{ "cells": [ { "cell_type": "markdown", "id": "0_xya1nyDHfY", "metadata": { "id": "0_xya1nyDHfY" }, "source": [ "\n", "\n", "\n", "\n", "\n", "
\n", "\n", "Supplementary code for the Build a Large Language Model From Scratch book by Sebastian Raschka
\n", "
Code repository: https://github.com/rasbt/LLMs-from-scratch\n", "\n", "		\n", "\n", "
" ] }, { "cell_type": "markdown", "id": "l62zIRRSBy_R", "metadata": { "id": "l62zIRRSBy_R" }, "source": [ "# Converting a From-Scratch GPT Architecture to Llama 2" ] }, { "cell_type": "markdown", "id": "aFmxTQbwCUMl", "metadata": { "id": "aFmxTQbwCUMl" }, "source": [ "- In this notebook, we convert the original GPT architecture into a Llama 2 model step by step (note the GPT and GPT-2 share the same architecture)\n", "- Why not Llama 1 or Llama 3?\n", " - The Llama 1 architecture is similar to Llama 2, except that Llama 2 has a larger context window (which is nice); the Llama 1 weights are not readily available and have more usage restrictions, so it makes more sense to focus on Llama 2\n", " - Regarding Llama 3, I will share a separate notebook to convert Llama 2 to Llama 3 (there are only a few small additional changes)\n", "- The explanations are purposefully kept minimal in this notebook not to bloat it unnecessarily and focus on the main code\n", "- For more information, please see the Llama 2 paper: [Llama 2: Open Foundation and Fine-Tuned Chat Models (2023)](https://arxiv.org/abs/2307.09288)" ] }, { "cell_type": "markdown", "id": "ohhMKUWvGm9z", "metadata": { "id": "ohhMKUWvGm9z" }, "source": [ "" ] }, { "cell_type": "markdown", "id": "JBpQwU89ETA1", "metadata": { "id": "JBpQwU89ETA1" }, "source": [ "- Packages that are being used in this notebook:" ] }, { "cell_type": "code", "execution_count": 1, "id": "34a9a440-84c2-42cc-808b-38677cb6af8a", "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "34a9a440-84c2-42cc-808b-38677cb6af8a", "outputId": "8118963b-3c72-43af-874b-439ffebdc94c" }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "huggingface_hub version: 0.33.0\n", "sentencepiece version: 0.2.0\n", "torch version: 2.6.0\n" ] } ], "source": [ "from importlib.metadata import version\n", "\n", "pkgs = [\n", " \"huggingface_hub\", # to download pretrained weights\n", " \"sentencepiece\", # to implement the tokenizer\n", " \"torch\", # to implement the model\n", "]\n", "for p in pkgs:\n", " print(f\"{p} version: {version(p)}\")" ] }, { "cell_type": "markdown", "id": "UJJneXpTEg4W", "metadata": { "id": "UJJneXpTEg4W" }, "source": [ " \n", "# 1. Convert the GPT model implementation step by step" ] }, { "cell_type": "markdown", "id": "v1zpfX2GHBKa", "metadata": { "id": "v1zpfX2GHBKa" }, "source": [ "- In this section, we go through the GPT model code from [chapter 4](../../ch04/01_main-chapter-code/ch04.ipynb) and modify it step by step to implement the Llama 2 architecture\n", "- Later, we load the original Llama 2 weights shared by Meta AI" ] }, { "cell_type": "markdown", "id": "979c7b6d-1370-4da1-8bfb-a2b27537bf2f", "metadata": { "id": "979c7b6d-1370-4da1-8bfb-a2b27537bf2f" }, "source": [ " \n", "## 1.1 Replace LayerNorm with RMSNorm layer" ] }, { "cell_type": "markdown", "id": "f8b27fc8-23a1-4e0e-a1ea-792e0428e5e6", "metadata": { "id": "f8b27fc8-23a1-4e0e-a1ea-792e0428e5e6" }, "source": [ "- First, we replace LayerNorm by Root Mean Square Layer Normalization (RMSNorm)\n", "- LayerNorm normalizes inputs using mean and variance, while RMSNorm uses only the root mean square, which improves computational efficiency\n", "- The RMSNorm operation is as follows, where $x$ is the input $\\gamma$ is a trainable parameter (vector), and $\\epsilon$ is a small constant to avoid zero-division errors:\n", "\n", "$$y_i = \\frac{x_i}{\\text{RMS}(x)} \\gamma_i, \\quad \\text{where} \\quad \\text{RMS}(x) = \\sqrt{\\epsilon + \\frac{1}{n} \\sum x_i^2}$$\n", "\n", "- For more details, please see the paper [Root Mean Square Layer Normalization (2019)](https://arxiv.org/abs/1910.07467)" ] }, { "cell_type": "code", "execution_count": 2, "id": "d7094381-9499-4e9e-93f9-b79470da3771", "metadata": { "id": "d7094381-9499-4e9e-93f9-b79470da3771" }, "outputs": [], "source": [ "import torch\n", "import torch.nn as nn\n", "\n", "\n", "#####################################\n", "# Chapter 4\n", "#####################################\n", "\n", "# class LayerNorm(nn.Module):\n", "# def __init__(self, emb_dim):\n", "# super().__init__()\n", "# self.eps = 1e-5\n", "# self.scale = nn.Parameter(torch.ones(emb_dim))\n", "# self.shift = nn.Parameter(torch.zeros(emb_dim))\n", "\n", "# def forward(self, x):\n", "# mean = x.mean(dim=-1, keepdim=True)\n", "# var = x.var(dim=-1, keepdim=True, unbiased=False)\n", "# norm_x = (x - mean) / torch.sqrt(var + self.eps)\n", "# return self.scale * norm_x + self.shift\n", "\n", "\n", "class RMSNorm(nn.Module):\n", " def __init__(self, emb_dim, eps=1e-5):\n", " super().__init__()\n", " self.eps = eps\n", " self.emb_dim = emb_dim\n", " self.weight = nn.Parameter(torch.ones(emb_dim)).float()\n", "\n", " def forward(self, x):\n", " means = x.pow(2).mean(dim=-1, keepdim=True)\n", " x_normed = x * torch.rsqrt(means + self.eps)\n", " return (x_normed * self.weight).to(dtype=x.dtype)" ] }, { "cell_type": "markdown", "id": "mtWC8DOmIu0F", "metadata": { "id": "mtWC8DOmIu0F" }, "source": [ "- The following code cell checks that this implementation works the same as PyTorch's built-in implementation:" ] }, { "cell_type": "code", "execution_count": 3, "id": "e41ade7a-bf06-48b1-8b7e-0e4037d5753f", "metadata": { "id": "e41ade7a-bf06-48b1-8b7e-0e4037d5753f" }, "outputs": [], "source": [ "torch.manual_seed(123)\n", "\n", "example_batch = torch.randn(2, 3, 4)\n", "\n", "rms_norm = RMSNorm(emb_dim=example_batch.shape[-1])\n", "rmsnorm_pytorch = torch.nn.RMSNorm(example_batch.shape[-1], eps=1e-5)\n", "\n", "assert torch.allclose(rms_norm(example_batch), rmsnorm_pytorch(example_batch))" ] }, { "cell_type": "markdown", "id": "5eb81f83-c38c-46a4-b763-aa630a32e357", "metadata": { "id": "5eb81f83-c38c-46a4-b763-aa630a32e357" }, "source": [ " \n", "## 1.2 Replace GELU with SiLU activation" ] }, { "cell_type": "markdown", "id": "0b8aa702-f118-4ff6-9135-90725ec8756c", "metadata": { "id": "0b8aa702-f118-4ff6-9135-90725ec8756c" }, "source": [ "- Llama uses the SiLU activation function (instead of GELU), which is also known as the Swish function:\n", "\n", "$$\n", "\\text{silu}(x) = x \\cdot \\sigma(x), \\quad \\text{where} \\quad \\sigma(x) \\text{ is the logistic sigmoid.}\n", "$$\n", "\n", "- For more information, see the SiLU paper: [Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning (2017)](https://arxiv.org/abs/1702.03118)" ] }, { "cell_type": "code", "execution_count": 4, "id": "a74f3757-c634-4a3a-a8f3-6334cde454fe", "metadata": { "id": "a74f3757-c634-4a3a-a8f3-6334cde454fe" }, "outputs": [], "source": [ "#####################################\n", "# Chapter 4\n", "#####################################\n", "\n", "# class GELU(nn.Module):\n", "# def __init__(self):\n", "# super().__init__()\n", "\n", "# def forward(self, x):\n", "# return 0.5 * x * (1 + torch.tanh(\n", "# torch.sqrt(torch.tensor(2.0 / torch.pi)) *\n", "# (x + 0.044715 * torch.pow(x, 3))\n", "# ))\n", "\n", "\n", "class SiLU(nn.Module):\n", " def __init__(self):\n", " super(SiLU, self).__init__()\n", "\n", " def forward(self, x):\n", " return x * torch.sigmoid(x)" ] }, { "cell_type": "code", "execution_count": 5, "id": "72ecbe2e-b6b7-4319-972b-1a7fefa3368c", "metadata": { "id": "72ecbe2e-b6b7-4319-972b-1a7fefa3368c" }, "outputs": [], "source": [ "silu = SiLU()\n", "\n", "assert torch.allclose(silu(example_batch), torch.nn.functional.silu(example_batch))" ] }, { "cell_type": "markdown", "id": "4f9b5167-1da9-46c8-9964-8036b3b1deb9", "metadata": { "id": "4f9b5167-1da9-46c8-9964-8036b3b1deb9" }, "source": [ " \n", "## 1.3 Update the FeedForward module" ] }, { "cell_type": "markdown", "id": "3a381e7a-b807-472e-91c9-3e4e3fc5ad91", "metadata": { "id": "3a381e7a-b807-472e-91c9-3e4e3fc5ad91" }, "source": [ "- In fact, Llama uses a \"Gates Linear Unit\" (GLU) variant of SiLU called SwiGLU, which essentially results in a slightly differently structured `FeedForward` module\n", "- SwiGLU uses a gating mechanism in the feedforward layer, with the formula:\n", "\n", "$$\\text{SwiGLU}(x) = \\text{SiLU}(\\text{Linear}_1(x)) * (\\text{Linear}_2(x))$$\n", "\n", "- Here, $\\text{Linear}_1$ and $\\text{Linear}_2$ are two linear layers, and $*$ denotes element-wise multiplication\n", "- The third linear layer, $\\text{Linear}_3$, is applied after this gated activation\n", "\n", "- For more information, see SwiGLU paper: [GLU Variants Improve Transformer (2020)](https://arxiv.org/abs/2002.05202)" ] }, { "cell_type": "code", "execution_count": 6, "id": "d25fbe3d-b7c9-4772-ad67-bc0527e4e20a", "metadata": { "id": "d25fbe3d-b7c9-4772-ad67-bc0527e4e20a" }, "outputs": [], "source": [ "#####################################\n", "# Chapter 4\n", "#####################################\n", "# class FeedForward(nn.Module):\n", "# def __init__(self, cfg):\n", "# super().__init__()\n", "# self.layers = nn.Sequential(\n", "# nn.Linear(cfg[\"emb_dim\"], 4 * cfg[\"emb_dim\"]),\n", "# GELU(),\n", "# nn.Linear(4 * cfg[\"emb_dim\"], cfg[\"emb_dim\"]),\n", "# )\n", "\n", "# def forward(self, x):\n", "# return self.layers(x)" ] }, { "cell_type": "code", "execution_count": 7, "id": "477568cb-03cd-4510-b663-a42ce3ec64a2", "metadata": { "id": "477568cb-03cd-4510-b663-a42ce3ec64a2" }, "outputs": [], "source": [ "class FeedForward(nn.Module):\n", " def __init__(self, cfg):\n", " super().__init__()\n", " self.fc1 = nn.Linear(cfg[\"emb_dim\"], cfg[\"hidden_dim\"], dtype=cfg[\"dtype\"], bias=False)\n", " self.fc2 = nn.Linear(cfg[\"emb_dim\"], cfg[\"hidden_dim\"], dtype=cfg[\"dtype\"], bias=False)\n", " self.fc3 = nn.Linear(cfg[\"hidden_dim\"], cfg[\"emb_dim\"], dtype=cfg[\"dtype\"], bias=False)\n", " self.silu = SiLU()\n", "\n", " def forward(self, x):\n", " x_fc1 = self.fc1(x)\n", " x_fc2 = self.fc2(x)\n", " x = self.silu(x_fc1) * x_fc2\n", " return self.fc3(x)" ] }, { "cell_type": "markdown", "id": "qcD8LSHNhBRW", "metadata": { "id": "qcD8LSHNhBRW" }, "source": [ "- Note that we also added a `dtype=cfg[\"dtype\"]` setting above, which will allow us to load the model directly in lower precision formats later to reduce memory usage (versus instantiating it in the original 32-bit precision format and then converting it)\n", "- We also set `bias=False` since Llama doesn't use any bias units" ] }, { "cell_type": "markdown", "id": "f6b7bf4f-99d0-42c1-807c-5074d2cc1949", "metadata": { "id": "f6b7bf4f-99d0-42c1-807c-5074d2cc1949" }, "source": [ " \n", "## 1.4 Implement RoPE" ] }, { "cell_type": "markdown", "id": "d3487a6f-0373-49d8-b2eb-f8ee05d42884", "metadata": { "id": "d3487a6f-0373-49d8-b2eb-f8ee05d42884" }, "source": [ "- In the GPT model, the positional embeddings are implemented as follows:\n", "\n", "```python\n", "self.pos_emb = nn.Embedding(cfg[\"context_length\"], cfg[\"emb_dim\"])\n", "```\n", "\n", "- Unlike traditional absolute positional embeddings, Llama uses rotary position embeddings (RoPE), which enable it to capture both absolute and relative positional information simultaneously\n", "- The reference paper for RoPE is [RoFormer: Enhanced Transformer with Rotary Position Embedding (2021)](https://arxiv.org/abs/2104.09864)" ] }, { "cell_type": "code", "execution_count": 8, "id": "a34180fb-448f-44e9-a244-0c736051687b", "metadata": { "id": "a34180fb-448f-44e9-a244-0c736051687b" }, "outputs": [], "source": [ "def precompute_rope_params(head_dim, theta_base=10_000, context_length=4096):\n", " assert head_dim % 2 == 0, \"Embedding dimension must be even\"\n", "\n", " # Compute the inverse frequencies\n", " inv_freq = 1.0 / (theta_base ** (torch.arange(0, head_dim, 2)[: (head_dim // 2)].float() / head_dim))\n", "\n", " # Generate position indices\n", " positions = torch.arange(context_length)\n", "\n", " # Compute the angles\n", " angles = positions[:, None] * inv_freq[None, :] # Shape: (context_length, head_dim // 2)\n", "\n", " # Expand angles to match the head_dim\n", " angles = torch.cat([angles, angles], dim=1) # Shape: (context_length, head_dim)\n", "\n", " # Precompute sine and cosine\n", " cos = torch.cos(angles)\n", " sin = torch.sin(angles)\n", "\n", " return cos, sin\n", "\n", "def compute_rope(x, cos, sin):\n", " # x: (batch_size, num_heads, seq_len, head_dim)\n", " batch_size, num_heads, seq_len, head_dim = x.shape\n", " assert head_dim % 2 == 0, \"Head dimension must be even\"\n", "\n", " # Split x into first half and second half\n", " x1 = x[..., : head_dim // 2] # First half\n", " x2 = x[..., head_dim // 2 :] # Second half\n", "\n", " # Adjust sin and cos shapes\n", " cos = cos[:seq_len, :].unsqueeze(0).unsqueeze(0) # Shape: (1, 1, seq_len, head_dim)\n", " sin = sin[:seq_len, :].unsqueeze(0).unsqueeze(0)\n", "\n", " # Apply the rotary transformation\n", " rotated = torch.cat((-x2, x1), dim=-1)\n", " x_rotated = (x * cos) + (rotated * sin)\n", "\n", " return x_rotated.to(dtype=x.dtype)" ] }, { "cell_type": "markdown", "id": "8e841b8e-75aa-49db-b1a7-d5c2116dc299", "metadata": { "id": "8e841b8e-75aa-49db-b1a7-d5c2116dc299" }, "source": [ "- The following is an example of applying RoPE to the `q` and `k` tensors:" ] }, { "cell_type": "code", "execution_count": 9, "id": "8c89f022-7167-4001-8c21-8e012878733f", "metadata": { "id": "8c89f022-7167-4001-8c21-8e012878733f" }, "outputs": [], "source": [ "# Settings\n", "batch_size = 2\n", "context_len = 5\n", "num_heads = 4\n", "head_dim = 16\n", "\n", "# Instantiate RoPE parameters\n", "cos, sin = precompute_rope_params(head_dim=head_dim, context_length=context_len)\n", "\n", "# Dummy query and key tensors\n", "torch.manual_seed(123)\n", "queries = torch.randn(batch_size, num_heads, context_len, head_dim)\n", "keys = torch.randn(batch_size, num_heads, context_len, head_dim)\n", "\n", "# Apply rotary position embeddings\n", "queries_rot = compute_rope(queries, cos, sin)\n", "keys_rot = compute_rope(keys, cos, sin)" ] }, { "cell_type": "markdown", "id": "f78127b0-dda2-4c5a-98dd-bae8f5fe8297", "metadata": { "id": "f78127b0-dda2-4c5a-98dd-bae8f5fe8297" }, "source": [ " \n", "## 1.5 Add RoPE to MultiHeadAttention module" ] }, { "cell_type": "markdown", "id": "RnmSHROLhhR3", "metadata": { "id": "RnmSHROLhhR3" }, "source": [ "- It's important to note that GPT applies the positional embeddings to the inputs, whereas Llama applies rotations to the query and key vectors in the self-attention mechanism itself\n", "- Here, we modify the `MultiHeadAttention` class with the appropriate RoPE code\n", "- In addition, we remove the `qkv_bias` option and hardcode the `bias=False` setting\n", "- Also, we add a dtype setting to be able to instantiate the model with a lower precision later\n", " - Tip: since the `TransformerBlock`s (in the next section) are repeated exactly, we could simplify the code and only initialize the buffers once instead for each `MultiHeadAttention` module; however, we add the precomputed RoPE parameters to the `MultiHeadAttention` class so that it can function as a standalone module" ] }, { "cell_type": "code", "execution_count": 10, "id": "d81a441e-0b79-4a8b-8291-ea7f55d58c84", "metadata": { "id": "d81a441e-0b79-4a8b-8291-ea7f55d58c84" }, "outputs": [], "source": [ "#####################################\n", "# Chapter 3\n", "#####################################\n", "class MultiHeadAttention(nn.Module):\n", " def __init__(self, d_in, d_out, context_length, num_heads, dtype=None): # ,dropout, num_heads, qkv_bias=False):\n", " super().__init__()\n", " assert d_out % num_heads == 0, \"d_out must be divisible by n_heads\"\n", "\n", " self.d_out = d_out\n", " self.num_heads = num_heads\n", " self.head_dim = d_out // num_heads # Reduce the projection dim to match desired output dim\n", "\n", " ################################### NEW ###################################\n", " # Set bias=False and dtype=dtype for all linear layers below\n", " ###########################################################################\n", " self.W_query = nn.Linear(d_in, d_out, bias=False, dtype=dtype)\n", " self.W_key = nn.Linear(d_in, d_out, bias=False, dtype=dtype)\n", " self.W_value = nn.Linear(d_in, d_out, bias=False, dtype=dtype)\n", " self.out_proj = nn.Linear(d_out, d_out, bias=False, dtype=dtype) # Linear layer to combine head outputs\n", " # self.dropout = nn.Dropout(dropout)\n", " self.register_buffer(\"mask\", torch.triu(torch.ones(context_length, context_length), diagonal=1))\n", "\n", " ################################### NEW ###################################\n", " cos, sin = precompute_rope_params(head_dim=self.head_dim, context_length=context_length)\n", " self.register_buffer(\"cos\", cos)\n", " self.register_buffer(\"sin\", sin)\n", " ###########################################################################\n", "\n", "\n", " def forward(self, x):\n", "\n", " b, num_tokens, d_in = x.shape\n", "\n", " keys = self.W_key(x) # Shape: (b, num_tokens, d_out)\n", " queries = self.W_query(x)\n", " values = self.W_value(x)\n", "\n", " # We implicitly split the matrix by adding a `num_heads` dimension\n", " # Unroll last dim: (b, num_tokens, d_out) -> (b, num_tokens, num_heads, head_dim)\n", " keys = keys.view(b, num_tokens, self.num_heads, self.head_dim)\n", " values = values.view(b, num_tokens, self.num_heads, self.head_dim)\n", " queries = queries.view(b, num_tokens, self.num_heads, self.head_dim)\n", "\n", " # Transpose: (b, num_tokens, num_heads, head_dim) -> (b, num_heads, num_tokens, head_dim)\n", " keys = keys.transpose(1, 2)\n", " queries = queries.transpose(1, 2)\n", " values = values.transpose(1, 2)\n", "\n", " ################################### NEW ###################################\n", " keys = compute_rope(keys, self.cos, self.sin)\n", " queries = compute_rope(queries, self.cos, self.sin)\n", " ###########################################################################\n", "\n", " # Compute scaled dot-product attention (aka self-attention) with a causal mask\n", " attn_scores = queries @ keys.transpose(2, 3) # Dot product for each head\n", "\n", " # Original mask truncated to the number of tokens and converted to boolean\n", " mask_bool = self.mask.bool()[:num_tokens, :num_tokens]\n", "\n", " # Use the mask to fill attention scores\n", " attn_scores.masked_fill_(mask_bool, -torch.inf)\n", "\n", " attn_weights = torch.softmax(attn_scores / keys.shape[-1]**0.5, dim=-1)\n", " # attn_weights = self.dropout(attn_weights)\n", "\n", " # Shape: (b, num_tokens, num_heads, head_dim)\n", " context_vec = (attn_weights @ values).transpose(1, 2)\n", "\n", " # Combine heads, where self.d_out = self.num_heads * self.head_dim\n", " context_vec = context_vec.reshape(b, num_tokens, self.d_out)\n", " context_vec = self.out_proj(context_vec) # optional projection\n", "\n", " return context_vec" ] }, { "cell_type": "markdown", "id": "-lt9SfnVioB3", "metadata": { "id": "-lt9SfnVioB3" }, "source": [ "- Below is an example using the `MultiHeadAttention` module on an example input:" ] }, { "cell_type": "code", "execution_count": 11, "id": "03f15755-0083-483f-963b-99b599651638", "metadata": { "id": "03f15755-0083-483f-963b-99b599651638" }, "outputs": [], "source": [ "# Settings\n", "batch_size = 1\n", "context_len = 100\n", "max_context_len = 4096\n", "embed_dim = 128\n", "num_heads = 4\n", "\n", "\n", "example_batch = torch.randn((batch_size, context_len, embed_dim))\n", "\n", "mha = MultiHeadAttention(\n", " d_in=embed_dim,\n", " d_out=embed_dim,\n", " context_length=max_context_len,\n", " num_heads=num_heads\n", ")\n", "\n", "mha(example_batch)\n", "\n", "del mha # delete to free up memory" ] }, { "cell_type": "markdown", "id": "e5a1a272-a038-4b8f-aaaa-f4b241e7f23f", "metadata": { "id": "e5a1a272-a038-4b8f-aaaa-f4b241e7f23f" }, "source": [ " \n", "## 1.6 Update the TransformerBlock module" ] }, { "cell_type": "markdown", "id": "255f70ac-9c2e-4328-8af7-1c298b8d4a18", "metadata": { "id": "255f70ac-9c2e-4328-8af7-1c298b8d4a18" }, "source": [ "- At this stage, most of the hard work is already done; we can now update the `TransformerBlock` to use the code we implemented above\n", "- This means we\n", " - replace LayerNorm with RMSNorm\n", " - remove dropout\n", " - remove the `qkv_bias` setting\n", " - add the `dtype` setting" ] }, { "cell_type": "code", "execution_count": 12, "id": "2e110721-bf2b-42b3-989a-1635b1658af0", "metadata": { "id": "2e110721-bf2b-42b3-989a-1635b1658af0" }, "outputs": [], "source": [ "class TransformerBlock(nn.Module):\n", " def __init__(self, cfg):\n", " super().__init__()\n", " self.att = MultiHeadAttention(\n", " d_in=cfg[\"emb_dim\"],\n", " d_out=cfg[\"emb_dim\"],\n", " context_length=cfg[\"context_length\"],\n", " num_heads=cfg[\"n_heads\"],\n", " dtype=cfg[\"dtype\"] # NEW\n", " # dropout=cfg[\"drop_rate\"],\n", " # qkv_bias=cfg[\"qkv_bias\"]\n", " )\n", " self.ff = FeedForward(cfg)\n", "\n", " ################################### NEW ###################################\n", " # self.norm1 = LayerNorm(cfg[\"emb_dim\"])\n", " # self.norm2 = LayerNorm(cfg[\"emb_dim\"])\n", " self.norm1 = RMSNorm(cfg[\"emb_dim\"])\n", " self.norm2 = RMSNorm(cfg[\"emb_dim\"])\n", " ###########################################################################\n", "\n", " # self.drop_shortcut = nn.Dropout(cfg[\"drop_rate\"])\n", "\n", " def forward(self, x):\n", " # Shortcut connection for attention block\n", " shortcut = x\n", " x = self.norm1(x)\n", " x = self.att(x) # Shape [batch_size, num_tokens, emb_size]\n", " # x = self.drop_shortcut(x)\n", " x = x + shortcut # Add the original input back\n", "\n", " # Shortcut connection for feed-forward block\n", " shortcut = x\n", " x = self.norm2(x)\n", " x = self.ff(x)\n", " # x = self.drop_shortcut(x)\n", " x = x + shortcut # Add the original input back\n", "\n", " return x" ] }, { "cell_type": "markdown", "id": "ada953bc-e2c0-4432-a32d-3f7efa3f6e0f", "metadata": { "id": "ada953bc-e2c0-4432-a32d-3f7efa3f6e0f" }, "source": [ " \n", "## 1.7 Update the model class" ] }, { "cell_type": "markdown", "id": "ba5d991a-559b-47be-96f4-31b881ab2da8", "metadata": { "id": "ba5d991a-559b-47be-96f4-31b881ab2da8" }, "source": [ "- As you may recall from [chapter 5](../01_main-chapter-code/ch05.ipynb), the `TransformerBlock` is a repeated block within the main model\n", "- Our Llama model is almost complete; we just have to update the model code surrounding the `TransformerBlock`\n", "- This means we\n", " - remove absolute positional embeddings since we have RoPE embeddings now\n", " - replace LayerNorm with RMSNorm\n", " - remove dropout\n", " - add the dtype setting" ] }, { "cell_type": "code", "execution_count": 13, "id": "cf8240fe-5d7f-4e7e-b1ac-e0755aab5e79", "metadata": { "id": "cf8240fe-5d7f-4e7e-b1ac-e0755aab5e79" }, "outputs": [], "source": [ "# class GPTModel(nn.Module):\n", "class Llama2Model(nn.Module):\n", " def __init__(self, cfg):\n", " super().__init__()\n", " self.tok_emb = nn.Embedding(cfg[\"vocab_size\"], cfg[\"emb_dim\"], dtype=cfg[\"dtype\"])\n", " # self.pos_emb = nn.Embedding(cfg[\"context_length\"], cfg[\"emb_dim\"])\n", " # self.drop_emb = nn.Dropout(cfg[\"drop_rate\"])\n", "\n", " self.trf_blocks = nn.Sequential(\n", " *[TransformerBlock(cfg) for _ in range(cfg[\"n_layers\"])])\n", "\n", " ################################### NEW ###################################\n", " # self.final_norm = LayerNorm(cfg[\"emb_dim\"])\n", " self.final_norm = RMSNorm(cfg[\"emb_dim\"])\n", " ###########################################################################\n", " self.out_head = nn.Linear(cfg[\"emb_dim\"], cfg[\"vocab_size\"], bias=False, dtype=cfg[\"dtype\"])\n", "\n", " def forward(self, in_idx):\n", " # batch_size, seq_len = in_idx.shape\n", " tok_embeds = self.tok_emb(in_idx)\n", " # pos_embeds = self.pos_emb(torch.arange(seq_len, device=in_idx.device))\n", " x = tok_embeds # + pos_embeds # Shape [batch_size, num_tokens, emb_size]\n", " # x = self.drop_emb(x)\n", " x = self.trf_blocks(x)\n", " x = self.final_norm(x)\n", " logits = self.out_head(x)\n", " return logits" ] }, { "cell_type": "markdown", "id": "4bc94940-aaeb-45b9-9399-3a69b8043e60", "metadata": { "id": "4bc94940-aaeb-45b9-9399-3a69b8043e60" }, "source": [ " \n", "## 2. Initialize model" ] }, { "cell_type": "markdown", "id": "bG--zY-Ljj1f", "metadata": { "id": "bG--zY-Ljj1f" }, "source": [ "- The model code is now complete, and we are ready to initialize it\n", "- In [chapter 5](../01_main-chapter-code/ch05.ipynb), we used the following config file to specify the 124M-parameter GPT model:" ] }, { "cell_type": "code", "execution_count": 14, "id": "4b7428df-3d02-4ccd-97b5-a629bdabbe8f", "metadata": { "id": "4b7428df-3d02-4ccd-97b5-a629bdabbe8f" }, "outputs": [], "source": [ "GPT_CONFIG_124M = {\n", " \"vocab_size\": 50257, # Vocabulary size\n", " \"context_length\": 1024, # Context length\n", " \"emb_dim\": 768, # Embedding dimension\n", " \"n_heads\": 12, # Number of attention heads\n", " \"n_layers\": 12, # Number of layers\n", " \"drop_rate\": 0.1, # Dropout rate\n", " \"qkv_bias\": False # Query-Key-Value bias\n", "}" ] }, { "cell_type": "markdown", "id": "8bVi8uiBjw2T", "metadata": { "id": "8bVi8uiBjw2T" }, "source": [ "- For reference, the 1.5B parameter GPT model config is shown below as well:" ] }, { "cell_type": "code", "execution_count": 15, "id": "tAOojV_mkEnd", "metadata": { "id": "tAOojV_mkEnd" }, "outputs": [], "source": [ "GPT_CONFIG_1558M = {\n", " \"vocab_size\": 50257, # Vocabulary size\n", " \"context_length\": 1024, # Context length\n", " \"emb_dim\": 1600, # Embedding dimension\n", " \"n_heads\": 25, # Number of attention heads\n", " \"n_layers\": 48, # Number of layers\n", " \"drop_rate\": 0.1, # Dropout rate\n", " \"qkv_bias\": False # Query-Key-Value bias\n", "}" ] }, { "cell_type": "markdown", "id": "HoGGRAGykQTE", "metadata": { "id": "HoGGRAGykQTE" }, "source": [ "- Similarly, we can define a Llama 2 config file for the 7B model (we ignore the other larger models for simplicity here):" ] }, { "cell_type": "code", "execution_count": 16, "id": "e0564727-2d35-4f0c-b0fc-cde1e9134a18", "metadata": { "id": "e0564727-2d35-4f0c-b0fc-cde1e9134a18" }, "outputs": [], "source": [ "LLAMA2_CONFIG_7B = {\n", " \"vocab_size\": 32000, # Vocabulary size\n", " \"context_length\": 4096, # Context length\n", " \"emb_dim\": 4096, # Embedding dimension\n", " \"n_heads\": 32, # Number of attention heads\n", " \"n_layers\": 32, # Number of layers\n", " \"hidden_dim\": 11008, # NEW: Size of the intermediate dimension in FeedForward\n", " \"dtype\": torch.bfloat16 # NEW: Lower-precision dtype to reduce memory usage\n", "}" ] }, { "cell_type": "markdown", "id": "FAP7fiBzkaBz", "metadata": { "id": "FAP7fiBzkaBz" }, "source": [ "- Using these settings, we can now initialize a Llama 2 7B model (note that this requires ~26 GB of memory)" ] }, { "cell_type": "code", "execution_count": 17, "id": "7004d785-ac9a-4df5-8760-6807fc604686", "metadata": { "id": "7004d785-ac9a-4df5-8760-6807fc604686" }, "outputs": [], "source": [ "model = Llama2Model(LLAMA2_CONFIG_7B)" ] }, { "cell_type": "code", "execution_count": 18, "id": "6079f747-8f20-4c6b-8d38-7156f1101729", "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "6079f747-8f20-4c6b-8d38-7156f1101729", "outputId": "0a0eb34b-1a21-4c11-804f-b40007bda5a3" }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Total number of parameters: 6,738,415,616\n" ] } ], "source": [ "total_params = sum(p.numel() for p in model.parameters())\n", "print(f\"Total number of parameters: {total_params:,}\")" ] }, { "cell_type": "markdown", "id": "Bx14NtzWk2wj", "metadata": { "id": "Bx14NtzWk2wj" }, "source": [ "- As shown above, the model contains 6.7 billion parameters (commonly rounded and referred to as a 7B model)\n", "- Additionally, we can calculate the memory requirements for this model using the code below:" ] }, { "cell_type": "code", "execution_count": 19, "id": "0df1c79e-27a7-4b0f-ba4e-167fe107125a", "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "0df1c79e-27a7-4b0f-ba4e-167fe107125a", "outputId": "11ced939-556d-4511-d5c0-10a94ed3df32" }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "float32 (PyTorch default): 52.33 GB\n", "bfloat16: 26.17 GB\n" ] } ], "source": [ "def model_memory_size(model, input_dtype=torch.float32):\n", " total_params = 0\n", " total_grads = 0\n", " for param in model.parameters():\n", " # Calculate total number of elements per parameter\n", " param_size = param.numel()\n", " total_params += param_size\n", " # Check if gradients are stored for this parameter\n", " if param.requires_grad:\n", " total_grads += param_size\n", "\n", " # Calculate buffer size (non-parameters that require memory)\n", " total_buffers = sum(buf.numel() for buf in model.buffers())\n", "\n", " # Size in bytes = (Number of elements) * (Size of each element in bytes)\n", " # We assume parameters and gradients are stored in the same type as input dtype\n", " element_size = torch.tensor(0, dtype=input_dtype).element_size()\n", " total_memory_bytes = (total_params + total_grads + total_buffers) * element_size\n", "\n", " # Convert bytes to gigabytes\n", " total_memory_gb = total_memory_bytes / (1024**3)\n", "\n", " return total_memory_gb\n", "\n", "print(f\"float32 (PyTorch default): {model_memory_size(model, input_dtype=torch.float32):.2f} GB\")\n", "print(f\"bfloat16: {model_memory_size(model, input_dtype=torch.bfloat16):.2f} GB\")" ] }, { "cell_type": "markdown", "id": "zudd-5PulKFL", "metadata": { "id": "zudd-5PulKFL" }, "source": [ "- Lastly, we can also transfer the model to an NVIDIA or Apple Silicon GPU if applicable:" ] }, { "cell_type": "code", "execution_count": 20, "id": "a4c50e19-1402-45b6-8ccd-9077b2ba836d", "metadata": { "id": "a4c50e19-1402-45b6-8ccd-9077b2ba836d" }, "outputs": [], "source": [ "if torch.cuda.is_available():\n", " device = torch.device(\"cuda\")\n", "elif torch.backends.mps.is_available():\n", " device = torch.device(\"mps\")\n", "else:\n", " device = torch.device(\"cpu\")\n", "\n", "model.to(device);" ] }, { "cell_type": "markdown", "id": "5dc64a06-27dc-46ec-9e6d-1700a8227d34", "metadata": { "id": "5dc64a06-27dc-46ec-9e6d-1700a8227d34" }, "source": [ " \n", "## 3. Load tokenizer" ] }, { "cell_type": "markdown", "id": "0eb30f0c-6144-4bed-87d9-6b2bac377005", "metadata": { "id": "0eb30f0c-6144-4bed-87d9-6b2bac377005" }, "source": [ "- In this section, we are going to load the tokenizer for the model\n", "- Llama 2 uses Google's [SentencePiece](https://github.com/google/sentencepiece) tokenizer instead of OpenAI's [Tiktoken](https://github.com/openai/tiktoken) (but Llama 3 uses Tiktoken)\n", "- Meta AI shared the original Llama 2 model weights and tokenizer vocabulary on the Hugging Face Hub\n", "- We will download the tokenizer vocabulary from the Hub and load it into SentencePiece\n", "- Uncomment and run the following code to install the required libraries:" ] }, { "cell_type": "code", "execution_count": 21, "id": "768989ea-dc60-4dc8-ae84-cbb3fd224422", "metadata": { "id": "768989ea-dc60-4dc8-ae84-cbb3fd224422" }, "outputs": [], "source": [ "# !pip install huggingface_hub sentencepiece" ] }, { "cell_type": "markdown", "id": "KbnlzsbYmJU6", "metadata": { "id": "KbnlzsbYmJU6" }, "source": [ "- Please note that Meta AI requires that you accept the Llama 2 licensing terms before you can download the files; to do this, you have to create a Hugging Face Hub account and visit the [meta-llama/Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b) repository to accept the terms\n", "- Next, you will need to create an access token; to generate an access token with READ permissions, click on the profile picture in the upper right and click on \"Settings\"\n", "\n", "\n", "\n", "\n", "- Then, create and copy the access token so you can copy & paste it into the next code cell\n", "\n", "" ] }, { "cell_type": "code", "execution_count": 22, "id": "3357a230-b678-4691-a238-257ee4e80185", "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "3357a230-b678-4691-a238-257ee4e80185", "outputId": "768ed6af-ce14-40bc-ca18-117b4b448269" }, "outputs": [], "source": [ "from huggingface_hub import login\n", "import json\n", "\n", "with open(\"config.json\", \"r\") as config_file:\n", " config = json.load(config_file)\n", " access_token = config[\"HF_ACCESS_TOKEN\"]\n", "\n", "login(token=access_token)" ] }, { "cell_type": "markdown", "id": "IxGh6ZYQo0VN", "metadata": { "id": "IxGh6ZYQo0VN" }, "source": [ "- After login via the access token, which is necessary to verify that we accepted the Llama 2 licensing terms, we can now download the tokenizer vocabulary:" ] }, { "cell_type": "code", "execution_count": 23, "id": "69714ea8-b9b8-4687-8392-f3abb8f93a32", "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 153, "referenced_widgets": [ "e6c75a6aa7b942fe84160e286e3acb3d", "08f0bf9459bd425498a5cb236f9d4a72", "10251d6f724e43788c41d4b7879cbfd3", "53a973c0853b44418698136bd04df039", "bdb071e7145a4007ae01599333e72612", "6b1821a7f4574e3aba09c1e410cc81e4", "8c2873eaec3445888ad3d54ad7387950", "0c8f7044966e4207b12352503c67dcbb", "8b5951213c9e4798a258146d61d02d11", "2c05df3f91e64df7b33905b1065a76f7", "742ae5487f2648fcae7ca8e22c7f8db9" ] }, "id": "69714ea8-b9b8-4687-8392-f3abb8f93a32", "outputId": "c230fec9-5c71-4a41-90ab-8a34d114ea01" }, "outputs": [], "source": [ "from huggingface_hub import hf_hub_download\n", "\n", "tokenizer_file = hf_hub_download(\n", " repo_id=\"meta-llama/Llama-2-7b\",\n", " filename=\"tokenizer.model\",\n", " local_dir=\"Llama-2-7b\"\n", ")" ] }, { "cell_type": "markdown", "id": "gp7iQ8cXAJLv", "metadata": { "id": "gp7iQ8cXAJLv" }, "source": [ "- To provide a more familiar interface for the tokenizer, we define a small `LlamaTokenizer` wrapper class:" ] }, { "cell_type": "code", "execution_count": 24, "id": "Ef4WxhjOBOOc", "metadata": { "id": "Ef4WxhjOBOOc" }, "outputs": [], "source": [ "import sentencepiece as spm\n", "\n", "\n", "class LlamaTokenizer:\n", " def __init__(self, tokenizer_file):\n", " sp = spm.SentencePieceProcessor()\n", " sp.load(tokenizer_file)\n", " self.tokenizer = sp\n", "\n", " def encode(self, text):\n", " return self.tokenizer.encode(text, out_type=int)\n", "\n", " def decode(self, ids):\n", " return self.tokenizer.decode(ids)\n", "\n", "\n", "tokenizer = LlamaTokenizer(tokenizer_file)" ] }, { "cell_type": "markdown", "id": "NVhmFeX3pT_M", "metadata": { "id": "NVhmFeX3pT_M" }, "source": [ "- We can now use the `generate` function to have the Llama 2 model generate new text:" ] }, { "cell_type": "code", "execution_count": 25, "id": "e0a2b5cd-6cba-4d72-b8ff-04d8315d483e", "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "e0a2b5cd-6cba-4d72-b8ff-04d8315d483e", "outputId": "acd5065d-8900-4ba8-ef85-968365f3a0cb" }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Output text:\n", " Every effort movesαllRadius deletingpretcc否']; future eer napulate lackус während inter DES издаSchéonkkarto Оryptato#{ningproof eerbye\n" ] } ], "source": [ "from previous_chapters import generate, text_to_token_ids, token_ids_to_text\n", "# If the `previous_chapters.py` file is not available locally,\n", "# you can import it from the `llms-from-scratch` PyPI package.\n", "# For details, see: https://github.com/rasbt/LLMs-from-scratch/tree/main/pkg\n", "# E.g.,\n", "# from llms_from_scratch.ch05 import generate, text_to_token_ids, token_ids_to_text\n", "\n", "\n", "\n", "torch.manual_seed(123)\n", "\n", "token_ids = generate(\n", " model=model,\n", " idx=text_to_token_ids(\"Every effort moves\", tokenizer).to(device),\n", " max_new_tokens=30,\n", " context_size=LLAMA2_CONFIG_7B[\"context_length\"],\n", " top_k=1,\n", " temperature=0.\n", ")\n", "\n", "print(\"Output text:\\n\", token_ids_to_text(token_ids, tokenizer))" ] }, { "cell_type": "markdown", "id": "93WTtAA5paYV", "metadata": { "id": "93WTtAA5paYV" }, "source": [ "- Of course, as we can see above, the text is nonsensical since we haven't trained the Llama 2 model yet\n", "- In the next section, instead of training it ourselves, which would cost tens to hundreds of thousands of dollars, we load the pretrained weights from Meta AI" ] }, { "cell_type": "markdown", "id": "f63cc248-1d27-4eb6-aa50-173b436652f8", "metadata": { "id": "f63cc248-1d27-4eb6-aa50-173b436652f8" }, "source": [ " \n", "## 4. Load pretrained weights" ] }, { "cell_type": "markdown", "id": "aKeN7rUfqZMI", "metadata": { "id": "aKeN7rUfqZMI" }, "source": [ "- We are loading the [\"meta-llama/Llama-2-7b\"](https://huggingface.co/meta-llama/Llama-2-7b) base model below, which is a simple text completion model before finetuning\n", "- Alternatively, you can load the instruction-finetuned and aligned [\"meta-llama/Llama-2-7b-chat\"](https://huggingface.co/meta-llama/Llama-2-7b-chat) model by modifying the string in the next code cell accordingly" ] }, { "cell_type": "code", "execution_count": 26, "id": "5fa9c06c-7a53-4b4d-9ce4-acc027322ee4", "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 49, "referenced_widgets": [ "66e777955e8748df878f118f07f38dab", "da89ae3ea4d2474e98f64ada608f3cea", "93e6da39c25f4edfaa72056c89df1f7f", "b628603e4cb0405398c916587ee96756", "93bedcb9245e44a0a1eb7e4155070f66", "0723f467d37b4904819a8bb33ebda10f", "e54928776bc649339002adced63738b0", "d8e0f42068af4cb094e2f115f76e06e0", "0a939565b6e94f08bee0a66e0f9827d4", "a5fedbb7ec2e43d99711bb4cd84b9486", "0c186f6539714d8eab023969ce47c500" ] }, "id": "5fa9c06c-7a53-4b4d-9ce4-acc027322ee4", "outputId": "0d8942cc-e5e2-4e77-ec41-1ac7bec7d94f" }, "outputs": [ { "data": { "application/vnd.jupyter.widget-view+json": { "model_id": "66e777955e8748df878f118f07f38dab", "version_major": 2, "version_minor": 0 }, "text/plain": [ "consolidated.00.pth: 0%| | 0.00/13.5G [00:00