Upload HauhauCS_Abliterated_Qwen_3_5_4B_LLM_model.ipynb

HauhauCS_Abliterated_Qwen_3_5_4B_LLM_model.ipynb

@@ -0,0 +1 @@
+{"cells":[{"cell_type":"markdown","metadata":{"id":"gZTGUb5FShVL"},"source":["# Qwen3.5-4B HauhauCS — Google Colab\n","\n","This notebook loads the DreamFast Qwen3.5-4B multimodal model in 4-bit NF4 and provides a Gradio chat interface for text and image conversations. Because the fine-tuned repository does not include its own chat template, the notebook downloads `chat_template.jinja` directly from the official `Qwen/Qwen3.5-4B` repository and assigns it to both the tokenizer and processor.\n","\n","**Important setup:** Run the dependency-install cell once, then restart the Colab runtime before running the remaining cells. In Colab, enable a GPU under **Runtime → Change runtime type → GPU** and add a Hugging Face token named `HF_TOKEN` under **Secrets**.\n"],"id":"gZTGUb5FShVL"},{"cell_type":"code","execution_count":null,"metadata":{"id":"nAUyO568ShVN"},"outputs":[],"source":["# CELL 1 — Install compatible dependencies\n","# IMPORTANT: Run this cell once. It will automatically restart the runtime if package updates are detected.\n","# Do not force-reinstall torch/torchvision: Colab supplies a matched CUDA build.\n","\n","import sys\n","import os\n","import time\n","import pkg_resources\n","\n","# Record the currently installed packages and versions\n","pre_installed = {pkg.key: pkg.version for pkg in pkg_resources.working_set}\n","\n","# Run the installation command\n","get_ipython().run_line_magic('pip', 'install -q --upgrade --no-cache-dir \"Pillow>=10.4.0\" \"transformers>=5.0.0\" \"accelerate>=1.2.0\" \"bitsandbytes>=0.45.0\" \"gradio>=5.0.0\" \"huggingface_hub>=0.27.0\"')\n","\n","# Reload working set to detect newly installed or modified packages\n","import importlib\n","importlib.reload(pkg_resources)\n","post_installed = {pkg.key: pkg.version for pkg in pkg_resources.working_set}\n","\n","# Check if anything changed\n","changes_detected = False\n","for pkg, version in post_installed.items():\n","    if pkg not in pre_installed or pre_installed[pkg] != version:\n","        changes_detected = True\n","        break\n","\n","if changes_detected:\n","    print(\"\\n[INFO] New dependencies installed or updated.\")\n","    print(\"Preparing to restart the Google Colab session to apply changes...\")\n","    print(\"NOTE: Google Colab might show a pop-up saying the runtime 'Crashed for an unknown reason'.\")\n","    print(\"This is completely normal when programmatically restarting. The session is simply refreshing.\")\n","\n","    for i in range(5, 0, -1):\n","        print(f\"Restarting in {i} seconds...\")\n","        time.sleep(1)\n","\n","    print(\"Restarting session now...\")\n","    # Programmatically restart the Colab runtime session\n","    os.kill(os.getpid(), 9)\n","else:\n","    print(\"\\nNo package updates detected. Runtime restart is not required.\")"],"id":"nAUyO568ShVN"},{"cell_type":"code","execution_count":null,"metadata":{"id":"c9VvMbzuShVO"},"outputs":[],"source":["# ============================================================\n","# QWEN3.5-4B HAUHAUCS — GOOGLE COLAB\n","# Text + image chat, 4-bit NF4, Gradio interface\n","# Recommended GPU: T4 (16 GB) or better\n","# ============================================================\n","\n","# CELL 2 — Imports and environment checks\n","import os\n","import re\n","import gc\n","import torch\n","import gradio as gr\n","from PIL import Image\n","from google.colab import userdata\n","from huggingface_hub import hf_hub_download\n","from transformers import (\n","    AutoProcessor,\n","    AutoTokenizer,\n","    AutoModelForMultimodalLM,\n","    BitsAndBytesConfig,\n",")\n","\n","MODEL_ID = \"DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark\"\n","BASE_MODEL_ID = \"Qwen/Qwen3.5-4B\"\n","\n","# Read the Hugging Face token from Colab Secrets.\n","# In Colab: left sidebar > Secrets (key icon) > add HF_TOKEN.\n","HF_TOKEN = userdata.get(\"HF_TOKEN\")\n","if not HF_TOKEN:\n","    raise ValueError(\n","        \"HF_TOKEN was not found in Colab Secrets. Add a secret named HF_TOKEN \"\n","        \"containing your Hugging Face access token, then rerun this cell.\"\n","    )\n","\n","if not torch.cuda.is_available():\n","    raise RuntimeError(\n","        \"No CUDA GPU is available. In Colab select Runtime > Change runtime type > GPU.\"\n","    )\n","\n","print(\"PyTorch:\", torch.__version__)\n","print(\"CUDA:\", torch.version.cuda)\n","print(\"Pillow:\", Image.__version__)\n","print(\"GPU:\", torch.cuda.get_device_name(0))\n","print(\"VRAM:\", round(torch.cuda.get_device_properties(0).total_memory / 1024**3, 2), \"GB\")\n","\n","# Helpful diagnostic: detect mismatched/broken Pillow installations early.\n","try:\n","    from PIL import ImageDraw\n","    from torchvision.transforms import InterpolationMode\n","    print(\"Pillow and torchvision imports: OK\")\n","except Exception as exc:\n","    raise RuntimeError(\n","        \"A vision dependency failed to import. Run CELL 1, restart the Colab \"\n","        \"runtime, and then run this notebook from CELL 2. Original error: \"\n","        + repr(exc)\n","    ) from exc\n","\n","torch.backends.cuda.matmul.allow_tf32 = True\n","\n","\n","# ============================================================\n","# CELL 3 — Load processor, tokenizer and 4-bit model\n","# ============================================================\n","\n","print(\"\\\\nLoading tokenizer...\")\n","tokenizer = AutoTokenizer.from_pretrained(\n","    BASE_MODEL_ID,\n","    token=HF_TOKEN,\n","    trust_remote_code=True,\n",")\n","\n","# The fine-tuned DreamFast repository may not contain a chat template.\n","# Always fetch the official Qwen3.5-4B template directly from the base repo.\n","print(\"Downloading the official Qwen3.5-4B chat template...\")\n","template_path = hf_hub_download(\n","    repo_id=BASE_MODEL_ID,\n","    filename=\"chat_template.jinja\",\n","    token=HF_TOKEN,\n",")\n","\n","with open(template_path, \"r\", encoding=\"utf-8\") as template_file:\n","    qwen_chat_template = template_file.read()\n","\n","if not qwen_chat_template.strip():\n","    raise RuntimeError(\"The chat_template.jinja downloaded from the base repo is empty.\")\n","\n","# Explicitly override any missing or inherited template on the tokenizer.\n","tokenizer.chat_template = qwen_chat_template\n","print(\"Official Qwen3.5-4B chat template loaded and assigned to tokenizer.\")\n","\n","print(\"Loading multimodal processor...\")\n","processor = AutoProcessor.from_pretrained(\n","    MODEL_ID,\n","    token=HF_TOKEN,\n","    trust_remote_code=True,\n",")\n","\n","# Ensure both tokenizer and processor use the exact same official template.\n","processor.tokenizer = tokenizer\n","processor.chat_template = qwen_chat_template\n","print(\"Official Qwen3.5-4B chat template assigned to processor.\")\n","\n","print(\"Configuring 4-bit NF4 quantization...\")\n","quantization_config = BitsAndBytesConfig(\n","    load_in_4bit=True,\n","    bnb_4bit_quant_type=\"nf4\",\n","    bnb_4bit_compute_dtype=torch.float16,\n","    bnb_4bit_use_double_quant=True,\n",")\n","\n","print(\"Loading model. This may take several minutes...\")\n","model = AutoModelForMultimodalLM.from_pretrained(\n","    MODEL_ID,\n","    token=HF_TOKEN,\n","    quantization_config=quantization_config,\n","    torch_dtype=torch.float16,\n","    device_map=\"auto\",\n","    low_cpu_mem_usage=True,\n","    attn_implementation=\"sdpa\",\n","    trust_remote_code=True,\n",")\n","model.eval()\n","\n","INPUT_DEVICE = model.get_input_embeddings().weight.device\n","print(\"Model loaded.\")\n","print(\"Input device:\", INPUT_DEVICE)\n","\n","\n","# ============================================================\n","# CELL 4 — Input preparation and response parsing\n","# ============================================================\n","\n","def prepare_inputs(messages):\n","    \"\"\"\n","    Apply the official Qwen3.5-4B chat template downloaded from the base repo,\n","    then let the multimodal processor tokenize text and preprocess images.\n","    \"\"\"\n","    # Keep the template assignment explicit in case another component changed it.\n","    processor.chat_template = qwen_chat_template\n","    processor.tokenizer.chat_template = qwen_chat_template\n","\n","    batch = processor.apply_chat_template(\n","        messages,\n","        chat_template=qwen_chat_template,\n","        tokenize=True,\n","        add_generation_prompt=True,\n","        return_dict=True,\n","        return_tensors=\"pt\",\n","    )\n","    return batch\n","\n","\n","def split_thinking_and_answer(raw_text):\n","    raw_text = re.sub(\n","        r\"<\\\\|im_end\\\\|>|<\\\\|endoftext\\\\|>|</s>\",\n","        \"\",\n","        raw_text,\n","    ).strip()\n","\n","    match = re.search(\n","        r\"<think>(.*?)</think>\",\n","        raw_text,\n","        flags=re.DOTALL | re.IGNORECASE,\n","    )\n","\n","    if match:\n","        thinking = match.group(1).strip()\n","        answer = (raw_text[:match.start()] + raw_text[match.end():]).strip()\n","    else:\n","        # Do not label an entire response as hidden reasoning if no think\n","        # section was emitted.\n","        thinking = \"No separate <think> section was returned by the model.\"\n","        answer = raw_text\n","\n","    answer = re.sub(r\"<\\\\|im_start\\\\|>|<\\\\|im_sep\\\\|>|<\\\\|im_end\\\\|>\", \"\", answer).strip()\n","    thinking = re.sub(r\"<\\\\|im_start\\\\|>|<\\\\|im_sep\\\\|>|<\\\\|im_end\\\\|>\", \"\", thinking).strip()\n","\n","    if not answer:\n","        answer = \"(No separate final answer was returned.)\"\n","    if not thinking:\n","        thinking = \"(Empty thinking section.)\"\n","\n","    return thinking, answer\n","\n","\n","# ============================================================\n","# CELL 5 — Generation\n","# ============================================================\n","\n","def generate_response(\n","    messages,\n","    max_new_tokens=1024,\n","    temperature=0.7,\n","    top_p=0.9,\n","):\n","    inputs = prepare_inputs(messages)\n","\n","    # The model is loaded with device_map=\"auto\". Move processor tensors to\n","    # the input embedding device; this is normally the GPU on a T4.\n","    inputs = {\n","        key: value.to(INPUT_DEVICE) if hasattr(value, \"to\") else value\n","        for key, value in inputs.items()\n","    }\n","\n","    input_ids = inputs[\"input_ids\"]\n","    input_length = input_ids.shape[-1]\n","\n","    do_sample = float(temperature) > 0\n","    generation_kwargs = {\n","        \"max_new_tokens\": int(max_new_tokens),\n","        \"do_sample\": do_sample,\n","        \"repetition_penalty\": 1.05,\n","        \"use_cache\": True,\n","    }\n","\n","    if tokenizer.eos_token_id is not None:\n","        generation_kwargs[\"eos_token_id\"] = tokenizer.eos_token_id\n","    if tokenizer.pad_token_id is not None:\n","        generation_kwargs[\"pad_token_id\"] = tokenizer.pad_token_id\n","    elif tokenizer.eos_token_id is not None:\n","        generation_kwargs[\"pad_token_id\"] = tokenizer.eos_token_id\n","\n","    if do_sample:\n","        generation_kwargs[\"temperature\"] = max(float(temperature), 1e-5)\n","        generation_kwargs[\"top_p\"] = float(top_p)\n","\n","    with torch.inference_mode():\n","        output_ids = model.generate(**inputs, **generation_kwargs)\n","\n","    new_tokens = output_ids[0, input_length:]\n","    raw_response = tokenizer.decode(\n","        new_tokens,\n","        skip_special_tokens=False,\n","        clean_up_tokenization_spaces=False,\n","    ).strip()\n","\n","    thinking, answer = split_thinking_and_answer(raw_response)\n","    return raw_response, thinking, answer\n","\n","\n","# ============================================================\n","# CELL 6 — Gradio chat callbacks\n","# ============================================================\n","\n","def chat(\n","    message,\n","    image,\n","    chat_history,\n","    conversation_history,\n","    max_new_tokens,\n","    temperature,\n","    top_p,\n","):\n","    chat_history = list(chat_history or [])\n","    conversation_history = list(conversation_history or [])\n","    message = (message or \"\").strip()\n","\n","    if not message and image is None:\n","        return (\n","            chat_history,\n","            conversation_history,\n","            message,\n","            image,\n","            \"\",\n","            \"Please enter a message or upload an image.\",\n","        )\n","\n","    user_content = []\n","    if message:\n","        user_content.append({\"type\": \"text\", \"text\": message})\n","\n","    if image is not None:\n","        if not isinstance(image, Image.Image):\n","            image = Image.fromarray(image)\n","        user_content.append({\"type\": \"image\", \"image\": image.convert(\"RGB\")})\n","\n","    conversation_history.append({\"role\": \"user\", \"content\": user_content})\n","\n","    display_text = message or \"Please describe this image.\"\n","    if image is not None:\n","        display_text += \"\\\\n\\\\n📷 Image attached\"\n","    chat_history.append({\"role\": \"user\", \"content\": display_text})\n","\n","    try:\n","        raw_response, thinking, answer = generate_response(\n","            conversation_history,\n","            max_new_tokens=max_new_tokens,\n","            temperature=temperature,\n","            top_p=top_p,\n","        )\n","\n","        conversation_history.append({\n","            \"role\": \"assistant\",\n","            \"content\": raw_response,\n","        })\n","\n","        chat_history.append({\n","            \"role\": \"assistant\",\n","            \"content\": \"### 💭 WHAT THE MODEL THINKS\\\\n\\\\n\" + thinking,\n","        })\n","        chat_history.append({\n","            \"role\": \"assistant\",\n","            \"content\": \"### 🗣️ WHAT THE MODEL SAYS\\\\n\\\\n\" + answer,\n","        })\n","\n","        return chat_history, conversation_history, \"\", None, thinking, answer\n","\n","    except Exception as exc:\n","        error_message = f\"Generation error:\\\\n\\\\n{type(exc).__name__}: {exc}\"\n","        chat_history.append({\"role\": \"assistant\", \"content\": error_message})\n","        # Remove the failed user turn from model context so retrying does not\n","        # leave a malformed/incomplete conversation.\n","        if conversation_history and conversation_history[-1].get(\"role\") == \"user\":\n","            conversation_history.pop()\n","        return chat_history, conversation_history, message, image, \"\", error_message\n","\n","\n","def clear_chat():\n","    return [], [], \"\", None, \"\", \"\"\n","\n","\n","# ============================================================\n","# CELL 7 — Gradio interface\n","# ============================================================\n","\n","CUSTOM_CSS = \"\"\"\n",".gradio-container { max-width: 1500px !important; margin: auto !important; }\n","#main-title { text-align: center; margin-bottom: 5px; }\n","#chat-window { min-height: 600px !important; }\n","#chat-window .message { font-size: 16px !important; line-height: 1.6 !important; }\n","#send-button { font-weight: bold !important; }\n","\"\"\"\n","\n","with gr.Blocks(\n","    theme=gr.themes.Base(\n","        primary_hue=\"violet\",\n","        secondary_hue=\"orange\",\n","        neutral_hue=\"slate\",\n","    ),\n","    css=CUSTOM_CSS,\n","    title=\"Qwen3.5-4B HauhauCS Chat\",\n",") as demo:\n","\n","    gr.Markdown(\n","        \"\"\"\n","        # Qwen3.5-4B HauhauCS Chat\n","        **DreamFast / HauhauCS · 4-bit NF4 · NVIDIA GPU**\n","\n","        Supports text and image input. The model's generated response is\n","        displayed in the conversation and split into separate panels when\n","        it returns a `<think>` section.\n","        \"\"\",\n","        elem_id=\"main-title\",\n","    )\n","\n","    conversation_state = gr.State([])\n","\n","    with gr.Row():\n","        with gr.Column(scale=7):\n","            chatbot = gr.Chatbot(\n","                label=\"Conversation\",\n","                height=620,\n","                autoscroll=True,\n","                elem_id=\"chat-window\",\n","            )\n","\n","            with gr.Row():\n","                message_input = gr.Textbox(\n","                    label=\"Your message\",\n","                    placeholder=\"Write your message here...\",\n","                    lines=4,\n","                    max_lines=10,\n","                    scale=5,\n","                )\n","                image_input = gr.Image(\n","                    label=\"Optional image\",\n","                    height=150,\n","                    sources=[\"upload\", \"clipboard\"],\n","                    scale=2,\n","                )\n","\n","            with gr.Row():\n","                send_button = gr.Button(\"Send\", variant=\"primary\", scale=3, elem_id=\"send-button\")\n","                clear_button = gr.Button(\"Clear conversation\", variant=\"secondary\", scale=1)\n","\n","        with gr.Column(scale=3):\n","            gr.Markdown(\"## Generation settings\")\n","            max_new_tokens = gr.Slider(\n","                minimum=128, maximum=4096, value=1024, step=128,\n","                label=\"Maximum new tokens\",\n","            )\n","            temperature = gr.Slider(\n","                minimum=0.0, maximum=1.5, value=0.7, step=0.05,\n","                label=\"Temperature (0 = deterministic)\",\n","            )\n","            top_p = gr.Slider(\n","                minimum=0.1, maximum=1.0, value=0.9, step=0.05,\n","                label=\"Top-p\",\n","            )\n","\n","    gr.Markdown(\"---\")\n","    gr.Markdown(\"## Latest model response\")\n","\n","    with gr.Row():\n","        with gr.Column():\n","            thinking_output = gr.Textbox(\n","                label=\"💭 What the model thinks\",\n","                lines=12,\n","                max_lines=25,\n","                interactive=False,\n","                placeholder=\"A separate think section, if returned, appears here.\",\n","            )\n","        with gr.Column():\n","            answer_output = gr.Textbox(\n","                label=\"🗣️ What the model says\",\n","                lines=12,\n","                max_lines=25,\n","                interactive=False,\n","                placeholder=\"The model's answer appears here.\",\n","            )\n","\n","    callback_inputs = [\n","        message_input,\n","        image_input,\n","        chatbot,\n","        conversation_state,\n","        max_new_tokens,\n","        temperature,\n","        top_p,\n","    ]\n","    callback_outputs = [\n","        chatbot,\n","        conversation_state,\n","        message_input,\n","        image_input,\n","        thinking_output,\n","        answer_output,\n","    ]\n","\n","    send_button.click(chat, inputs=callback_inputs, outputs=callback_outputs)\n","    message_input.submit(chat, inputs=callback_inputs, outputs=callback_outputs)\n","    clear_button.click(clear_chat, inputs=[], outputs=callback_outputs)\n","\n","\n","# ============================================================\n","# CELL 8 — Launch\n","# ============================================================\n","\n","print(\"Starting Gradio. The public share link will appear below.\")\n","demo.queue().launch(\n","    share=True,\n","    debug=True,\n","    show_error=True,\n",")\n"],"id":"c9VvMbzuShVO"}],"metadata":{"colab":{"provenance":[],"gpuType":"T4"},"kernelspec":{"display_name":"Python 3","name":"python3"},"language_info":{"name":"python"},"accelerator":"GPU"},"nbformat":4,"nbformat_minor":5}