File size: 7,688 Bytes
9eb0412 | 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 | """
MicroGLM: Tiny GLM-style model in PyTorch.
GLM (General Language Model) uses a prefix-LM architecture:
bidirectional attention on a prefix span + autoregressive generation on the rest.
This is a minimal implementation that fits in 6GB VRAM.
Reference: "GLM: General Language Model Pretraining with Autoregressive Blank Infilling" (Du et al., 2021)
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class RMSNorm(nn.Module):
def __init__(self, dim, eps=1e-5):
super().__init__()
self.weight = nn.Parameter(torch.ones(dim))
self.eps = eps
def forward(self, x):
return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps) * self.weight
class GLMAttention(nn.Module):
"""
Attention with support for 2D attention masks (for prefix-LM).
The mask has shape (B, 1, T, T) where:
- 0 means attend (bidirectional / causal allowed)
- -inf means blocked
"""
def __init__(self, n_embd, n_head, block_size, dropout):
super().__init__()
assert n_embd % n_head == 0
self.n_head = n_head
self.head_dim = n_embd // n_head
self.qkv = nn.Linear(n_embd, n_embd * 3, bias=False)
self.proj = nn.Linear(n_embd, n_embd, bias=False)
self.attn_drop = nn.Dropout(dropout)
self.resid_drop = nn.Dropout(dropout)
def forward(self, x, attn_mask=None):
"""
x: (B, T, C)
attn_mask: (B, 1, T, T) or None (uses default causal mask)
"""
B, T, C = x.shape
qkv = self.qkv(x).reshape(B, T, 3, self.n_head, self.head_dim).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
att = (q @ k.transpose(-2, -1)) / math.sqrt(self.head_dim)
if attn_mask is not None:
# attn_mask: 0 = attend, -inf = block
att = att + attn_mask
else:
# Default causal mask
mask = torch.tril(torch.ones(T, T, device=x.device)).view(1, 1, T, T)
att = att.masked_fill(mask == 0, float('-inf'))
att = F.softmax(att, dim=-1)
att = self.attn_drop(att)
y = att @ v
y = y.transpose(1, 2).contiguous().view(B, T, C)
return self.resid_drop(self.proj(y))
class GLMMLP(nn.Module):
def __init__(self, n_embd, dropout):
super().__init__()
self.fc = nn.Linear(n_embd, 4 * n_embd, bias=False)
self.proj = nn.Linear(4 * n_embd, n_embd, bias=False)
self.drop = nn.Dropout(dropout)
def forward(self, x):
x = F.gelu(self.fc(x))
x = self.proj(x)
return self.drop(x)
class GLMBlock(nn.Module):
def __init__(self, n_embd, n_head, block_size, dropout):
super().__init__()
self.ln1 = RMSNorm(n_embd)
self.attn = GLMAttention(n_embd, n_head, block_size, dropout)
self.ln2 = RMSNorm(n_embd)
self.mlp = GLMMLP(n_embd, dropout)
def forward(self, x, attn_mask=None):
x = x + self.attn(self.ln1(x), attn_mask)
x = x + self.mlp(self.ln2(x))
return x
class MicroGLM(nn.Module):
"""
Micro GLM: Prefix-LM decoder.
Supports two attention modes:
- Causal LM (default): standard autoregressive generation
- Prefix LM: bidirectional attention on first `prefix_len` tokens, causal on the rest
This is controlled by passing a custom attention mask during forward().
For training, we create a 2D mask where the prefix region is bidirectional
and the suffix region is causal.
"""
def __init__(self, vocab_size, block_size, n_layer=2, n_head=4, n_embd=128, dropout=0.1):
super().__init__()
self.block_size = block_size
self.wte = nn.Embedding(vocab_size, n_embd)
self.wpe = nn.Embedding(block_size, n_embd)
self.blocks = nn.ModuleList([GLMBlock(n_embd, n_head, block_size, dropout) for _ in range(n_layer)])
self.ln_f = RMSNorm(n_embd)
self.lm_head = nn.Linear(n_embd, vocab_size, bias=False)
self.lm_head.weight = self.wte.weight
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def _build_prefix_lm_mask(self, T, prefix_len, device):
"""
Build a 2D attention mask for prefix-LM.
- For positions i < prefix_len and j < prefix_len: bidirectional (0)
- For positions i >= prefix_len: causal (attend only to j <= i)
- All other positions: -inf
Returns: (1, 1, T, T) mask where 0 = allowed, -inf = blocked
"""
# Start with causal mask
mask = torch.tril(torch.ones(T, T, device=device))
# Set the prefix block to be fully connected (bidirectional)
mask[:, :prefix_len] = 1.0
# Convert to float mask: 0 = attend, -inf = blocked
mask = mask.view(1, 1, T, T)
mask = mask.masked_fill(mask == 0, float('-inf'))
mask = mask.masked_fill(mask == 1.0, 0.0)
return mask
def forward(self, idx, targets=None, prefix_len=0):
"""
idx: (B, T) input tokens
targets: (B, T) target tokens (shifted for loss)
prefix_len: number of prefix tokens to use bidirectional attention
"""
B, T = idx.shape
if T > self.block_size:
raise ValueError(f'block size exceeded: {T} > {self.block_size}')
pos = torch.arange(T, device=idx.device)
x = self.wte(idx) + self.wpe(pos)
if prefix_len > 0:
attn_mask = self._build_prefix_lm_mask(T, prefix_len, idx.device)
else:
attn_mask = None
for block in self.blocks:
x = block(x, attn_mask)
x = self.ln_f(x)
logits = self.lm_head(x)
loss = None
if targets is not None:
loss = F.cross_entropy(logits.reshape(-1, logits.size(-1)), targets.reshape(-1))
return logits, loss
@torch.no_grad()
def generate(self, idx, max_new_tokens, temperature=1.0, top_k=40):
"""Standard causal generation (no prefix)."""
self.eval()
for _ in range(max_new_tokens):
idx_cond = idx[:, -self.block_size:]
logits, _ = self(idx_cond, prefix_len=0)
logits = logits[:, -1, :] / 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)
return idx
@torch.no_grad()
def generate_with_prefix(self, idx, prefix_len, max_new_tokens, temperature=1.0, top_k=40):
"""
Generate tokens where the first `prefix_len` tokens use bidirectional attention
and the rest are autoregressive.
"""
self.eval()
for _ in range(max_new_tokens):
idx_cond = idx[:, -self.block_size:]
T = idx_cond.shape[1]
prefix = min(prefix_len, T)
logits, _ = self(idx_cond, prefix_len=prefix)
logits = logits[:, -1, :] / 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)
return idx |