Create check_perplexity.ipynb

check_perplexity.ipynb

@@ -0,0 +1,691 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Original GPT-J perlexity"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import torch.nn as nn\n",
+    "import torch.nn.functional as F\n",
+    "\n",
+    "from bitsandbytes.functional import quantize_blockwise, dequantize_blockwise\n",
+    "import transformers\n",
+    "from tqdm.auto import tqdm\n",
+    "\n",
+    "\n",
+    "\n",
+    "model_name = \"EleutherAI/gpt-j-6B\"\n",
+    "gpt = transformers.AutoModelForCausalLM.from_pretrained(model_name)\n",
+    "tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = 'cuda' if torch.cuda.is_available else 'cpu'\n",
+    "gpt.to(device).train(False);"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Reusing dataset wikitext (/home/jheuristic/.cache/huggingface/datasets/wikitext/wikitext-2-v1/1.0.0/a241db52902eaf2c6aa732210bead40c090019a499ceb13bcbfa3f8ab646a126)\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "47f0459174da4ee2bf064c9ae81fdecd",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "  0%|          | 0/3 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from datasets import load_dataset\n",
+    "data = load_dataset('wikitext', 'wikitext-2-v1')['test']"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 62,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "26cca02205624aafa740e55542ca2e6c",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "  0%|          | 0/4358 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "\n",
+    "numerator, denominator = 0, 0\n",
+    "collator = transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False)\n",
+    "loader = torch.utils.data.DataLoader(data, batch_size=1, num_workers=0, shuffle=False)\n",
+    "\n",
+    "\n",
+    "with torch.no_grad(), torch.cuda.amp.autocast(), tqdm(loader) as progressbar:\n",
+    "    for i, row in enumerate(progressbar):\n",
+    "        if max(map(len, row['text'])) <= 1:\n",
+    "            continue\n",
+    "        batch = tokenizer(**row, truncation=False, return_tensors='pt')\n",
+    "        batch = {k: v.cuda() for k, v in batch.items()}\n",
+    "\n",
+    "        out = gpt.forward(**batch,)\n",
+    "\n",
+    "        loss = F.cross_entropy(out.logits[:, :-1, :].flatten(0, -2), batch['input_ids'][:, 1:].flatten(),\n",
+    "                               reduction='none')\n",
+    "\n",
+    "        numerator += loss.sum().item()\n",
+    "        denominator += len(loss)\n",
+    "        progressbar.desc = f\"{numerator/denominator:.3f}\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 63,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "18.435175441788164"
+      ]
+     },
+     "execution_count": 63,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# test perplexity\n",
+    "import math\n",
+    "math.exp(numerator/denominator)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Quantized GPT-J Perplexity"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 64,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "\n",
+    "import torch\n",
+    "import torch.nn as nn\n",
+    "from torch.cuda.amp import custom_fwd, custom_bwd\n",
+    " \n",
+    "from bitsandbytes.functional import quantize_blockwise, dequantize_blockwise\n",
+    "import transformers\n",
+    "\n",
+    "\n",
+    "class DequantizeAndLinear(torch.autograd.Function):\n",
+    " \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 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.bias = bias\n",
+    " \n",
+    "    def forward(self, input):\n",
+    "        return DequantizeAndLinear.apply(input, self.weight, self.absmax, self.code, self.bias)\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 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",
+    " \n",
+    "    def forward(self, x, **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",
+    "            return F.embedding(x, weight_deq, **kwargs)\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 dummify(model, adapter_dim: int = 0):\n",
+    "    for module in list(model.modules()):\n",
+    "        for name, child in module.named_children():\n",
+    "            if isinstance(child, nn.Linear):\n",
+    "                print(name, child)\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",
+    "\n",
+    "\n",
+    "def bnbfy_(model, adapter_dim: int = 0):\n",
+    "    for module in list(model.modules()):\n",
+    "        for name, child in module.named_children():\n",
+    "            if isinstance(child, nn.Linear):\n",
+    "                print(name, child)\n",
+    "                setattr(module, name, FrozenBNBLinear.from_linear(child))\n",
+    "                \n",
+    "            elif isinstance(child, nn.Embedding):\n",
+    "                print(name, child)\n",
+    "                setattr(module, name, FrozenBNBEmbedding.from_embedding(child))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 66,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class GPTJBlock(transformers.models.gptj.modeling_gptj.GPTJBlock):\n",
+    "    def __init__(self, config):\n",
+    "        print(\"MONKEYPATCH BLOCK\")\n",
+    "        super().__init__(config)\n",
+    "\n",
+    "        dummify(self.attn)\n",
+    "        dummify(self.mlp)\n",
+    "\n",
+    "transformers.models.gptj.modeling_gptj.GPTJBlock = GPTJBlock\n",
+    "\n",
+    "\n",
+    "class GPTJModel(transformers.models.gptj.modeling_gptj.GPTJModel):\n",
+    "    def __init__(self, config):\n",
+    "        super().__init__(config)\n",
+    "        dummify(self)\n",
+    "class GPTJForCausalLM(transformers.models.gptj.modeling_gptj.GPTJForCausalLM):\n",
+    "    def __init__(self, config):\n",
+    "        super().__init__(config)\n",
+    "        dummify(self)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 67,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "1c98b9ebbf8d44d8b0bc422d4bfce21f",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Downloading:   0%|          | 0.00/0.98k [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "04bc6b612ff146308ec0b63fc15640f8",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Downloading:   0%|          | 0.00/5.75G [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "MONKEYPATCH BLOCK\n",
+      "k_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "v_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "q_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "out_proj Linear(in_features=4096, out_features=4096, bias=False)\n",
+      "fc_in Linear(in_features=4096, out_features=16384, bias=True)\n",
+      "fc_out Linear(in_features=16384, out_features=4096, bias=True)\n",
+      "lm_head Linear(in_features=4096, out_features=50400, bias=True)\n"
+     ]
+    }
+   ],
+   "source": [
+    "config = transformers.GPTJConfig.from_pretrained(\"EleutherAI/gpt-j-6B\")\n",
+    "tokenizer = transformers.AutoTokenizer.from_pretrained(\"EleutherAI/gpt-j-6B\")\n",
+    "gpt = GPTJForCausalLM.from_pretrained(\"hivemind/gpt-j-6B-8bit\", low_cpu_mem_usage=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 68,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = 'cuda' if torch.cuda.is_available else 'cpu'\n",
+    "gpt.to(device).train(False);"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 69,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Reusing dataset wikitext (/home/jheuristic/.cache/huggingface/datasets/wikitext/wikitext-2-v1/1.0.0/a241db52902eaf2c6aa732210bead40c090019a499ceb13bcbfa3f8ab646a126)\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "bfbf0e20ed194d679d2f877085f679cb",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "  0%|          | 0/3 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from datasets import load_dataset\n",
+    "data = load_dataset('wikitext', 'wikitext-2-v1')['test']"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 70,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "53d7e76934de4a1498306d49e4f41ad2",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "  0%|          | 0/4358 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "\n",
+    "numerator, denominator = 0, 0\n",
+    "collator = transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False)\n",
+    "loader = torch.utils.data.DataLoader(data, batch_size=1, num_workers=0, shuffle=False)\n",
+    "\n",
+    "\n",
+    "with torch.no_grad(), torch.cuda.amp.autocast(), tqdm(loader) as progressbar:\n",
+    "    for i, row in enumerate(progressbar):\n",
+    "        if max(map(len, row['text'])) <= 1:\n",
+    "            continue\n",
+    "        batch = tokenizer(**row, truncation=False, return_tensors='pt')\n",
+    "        batch = {k: v.cuda() for k, v in batch.items()}\n",
+    "\n",
+    "        out = gpt.forward(**batch,)\n",
+    "\n",
+    "        loss = F.cross_entropy(out.logits[:, :-1, :].flatten(0, -2), batch['input_ids'][:, 1:].flatten(),\n",
+    "                               reduction='none')\n",
+    "\n",
+    "        numerator += loss.sum().item()\n",
+    "        denominator += len(loss)\n",
+    "        progressbar.desc = f\"{numerator/denominator:.3f}\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 71,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "18.427138288946292"
+      ]
+     },
+     "execution_count": 71,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# test perplexity\n",
+    "import math\n",
+    "math.exp(numerator/denominator)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "py38",
+   "language": "python",
+   "name": "py38"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.8.1"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}