{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "inference",
      "provenance": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU",
    "gpuClass": "standard"
  },
  "cells": [
    {
      "cell_type": "code",
      "source": [
        "!pip install transformers==4.14.1\n",
        "!pip install bitsandbytes-cuda111==0.26.0\n",
        "\n",
        "from IPython import display \n",
        "display.clear_output()"
      ],
      "metadata": {
        "id": "Q8cuAdVDGXR6"
      },
      "execution_count": 1,
      "outputs": []
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "8mkqaWlNGLKn"
      },
      "outputs": [],
      "source": [
        "import transformers\n",
        "import torch\n",
        "import torch.nn.functional as F\n",
        "from torch import nn\n",
        "from torch.cuda.amp import custom_fwd, custom_bwd\n",
        "from bitsandbytes.functional import quantize_blockwise, dequantize_blockwise\n",
        "from tqdm.auto import tqdm"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#@title convert to 8bit\n",
        "class FrozenBNBLinear(nn.Module):\n",
        "    def __init__(self, weight, absmax, code, bias=None):\n",
        "        assert isinstance(bias, nn.Parameter) or bias is None\n",
        "        super().__init__()\n",
        "        self.out_features, self.in_features = weight.shape\n",
        "        self.register_buffer(\"weight\", weight.requires_grad_(False))\n",
        "        self.register_buffer(\"absmax\", absmax.requires_grad_(False))\n",
        "        self.register_buffer(\"code\", code.requires_grad_(False))\n",
        "        self.adapter = None\n",
        "        self.bias = bias\n",
        " \n",
        "    def forward(self, input):\n",
        "        output = DequantizeAndLinear.apply(input, self.weight, self.absmax, self.code, self.bias)\n",
        "        if self.adapter:\n",
        "            output += self.adapter(input)\n",
        "        return output\n",
        " \n",
        "    @classmethod\n",
        "    def from_linear(cls, linear: nn.Linear) -> \"FrozenBNBLinear\":\n",
        "        weights_int8, state = quantize_blockise_lowmemory(linear.weight)\n",
        "        return cls(weights_int8, *state, linear.bias)\n",
        " \n",
        "    def __repr__(self):\n",
        "        return f\"{self.__class__.__name__}({self.in_features}, {self.out_features})\"\n",
        " \n",
        " \n",
        "class DequantizeAndLinear(torch.autograd.Function): \n",
        "    @staticmethod\n",
        "    @custom_fwd\n",
        "    def forward(ctx, input: torch.Tensor, weights_quantized: torch.ByteTensor,\n",
        "                absmax: torch.FloatTensor, code: torch.FloatTensor, bias: torch.FloatTensor):\n",
        "        weights_deq = dequantize_blockwise(weights_quantized, absmax=absmax, code=code)\n",
        "        ctx.save_for_backward(input, weights_quantized, absmax, code)\n",
        "        ctx._has_bias = bias is not None\n",
        "        return F.linear(input, weights_deq, bias)\n",
        " \n",
        "    @staticmethod\n",
        "    @custom_bwd\n",
        "    def backward(ctx, grad_output: torch.Tensor):\n",
        "        assert not ctx.needs_input_grad[1] and not ctx.needs_input_grad[2] and not ctx.needs_input_grad[3]\n",
        "        input, weights_quantized, absmax, code = ctx.saved_tensors\n",
        "        # grad_output: [*batch, out_features]\n",
        "        weights_deq = dequantize_blockwise(weights_quantized, absmax=absmax, code=code)\n",
        "        grad_input = grad_output @ weights_deq\n",
        "        grad_bias = grad_output.flatten(0, -2).sum(dim=0) if ctx._has_bias else None\n",
        "        return grad_input, None, None, None, grad_bias\n",
        " \n",
        " \n",
        "class FrozenBNBEmbedding(nn.Module):\n",
        "    def __init__(self, weight, absmax, code):\n",
        "        super().__init__()\n",
        "        self.num_embeddings, self.embedding_dim = weight.shape\n",
        "        self.register_buffer(\"weight\", weight.requires_grad_(False))\n",
        "        self.register_buffer(\"absmax\", absmax.requires_grad_(False))\n",
        "        self.register_buffer(\"code\", code.requires_grad_(False))\n",
        "        self.adapter = None\n",
        " \n",
        "    def forward(self, input, **kwargs):\n",
        "        with torch.no_grad():\n",
        "            # note: both quantuized weights and input indices are *not* differentiable\n",
        "            weight_deq = dequantize_blockwise(self.weight, absmax=self.absmax, code=self.code)\n",
        "            output = F.embedding(input, weight_deq, **kwargs)\n",
        "        if self.adapter:\n",
        "            output += self.adapter(input)\n",
        "        return output \n",
        " \n",
        "    @classmethod\n",
        "    def from_embedding(cls, embedding: nn.Embedding) -> \"FrozenBNBEmbedding\":\n",
        "        weights_int8, state = quantize_blockise_lowmemory(embedding.weight)\n",
        "        return cls(weights_int8, *state)\n",
        " \n",
        "    def __repr__(self):\n",
        "        return f\"{self.__class__.__name__}({self.num_embeddings}, {self.embedding_dim})\"\n",
        " \n",
        " \n",
        "def quantize_blockise_lowmemory(matrix: torch.Tensor, chunk_size: int = 2 ** 20):\n",
        "    assert chunk_size % 4096 == 0\n",
        "    code = None\n",
        "    chunks = []\n",
        "    absmaxes = []\n",
        "    flat_tensor = matrix.view(-1)\n",
        "    for i in range((matrix.numel() - 1) // chunk_size + 1):\n",
        "        input_chunk = flat_tensor[i * chunk_size: (i + 1) * chunk_size].clone()\n",
        "        quantized_chunk, (absmax_chunk, code) = quantize_blockwise(input_chunk, code=code)\n",
        "        chunks.append(quantized_chunk)\n",
        "        absmaxes.append(absmax_chunk)\n",
        " \n",
        "    matrix_i8 = torch.cat(chunks).reshape_as(matrix)\n",
        "    absmax = torch.cat(absmaxes)\n",
        "    return matrix_i8, (absmax, code)\n",
        " \n",
        " \n",
        "def convert_to_int8(model):\n",
        "    \"\"\"Convert linear and embedding modules to 8-bit with optional adapters\"\"\"\n",
        "    for module in list(model.modules()):\n",
        "        for name, child in module.named_children():\n",
        "            if isinstance(child, nn.Linear):\n",
        "                setattr( \n",
        "                    module,\n",
        "                    name,\n",
        "                    FrozenBNBLinear(\n",
        "                        weight=torch.zeros(child.out_features, child.in_features, dtype=torch.uint8),\n",
        "                        absmax=torch.zeros((child.weight.numel() - 1) // 4096 + 1),\n",
        "                        code=torch.zeros(256),\n",
        "                        bias=child.bias,\n",
        "                    ),\n",
        "                )\n",
        "            elif isinstance(child, nn.Embedding):\n",
        "                setattr(\n",
        "                    module,\n",
        "                    name,\n",
        "                    FrozenBNBEmbedding(\n",
        "                        weight=torch.zeros(child.num_embeddings, child.embedding_dim, dtype=torch.uint8),\n",
        "                        absmax=torch.zeros((child.weight.numel() - 1) // 4096 + 1),\n",
        "                        code=torch.zeros(256),\n",
        "                    )\n",
        "                )\n",
        "class GPTJBlock(transformers.models.gptj.modeling_gptj.GPTJBlock):\n",
        "    def __init__(self, config):\n",
        "        super().__init__(config)\n",
        "\n",
        "        convert_to_int8(self.attn)\n",
        "        convert_to_int8(self.mlp)\n",
        "\n",
        "class GPTJModel(transformers.models.gptj.modeling_gptj.GPTJModel):\n",
        "    def __init__(self, config):\n",
        "        super().__init__(config)\n",
        "        convert_to_int8(self)\n",
        "        \n",
        "class GPTJForCausalLM(transformers.models.gptj.modeling_gptj.GPTJForCausalLM):\n",
        "    def __init__(self, config):\n",
        "        super().__init__(config)\n",
        "        convert_to_int8(self)\n",
        "\n",
        "\n",
        "transformers.models.gptj.modeling_gptj.GPTJBlock = GPTJBlock  # monkey-patch GPT-J"
      ],
      "metadata": {
        "cellView": "form",
        "id": "fmpdVvfVG7Pc"
      },
      "execution_count": 3,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "tokenizer = transformers.AutoTokenizer.from_pretrained(\"EleutherAI/gpt-j-6B\")\n",
        "gpt = GPTJForCausalLM.from_pretrained(\"crumb/gpt-j-6b-shakespeare\", low_cpu_mem_usage=True)\n",
        "\n",
        "device = 'cuda' if torch.cuda.is_available() else 'cpu'\n",
        "gpt = gpt.to(device)"
      ],
      "metadata": {
        "id": "ttKTRoUlG5YM"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "prompt = \"\"\"ROMEO: I would I were thy bird. \n",
        "JULIET: Sweet, so would I, Yet I should kill thee with much cherishing. Good night, good night! Parting is such sweet\"\"\"\n",
        "prompt = tokenizer(prompt, return_tensors='pt')\n",
        "prompt = {key: value.to(device) for key, value in prompt.items()}\n",
        "out = gpt.generate(**prompt, min_length=32, max_length=64, do_sample=True)\n",
        "out = tokenizer.decode(out[0])\n",
        "print(out)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "kSXSZz_kGcfm",
        "outputId": "d91dda66-88ab-4e52-bfb9-3df8092abe2f"
      },
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "ROMEO: I would I were thy bird. \n",
            "JULIET: Sweet, so would I, Yet I should kill thee with much cherishing. Good night, good night! Parting is such sweet sorrow, As a lost angel's song, in answer to An evil dream.\n",
            "\n",
            "ROMEO\n"
          ]
        }
      ]
    }
  ]
}