File size: 8,082 Bytes
d993048 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 | """
ArmGPT Model - A modern GPT with RMSNorm, SwiGLU, and RoPE.
Architecture:
1. Token Embedding: convert token IDs to vectors
2. RoPE: rotary position embeddings (no learned position table)
3. Transformer Blocks: RMSNorm + Attention + SwiGLU MLP
4. Output Head: predict the next token
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class RMSNorm(nn.Module):
"""Root Mean Square Layer Normalization — faster than LayerNorm, no bias."""
def __init__(self, dim, eps=1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x):
rms = torch.rsqrt(x.float().pow(2).mean(-1, keepdim=True) + self.eps)
return (x.float() * rms).type_as(x) * self.weight
def precompute_rope(dim, max_seq_len, theta=10000.0):
"""Precompute rotary position embedding frequencies."""
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
t = torch.arange(max_seq_len).float()
freqs = torch.outer(t, freqs)
cos = freqs.cos()
sin = freqs.sin()
return cos, sin
def apply_rope(x, cos, sin):
"""Apply rotary position embeddings to query/key tensors."""
B, n_head, T, head_dim = x.shape
cos = cos[:T].unsqueeze(0).unsqueeze(0) # (1, 1, T, head_dim//2)
sin = sin[:T].unsqueeze(0).unsqueeze(0)
# Split into pairs and rotate
x1 = x[..., :head_dim // 2]
x2 = x[..., head_dim // 2:]
return torch.cat([x1 * cos - x2 * sin, x2 * cos + x1 * sin], dim=-1)
class CausalSelfAttention(nn.Module):
"""Self-attention with RoPE (no causal mask buffer needed — using F.scaled_dot_product_attention)."""
def __init__(self, n_embd, n_head, block_size, dropout):
super().__init__()
assert n_embd % n_head == 0
self.c_attn = nn.Linear(n_embd, 3 * n_embd, bias=False)
self.c_proj = nn.Linear(n_embd, n_embd, bias=False)
self.n_head = n_head
self.n_embd = n_embd
self.head_dim = n_embd // n_head
self.dropout = dropout
# Precompute RoPE
cos, sin = precompute_rope(self.head_dim, block_size)
self.register_buffer("rope_cos", cos)
self.register_buffer("rope_sin", sin)
def forward(self, x):
B, T, C = x.size()
qkv = self.c_attn(x)
q, k, v = qkv.split(self.n_embd, dim=2)
q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
# Apply RoPE to queries and keys
q = apply_rope(q, self.rope_cos, self.rope_sin)
k = apply_rope(k, self.rope_cos, self.rope_sin)
# Use PyTorch's efficient attention (handles causal mask internally)
y = F.scaled_dot_product_attention(
q, k, v,
is_causal=True,
dropout_p=self.dropout if self.training else 0.0,
)
y = y.transpose(1, 2).contiguous().view(B, T, C)
y = self.c_proj(y)
return y
class SwiGLUMLP(nn.Module):
"""SwiGLU feed-forward network — better than GELU, used by LLaMA/Mistral."""
def __init__(self, n_embd, dropout):
super().__init__()
# SwiGLU uses 8/3 * n_embd hidden dim (rounded to multiple of 64 for efficiency)
hidden = int(8 / 3 * n_embd)
hidden = ((hidden + 63) // 64) * 64 # round up to multiple of 64
self.w1 = nn.Linear(n_embd, hidden, bias=False) # gate
self.w2 = nn.Linear(hidden, n_embd, bias=False) # down
self.w3 = nn.Linear(n_embd, hidden, bias=False) # up
self.dropout = nn.Dropout(dropout)
def forward(self, x):
return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x)))
class Block(nn.Module):
"""Transformer block: RMSNorm + Attention + SwiGLU MLP."""
def __init__(self, n_embd, n_head, block_size, dropout):
super().__init__()
self.ln_1 = RMSNorm(n_embd)
self.attn = CausalSelfAttention(n_embd, n_head, block_size, dropout)
self.ln_2 = RMSNorm(n_embd)
self.mlp = SwiGLUMLP(n_embd, dropout)
def forward(self, x):
x = x + self.attn(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
class GPT(nn.Module):
"""GPT language model with RMSNorm, RoPE, and SwiGLU."""
def __init__(self, vocab_size, n_layer, n_head, n_embd, block_size, dropout):
super().__init__()
self.block_size = block_size
self.transformer = nn.ModuleDict(dict(
wte=nn.Embedding(vocab_size, n_embd),
drop=nn.Dropout(dropout),
blocks=nn.ModuleList([
Block(n_embd, n_head, block_size, dropout)
for _ in range(n_layer)
]),
ln_f=RMSNorm(n_embd),
))
self.lm_head = nn.Linear(n_embd, vocab_size, bias=False)
self.transformer.wte.weight = self.lm_head.weight
self.apply(self._init_weights)
n_params = sum(p.numel() for p in self.parameters())
print(f"GPT model initialized: {n_params:,} parameters")
def _init_weights(self, module):
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, idx, targets=None):
B, T = idx.size()
assert T <= self.block_size, f"Sequence length {T} exceeds block_size {self.block_size}"
# Token embeddings only — RoPE handles positions inside attention
x = self.transformer.drop(self.transformer.wte(idx))
for block in self.transformer.blocks:
x = block(x)
x = self.transformer.ln_f(x)
logits = self.lm_head(x)
loss = None
if targets is not None:
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
return logits, loss
@torch.no_grad()
def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None,
stop_tokens=None, repetition_penalty=1.0):
"""Generate tokens autoregressively.
Args:
repetition_penalty: 1.0 = no penalty (off). >1.0 discourages
repeating tokens already in the context (CTRL-style penalty).
Typical values: 1.1–1.3. Helps small LMs escape repetition loops.
"""
for _ in range(max_new_tokens):
idx_cond = idx[:, -self.block_size:]
logits, _ = self(idx_cond)
logits = logits[:, -1, :]
# CTRL-style repetition penalty: scale logits for tokens already seen.
# Positive logits get divided (made smaller); negative logits get
# multiplied (made more negative). Applied before temperature.
if repetition_penalty != 1.0:
seen = torch.unique(idx_cond)
seen_logits = logits[:, seen]
seen_logits = torch.where(
seen_logits > 0,
seen_logits / repetition_penalty,
seen_logits * repetition_penalty,
)
logits[:, seen] = seen_logits
logits = logits / temperature
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < v[:, [-1]]] = float("-inf")
probs = F.softmax(logits, dim=-1)
idx_next = torch.multinomial(probs, num_samples=1)
idx = torch.cat((idx, idx_next), dim=1)
if stop_tokens and idx_next.item() in stop_tokens:
break
return idx
|