File size: 9,955 Bytes
1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 be1bcbb 1973cf0 d7d1235 be1bcbb 1973cf0 d7d1235 1973cf0 be1bcbb 1973cf0 be1bcbb 1973cf0 be1bcbb 1973cf0 d7d1235 be1bcbb d7d1235 1973cf0 d7d1235 be1bcbb 1973cf0 d7d1235 be1bcbb 1973cf0 be1bcbb 1973cf0 be1bcbb d7d1235 be1bcbb d7d1235 be1bcbb d7d1235 be1bcbb d7d1235 1973cf0 d7d1235 be1bcbb d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 be1bcbb d7d1235 1973cf0 be1bcbb 1973cf0 d7d1235 1973cf0 d7d1235 be1bcbb d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 be1bcbb d7d1235 be1bcbb d7d1235 be1bcbb d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 1973cf0 d7d1235 be1bcbb 1973cf0 be1bcbb d7d1235 1973cf0 be1bcbb 1973cf0 d7d1235 1973cf0 d7d1235 be1bcbb d7d1235 1973cf0 be1bcbb d7d1235 be1bcbb | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | """
Mamba-2 SSD — OPTIMIZED: intra-chunk parallelism via matrix multiply.
The key Mamba-2 insight (State Space Duality):
Within each chunk of size T, the SSM can be computed as a MATRIX MULTIPLY:
Y_chunk = (L ⊙ (C B^T)) @ (Δ ⊙ X)
Where L is a lower-triangular mask with cumulative A products.
This replaces the T sequential steps with a single matmul of size T×T.
For L=256, T=16, num_chunks=16:
- Within chunk: parallel matmul (T×T = 16×16)
- Across chunks: 16 sequential state carries (unavoidable, but trivial)
Total: 16 sequential state carries + 16 parallel matmuls = FAST.
NO in-place ops. Fully autograd safe. Works on CPU and GPU.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
class Mamba2SSD(nn.Module):
"""
Mamba-2 SSD with intra-chunk matrix-multiply parallelism.
Args:
dim: Input/output dimension
d_state: SSM state dimension (default 16)
d_conv: Conv1d kernel size (default 4)
expand: Inner dimension expansion (default 2)
chunk_size: Chunk size for scan (default 64 — larger = more parallel)
"""
def __init__(self, dim, d_state=16, d_conv=4, expand=2, chunk_size=64):
super().__init__()
self.dim = dim
self.d_state = d_state
self.chunk_size = chunk_size
self.inner_dim = dim * expand
# Input projection: x and gate
self.in_proj = nn.Linear(dim, self.inner_dim * 2, bias=False)
# Short causal conv for local context
self.conv1d = nn.Conv1d(
self.inner_dim, self.inner_dim,
kernel_size=d_conv, padding=d_conv - 1,
groups=self.inner_dim, bias=True
)
# SSM parameter projections
self.dt_proj = nn.Linear(self.inner_dim, self.inner_dim, bias=True)
self.B_proj = nn.Linear(self.inner_dim, d_state, bias=False)
self.C_proj = nn.Linear(self.inner_dim, d_state, bias=False)
# A: fixed decay rates (log-space, negative for stability)
A = torch.arange(1, d_state + 1, dtype=torch.float32)
self.A_log = nn.Parameter(torch.log(A))
# D: residual skip
self.D = nn.Parameter(torch.ones(self.inner_dim))
# Output
self.norm = nn.LayerNorm(self.inner_dim)
self.out_proj = nn.Linear(self.inner_dim, dim, bias=False)
self._init_weights()
def _init_weights(self):
nn.init.constant_(self.dt_proj.bias, -4.0) # softplus(-4) ≈ 0.018
nn.init.xavier_uniform_(self.in_proj.weight, gain=0.1)
nn.init.xavier_uniform_(self.out_proj.weight, gain=0.1)
def forward(self, x):
"""x: [B, L, dim] → [B, L, dim]"""
return self._process(x)
def _process(self, x):
B, L, D = x.shape
# Input projection
xz = self.in_proj(x)
x_inner, z = xz.chunk(2, dim=-1)
# Causal conv
x_conv = self.conv1d(x_inner.transpose(1, 2))[:, :, :L].transpose(1, 2)
x_conv = F.silu(x_conv)
# SSM params
dt = F.softplus(self.dt_proj(x_conv)) # [B, L, inner_dim], positive
B_mat = self.B_proj(x_conv) # [B, L, d_state]
C_mat = self.C_proj(x_conv) # [B, L, d_state]
A = -torch.exp(self.A_log) # [d_state], negative
# Chunk-parallel scan
y = self._chunk_ssm(x_conv, dt, A, B_mat, C_mat)
# Skip + norm + gate
y = y + x_conv * self.D.unsqueeze(0).unsqueeze(0)
y = self.norm(y) * F.silu(z)
return self.out_proj(y)
def _chunk_ssm(self, u, dt, A, B, C):
"""
Chunk-parallel SSM computation.
Within each chunk: compute via cumulative decay matrix (parallel).
Across chunks: propagate final state (sequential, only num_chunks steps).
The intra-chunk computation uses the identity:
h_t = sum_{s=0}^{t} (prod_{k=s+1}^{t} dA_k) * dB_s * u_s
This is a lower-triangular matrix-vector product, computable in parallel.
"""
batch, L, d_inner = u.shape
d_state = A.shape[0]
T = min(self.chunk_size, L)
# Pad to multiple of T
pad = (T - L % T) % T
if pad > 0:
u = F.pad(u, (0, 0, 0, pad))
dt = F.pad(dt, (0, 0, 0, pad))
B = F.pad(B, (0, 0, 0, pad))
C = F.pad(C, (0, 0, 0, pad))
L_pad = u.shape[1]
n_chunks = L_pad // T
# Reshape: [B, n_chunks, T, ...]
u_c = u.reshape(batch, n_chunks, T, d_inner)
dt_c = dt.reshape(batch, n_chunks, T, d_inner)
B_c = B.reshape(batch, n_chunks, T, d_state)
C_c = C.reshape(batch, n_chunks, T, d_state)
# Mean dt per position for state decay (simplification for scalar-A)
dt_mean = dt_c.mean(dim=-1) # [B, n_chunks, T]
# Compute log(dA) per position: log_dA = dt_mean * A
# A is [d_state], dt_mean is [B, nc, T]
log_dA = dt_mean.unsqueeze(-1) * A.unsqueeze(0).unsqueeze(0).unsqueeze(0)
# log_dA: [B, nc, T, d_state]
# Cumulative sum for decay within chunk: cumsum along T dimension
# For position t, decay from position s is: exp(sum_{k=s+1}^{t} log_dA_k)
log_dA_cumsum = torch.cumsum(log_dA, dim=2) # [B, nc, T, d_state]
# Lower-triangular decay matrix: L[t,s] = exp(cumsum[t] - cumsum[s])
# L[t,s,n] = exp(sum_{k=s+1}^{t} log_dA_k_n) for t >= s, else 0
# Shape: [B, nc, T, T, d_state]
decay_matrix = log_dA_cumsum.unsqueeze(3) - log_dA_cumsum.unsqueeze(2)
# decay_matrix[..., t, s, :] = cumsum[t] - cumsum[s]
# Apply causal mask (t >= s only)
causal_mask = torch.tril(torch.ones(T, T, device=u.device)) # [T, T]
decay_matrix = decay_matrix * causal_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1)
decay_matrix = torch.exp(decay_matrix) * causal_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1)
# [B, nc, T, T, d_state]
# Compute dBu: dt * B * u → state input at each position
# dt_c: [B, nc, T, d_inner], B_c: [B, nc, T, d_state], u_c: [B, nc, T, d_inner]
# We need [B, nc, T, d_state, d_inner]
dBu = dt_c.unsqueeze(-2) * B_c.unsqueeze(-1) * u_c.unsqueeze(-2)
# dBu: [B, nc, T, d_state, d_inner]
# Intra-chunk SSM via matrix multiply:
# h[t] = sum_s decay[t,s] * dBu[s]
# h: [B, nc, T, d_state, d_inner]
# decay_matrix: [B, nc, T, T, d_state]
# dBu: [B, nc, T, d_state, d_inner]
# Einsum: h[b,c,t,n,d] = sum_s decay[b,c,t,s,n] * dBu[b,c,s,n,d]
h_intra = torch.einsum('bctsn,bcsnd->bctnd', decay_matrix, dBu)
# h_intra: [B, nc, T, d_state, d_inner]
# Inter-chunk state propagation
# Decay of previous chunk's final state into current chunk
# Total decay for a full chunk: exp(sum of all T log_dA values)
chunk_decay = torch.exp(log_dA_cumsum[:, :, -1, :]) # [B, nc, d_state]
# Decay from chunk start to each position within chunk:
# position_decay[t] = exp(cumsum[t]) (from position 0)
position_decay = torch.exp(log_dA_cumsum) # [B, nc, T, d_state]
# Propagate states across chunks
h_carry = torch.zeros(batch, d_state, d_inner, device=u.device)
h_chunks = []
for c_idx in range(n_chunks):
# Decay carry state to each position in this chunk
# h_from_prev[t] = position_decay[t] * h_carry
h_from_prev = position_decay[:, c_idx, :, :].unsqueeze(-1) * h_carry.unsqueeze(1)
# h_from_prev: [B, T, d_state, d_inner]
# Total hidden state
h_total = h_intra[:, c_idx] + h_from_prev # [B, T, d_state, d_inner]
h_chunks.append(h_total)
# Update carry: final state of this chunk
h_carry = h_total[:, -1, :, :] # [B, d_state, d_inner]
# Stack chunks: [B, nc, T, d_state, d_inner]
h_all = torch.stack(h_chunks, dim=1)
# Output: y[t] = C[t]^T @ h[t]
# C_c: [B, nc, T, d_state], h_all: [B, nc, T, d_state, d_inner]
y = torch.einsum('bctn,bctnd->bctd', C_c, h_all)
# y: [B, nc, T, d_inner]
# Reshape back
y = y.reshape(batch, L_pad, d_inner)
return y[:, :L, :]
class Mamba2Block(nn.Module):
"""
Mamba-2 block with bidirectional scanning for 2D images.
Forward + backward raster scan, merged via learned projection.
"""
def __init__(self, dim, d_state=16, d_conv=4, expand=2, dropout=0.0):
super().__init__()
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.ssd_fwd = Mamba2SSD(dim, d_state, d_conv, expand)
self.ssd_bwd = Mamba2SSD(dim, d_state, d_conv, expand)
self.merge = nn.Linear(dim * 2, dim, bias=False)
ff_dim = dim * expand
self.ff = nn.Sequential(
nn.Linear(dim, ff_dim), nn.GELU(), nn.Dropout(dropout),
nn.Linear(ff_dim, dim), nn.Dropout(dropout),
)
def forward(self, x):
"""x: [B, C, H, W] or [B, L, C]"""
is_2d = x.dim() == 4
if is_2d:
B, C, H, W = x.shape
x = x.flatten(2).transpose(1, 2)
residual = x
x_norm = self.norm1(x)
fwd = self.ssd_fwd(x_norm)
bwd = torch.flip(self.ssd_bwd(torch.flip(x_norm, [1])), [1])
merged = self.merge(torch.cat([fwd, bwd], dim=-1))
x = residual + merged
x = x + self.ff(self.norm2(x))
if is_2d:
x = x.transpose(1, 2).reshape(B, C, H, W)
return x
|