File size: 18,046 Bytes
34e468d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 | """
Mamba-2 (SSD-style) language model -- "semi-official", mirroring model.mamba.
Like `model.mamba.Mamba`, this keeps a hand-written block (in_proj / depthwise
conv1d / gated RMSNorm / out_proj) but delegates the heavy state-space scan to
the OFFICIAL fused Triton kernel `mamba_chunk_scan_combined` from the
`mamba_ssm` package WHEN it is available and `config.use_cuda=True`. If the
package (mamba_ssm >= 2.2 with a `Mamba2`) is not installed, it transparently
falls back to the self-contained pure-PyTorch chunked SSD scan below, so the
exact same code runs in either environment.
It mirrors the public interface of `model.mamba.Mamba` (forward(idx, targets)
-> (logits, loss), .generate(), .configure_optimizers(), .get_num_params(), and
a `.layers` ModuleList) so it plugs directly into train_maze.py / test_maze.py
/ maze_kstep_detour_test.py.
Key differences from Mamba-1 (model.mamba):
* Multi-head structure: d_inner is split into `nheads` heads of `headdim`.
* The state-transition A is a single scalar PER HEAD (A = -exp(A_log),
A_log shape (nheads,)) instead of a full (d_inner, d_state) matrix. This is
the Mamba-2 "scalar-times-identity" SSD simplification.
* dt (the discretization step) is per-head, shape (B, L, nheads).
* A single depthwise conv is applied jointly to (x, B, C).
* Output is gated-RMSNorm'd with the z branch before the final projection.
The linear recurrence H[t] = a[t] * H[t-1] + X[t] is computed with the
chunked SSD ("state-space duality") algorithm from the Mamba-2 paper. Instead
of materializing the full (B, L, d_inner, d_state) per-timestep state for the
whole sequence (as a naive parallel scan would), the sequence is split into
chunks of `chunk_size`: outputs within a chunk are computed with an
attention-like matmul, and only small chunk-level states (B, nheads, headdim,
d_state) are carried across chunks. This is dramatically faster and lighter on
memory than a full scan, which is the whole point of Mamba-2.
"""
import math
import inspect
from dataclasses import dataclass
from typing import Union
import torch
import torch.nn as nn
import torch.nn.functional as F
# Official fused Triton SSD kernel (mamba_ssm >= 2.2). Optional: if it is not
# installed we fall back to the pure-PyTorch chunked scan in `_ssd` below.
try:
from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined
_HAS_MAMBA2_KERNEL = True
except Exception: # pragma: no cover - import guard
mamba_chunk_scan_combined = None
_HAS_MAMBA2_KERNEL = False
class RMSNorm(nn.Module):
def __init__(self, n_embd: int, eps: float = 1e-5):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(n_embd))
def forward(self, x):
output = (
x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight
)
return output
@dataclass
class Mamba2Config:
n_embd: int # D (model dimension)
n_layer: int
d_state: int = 64 # N: SSM state size per head
expand_factor: int = 2 # E: d_inner = E * D
headdim: int = 64 # P: dimension per head (d_inner must be divisible by it)
ngroups: int = 1 # number of (B, C) groups (1 == shared across all heads)
d_conv: int = 4
chunk_size: int = 64 # SSD chunk length for the chunked scan
vocab_size: int = 64
dt_min: float = 0.001
dt_max: float = 0.1
dt_init_floor: float = 1e-4
A_init_min: float = 1.0 # A_log initialized from U(log(A_init_min), log(A_init_max))
A_init_max: float = 16.0
rms_norm_eps: float = 1e-5
bias: bool = False
conv_bias: bool = True
pscan: bool = True # kept for config compatibility (always parallel scan here)
use_cuda: bool = False # use the official fused Triton kernel when mamba_ssm>=2.2 is installed (auto-fallback to pure PyTorch otherwise)
model_type: str = "mamba2"
pad_id: int = -1
def __post_init__(self):
self.d_inner = self.expand_factor * self.n_embd
assert self.d_inner % self.headdim == 0, (
f"d_inner ({self.d_inner}) must be divisible by headdim ({self.headdim})")
self.nheads = self.d_inner // self.headdim
class RMSNormGated(nn.Module):
"""RMSNorm with a SiLU gate (Mamba-2 output normalization): normalize x * silu(z)."""
def __init__(self, n_embd: int, eps: float = 1e-5):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(n_embd))
def forward(self, x, z):
x = x * F.silu(z)
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight
def _segsum(x):
"""Stable segment-sum: x (..., T) -> (..., T, T) lower-triangular cumulative sums."""
T = x.size(-1)
x = x.unsqueeze(-1).expand(*x.shape, T)
mask = torch.tril(torch.ones(T, T, device=x.device, dtype=torch.bool), diagonal=-1)
x = x.masked_fill(~mask, 0)
x_segsum = torch.cumsum(x, dim=-2)
mask = torch.tril(torch.ones(T, T, device=x.device, dtype=torch.bool), diagonal=0)
x_segsum = x_segsum.masked_fill(~mask, float('-inf'))
return x_segsum
def _ssd(X, A, B, C, chunk_size):
"""Chunked state-space-duality scan (Mamba-2).
Args:
X : (b, l, h, p) input (already scaled by dt)
A : (b, l, h) per-step log-decay (= dt * A_scalar)
B : (b, l, h, n)
C : (b, l, h, n)
chunk_size : int, must divide l
Returns:
Y : (b, l, h, p)
"""
b, l, h, p = X.shape
c = l // chunk_size
X = X.reshape(b, c, chunk_size, h, p)
A = A.reshape(b, c, chunk_size, h).permute(0, 3, 1, 2) # (b, h, c, k)
B = B.reshape(b, c, chunk_size, h, B.size(-1))
C = C.reshape(b, c, chunk_size, h, C.size(-1))
A_cumsum = torch.cumsum(A, dim=-1) # (b, h, c, k)
# 1. intra-chunk (diagonal) outputs
Lmat = torch.exp(_segsum(A)) # (b, h, c, k, k)
Y_diag = torch.einsum("bclhn,bcshn,bhcls,bcshp->bclhp", C, B, Lmat, X)
# 2. each chunk's end state
decay_states = torch.exp(A_cumsum[..., -1:] - A_cumsum) # (b, h, c, k)
states = torch.einsum("bclhn,bhcl,bclhp->bchpn", B, decay_states, X)
# 3. inter-chunk recurrence (scan over chunks only)
init = torch.zeros_like(states[:, :1])
states = torch.cat([init, states], dim=1) # (b, c+1, h, p, n)
decay_chunk = torch.exp(_segsum(F.pad(A_cumsum[..., -1], (1, 0)))) # (b, h, c+1, c+1)
new_states = torch.einsum("bhzc,bchpn->bzhpn", decay_chunk, states)
states = new_states[:, :-1] # (b, c, h, p, n)
# 4. add the contribution of each chunk's initial state to its outputs
state_decay_out = torch.exp(A_cumsum) # (b, h, c, k)
Y_off = torch.einsum("bclhn,bchpn,bhcl->bclhp", C, states, state_decay_out)
return (Y_diag + Y_off).reshape(b, l, h, p)
class Mamba2Block(nn.Module):
def __init__(self, config: Mamba2Config):
super().__init__()
self.config = config
d_inner = config.d_inner
nheads = config.nheads
ngroups = config.ngroups
d_state = config.d_state
# in_proj produces [z, x, B, C, dt]
conv_dim = d_inner + 2 * ngroups * d_state
d_in_proj = 2 * d_inner + 2 * ngroups * d_state + nheads
self.in_proj = nn.Linear(config.n_embd, d_in_proj, bias=config.bias)
# depthwise conv over the (x, B, C) channels
self.conv_dim = conv_dim
self.conv1d = nn.Conv1d(
in_channels=conv_dim,
out_channels=conv_dim,
kernel_size=config.d_conv,
groups=conv_dim,
bias=config.conv_bias,
padding=config.d_conv - 1,
)
# dt bias (per head); softplus(dt + dt_bias) at runtime
dt = torch.exp(
torch.rand(nheads) * (math.log(config.dt_max) - math.log(config.dt_min))
+ math.log(config.dt_min)
).clamp(min=config.dt_init_floor)
inv_dt = dt + torch.log(-torch.expm1(-dt)) # inverse softplus
self.dt_bias = nn.Parameter(inv_dt)
# per-head scalar A (stored in log space to keep A < 0 via A = -exp(A_log))
A = torch.empty(nheads).uniform_(config.A_init_min, config.A_init_max)
self.A_log = nn.Parameter(torch.log(A))
self.A_log._no_weight_decay = True
# per-head skip connection D
self.D = nn.Parameter(torch.ones(nheads))
self.D._no_weight_decay = True
self.norm = RMSNormGated(d_inner, eps=config.rms_norm_eps)
self.out_proj = nn.Linear(d_inner, config.n_embd, bias=config.bias)
# Use the official fused Triton kernel only if requested AND available.
self.use_kernel = bool(config.use_cuda) and _HAS_MAMBA2_KERNEL
def forward(self, u):
# u : (B, L, D)
B_, L, _ = u.shape
cfg = self.config
d_inner, nheads, headdim = cfg.d_inner, cfg.nheads, cfg.headdim
ngroups, d_state = cfg.ngroups, cfg.d_state
zxbcdt = self.in_proj(u) # (B, L, d_in_proj)
z, xBC, dt = torch.split(
zxbcdt, [d_inner, self.conv_dim, nheads], dim=-1)
# depthwise conv (causal) over (x, B, C)
xBC = xBC.transpose(1, 2) # (B, conv_dim, L)
xBC = self.conv1d(xBC)[:, :, :L] # causal: drop the right padding
xBC = xBC.transpose(1, 2) # (B, L, conv_dim)
xBC = F.silu(xBC)
x, Bmat, Cmat = torch.split(
xBC, [d_inner, ngroups * d_state, ngroups * d_state], dim=-1)
A = -torch.exp(self.A_log.float()) # (nheads,)
x = x.view(B_, L, nheads, headdim) # (B, L, H, P)
Bmat = Bmat.view(B_, L, ngroups, d_state)
Cmat = Cmat.view(B_, L, ngroups, d_state)
if self.use_kernel:
# --- official fused Triton SSD kernel ---
# It applies softplus(dt + dt_bias), the chunked scan, and the
# per-head D skip connection internally. B/C keep the ngroups dim.
y = mamba_chunk_scan_combined(
x, dt, A, Bmat, Cmat,
chunk_size=cfg.chunk_size,
D=self.D,
z=None,
dt_bias=self.dt_bias,
dt_softplus=True,
) # (B, L, H, P)
y = y.reshape(B_, L, d_inner).to(z.dtype)
else:
# --- pure-PyTorch chunked SSD scan (float32 for stability) ---
dt = F.softplus(dt + self.dt_bias) # (B, L, nheads)
# heads per group (ngroups==1 -> shared across all heads)
rep = nheads // ngroups
Bh = Bmat.repeat_interleave(rep, dim=2) # (B, L, H, N)
Ch = Cmat.repeat_interleave(rep, dim=2) # (B, L, H, N)
dt_f = dt.float()
X_in = x.float() * dt_f.unsqueeze(-1) # (B, L, H, P)
A_in = A * dt_f # (B, L, H)
Bf = Bh.float()
Cf = Ch.float()
chunk = min(self.config.chunk_size, L)
pad_len = (chunk - L % chunk) % chunk # pad L up to a multiple of chunk
if pad_len:
X_in = F.pad(X_in, (0, 0, 0, 0, 0, pad_len))
A_in = F.pad(A_in, (0, 0, 0, pad_len))
Bf = F.pad(Bf, (0, 0, 0, 0, 0, pad_len))
Cf = F.pad(Cf, (0, 0, 0, 0, 0, pad_len))
y = _ssd(X_in, A_in, Bf, Cf, chunk) # (B, L_pad, H, P)
y = y[:, :L] # drop the padded steps
y = y + self.D.float().view(1, 1, nheads, 1) * x.float() # per-head skip
y = y.reshape(B_, L, d_inner).to(z.dtype) # back to input dtype
y = self.norm(y, z) # gated RMSNorm
return self.out_proj(y) # (B, L, D)
class ResidualBlock2(nn.Module):
def __init__(self, config: Mamba2Config):
super().__init__()
self.mixer = Mamba2Block(config)
self.norm = RMSNorm(config.n_embd, config.rms_norm_eps)
def forward(self, x):
return self.mixer(self.norm(x)) + x
class Mamba2(nn.Module):
def __init__(self, config: Mamba2Config):
super().__init__()
self.config = config
self.embedding = nn.Embedding(config.vocab_size, config.n_embd, padding_idx=0)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.lm_head.weight = self.embedding.weight # weight tying
self.layers = nn.ModuleList([ResidualBlock2(config) for _ in range(config.n_layer)])
self.out_norm = RMSNorm(config.n_embd, config.rms_norm_eps)
self.apply(self._init_weights)
for pn, p in self.named_parameters():
if pn.endswith("out_proj.weight"):
torch.nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * config.n_layer))
print(f"number of parameters: {self.get_num_params() / 1e6:.2f}M")
if config.use_cuda and _HAS_MAMBA2_KERNEL:
print("[mamba2] using official fused Triton kernel (mamba_chunk_scan_combined)")
elif config.use_cuda and not _HAS_MAMBA2_KERNEL:
print("[mamba2] use_cuda=True but mamba_ssm kernel not found -> "
"falling back to pure-PyTorch chunked SSD")
else:
print("[mamba2] using pure-PyTorch chunked SSD scan")
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):
x = self.embedding(idx)
for layer in self.layers:
x = layer(x)
x = self.out_norm(x)
if targets is not None:
logits = self.lm_head(x)
loss = F.cross_entropy(
logits.view(-1, logits.size(-1)),
targets.view(-1),
ignore_index=self.config.pad_id,
)
else:
logits = self.lm_head(x[:, [-1], :])
loss = None
return logits, loss
@torch.no_grad()
def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None, return_confidence=False):
"""Autoregressive generation matching model.mamba.Mamba.generate's contract."""
confidences = [] if return_confidence else None
top3_tokens = [] if return_confidence else None
top3_probs = [] if return_confidence else None
B = idx.size(0)
for _ in range(max_new_tokens):
logits, _ = self(idx)
if temperature <= 0:
probs = F.softmax(logits[:, -1, :], dim=-1)
idx_next = probs.argmax(dim=-1, keepdim=True)
else:
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)
if return_confidence:
sampled_probs = probs.gather(1, idx_next).squeeze(-1)
confidences.append(sampled_probs.cpu().tolist())
top3_prob_vals, top3_token_ids = torch.topk(probs, 3, dim=-1)
top3_tokens.append(top3_token_ids.cpu().tolist())
top3_probs.append(top3_prob_vals.cpu().tolist())
idx = torch.cat((idx, idx_next), dim=1)
if return_confidence:
if B == 1:
return (idx,
[c[0] for c in confidences],
[t[0] for t in top3_tokens],
[p[0] for p in top3_probs])
T = len(confidences)
conf_bs = [[confidences[t][b] for t in range(T)] for b in range(B)]
tok_bs = [[top3_tokens[t][b] for t in range(T)] for b in range(B)]
prob_bs = [[top3_probs[t][b] for t in range(T)] for b in range(B)]
return idx, conf_bs, tok_bs, prob_bs
return idx
def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
param_dict = {pn: p for pn, p in self.named_parameters() if p.requires_grad}
decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
optim_groups = [
{"params": decay_params, "weight_decay": weight_decay},
{"params": nodecay_params, "weight_decay": 0.0},
]
num_decay = sum(p.numel() for p in decay_params)
num_nodecay = sum(p.numel() for p in nodecay_params)
print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay:,} parameters")
print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay:,} parameters")
fused_available = "fused" in inspect.signature(torch.optim.AdamW).parameters
use_fused = fused_available and device_type == "cuda"
extra_args = dict(fused=True) if use_fused else {}
optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, **extra_args)
print(f"using fused AdamW: {use_fused}")
return optimizer
def estimate_mfu(self, fwdbwd_per_iter, dt):
return -1
def get_num_params(self, non_embedding=True):
n_params = sum(p.numel() for p in self.parameters())
if non_embedding:
n_params -= self.embedding.weight.numel()
return n_params
|