Upload Alpaca_+_Mistral_7b_full_example.ipynb

Alpaca_+_Mistral_7b_full_example.ipynb

@@ -5,9 +5,9 @@
       "source": [
         "To run this, press \"Runtime\" and press \"Run all\" on a **free** Tesla T4 Google Colab instance!\n",
         "<div class=\"align-center\">\n",
-        "  <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"110\"></a>\n",
-        "  <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"150\"></a>\n",
-        "  <a href=\"https://huggingface.co/docs/trl/main/en/index\"><img src=\"https://github.com/huggingface/blog/blob/main/assets/133_trl_peft/thumbnail.png?raw=true\" width=\"100\"></a> Join our Discord if you need help!\n",
+        "  <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" height=\"45\"></a>\n",
+        "  <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" height=\"45\"></a>\n",
+        "  <a href=\"https://ko-fi.com/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Kofi button.png\" height=\"45\"></a></a> Join our Discord if you need help!\n",
         "</div>\n",
         "\n",
         "To install Unsloth on your own computer, follow the installation instructions on our Github page [here](https://github.com/unslothai/unsloth#installation-instructions---conda).\n",
@@ -272,6 +272,15 @@
         "dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+\n",
         "load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.\n",
         "\n",
+        "# 4bit pre quantized models we support for 4x faster downloading!\n",
+        "fourbit_models = [\n",
+        "    \"unsloth/mistral-7b-bnb-4bit\",\n",
+        "    \"unsloth/llama-2-7b-bnb-4bit\",\n",
+        "    \"unsloth/llama-2-13b-bnb-4bit\",\n",
+        "    \"unsloth/codellama-34b-bnb-4bit\",\n",
+        "    \"unsloth/tinyllama-bnb-4bit\",\n",
+        "]\n",
+        "\n",
         "model, tokenizer = FastLanguageModel.from_pretrained(\n",
         "    model_name = \"unsloth/mistral-7b-bnb-4bit\", # \"unsloth/mistral-7b\" for 16bit loading\n",
         "    max_seq_length = max_seq_length,\n",
@@ -316,8 +325,8 @@
         "    target_modules = [\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n",
         "                      \"gate_proj\", \"up_proj\", \"down_proj\",],\n",
         "    lora_alpha = 16,\n",
-        "    lora_dropout = 0, # Currently only supports dropout = 0\n",
-        "    bias = \"none\",    # Currently only supports bias = \"none\"\n",
+        "    lora_dropout = 0, # Supports any, but = 0 is optimized\n",
+        "    bias = \"none\",    # Supports any, but = \"none\" is optimized\n",
         "    use_gradient_checkpointing = True,\n",
         "    random_state = 3407,\n",
         "    max_seq_length = max_seq_length,\n",
@@ -999,39 +1008,6 @@
       "execution_count": null,
       "outputs": []
     },
-    {
-      "cell_type": "markdown",
-      "source": [
-        "To save to `GGUF` / `llama.cpp`, or for model merging, use `model.merge_and_unload` first, then save the model. See this [issue](https://github.com/ggerganov/llama.cpp/issues/3097) on llama.cpp for more info."
-      ],
-      "metadata": {
-        "id": "TCv4vXHd61i7"
-      }
-    },
-    {
-      "cell_type": "code",
-      "source": [
-        "model = model.merge_and_unload()"
-      ],
-      "metadata": {
-        "id": "xcRjsZe0RK1b",
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "outputId": "b40d4730-b9dc-4a4b-e42c-c191060ff66c"
-      },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stderr",
-          "text": [
-            "/usr/local/lib/python3.10/dist-packages/peft/tuners/lora/bnb.py:229: UserWarning: Merge lora module to 4-bit linear may get different generations due to rounding errors.\n",
-            "  warnings.warn(\n"
-          ]
-        }
-      ]
-    },
     {
       "cell_type": "markdown",
       "source": [
@@ -1098,14 +1074,164 @@
         }
       ]
     },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "To save to `GGUF` / `llama.cpp`, we support it natively now! You can also go to our dedicated GGUF notebook [here](https://colab.research.google.com/drive/14DW0VwuqL2O3tqGlX7aUF6TOBA8S59M4?usp=sharing). Select either `save locally` for local saving or `save locally and quantize to 4bit` for 4bit quantization for llama.cpp / GGUF."
+      ],
+      "metadata": {
+        "id": "TCv4vXHd61i7"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title Code for conversion to GGUF\n",
+        "def colab_quantize_to_gguf(save_directory, quantization_method = \"q4_k_m\"):\n",
+        "    from transformers.models.llama.modeling_llama import logger\n",
+        "    import os\n",
+        "\n",
+        "    logger.warning_once(\n",
+        "        \"Unsloth: `colab_quantize_to_gguf` is still in development mode.\\n\"\\\n",
+        "        \"If anything errors or breaks, please file a ticket on Github.\\n\"\\\n",
+        "        \"Also, if you used this successfully, please tell us on Discord!\"\n",
+        "    )\n",
+        "\n",
+        "    # From https://mlabonne.github.io/blog/posts/Quantize_Llama_2_models_using_ggml.html\n",
+        "    ALLOWED_QUANTS = \\\n",
+        "    {\n",
+        "        \"q2_k\"   : \"Uses Q4_K for the attention.vw and feed_forward.w2 tensors, Q2_K for the other tensors.\",\n",
+        "        \"q3_k_l\" : \"Uses Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K\",\n",
+        "        \"q3_k_m\" : \"Uses Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K\",\n",
+        "        \"q3_k_s\" : \"Uses Q3_K for all tensors\",\n",
+        "        \"q4_0\"   : \"Original quant method, 4-bit.\",\n",
+        "        \"q4_1\"   : \"Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models.\",\n",
+        "        \"q4_k_m\" : \"Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K\",\n",
+        "        \"q4_k_s\" : \"Uses Q4_K for all tensors\",\n",
+        "        \"q5_0\"   : \"Higher accuracy, higher resource usage and slower inference.\",\n",
+        "        \"q5_1\"   : \"Even higher accuracy, resource usage and slower inference.\",\n",
+        "        \"q5_k_m\" : \"Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K\",\n",
+        "        \"q5_k_s\" : \"Uses Q5_K for all tensors\",\n",
+        "        \"q6_k\"   : \"Uses Q8_K for all tensors\",\n",
+        "        \"q8_0\"   : \"Almost indistinguishable from float16. High resource use and slow. Not recommended for most users.\",\n",
+        "    }\n",
+        "\n",
+        "    if quantization_method not in ALLOWED_QUANTS.keys():\n",
+        "        error = f\"Unsloth: Quant method = [{quantization_method}] not supported. Choose from below:\\n\"\n",
+        "        for key, value in ALLOWED_QUANTS.items():\n",
+        "            error += f\"[{key}] => {value}\\n\"\n",
+        "        raise RuntimeError(error)\n",
+        "    pass\n",
+        "\n",
+        "    print_info = \\\n",
+        "        f\"==((====))==  Unsloth: Conversion from QLoRA to GGUF information\\n\"\\\n",
+        "        f\"   \\\\\\   /|    [0] Installing llama.cpp will take 3 minutes.\\n\"\\\n",
+        "        f\"O^O/ \\_/ \\\\    [1] Converting HF to GUUF 16bits will take 3 minutes.\\n\"\\\n",
+        "        f\"\\        /    [2] Converting GGUF 16bits to q4_k_m will take 20 minutes.\\n\"\\\n",
+        "        f' \"-____-\"     In total, you will have to wait around 26 minutes.\\n'\n",
+        "    print(print_info)\n",
+        "\n",
+        "    if not os.path.exists(\"llama.cpp\"):\n",
+        "        print(\"Unsloth: [0] Installing llama.cpp. This will take 3 minutes...\")\n",
+        "        !git clone https://github.com/ggerganov/llama.cpp\n",
+        "        !cd llama.cpp && make clean && LLAMA_CUBLAS=1 make -j\n",
+        "        !pip install gguf protobuf\n",
+        "    pass\n",
+        "\n",
+        "    print(\"Unsloth: [1] Converting HF into GGUF 16bit. This will take 3 minutes...\")\n",
+        "    !python llama.cpp/convert.py {save_directory} \\\n",
+        "        --outfile {save_directory}-unsloth.gguf \\\n",
+        "        --outtype f16\n",
+        "\n",
+        "    print(\"Unsloth: [2] Converting GGUF 16bit into q4_k_m. This will take 20 minutes...\")\n",
+        "    final_location = f\"./{save_directory}-{quantization_method}-unsloth.gguf\"\n",
+        "    !./llama.cpp/quantize ./{save_directory}-unsloth.gguf \\\n",
+        "        {final_location} {quantization_method}\n",
+        "\n",
+        "    print(f\"Unsloth: Output location: {final_location}\")\n",
+        "pass\n"
+      ],
+      "metadata": {
+        "cellView": "form",
+        "id": "nCVtR2ElF1GX"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "from unsloth import unsloth_save_model\n",
+        "\n",
+        "# Change to `save locally` to save a float16 GGUF file or `\"save locally and quantize to 4bit\"`\n",
+        "# to quantize down to 4bit\n",
+        "SAVE_STRATEGY = \"none\"\n",
+        "\n",
+        "if SAVE_STRATEGY == \"save locally\":\n",
+        "\n",
+        "    unsloth_save_model(model, tokenizer, \"output_model\")\n",
+        "\n",
+        "elif SAVE_STRATEGY == \"save locally and quantize to 4bit\":\n",
+        "\n",
+        "    unsloth_save_model(model, tokenizer, \"output_model\")\n",
+        "    colab_quantize_to_gguf(\"output_model\", quantization_method = \"q4_k_m\")"
+      ],
+      "metadata": {
+        "id": "FqfebeAdT073"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "Otherwise, to merge the LoRA adapters into the 4bit model:"
+      ],
+      "metadata": {
+        "id": "acUVCgzzU1Wv"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "model = model.merge_and_unload()"
+      ],
+      "metadata": {
+        "id": "xcRjsZe0RK1b",
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "outputId": "b40d4730-b9dc-4a4b-e42c-c191060ff66c"
+      },
+      "execution_count": null,
+      "outputs": [
+        {
+          "output_type": "stream",
+          "name": "stderr",
+          "text": [
+            "/usr/local/lib/python3.10/dist-packages/peft/tuners/lora/bnb.py:229: UserWarning: Merge lora module to 4-bit linear may get different generations due to rounding errors.\n",
+            "  warnings.warn(\n"
+          ]
+        }
+      ]
+    },
     {
       "cell_type": "markdown",
       "source": [
         "And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/u54VK8m8tk) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
+        "\n",
+        "Some other links:\n",
+        "1. Zephyr DPO 2x faster [free Colab](https://colab.research.google.com/drive/15vttTpzzVXv_tJwEk-hIcQ0S9FcEWvwP?usp=sharing)\n",
+        "2. Llama 7b 2x faster [free Colab](https://colab.research.google.com/drive/1lBzz5KeZJKXjvivbYvmGarix9Ao6Wxe5?usp=sharing)\n",
+        "3. TinyLlama 4x faster full Alpaca 52K in 1 hour [free Colab](https://colab.research.google.com/drive/1AZghoNBQaMDgWJpi4RbffGM1h6raLUj9?usp=sharing)\n",
+        "4. CodeLlama 34b 2x faster [A100 on Colab](https://colab.research.google.com/drive/1y7A0AxE3y8gdj4AVkl2aZX47Xu3P1wJT?usp=sharing)\n",
+        "5. Llama 7b [free Kaggle](https://www.kaggle.com/danielhanchen/unsloth-alpaca-t4-ddp)\n",
+        "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
+        "\n",
         "<div class=\"align-center\">\n",
-        "  <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"110\"></a>\n",
-        "  <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"150\"></a>\n",
-        "  <a href=\"https://huggingface.co/docs/trl/main/en/index\"><img src=\"https://github.com/huggingface/blog/blob/main/assets/133_trl_peft/thumbnail.png?raw=true\" width=\"100\"></a>\n",
+        "  <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" height=\"45\"></a>\n",
+        "  <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" height=\"45\"></a>\n",
+        "  <a href=\"https://ko-fi.com/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Kofi button.png\" height=\"45\"></a></a> Support our work if you can!!\n",
         "</div>"
       ],
       "metadata": {