File size: 8,325 Bytes
6f645f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 | from __future__ import annotations
import math
from dataclasses import asdict, dataclass
import torch
import torch.nn as nn
from torch.nn import functional as F
@dataclass
class GPTConfig:
block_size: int = 1024
vocab_size: int = 8192
n_layer: int = 12
n_head: int = 12
n_embd: int = 768
dropout: float = 0.0
bias: bool = False
class CausalSelfAttention(nn.Module):
def __init__(self, config: GPTConfig) -> None:
super().__init__()
if config.n_embd % config.n_head:
raise ValueError("embedding dimension must be divisible by number of heads")
self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
self.attn_dropout = nn.Dropout(config.dropout)
self.resid_dropout = nn.Dropout(config.dropout)
self.n_head = config.n_head
self.n_embd = config.n_embd
self.dropout = config.dropout
def forward(self, value: torch.Tensor) -> torch.Tensor:
batch, time, channels = value.size()
query, key, val = self.c_attn(value).split(self.n_embd, dim=2)
head_size = channels // self.n_head
query = query.view(batch, time, self.n_head, head_size).transpose(1, 2)
key = key.view(batch, time, self.n_head, head_size).transpose(1, 2)
val = val.view(batch, time, self.n_head, head_size).transpose(1, 2)
attended = F.scaled_dot_product_attention(
query,
key,
val,
attn_mask=None,
dropout_p=self.dropout if self.training else 0,
is_causal=True,
)
attended = attended.transpose(1, 2).contiguous().view(batch, time, channels)
return self.resid_dropout(self.c_proj(attended))
class MLP(nn.Module):
def __init__(self, config: GPTConfig) -> None:
super().__init__()
self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
self.gelu = nn.GELU()
self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
self.dropout = nn.Dropout(config.dropout)
def forward(self, value: torch.Tensor) -> torch.Tensor:
return self.dropout(self.c_proj(self.gelu(self.c_fc(value))))
class Block(nn.Module):
def __init__(self, config: GPTConfig) -> None:
super().__init__()
self.ln_1 = nn.LayerNorm(config.n_embd, bias=config.bias)
self.attn = CausalSelfAttention(config)
self.ln_2 = nn.LayerNorm(config.n_embd, bias=config.bias)
self.mlp = MLP(config)
def forward(self, value: torch.Tensor) -> torch.Tensor:
value = value + self.attn(self.ln_1(value))
return value + self.mlp(self.ln_2(value))
class GPT(nn.Module):
def __init__(self, config: GPTConfig) -> None:
super().__init__()
self.config = config
self.transformer = nn.ModuleDict(
{
"wte": nn.Embedding(config.vocab_size, config.n_embd),
"wpe": nn.Embedding(config.block_size, config.n_embd),
"drop": nn.Dropout(config.dropout),
"h": nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
"ln_f": nn.LayerNorm(config.n_embd, bias=config.bias),
}
)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.transformer.wte.weight = self.lm_head.weight
self.apply(self._init_weights)
for name, parameter in self.named_parameters():
if name.endswith("c_proj.weight"):
torch.nn.init.normal_(
parameter, mean=0.0, std=0.02 / math.sqrt(2 * config.n_layer)
)
@staticmethod
def _init_weights(module: nn.Module) -> None:
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(
self, index: torch.Tensor, targets: torch.Tensor | None = None
) -> tuple[torch.Tensor, torch.Tensor | None]:
_, time = index.shape
if time > self.config.block_size:
raise ValueError("sequence exceeds model block size")
positions = torch.arange(0, time, dtype=torch.long, device=index.device)
value = self.transformer.drop(self.transformer.wte(index) + self.transformer.wpe(positions))
for block in self.transformer.h:
value = block(value)
value = self.transformer.ln_f(value)
logits = self.lm_head(value)
loss = (
F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
if targets is not None
else None
)
return logits, loss
@torch.no_grad()
def generate(
self,
index: torch.Tensor,
max_new_tokens: int,
temperature: float = 0.8,
top_k: int | None = 200,
top_p: float | None = None,
repetition_penalty: float = 1.0,
no_repeat_ngram_size: int | None = None,
) -> torch.Tensor:
"""Sample a continuation.
The defaults reproduce the temperature/top-k-only sampler used for the
phase 1-5 evaluations, so held-out numbers stay comparable. The
additional knobs are opt-in: phase-5 generation samples showed the
low-temperature repetition loop surviving the 345M -> 730M scale-up
(reports/phase5_generation_samples.md), and the sampler had no
repetition control of any kind to blame it on.
"""
for _ in range(max_new_tokens):
cropped = index[:, -self.config.block_size :]
logits, _ = self(cropped)
logits = logits[:, -1, :]
if repetition_penalty != 1.0:
for row, sequence in enumerate(index):
seen = torch.unique(sequence)
scores = logits[row, seen]
logits[row, seen] = torch.where(
scores > 0, scores / repetition_penalty, scores * repetition_penalty
)
logits = logits / temperature
if no_repeat_ngram_size:
for row, sequence in enumerate(index):
for token in self._banned_ngram_tokens(sequence, no_repeat_ngram_size):
logits[row, token] = -float("Inf")
if top_k is not None:
values, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < values[:, [-1]]] = -float("Inf")
if top_p is not None:
ordered, order = torch.sort(logits, descending=True, dim=-1)
ranked = F.softmax(ordered, dim=-1)
# Drop a token once the mass ahead of it already covers top_p,
# which always keeps at least the most likely token.
remove = ranked.cumsum(dim=-1) - ranked >= top_p
logits = logits.masked_fill(
torch.zeros_like(remove).scatter(1, order, remove), -float("Inf")
)
probabilities = F.softmax(logits, dim=-1)
index = torch.cat((index, torch.multinomial(probabilities, num_samples=1)), dim=1)
return index
@staticmethod
def _banned_ngram_tokens(sequence: torch.Tensor, size: int) -> list[int]:
"""Tokens that would repeat an n-gram already present in ``sequence``."""
if size < 2 or len(sequence) < size:
return []
tokens = sequence.tolist()
prefix = tuple(tokens[-(size - 1) :])
banned = [
tokens[start + size - 1]
for start in range(len(tokens) - size + 1)
if tuple(tokens[start : start + size - 1]) == prefix
]
return banned
def parameter_count(self, non_embedding: bool = False) -> int:
count = sum(parameter.numel() for parameter in self.parameters())
if non_embedding:
count -= self.transformer.wpe.weight.numel()
return count
def config_dict(self) -> dict[str, object]:
return asdict(self.config)
|