path.basename(image_path)}\")\n","\n","        # Build user content: image (if available) + text (always included)\n","        user_content = []\n","        if image_path is not None:\n","            user_content.append({\"type\": \"image\", \"url\": image_path})\n","        user_content.append({\"type\": \"text\", \"text\": user_message})\n","\n","        messages.append({\"role\": \"user\", \"content\": user_content})\n","\n","        print(f\"\\n👤 You: {user_message}\")\n","        if image_path is not None:\n","            print(f\"   📸 (with latest image: {os.path.basename(image_path)})\")\n","\n","        # === KEY CHANGE: Full history (including all LLM replies) is always provided ===\n","        print(f\"   🔄 Sending FULL chat history ({len(messages)} total messages) to Gemma-4...\")\n","\n","        # Prepare inputs using the exact same chat template style as Cell 2\n","        inputs = processor.apply_chat_template(\n","            messages,                    # ← Entire history (user + assistant turns) is passed here\n","            add_generation_prompt=True,\n","            tokenize=True,\n","            return_tensors=\"pt\",\n","            return_dict=True\n","        ).to(model.device)\n","\n","        print(\"🤖 Generating response... (this may take a few seconds)\")\n","\n","        with torch.no_grad():\n","            outputs = model.generate(\n","                **inputs,\n","                max_new_tokens=512,\n","                do_sample=True,\n","                temperature=0.7,\n","                top_p=0.9,\n","            )\n","\n","        # Decode only the new tokens (same logic as Cell 2)\n","        input_len = inputs[\"input_ids\"].shape[-1]\n","        raw = processor.decode(outputs[0][input_len:], skip_special_tokens=False)\n","\n","        # Use the same parsing as Cell 2 (handles Gemma-4 structured output)\n","        if hasattr(processor, \"parse_response\"):\n","            parsed = processor.parse_response(raw)\n","        else:\n","            parsed = raw.strip()\n","\n","        reply = parsed[\"content\"] if isinstance(parsed, dict) and \"content\" in parsed else str(parsed)\n","\n","        # Add assistant reply to history (so next turn keeps the full conversation)\n","        assistant_content = [{\"type\": \"text\", \"text\": reply}]\n","        messages.append({\"role\": \"assistant\", \"content\": assistant_content})\n","\n","        # Print the LLM reply\n","        print(f\"\\n🤖 Gemma-4-Heretic: {reply}\\n\")\n","\n","        # Save structured history (JSON) for perfect persistence across runs\n","        with open(chat_history_path, \"w\", encoding=\"utf-8\") as f:\n","            json.dump(messages, f, ensure_ascii=False, indent=2)\n","\n","        # Rebuild human-readable chat log TXT (full conversation every time)\n","        with open(chat_log_path, \"w\", encoding=\"utf-8\") as f:\n","            for msg in messages:\n","                role = \"👤 You\" if msg[\"role\"] == \"user\" else \"🤖 Gemma-4-Heretic\"\n","                content = msg.get(\"content\", [])\n","                text_part = \"\"\n","                image_info = \"\"\n","\n","                if isinstance(content, list):\n","                    for item in content:\n","                        if item.get(\"type\") == \"image\":\n","                            url = item.get(\"url\", \"\")\n","                            filename = os.path.basename(url) if url else \"unknown_image\"\n","                            image_info = f\"[📸 Latest image: {filename}] \"\n","                            break\n","                    text_part = next((item.get(\"text\", \"\") for item in content if item.get(\"type\") == \"text\"), \"\")\n","                else:\n","                    text_part = str(content or \"\")\n","\n","                if image_info:\n","                    text_part = image_info + (text_part if text_part else \"\")\n","\n","                f.write(f\"{role}:\\n{text_part}\\n\\n{'─' * 60}\\n\\n\")\n","\n","        print(f\"📝 Chat log updated → {chat_log_path}\")\n","        print(\"   (You can download this TXT anytime from the file browser on the left)\")\n","\n","    # ==================== UPDATED DISPLAY LOGIC ====================\n","    print(\"\\n\" + \"=\"*80)\n","\n","    if only_show_latest_reply:\n","        print(\"📜 ONLY SHOWING LATEST REPLY (checkbox enabled)\")\n","        latest_reply = None\n","        for msg in reversed(messages):\n","            if msg[\"role\"] == \"assistant\":\n","                content = msg.get(\"content\", [])\n","                if isinstance(content, list):\n","                    latest_reply = next((item.get(\"text\", \"\") for item in content if item.get(\"type\") == \"text\"), \"\")\n","                else:\n","                    latest_reply = str(content or \"\")\n","                break\n","\n","        if latest_reply:\n","            print(\"🤖 Latest Gemma-4-Heretic Reply:\")\n","            print(latest_reply)\n","\n","            if show_image_with_reply and latest_image_path and os.path.exists(latest_image_path):\n","                print(\"\\n📸 Currently loaded image (attached to this reply):\")\n","                display(Image(filename=latest_image_path, width=400))  # small image preview\n","                print(\"   (↑ Small image displayed above)\")\n","            elif show_image_with_reply:\n","                print(\"   (No image currently loaded)\")\n","        else:\n","            print(\"💬 No LLM reply yet in this conversation.\")\n","\n","    else:\n","        # Original full-history behavior (unchanged)\n","        if os.path.exists(chat_log_path):\n","            print(\"📜 FULL CHAT HISTORY (from chat_log.txt):\")\n","            with open(chat_log_path, \"r\", encoding=\"utf-8\") as f:\n","                print(f.read())\n","        elif not reset_chat:\n","            print(\"💬 No messages yet. Type something in the box above and run the cell!\")\n","\n","    print(\"=\"*80)\n","\n","    print(\"\\n✅ Ready for next message!\")\n","    print(\"   → The **entire chatlog** (all your messages + all Gemma-4 replies) is sent every time.\")\n","    print(\"   → Latest uploaded image is automatically attached to every query.\")\n","    print(\"   → Your text input is always sent together with the image.\")\n","    print(\"   Check 'attach_image' only when you want to upload a different image.\")\n","    print(\"   (Check 'Reset Chat' if you want to start a completely new conversation)\")\n","    print(\"\\n   🆕 NEW FEATURES:\")\n","    print(\"   → Toggle 'only_show_latest_reply' to show ONLY the latest LLM reply (cleaner output).\")\n","    print(\"   → Toggle 'show_image_with_reply' to display the currently loaded image as a small preview with the latest reply.\")"],"metadata":{"cellView":"form","id":"z30ARmSDYTJX"},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":["LoRa training"],"metadata":{"id":"m8W_d-1iRrr2"}},{"cell_type":"code","source":["import os\n","import zipfile\n","from pathlib import Path\n","import torch\n","import gc\n","from google.colab import files\n","from PIL import Image\n","from google.colab import drive\n","#@markdown Train LoRa on LLM using marching filenames in zip.  In the txt files found in zip file , you can divide the Query and Reply desired with '//--//'\n","# ==================== MOUNT GOOGLE DRIVE FIRST ====================\n","print(\"🔌 Checking Google Drive...\")\n","if not os.path.exists('/content/drive/MyDrive'):\n","    print(\"📂 Mounting Google Drive (this may ask for permission)...\")\n","    drive.mount('/content/drive', force_remount=False)\n","else:\n","    print(\"✅ Google Drive already mounted.\")\n","\n","drive_adapter_path = \"/content/drive/MyDrive/gemma_lora_adapter\"\n","os.makedirs(drive_adapter_path, exist_ok=True)\n","print(f\"💾 LoRA will be saved directly to: {drive_adapter_path}\")\n","\n","# ==================== TRAINING PARAMETERS ====================\n","zip_path = \"\" #@param {type:\"string\"}\n","upload_new_zip = True #@param {type:\"boolean\"}\n","num_epochs = 3 #@param {type:\"slider\", min:1, max:10, step:1}\n","lora_rank = 16 #@param {type:\"integer\"}\n","lora_alpha = 32 #@param {type:\"integer\"}\n","learning_rate = \"2e-4\" #@param {type:\"string\"}\n","batch_size = 1 #@param {type:\"integer\"}\n","gradient_accumulation_steps = 4 #@param {type:\"integer\"}\n","# ===========================================================\n","\n","# Install required packages for LoRA training\n","!pip install -q --upgrade peft trl\n","\n","# Get HF token (reuse from previous cells)\n","from google.colab import userdata\n","hf_token = userdata.get('HF_TOKEN')\n","if not hf_token:\n","    print(\"⚠️ Please add HF_TOKEN to Colab secrets first!\")\n","\n","# Model ID (same as Cell 1a)\n","model_id = \"coder3101/gemma-4-E2B-it-heretic\"\n","\n","# ====================== UNLOAD VRAM FIRST ======================\n","print(\"\\n🧹 Unloading VRAM prior to LoRA training...\")\n","torch.cuda.empty_cache()\n","gc.collect()\n","\n","reuse_model = False\n","if 'model' in globals() and 'processor' in globals():\n","    current_device = getattr(model, 'device', None)\n","    print(f\"   Detected existing model on device: {current_device}\")\n","    if current_device is not None and current_device.type == 'cpu':\n","        print(\"✅ Transformer model found on CPU – reusing existing model (no re-download from HF)\")\n","        reuse_model = True\n","    else:\n","        print(\"🔄 Existing model detected (likely on GPU). Unloading to load efficient 4-bit version...\")\n","        del globals()['model']\n","        if 'processor' in globals():\n","            del globals()['processor']\n","        torch.cuda.empty_cache()\n","        gc.collect()\n","else:\n","    print(\"🆕 No existing model found – will load from Hugging Face.\")\n","\n","# ====================== LOAD / REUSE MODEL ======================\n","print(\"\\n🔄 Preparing model for LoRA training...\")\n","\n","from transformers import BitsAndBytesConfig, TrainingArguments, Trainer\n","from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training\n","from trl import SFTTrainer  # kept for future compatibility\n","\n","if reuse_model:\n","    # Reuse the CPU-loaded model (full precision)\n","    print(\"   Reusing existing model + processor\")\n","    # Prepare for kbit training (works on already-loaded model)\n","    model = prepare_model_for_kbit_training(model)\n","else:\n","    # Standard 4-bit load (most common case)\n","    print(\"   Loading model in 4-bit for efficient training...\")\n","    quant_config = BitsAndBytesConfig(\n","        load_in_4bit=True,\n","        bnb_4bit_quant_type=\"nf4\",\n","        bnb_4bit_compute_dtype=torch.bfloat16,\n","        bnb_4bit_use_double_quant=True,\n","    )\n","\n","    processor = AutoProcessor.from_pretrained(model_id, token=hf_token)\n","\n","    model = AutoModelForMultimodalLM.from_pretrained(\n","        model_id,\n","        token=hf_token,\n","        quantization_config=quant_config,\n","        device_map=\"auto\",\n","        low_cpu_mem_usage=True,\n","    )\n","\n","    model = prepare_model_for_kbit_training(model)\n","\n","# Apply LoRA (targets all linear layers – perfect for Gemma-4)\n","lora_config = LoraConfig(\n","    r=lora_rank,\n","    lora_alpha=lora_alpha,\n","    target_modules=\"all-linear\",\n","    lora_dropout=0.05,\n","    bias=\"none\",\n","    task_type=\"CAUSAL_LM\",\n",")\n","\n","model = get_peft_model(model, lora_config)\n","model.print_trainable_parameters()\n","\n","print(\"✅ Model + LoRA adapter ready (VRAM optimized)\")\n","\n","# ====================== DATASET PREPARATION ======================\n","print(\"\\n📦 Preparing dataset from ZIP...\")\n","\n","# Handle ZIP upload or existing path\n","if upload_new_zip or not zip_path or not os.path.exists(zip_path):\n","    print(\"📤 Upload your image-text pairs ZIP file (must contain matching .jpg/.png + .txt files)\")\n","    uploaded = files.upload()\n","    if uploaded:\n","        zip_filename = list(uploaded.keys())[0]\n","        zip_path = f\"/content/{zip_filename}\"\n","        print(f\"✅ Using uploaded file: {zip_filename}\")\n","    else:\n","        print(\"❌ No ZIP uploaded. Aborting.\")\n","        raise SystemExit\n","\n","if not os.path.exists(zip_path):\n","    print(f\"❌ ZIP file not found: {zip_path}\")\n","    raise SystemExit\n","\n","# Unpack ZIP\n","dataset_dir = \"/content/lora_training_data\"\n","os.makedirs(dataset_dir, exist_ok=True)\n","\n","with zipfile.ZipFile(zip_path, 'r') as z:\n","    z.extractall(dataset_dir)\n","\n","print(f\"✅ Extracted dataset to {dataset_dir}\")\n","\n","# Find all matching image + txt pairs\n","image_exts = ('.png', '.jpg', '.jpeg', '.webp', '.bmp')\n","pairs = []\n","\n","for txt_file in Path(dataset_dir).glob(\"**/*.txt\"):\n","    base_name = txt_file.stem\n","    image_path = None\n","    for ext in image_exts:\n","        candidate = txt_file.with_name(base_name + ext)\n","        if candidate.exists():\n","            image_path = str(candidate)\n","            break\n","    if not image_path:\n","        continue\n","\n","    with open(txt_file, \"r\", encoding=\"utf-8\") as f:\n","        text_content = f.read().strip()\n","\n","    # Split into Query and Reply using '//--//'\n","    if '//--//' in text_content:\n","        query_part, reply_part = [x.strip() for x in text_content.split('//--//', 1)]\n","    else:\n","        query_part = \"Describe this image in detail, including key objects, scene, colors, mood, lighting, and any visible text. Be accurate and comprehensive.\"\n","        reply_part = text_content\n","\n","    pairs.append({\n","        \"image_path\": image_path,\n","        \"query\": query_part,\n","        \"reply\": reply_part\n","    })\n","\n","print(f\"✅ Found {len(pairs)} valid image-text training pairs.\")\n","\n","if len(pairs) == 0:\n","    print(\"❌ No matching image+txt pairs found in the ZIP. Aborting.\")\n","    raise SystemExit\n","\n","# ====================== CUSTOM DATASET + COLLATOR ======================\n","class MultimodalTrainingDataset(torch.utils.data.Dataset):\n","    def __init__(self, pairs, processor):\n","        self.pairs = pairs\n","        self.processor = processor\n","\n","    def __len__(self):\n","        return len(self.pairs)\n","\n","    def __getitem__(self, idx):\n","        item = self.pairs[idx]\n","        messages = [\n","            {\"role\": \"user\", \"content\": [\n","                {\"type\": \"image\", \"url\": item[\"image_path\"]},\n","                {\"type\": \"text\", \"text\": item[\"query\"]}\n","            ]},\n","            {\"role\": \"assistant\", \"content\": [{\"type\": \"text\", \"text\": item[\"reply\"]}]}\n","        ]\n","\n","        inputs = self.processor.apply_chat_template(\n","            messages,\n","            add_generation_prompt=False,\n","            tokenize=True,\n","            return_tensors=\"pt\",\n","            return_dict=True\n","        )\n","        inputs[\"labels\"] = inputs[\"input_ids\"].clone()\n","\n","        # Remove batch dim\n","        for k in list(inputs.keys()):\n","            if isinstance(inputs[k], torch.Tensor):\n","                inputs[k] = inputs[k].squeeze(0)\n","        return inputs\n","\n","train_dataset = MultimodalTrainingDataset(pairs, processor)\n","\n","def multimodal_collator(features):\n","    batch = {}\n","    input_ids = [f[\"input_ids\"] for f in features]\n","    attention_mask = [f[\"attention_mask\"] for f in features]\n","    labels = [f[\"labels\"] for f in features]\n","\n","    text_padded = processor.tokenizer.pad(\n","        {\"input_ids\": input_ids, \"attention_mask\": attention_mask},\n","        padding=True,\n","        return_tensors=\"pt\"\n","    )\n","    batch[\"input_ids\"] = text_padded[\"input_ids\"]\n","    batch[\"attention_mask\"] = text_padded[\"attention_mask\"]\n","\n","    labels_padded = processor.tokenizer.pad(\n","        {\"input_ids\": labels},\n","        padding=True,\n","        return_tensors=\"pt\"\n","    )[\"input_ids\"]\n","    labels_padded[labels_padded == processor.tokenizer.pad_token_id] = -100\n","    batch[\"labels\"] = labels_padded\n","\n","    # Vision tensors (pixel_values etc.)\n","    first_item = features[0]\n","    for key in first_item.keys():\n","        if key in [\"input_ids\", \"attention_mask\", \"labels\"]:\n","            continue\n","        if isinstance(first_item[key], torch.Tensor):\n","            vals = [f[key] for f in features]\n","            batch[key] = torch.stack(vals) if vals[0].dim() == 3 else torch.cat(vals)\n","\n","    return batch\n","\n","# ====================== TRAINING ======================\n","print(f\"\\n🚀 Starting LoRA training on {len(pairs)} pairs for {num_epochs} epochs...\")\n","\n","training_args = TrainingArguments(\n","    output_dir=drive_adapter_path,   # ← Saved directly to Google Drive\n","    num_train_epochs=num_epochs,\n","    per_device_train_batch_size=batch_size,\n","    gradient_accumulation_steps=gradient_accumulation_steps,\n","    learning_rate=float(learning_rate),\n","    fp16=True,\n","    optim=\"paged_adamw_8bit\",\n","    logging_steps=5,\n","    save_steps=50,\n","    save_total_limit=2,\n","    report_to=\"none\",\n","    remove_unused_columns=False,\n",")\n","\n","trainer = Trainer(\n","    model=model,\n","    args=training_args,\n","    train_dataset=train_dataset,\n","    data_collator=multimodal_collator,\n",")\n","\n","trainer.train()\n","\n","# ====================== SAVE TO DRIVE + ZIP FOR DOWNLOAD ======================\n","print(\"\\n💾 Saving final LoRA adapter to Google Drive...\")\n","\n","model.save_pretrained(drive_adapter_path)\n","processor.save_pretrained(drive_adapter_path)\n","\n","# Create downloadable ZIP (from Drive folder)\n","zip_adapter_path = \"/content/gemma_lora_adapter.zip\"\n","with zipfile.ZipFile(zip_adapter_path, 'w') as z:\n","    for root, _, files_list in os.walk(drive_adapter_path):\n","        for file in files_list:\n","            file_path = os.path.join(root, file)\n","            arcname = os.path.relpath(file_path, drive_adapter_path)\n","            z.write(file_path, arcname)\n","\n","print(f\"🎉 LoRA training complete!\")\n","print(f\"   ✅ Saved to Google Drive: {drive_adapter_path}\")\n","print(f\"   📦 Zipped copy ready for download → {zip_adapter_path}\")\n","files.download(zip_adapter_path)\n","\n","print(\"\\n✅ You can now use this LoRA adapter from Google Drive in future inference cells.\")\n","print(\"   Example: peft_model = PeftModel.from_pretrained(base_model, drive_adapter_path)\")\n","```​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​"],"metadata":{"cellView":"form","id":"4cCNLXWsKHe_"},"execution_count":null,"outputs":[]}],"metadata":{"accelerator":"GPU","colab":{"gpuType":"T4","provenance":[{"file_id":"1IFAN1z44DC5xcoAdXNeupO-nUTedA2uG","timestamp":1775323816042},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775308348942},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775233835241},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775233735620},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775159289229},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774964848181},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774906863352},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774901938471},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774884683620},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774883563998},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774882033159},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774881268156},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Drive to WebP.ipynb","timestamp":1774879795776},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Drive to WebP.ipynb","timestamp":1763646205520},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Drive to WebP.ipynb","timestamp":1760993725927},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1760450712160},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1756712618300},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1747490904984},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1740037333374},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1736477078136},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1725365086834}]},"kernelspec":{"display_name":"Python 3","name":"python3"},"language_info":{"name":"python"}},"nbformat":4,"nbformat_minor":0}
+{"cells":[{"cell_type":"code","source":["#@markdown # Cell 1: Connect Google Drive + Restore HF_TOKEN (Run this FIRST)\n","\n","from google.colab import drive\n","from google.colab import userdata\n","import os\n","from pathlib import Path\n","\n","print(\"🔄 Connecting to Google Drive and setting up HF_TOKEN...\")\n","\n","# Mount Drive\n","drive.mount('/content/drive', force_remount=True)\n","\n","# Create working folder (everything will be saved here)\n","WORKING_DIR = \"/content/drive/MyDrive/Gemma4Captioner\"\n","os.makedirs(WORKING_DIR, exist_ok=True)\n","print(f\"✅ Working directory ready: {WORKING_DIR}\")\n","\n","# Restore HF_TOKEN from Colab secrets\n","hf_token = userdata.get('HF_TOKEN')\n","if hf_token:\n","    os.environ[\"HF_TOKEN\"] = hf_token\n","    print(\"✅ HF_TOKEN restored from Colab Secrets.\")\n","else:\n","    print(\"⚠️ HF_TOKEN not found. Please click the key icon on the left sidebar, add 'HF_TOKEN', and run this cell again.\")\n","\n","print(\"\\n✅ Cell 1 complete! You can now run Cell 2 to load the model.\")"],"metadata":{"cellView":"form","id":"oQRO5Wo8DXT2"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["#@markdown # Cell 2: Load Gemma-4 Model (Heretic by default + Vision Latent Helpers)\n","\n","!pip install -q --upgrade --force-reinstall \"pillow<12.0\"\n","!pip install -q --upgrade transformers accelerate bitsandbytes\n","\n","import torch\n","from transformers import AutoProcessor, AutoModelForMultimodalLM\n","from google.colab import userdata\n","from PIL import Image\n","import base64\n","import io\n","\n","# ====================== VISION LATENT IMPROVEMENT HELPERS ======================\n","def get_optimal_vision_size(processor):\n","    \"\"\"Automatically detects the best resolution for Gemma-4's vision encoder (improves latent features).\"\"\"\n","    target_max_dim = 1024\n","    if hasattr(processor, \"image_processor\") and hasattr(processor.image_processor, \"size\"):\n","        size_cfg = processor.image_processor.size\n","        if isinstance(size_cfg, dict):\n","            dims = [v for v in size_cfg.values() if isinstance(v, (int, float))]\n","            if dims:\n","                target_max_dim = int(max(dims))\n","        elif isinstance(size_cfg, (int, float)):\n","            target_max_dim = int(size_cfg)\n","    return target_max_dim\n","\n","def image_to_data_url(img: Image.Image) -> str:\n","    \"\"\"Lossless PNG data URL → cleanest latent representation for the vision encoder.\"\"\"\n","    buffered = io.BytesIO()\n","    img.save(buffered, format=\"PNG\")\n","    img_str = base64.b64encode(buffered.getvalue()).decode(\"utf-8\")\n","    return f\"data:image/png;base64,{img_str}\"\n","\n","#@markdown **Choose your model version:**\n","use_heretic_version = True #@param {type:\"boolean\"}\n","\n","hf_token = os.environ.get(\"HF_TOKEN\") or userdata.get('HF_TOKEN')\n","\n","if use_heretic_version:\n","    model_id = \"coder3101/gemma-4-E2B-it-heretic\"\n","    print(\"🔽 Loading **Gemma-4-E2B-Heretic** (your preferred uncensored version)...\")\n","    processor = AutoProcessor.from_pretrained(model_id, token=hf_token)\n","    model = AutoModelForMultimodalLM.from_pretrained(\n","        model_id,\n","        token=hf_token,\n","        dtype=\"auto\",\n","        device_map=\"auto\",\n","        low_cpu_mem_usage=True,\n","    )\n","else:\n","    model_id = \"google/gemma-4-E2B-it\"\n","    print(\"🔽 Loading **Official Google Gemma-4-E2B-it**...\")\n","    processor = AutoProcessor.from_pretrained(model_id, token=hf_token)\n","    model = AutoModelForMultimodalLM.from_pretrained(\n","        model_id,\n","        token=hf_token,\n","        torch_dtype=\"auto\",\n","        device_map=\"auto\",\n","        low_cpu_mem_usage=True,\n","    )\n","\n","print(\"✅ Model loaded successfully!\")\n","print(f\"   Model: {model_id}\")\n","print(f\"   Device: {model.device}\")\n","print(f\"   VRAM used: {torch.cuda.memory_allocated() / 1024**3:.2f} GB\")\n","print(\"   ✅ Vision latent helpers are now active (optimal resize + lossless PNG)\")\n","print(\"\\nCell 2 complete. Run Cell 3 or Cell 4 next.\")"],"metadata":{"cellView":"form","id":"4btb30LwDZ2r"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["#@markdown # Cell 3: Interactive Chat – Talk to Gemma-4 with Images (Full History + Enhanced Vision)\n","\n","import json\n","import os\n","import torch\n","from google.colab import files\n","from IPython.display import Image, display\n","\n","#@markdown\n","user_message = \"Describe this image in detail, including key objects, scene, colors, mood, lighting, and any visible text. Be accurate and comprehensive. you may not use newline or **bold font**.  You may not itemize. I want a single block of text , 400 words long. \" #@param {type:\"string\"}\n","attach_image = True #@param {type:\"boolean\"}\n","reset_chat = False #@param {type:\"boolean\"}\n","only_show_latest_reply = True #@param {type:\"boolean\"}\n","show_image_with_reply = True #@param {type:\"boolean\"}\n","\n","chat_history_path = \"/content/chat_history.json\"\n","chat_log_path = \"/content/chat_log.txt\"\n","latest_image_file = \"/content/latest_image_path.txt\"\n","\n","if 'model' not in globals() or 'processor' not in globals():\n","    print(\"⚠️ Please run Cell 1 → Cell 2 first!\")\n","else:\n","    print(\"✅ Ready for chat with **improved vision latent representation**.\")\n","\n","    if reset_chat:\n","        for p in [chat_history_path, chat_log_path, latest_image_file]:\n","            if os.path.exists(p): os.remove(p)\n","        print(\"🗑️ Chat reset.\")\n","\n","    messages = []\n","    if os.path.exists(chat_history_path) and not reset_chat:\n","        with open(chat_history_path, \"r\", encoding=\"utf-8\") as f:\n","            messages = json.load(f)\n","\n","    latest_image_path = None\n","    if os.path.exists(latest_image_file):\n","        with open(latest_image_file, \"r\", encoding=\"utf-8\") as f:\n","            latest_image_path = f.read().strip()\n","\n","    if user_message.strip():\n","        image_path = None\n","        if attach_image:\n","            print(\"📤 Upload image...\")\n","            uploaded = files.upload()\n","            if uploaded:\n","                filename = next(iter(uploaded.keys()))\n","                if filename.lower().endswith(('.png','.jpg','.jpeg','.webp','.bmp')):\n","                    image_path = f\"/content/{filename}\"\n","                    latest_image_path = image_path\n","                    with open(latest_image_file, \"w\", encoding=\"utf-8\") as f:\n","                        f.write(image_path)\n","\n","        if image_path is None and latest_image_path:\n","            image_path = latest_image_path\n","\n","        user_content = []\n","        if image_path:\n","            raw_image = Image.open(image_path).convert(\"RGB\")\n","            target = get_optimal_vision_size(processor)\n","            if max(raw_image.size) != target or min(raw_image.size) < 512:\n","                scale = target / max(raw_image.size)\n","                raw_image = raw_image.resize((int(raw_image.width*scale), int(raw_image.height*scale)), Image.LANCZOS)\n","            image_url = image_to_data_url(raw_image)\n","            user_content.append({\"type\": \"image\", \"url\": image_url})\n","            print(f\"   🔬 Image prepared with enhanced latent features ({target}px)\")\n","\n","        user_content.append({\"type\": \"text\", \"text\": user_message})\n","        messages.append({\"role\": \"user\", \"content\": user_content})\n","\n","        inputs = processor.apply_chat_template(\n","            messages, add_generation_prompt=True, tokenize=True, return_tensors=\"pt\", return_dict=True\n","        ).to(model.device)\n","\n","        with torch.no_grad():\n","            outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.9)\n","\n","        input_len = inputs[\"input_ids\"].shape[-1]\n","        raw = processor.decode(outputs[0][input_len:], skip_special_tokens=False)\n","        parsed = processor.parse_response(raw) if hasattr(processor, \"parse_response\") else raw.strip()\n","        reply = parsed[\"content\"] if isinstance(parsed, dict) and \"content\" in parsed else str(parsed)\n","\n","        messages.append({\"role\": \"assistant\", \"content\": [{\"type\": \"text\", \"text\": reply}]})\n","\n","        with open(chat_history_path, \"w\", encoding=\"utf-8\") as f:\n","            json.dump(messages, f, ensure_ascii=False, indent=2)\n","\n","        print(f\"\\n🤖 Gemma-4: {reply}\\n\")\n","\n","    # Display logic (same as before)\n","    print(\"\\n\" + \"=\"*80)\n","    if only_show_latest_reply:\n","        latest_reply = None\n","        for msg in reversed(messages):\n","            if msg[\"role\"] == \"assistant\":\n","                content = msg.get(\"content\", [])\n","                latest_reply = next((item.get(\"text\", \"\") for item in content if item.get(\"type\") == \"text\"), \"\") if isinstance(content, list) else str(content)\n","                break\n","        if latest_reply:\n","            print(\"🤖 Latest Reply:\\n\" + latest_reply)\n","            if show_image_with_reply and latest_image_path and os.path.exists(latest_image_path):\n","                display(Image(filename=latest_image_path, width=400))\n","    print(\"=\"*80)"],"metadata":{"cellView":"form","id":"SShL3oqBDdhE"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["#@markdown # Cell 4: Batch Caption Generator – Upload Multiple Images → 400-Word Captions + ZIP Download\n","\n","from google.colab import files\n","import zipfile\n","import os\n","from PIL import Image\n","\n","output_dir = \"/content/image_text_pairs\"\n","os.makedirs(output_dir, exist_ok=True)\n","\n","print(\"📤 Upload your image(s) (you can select multiple)\")\n","uploaded = files.upload()\n","\n","prompt_style = \"Describe this image in detail, including key objects, scene, colors, mood, lighting, and any visible text. Be accurate and comprehensive. you may not use newline or **bold font**.  You may not itemize. I want a single block of text , 400 words long. \"  #@param {type:\"string\"}\n","\n","generated_count = 0\n","\n","for filename in list(uploaded.keys()):\n","    if filename.lower().endswith(('.png', '.jpg', '.jpeg', '.webp', '.bmp')):\n","        image_path = f\"/content/{filename}\"\n","\n","        raw_image = Image.open(image_path).convert(\"RGB\")\n","        target_max_dim = get_optimal_vision_size(processor)\n","        if max(raw_image.size) != target_max_dim or min(raw_image.size) < 512:\n","            scale = target_max_dim / max(raw_image.size)\n","            new_size = (int(raw_image.width * scale), int(raw_image.height * scale))\n","            raw_image = raw_image.resize(new_size, Image.LANCZOS)\n","        image_url = image_to_data_url(raw_image)\n","\n","        print(f\"📸 Processing {filename} (enhanced latent representation)...\")\n","\n","        messages = [{\"role\": \"user\", \"content\": [\n","            {\"type\": \"image\", \"url\": image_url},\n","            {\"type\": \"text\", \"text\": prompt_style}\n","        ]}]\n","\n","        inputs = processor.apply_chat_template(\n","            messages, add_generation_prompt=True, tokenize=True, return_tensors=\"pt\", return_dict=True\n","        ).to(model.device)\n","\n","        with torch.no_grad():\n","            outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.9)\n","\n","        input_len = inputs[\"input_ids\"].shape[-1]\n","        raw = processor.decode(outputs[0][input_len:], skip_special_tokens=False)\n","        parsed = processor.parse_response(raw) if hasattr(processor, \"parse_response\") else raw.strip()\n","        caption = parsed[\"content\"] if isinstance(parsed, dict) and \"content\" in parsed else str(parsed)\n","\n","        base_name = os.path.splitext(filename)[0]\n","        with open(os.path.join(output_dir, f\"{base_name}.txt\"), \"w\", encoding=\"utf-8\") as f:\n","            f.write(caption)\n","        Image.open(image_path).save(os.path.join(output_dir, filename))\n","        generated_count += 1\n","\n","zip_path = \"/content/image_text_pairs.zip\"\n","with zipfile.ZipFile(zip_path, 'w') as zipf:\n","    for root, _, files_list in os.walk(output_dir):\n","        for file in files_list:\n","            zipf.write(os.path.join(root, file), file)\n","\n","print(f\"\\n🎉 Done! Generated {generated_count} captions with enhanced vision.\")\n","files.download(zip_path)"],"metadata":{"cellView":"form","id":"vH512u3QDhzF"},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":["✅ Welcome to your new, beginner-friendly Gemma-4 Image Captioning Notebook!\n","\n","This notebook lets you generate long, detailed 400-word captions for any images using Gemma-4-E2B-Heretic (the uncensored version you prefer).  \n","It automatically improves the latent image representation (the vision features the model actually “sees”) so captions are much more accurate and vivid.\n","\n","### What’s special about this notebook?\n","- Heretic version (default): Removes safety restrictions → richer, more creative, and fully compliant captions.  \n","- Official Google version (optional via checkbox): Same vision quality but slightly more cautious.  \n","- Vision latent improvements (built-in): Every image is resized to the model’s exact preferred resolution + sent as a lossless PNG data URL. This gives the vision encoder the cleanest possible pixel data → stronger latent features → noticeably better captions.  \n","- No more “sort-of correct” descriptions — the model now truly sees the details.\n","\n","### How to use (super simple for new users):\n","1. Run Cell 1 first (connects Drive + restores your HF_TOKEN).  \n","2. Run Cell 2 (loads the model — choose Heretic or Official).  \n","3. Use Cell 3 for interactive chat with images.  \n","4. Use Cell 4 for batch-processing many images + ZIP download.  \n","\n","Always run cells in order (top to bottom).  \n","Everything is saved to Google Drive automatically if you want persistence.\n"],"metadata":{"id":"THiwFxj3D6Xx"}},{"cell_type":"markdown","source":["LoRa training (Untested)"],"metadata":{"id":"m8W_d-1iRrr2"}},{"cell_type":"code","source":["import os\n","import zipfile\n","from pathlib import Path\n","import torch\n","import gc\n","from google.colab import files\n","from PIL import Image\n","from google.colab import drive\n","#@markdown Train LoRa on LLM using marching filenames in zip.  In the txt files found in zip file , you can divide the Query and Reply desired with '//--//'\n","# ==================== MOUNT GOOGLE DRIVE FIRST ====================\n","print(\"🔌 Checking Google Drive...\")\n","if not os.path.exists('/content/drive/MyDrive'):\n","    print(\"📂 Mounting Google Drive (this may ask for permission)...\")\n","    drive.mount('/content/drive', force_remount=False)\n","else:\n","    print(\"✅ Google Drive already mounted.\")\n","\n","drive_adapter_path = \"/content/drive/MyDrive/gemma_lora_adapter\"\n","os.makedirs(drive_adapter_path, exist_ok=True)\n","print(f\"💾 LoRA will be saved directly to: {drive_adapter_path}\")\n","\n","# ==================== TRAINING PARAMETERS ====================\n","zip_path = \"\" #@param {type:\"string\"}\n","upload_new_zip = True #@param {type:\"boolean\"}\n","num_epochs = 3 #@param {type:\"slider\", min:1, max:10, step:1}\n","lora_rank = 16 #@param {type:\"integer\"}\n","lora_alpha = 32 #@param {type:\"integer\"}\n","learning_rate = \"2e-4\" #@param {type:\"string\"}\n","batch_size = 1 #@param {type:\"integer\"}\n","gradient_accumulation_steps = 4 #@param {type:\"integer\"}\n","# ===========================================================\n","\n","# Install required packages for LoRA training\n","!pip install -q --upgrade peft trl\n","\n","# Get HF token (reuse from previous cells)\n","from google.colab import userdata\n","hf_token = userdata.get('HF_TOKEN')\n","if not hf_token:\n","    print(\"⚠️ Please add HF_TOKEN to Colab secrets first!\")\n","\n","# Model ID (same as Cell 1a)\n","model_id = \"coder3101/gemma-4-E2B-it-heretic\"\n","\n","# ====================== UNLOAD VRAM FIRST ======================\n","print(\"\\n🧹 Unloading VRAM prior to LoRA training...\")\n","torch.cuda.empty_cache()\n","gc.collect()\n","\n","reuse_model = False\n","if 'model' in globals() and 'processor' in globals():\n","    current_device = getattr(model, 'device', None)\n","    print(f\"   Detected existing model on device: {current_device}\")\n","    if current_device is not None and current_device.type == 'cpu':\n","        print(\"✅ Transformer model found on CPU – reusing existing model (no re-download from HF)\")\n","        reuse_model = True\n","    else:\n","        print(\"🔄 Existing model detected (likely on GPU). Unloading to load efficient 4-bit version...\")\n","        del globals()['model']\n","        if 'processor' in globals():\n","            del globals()['processor']\n","        torch.cuda.empty_cache()\n","        gc.collect()\n","else:\n","    print(\"🆕 No existing model found – will load from Hugging Face.\")\n","\n","# ====================== LOAD / REUSE MODEL ======================\n","print(\"\\n🔄 Preparing model for LoRA training...\")\n","\n","from transformers import BitsAndBytesConfig, TrainingArguments, Trainer\n","from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training\n","from trl import SFTTrainer  # kept for future compatibility\n","\n","if reuse_model:\n","    # Reuse the CPU-loaded model (full precision)\n","    print(\"   Reusing existing model + processor\")\n","    # Prepare for kbit training (works on already-loaded model)\n","    model = prepare_model_for_kbit_training(model)\n","else:\n","    # Standard 4-bit load (most common case)\n","    print(\"   Loading model in 4-bit for efficient training...\")\n","    quant_config = BitsAndBytesConfig(\n","        load_in_4bit=True,\n","        bnb_4bit_quant_type=\"nf4\",\n","        bnb_4bit_compute_dtype=torch.bfloat16,\n","        bnb_4bit_use_double_quant=True,\n","    )\n","\n","    processor = AutoProcessor.from_pretrained(model_id, token=hf_token)\n","\n","    model = AutoModelForMultimodalLM.from_pretrained(\n","        model_id,\n","        token=hf_token,\n","        quantization_config=quant_config,\n","        device_map=\"auto\",\n","        low_cpu_mem_usage=True,\n","    )\n","\n","    model = prepare_model_for_kbit_training(model)\n","\n","# Apply LoRA (targets all linear layers – perfect for Gemma-4)\n","lora_config = LoraConfig(\n","    r=lora_rank,\n","    lora_alpha=lora_alpha,\n","    target_modules=\"all-linear\",\n","    lora_dropout=0.05,\n","    bias=\"none\",\n","    task_type=\"CAUSAL_LM\",\n",")\n","\n","model = get_peft_model(model, lora_config)\n","model.print_trainable_parameters()\n","\n","print(\"✅ Model + LoRA adapter ready (VRAM optimized)\")\n","\n","# ====================== DATASET PREPARATION ======================\n","print(\"\\n📦 Preparing dataset from ZIP...\")\n","\n","# Handle ZIP upload or existing path\n","if upload_new_zip or not zip_path or not os.path.exists(zip_path):\n","    print(\"📤 Upload your image-text pairs ZIP file (must contain matching .jpg/.png + .txt files)\")\n","    uploaded = files.upload()\n","    if uploaded:\n","        zip_filename = list(uploaded.keys())[0]\n","        zip_path = f\"/content/{zip_filename}\"\n","        print(f\"✅ Using uploaded file: {zip_filename}\")\n","    else:\n","        print(\"❌ No ZIP uploaded. Aborting.\")\n","        raise SystemExit\n","\n","if not os.path.exists(zip_path):\n","    print(f\"❌ ZIP file not found: {zip_path}\")\n","    raise SystemExit\n","\n","# Unpack ZIP\n","dataset_dir = \"/content/lora_training_data\"\n","os.makedirs(dataset_dir, exist_ok=True)\n","\n","with zipfile.ZipFile(zip_path, 'r') as z:\n","    z.extractall(dataset_dir)\n","\n","print(f\"✅ Extracted dataset to {dataset_dir}\")\n","\n","# Find all matching image + txt pairs\n","image_exts = ('.png', '.jpg', '.jpeg', '.webp', '.bmp')\n","pairs = []\n","\n","for txt_file in Path(dataset_dir).glob(\"**/*.txt\"):\n","    base_name = txt_file.stem\n","    image_path = None\n","    for ext in image_exts:\n","        candidate = txt_file.with_name(base_name + ext)\n","        if candidate.exists():\n","            image_path = str(candidate)\n","            break\n","    if not image_path:\n","        continue\n","\n","    with open(txt_file, \"r\", encoding=\"utf-8\") as f:\n","        text_content = f.read().strip()\n","\n","    # Split into Query and Reply using '//--//'\n","    if '//--//' in text_content:\n","        query_part, reply_part = [x.strip() for x in text_content.split('//--//', 1)]\n","    else:\n","        query_part = \"Describe this image in detail, including key objects, scene, colors, mood, lighting, and any visible text. Be accurate and comprehensive.\"\n","        reply_part = text_content\n","\n","    pairs.append({\n","        \"image_path\": image_path,\n","        \"query\": query_part,\n","        \"reply\": reply_part\n","    })\n","\n","print(f\"✅ Found {len(pairs)} valid image-text training pairs.\")\n","\n","if len(pairs) == 0:\n","    print(\"❌ No matching image+txt pairs found in the ZIP. Aborting.\")\n","    raise SystemExit\n","\n","# ====================== CUSTOM DATASET + COLLATOR ======================\n","class MultimodalTrainingDataset(torch.utils.data.Dataset):\n","    def __init__(self, pairs, processor):\n","        self.pairs = pairs\n","        self.processor = processor\n","\n","    def __len__(self):\n","        return len(self.pairs)\n","\n","    def __getitem__(self, idx):\n","        item = self.pairs[idx]\n","        messages = [\n","            {\"role\": \"user\", \"content\": [\n","                {\"type\": \"image\", \"url\": item[\"image_path\"]},\n","                {\"type\": \"text\", \"text\": item[\"query\"]}\n","            ]},\n","            {\"role\": \"assistant\", \"content\": [{\"type\": \"text\", \"text\": item[\"reply\"]}]}\n","        ]\n","\n","        inputs = self.processor.apply_chat_template(\n","            messages,\n","            add_generation_prompt=False,\n","            tokenize=True,\n","            return_tensors=\"pt\",\n","            return_dict=True\n","        )\n","        inputs[\"labels\"] = inputs[\"input_ids\"].clone()\n","\n","        # Remove batch dim\n","        for k in list(inputs.keys()):\n","            if isinstance(inputs[k], torch.Tensor):\n","                inputs[k] = inputs[k].squeeze(0)\n","        return inputs\n","\n","train_dataset = MultimodalTrainingDataset(pairs, processor)\n","\n","def multimodal_collator(features):\n","    batch = {}\n","    input_ids = [f[\"input_ids\"] for f in features]\n","    attention_mask = [f[\"attention_mask\"] for f in features]\n","    labels = [f[\"labels\"] for f in features]\n","\n","    text_padded = processor.tokenizer.pad(\n","        {\"input_ids\": input_ids, \"attention_mask\": attention_mask},\n","        padding=True,\n","        return_tensors=\"pt\"\n","    )\n","    batch[\"input_ids\"] = text_padded[\"input_ids\"]\n","    batch[\"attention_mask\"] = text_padded[\"attention_mask\"]\n","\n","    labels_padded = processor.tokenizer.pad(\n","        {\"input_ids\": labels},\n","        padding=True,\n","        return_tensors=\"pt\"\n","    )[\"input_ids\"]\n","    labels_padded[labels_padded == processor.tokenizer.pad_token_id] = -100\n","    batch[\"labels\"] = labels_padded\n","\n","    # Vision tensors (pixel_values etc.)\n","    first_item = features[0]\n","    for key in first_item.keys():\n","        if key in [\"input_ids\", \"attention_mask\", \"labels\"]:\n","            continue\n","        if isinstance(first_item[key], torch.Tensor):\n","            vals = [f[key] for f in features]\n","            batch[key] = torch.stack(vals) if vals[0].dim() == 3 else torch.cat(vals)\n","\n","    return batch\n","\n","# ====================== TRAINING ======================\n","print(f\"\\n🚀 Starting LoRA training on {len(pairs)} pairs for {num_epochs} epochs...\")\n","\n","training_args = TrainingArguments(\n","    output_dir=drive_adapter_path,   # ← Saved directly to Google Drive\n","    num_train_epochs=num_epochs,\n","    per_device_train_batch_size=batch_size,\n","    gradient_accumulation_steps=gradient_accumulation_steps,\n","    learning_rate=float(learning_rate),\n","    fp16=True,\n","    optim=\"paged_adamw_8bit\",\n","    logging_steps=5,\n","    save_steps=50,\n","    save_total_limit=2,\n","    report_to=\"none\",\n","    remove_unused_columns=False,\n",")\n","\n","trainer = Trainer(\n","    model=model,\n","    args=training_args,\n","    train_dataset=train_dataset,\n","    data_collator=multimodal_collator,\n",")\n","\n","trainer.train()\n","\n","# ====================== SAVE TO DRIVE + ZIP FOR DOWNLOAD ======================\n","print(\"\\n💾 Saving final LoRA adapter to Google Drive...\")\n","\n","model.save_pretrained(drive_adapter_path)\n","processor.save_pretrained(drive_adapter_path)\n","\n","# Create downloadable ZIP (from Drive folder)\n","zip_adapter_path = \"/content/gemma_lora_adapter.zip\"\n","with zipfile.ZipFile(zip_adapter_path, 'w') as z:\n","    for root, _, files_list in os.walk(drive_adapter_path):\n","        for file in files_list:\n","            file_path = os.path.join(root, file)\n","            arcname = os.path.relpath(file_path, drive_adapter_path)\n","            z.write(file_path, arcname)\n","\n","print(f\"🎉 LoRA training complete!\")\n","print(f\"   ✅ Saved to Google Drive: {drive_adapter_path}\")\n","print(f\"   📦 Zipped copy ready for download → {zip_adapter_path}\")\n","files.download(zip_adapter_path)\n","\n","print(\"\\n✅ You can now use this LoRA adapter from Google Drive in future inference cells.\")\n","print(\"   Example: peft_model = PeftModel.from_pretrained(base_model, drive_adapter_path)\")\n","```​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​"],"metadata":{"cellView":"form","id":"4cCNLXWsKHe_"},"execution_count":null,"outputs":[]}],"metadata":{"accelerator":"GPU","colab":{"gpuType":"T4","provenance":[{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/gemma4_heretic_vlm.ipynb","timestamp":1775331762109},{"file_id":"1IFAN1z44DC5xcoAdXNeupO-nUTedA2uG","timestamp":1775323816042},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775308348942},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775233835241},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775233735620},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1775159289229},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774964848181},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774906863352},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774901938471},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774884683620},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774883563998},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774882033159},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/dino_tagger.ipynb","timestamp":1774881268156},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Drive to WebP.ipynb","timestamp":1774879795776},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Drive to WebP.ipynb","timestamp":1763646205520},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Drive to WebP.ipynb","timestamp":1760993725927},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1760450712160},{"file_id":"https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1756712618300},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1747490904984},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1740037333374},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1736477078136},{"file_id":"https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb","timestamp":1725365086834}]},"kernelspec":{"display_name":"Python 3","name":"python3"},"language_info":{"name":"python"}},"nbformat":4,"nbformat_minor":0}