File size: 8,041 Bytes
2ca7d54 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 | """
Resonance 200M — Content + RRPRAM dual attention transformer.
Low-rank RRPRAM (Wr = Wr_a @ Wr_b), SwiGLU MLP, RMSNorm, RoPE.
Content attention uses FlashAttention via F.scaled_dot_product_attention.
Architecture matches resonance-bpe.c (with low-rank extension).
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
class RMSNorm(nn.Module):
def __init__(self, dim, eps=1e-5):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight
class ResonanceBlock(nn.Module):
"""
Dual attention block: Content (QKV + RoPE + FlashAttn) + RRPRAM (low-rank Wr) + SwiGLU MLP.
"""
def __init__(self, config):
super().__init__()
E = config['n_embd']
H = config['n_head']
D = config['head_dim']
R = config['rrpram_rank']
T = config['context_len']
M = config['ffn_dim']
self.n_head = H
self.head_dim = D
self.n_embd = E
# Pre-attention norm
self.norm1 = RMSNorm(E)
# Content attention (MHA): Q, K, V
self.wq = nn.Linear(E, H * D, bias=False)
self.wk = nn.Linear(E, H * D, bias=False)
self.wv = nn.Linear(E, H * D, bias=False)
# RRPRAM attention (low-rank): Wr_a[H, E, R] @ Wr_b[H, R, T] = Wr[H, E, T]
self.wr_a = nn.Parameter(torch.randn(H, E, R) * (2.0 / E) ** 0.5)
self.wr_b = nn.Parameter(torch.randn(H, R, T) * (2.0 / R) ** 0.5)
# Per-head gate: sigmoid(gate) blends content vs RRPRAM
self.gate = nn.Parameter(torch.zeros(H)) # init 0 → sigmoid(0) = 0.5 = balanced
# Output projection
self.wo = nn.Linear(E, E, bias=False)
# Pre-MLP norm
self.norm2 = RMSNorm(E)
# SwiGLU MLP
self.mlp_gate = nn.Linear(E, M, bias=False)
self.mlp_up = nn.Linear(E, M, bias=False)
self.mlp_down = nn.Linear(M, E, bias=False)
# Init output projections with smaller std (GPT-2 convention)
n_layer = config['n_layer']
nn.init.normal_(self.wo.weight, std=0.02 / math.sqrt(2 * n_layer))
nn.init.normal_(self.mlp_down.weight, std=0.02 / math.sqrt(2 * n_layer))
def forward(self, x, rope_cos, rope_sin, mask):
B, T, E = x.shape
H = self.n_head
D = self.head_dim
# Pre-norm
xn = self.norm1(x)
# === Content attention with RoPE + FlashAttention ===
q = self.wq(xn).view(B, T, H, D).transpose(1, 2) # [B, H, T, D]
k = self.wk(xn).view(B, T, H, D).transpose(1, 2)
v = self.wv(xn).view(B, T, H, D).transpose(1, 2)
# Apply RoPE to Q and K
q = _apply_rope(q, rope_cos, rope_sin)
k = _apply_rope(k, rope_cos, rope_sin)
# FlashAttention — O(T) memory instead of O(T²)
c_out = F.scaled_dot_product_attention(q, k, v, is_causal=True) # [B, H, T, D]
# === RRPRAM attention (low-rank) ===
# Wr = Wr_a @ Wr_b: [H, E, R] @ [H, R, T] = [H, E, T]
# Score: xn @ Wr → [B, T, E] @ [H, E, T] → [B, H, T, T]
xn_h = xn.unsqueeze(1).expand(-1, H, -1, -1) # [B, H, T, E]
# Low-rank: (xn @ Wr_a) @ Wr_b
temp = torch.einsum('bhie,her->bhir', xn_h, self.wr_a) # [B, H, T, R]
r_attn = torch.einsum('bhir,hrj->bhij', temp, self.wr_b[:, :, :T]) # [B, H, T, T]
r_attn = r_attn * (D ** -0.5)
r_attn = r_attn.masked_fill(mask, float('-inf'))
r_attn = F.softmax(r_attn, dim=-1)
r_out = r_attn @ v # [B, H, T, D] — shared V with content
# === Gate: blend content and RRPRAM ===
g = torch.sigmoid(self.gate).view(1, H, 1, 1) # [1, H, 1, 1]
attn_out = g * c_out + (1 - g) * r_out # [B, H, T, D]
# Output projection + residual
attn_out = attn_out.transpose(1, 2).contiguous().view(B, T, E)
x = x + self.wo(attn_out)
# === SwiGLU MLP ===
xn = self.norm2(x)
gate = F.silu(self.mlp_gate(xn))
up = self.mlp_up(xn)
x = x + self.mlp_down(gate * up)
return x
def _apply_rope(x, cos, sin):
"""Apply RoPE to tensor x: [B, H, T, D]."""
x1 = x[..., ::2] # even dims
x2 = x[..., 1::2] # odd dims
out = torch.stack([
x1 * cos - x2 * sin,
x1 * sin + x2 * cos,
], dim=-1).flatten(-2)
return out
class Resonance(nn.Module):
"""
Resonance 200M: dual attention (Content + RRPRAM) transformer.
"""
def __init__(self, config):
super().__init__()
self.config = config
V = config['vocab_size']
E = config['n_embd']
T = config['context_len']
D = config['head_dim']
# Token embedding (no position — RoPE handles it)
self.tok_emb = nn.Embedding(V, E)
nn.init.normal_(self.tok_emb.weight, std=0.02)
# Transformer blocks
self.blocks = nn.ModuleList([
ResonanceBlock(config) for _ in range(config['n_layer'])
])
# Final norm + output head (untied from embedding)
self.norm_f = RMSNorm(E)
self.out_head = nn.Linear(E, V, bias=False)
nn.init.normal_(self.out_head.weight, std=0.02)
# Precompute RoPE
freqs = 1.0 / (10000.0 ** (torch.arange(0, D, 2).float() / D))
t = torch.arange(T).float()
angles = torch.outer(t, freqs)
self.register_buffer('rope_cos', angles.cos().unsqueeze(0).unsqueeze(0)) # [1,1,T,D//2]
self.register_buffer('rope_sin', angles.sin().unsqueeze(0).unsqueeze(0))
# Causal mask (for RRPRAM — content uses is_causal=True in SDPA)
mask = torch.triu(torch.ones(T, T, dtype=torch.bool), diagonal=1)
self.register_buffer('causal_mask', mask)
n_params = sum(p.numel() for p in self.parameters())
print(f" [Resonance] {n_params:,} parameters")
self._report_balance()
def _report_balance(self):
"""Report parameter budget distribution."""
cfg = self.config
E, H, D = cfg['n_embd'], cfg['n_head'], cfg['head_dim']
R, T, M = cfg['rrpram_rank'], cfg['context_len'], cfg['ffn_dim']
V, L = cfg['vocab_size'], cfg['n_layer']
emb = V * E * 2 # tok_emb + out_head (untied)
qkv = L * (3 * E * H * D)
rrpram = L * (H * E * R + H * R * T + H) # wr_a + wr_b + gate
wo = L * E * E
mlp = L * (3 * E * M)
norms = L * 2 * E + E # per-block norms + final
total = emb + qkv + rrpram + wo + mlp + norms
print(f" [Resonance] Budget: emb={emb/total*100:.1f}% qkv={qkv/total*100:.1f}% "
f"rrpram={rrpram/total*100:.1f}% wo={wo/total*100:.1f}% "
f"mlp={mlp/total*100:.1f}% norms={norms/total*100:.1f}%")
def set_gradient_checkpointing(self, enable=True):
self._grad_ckpt = enable
def forward(self, idx, targets=None):
B, T = idx.shape
x = self.tok_emb(idx)
cos = self.rope_cos[:, :, :T, :]
sin = self.rope_sin[:, :, :T, :]
mask = self.causal_mask[:T, :T]
for block in self.blocks:
if getattr(self, '_grad_ckpt', False) and self.training:
x = torch.utils.checkpoint.checkpoint(
block, x, cos, sin, mask, use_reentrant=False)
else:
x = block(x, cos, sin, mask)
logits = self.out_head(self.norm_f(x))
loss = None
if targets is not None:
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
return logits, loss
# === Default config: ~200M params ===
RESONANCE_200M = {
'n_embd': 768,
'n_head': 12,
'head_dim': 64, # n_embd // n_head
'n_layer': 20,
'rrpram_rank': 48, # low-rank R
'context_len': 2048,
'ffn_dim': 2048, # round(8*768/3, 256)
'vocab_size': 16384, # 256 + 16128 BPE merges
}
|