File size: 18,857 Bytes
b428368 | 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 | """Reference Toto2-style backbone β a patch transformer trained with
contiguous patch masking (CPM) and a multi-quantile head, from random init.
This module is **self-contained torch** and is *copied into every checkpoint*
(as ``model.py``) so the validator's ``forecast_wrapper.py`` can rebuild the
exact architecture to load the weights. Keep it dependency-light (torch only)
and free of cascade imports for that reason.
It follows the Toto 2.0 recipe (arXiv:2605.20119):
* **CPM** β a per-entry binary mask channel; training masks contiguous spans,
inference fills the horizon with mask patches and decodes it in **one
forward pass** (no autoregressive sampling).
* **Grouped time/variate attention** β the last layer of each group of 4
attends over variates (full), the rest over time (causal, rotary positions);
this matches ``Datadog/Toto-2.0-4m``'s ``layer_group_size=4`` /
``num_variate_layers_per_group=1`` / ``variate_layer_first=false``. cascade
currently trains and scores univariate (``OPEN_QUESTIONS.md`` Β§8), so the
variate layers run at ``C = 1`` β present and trainable, dormant until
multivariate corpora flip on.
* **Attention details** β PerDimScale (learned per-dimension query scaling)
with ``1/d_k`` attention scaling, biases on attention projections but not
MLPs, ``head_dim`` fixed at 64 across the family.
* **Robust causal scaler** β per-step causal location/scale (mask-aware, with
leading-patch backfill) under an arcsinh transform; targets are anchored at
each patch boundary so no future value leaks into its own scaling.
* **Residual SiLU patch projections** at both ends, and a 9-level
pinball/quantile head whose levels are exactly cascade's eval objective.
Shape and detail integers are pinned to the released ``Datadog/Toto-2.0-4m``
``config.json``: ``d_model=256``, ``num_layers=4``, ``num_heads=4``,
``qk/v_dim=64``, ``patch_size=32``, ``d_ff=688``, ``attn_bias``/no
``mlp_bias``, ``per_dim_scale``, ``use_xpos`` (Ξ³=0.4 decay on rotary),
``norm_eps=1e-4`` with weightless norms, layer grouping, and the u-ΞΌP residual
scheme (``residual_mult=0.75``, ``residual_attn_ratio=sqrt(S/log S)β5.14``,
applied via the unit-scaled a/b residual weights of u-ΞΌP eq. 25β31). The
optimiser orthogonalizes with Polar Express (see ``toto2_trainer.py``).
Remaining known approximations vs the release: the exact FFN inner structure
(param count 3.3M vs 4.1M) and the full u-ΞΌP init/LR width-scaling rules
(we keep fan-in init and a uniform LR). Pin ``base_arch_digest`` to whatever
you launch with.
"""
from __future__ import annotations
import math
from dataclasses import dataclass
import torch
import torch.nn as nn
import torch.nn.functional as F
# The 9 quantile levels 0.1..0.9 β identical to cascade's eval grid so the
# train objective equals the score objective.
QUANTILE_LEVELS = (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9)
@dataclass
class Toto2Config:
d_model: int = 256
num_layers: int = 4
num_heads: int = 4
head_dim: int = 64
patch_size: int = 32
mlp_expansion: int = 2
d_ff: int = 0 # exact FFN hidden width (0 β d_model Γ mlp_expansion); 4m ships 688
num_quantiles: int = 9
context_length: int = 4096
horizon: int = 64
max_patches: int = 256 # decode window capacity in patches (context + masked horizon)
# layer grouping (Toto-2.0 config.json: layer_group_size=4,
# num_variate_layers_per_group=1, variate_layer_first=false) β the last
# layer of each group of 4 attends over variates, the rest over time.
layer_group_size: int = 4
# CPM training-mask distribution (Toto 2.0 Β§2.1 sweep optima).
cpm_c_max: int = 16
cpm_p_max: float = 0.4
# u-ΞΌP residual scale Ξ±_res (released config: residual_mult = 0.75; the
# attention/FFN ratio is derived as sqrt(S/log S) from context/patch).
residual_mult: float = 0.75
@property
def ffn_hidden(self) -> int:
return self.d_ff if self.d_ff > 0 else self.d_model * self.mlp_expansion
@classmethod
def from_contract(cls, c: object) -> Toto2Config:
"""Build from a cascade ``TrainingContractConfig`` (duck-typed)."""
ctx = int(getattr(c, "context_length", 4096))
hz = int(getattr(c, "horizon", 64))
ps = int(getattr(c, "patch_size", 32))
return cls(
d_model=int(getattr(c, "d_model", 256)),
num_layers=int(getattr(c, "num_layers", 4)),
num_heads=int(getattr(c, "num_heads", 4)),
head_dim=int(getattr(c, "head_dim", 64)),
patch_size=ps,
mlp_expansion=int(getattr(c, "mlp_expansion", 2)),
d_ff=int(getattr(c, "d_ff", 0)),
num_quantiles=int(getattr(c, "num_quantiles", 9)),
context_length=ctx,
horizon=hz,
max_patches=max(8, (ctx + hz) // ps + 4),
cpm_c_max=int(getattr(c, "cpm_c_max", 16)),
cpm_p_max=float(getattr(c, "cpm_p_max", 0.4)),
)
def to_dict(self) -> dict:
return {k: getattr(self, k) for k in self.__dataclass_fields__}
def layer_axis(self, i: int) -> str:
"""Attention axis of layer ``i``: the last layer of each group of
``layer_group_size`` attends over variates, the rest over time."""
g = max(1, self.layer_group_size)
return "variate" if i % g == g - 1 else "time"
# ββ robust causal scaler ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Saturation bound on the standardized representation z = asinh((x-loc)/scale).
# Realistic data lives at |z| of a few (even a 1000Ο event is asinh(1000)β7.6), so
# this bound is never reached by honest corpora β clamp is the identity there. It
# exists purely as a backstop: it guarantees a finite, bounded z (and asinh target,
# clamped identically in the trainer) even for a pathological jump after an
# eps-clamped prefix, so the loss can never NaN or spike the shared training step.
# 64 leaves ~25 orders of asinh dynamic range above anything real; do NOT tighten
# it toward single digits without intent β that would start compressing legitimate
# heavy tails and change the scoring surface, not just add safety.
Z_CLAMP = 64.0
def causal_standardize(
x: torch.Tensor,
mask: torch.Tensor | None = None,
*,
min_obs: int = 8,
eps: float = 1e-5,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Toto 2.0's robust causal scaler: per-step causal location/scale under an
arcsinh transform.
``x`` is ``(B, L)``; ``mask`` is an optional binary ``(B, L)`` with 1 =
unobserved β masked entries are excluded from the statistics, so the stats
carry forward unchanged across masked spans (matching inference, where
horizon mask patches contribute nothing). Steps whose causal window holds
fewer than ``min_obs`` observations are backfilled with the first stable
stats (the paper's leading-patch backfill). Returns ``(z, loc, scale)``
where ``z = arcsinh((x - loc) / scale)``; all three are ``(B, L)``.
"""
B, L = x.shape
keep = torch.ones_like(x) if mask is None else 1.0 - mask.to(x.dtype)
# The cumulative E[xΒ²]βE[x]Β² form cancels catastrophically once
# meanΒ²/var exceeds the dtype's precision (~1e7 in float32 β routine for
# counter/gauge-style series at large levels with small fluctuations),
# collapsing scale to eps. Accumulate in float64 and shift each row to its
# first observation so the moments stay small regardless of series level.
x64 = x.double()
k64 = keep.double()
ref = x64.gather(-1, (k64 > 0).to(torch.int64).argmax(dim=-1, keepdim=True))
xk = (x64 - ref) * k64
n = k64.cumsum(dim=-1)
cnt = n.clamp_min(1.0)
loc = xk.cumsum(dim=-1) / cnt
var = (xk * xk).cumsum(dim=-1) / cnt - loc * loc
loc = loc + ref
scale = var.clamp_min(0.0).sqrt().clamp_min(eps)
ok = n >= float(min_obs)
has = ok.any(dim=-1)
first = torch.where(
has, ok.to(torch.int64).argmax(dim=-1), torch.full((B,), L - 1, device=x.device)
)[:, None]
loc = torch.where(ok, loc, loc.gather(-1, first))
scale = torch.where(ok, scale, scale.gather(-1, first))
z = torch.asinh((x64 - loc) / scale).clamp_(-Z_CLAMP, Z_CLAMP)
return z.to(x.dtype), loc.to(x.dtype), scale.to(x.dtype)
def patch_anchors(loc: torch.Tensor, scale: torch.Tensor, patch_size: int) -> tuple[torch.Tensor, torch.Tensor]:
"""Causal stats at the last step of each patch β the scaling a forecast of
the *next* patch is anchored to. ``(B, L)`` β ``(B, P)`` each."""
B, L = loc.shape
P = L // patch_size
return (
loc.view(B, P, patch_size)[:, :, -1],
scale.view(B, P, patch_size)[:, :, -1],
)
def invert_standardize(z: torch.Tensor, loc: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
"""Inverse of :func:`causal_standardize` at a fixed anchor:
``x = sinh(z) * scale + loc``."""
return torch.sinh(z) * scale + loc
# ββ building blocks βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class _ResidualMLP(nn.Module):
"""Two-layer SiLU MLP with a residual connection β Toto 2.0's nonlinear
patch projection, used at both ends of the transformer. Bias-free (biases
live on attention projections, not MLPs)."""
def __init__(self, dim: int, hidden: int):
super().__init__()
self.net = nn.Sequential(
nn.Linear(dim, hidden, bias=False),
nn.SiLU(),
nn.Linear(hidden, dim, bias=False),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x + self.net(x)
def _xpos(
q: torch.Tensor, k: torch.Tensor, inv_freq: torch.Tensor, zeta: torch.Tensor,
scale_base: float = 512.0,
) -> tuple[torch.Tensor, torch.Tensor]:
"""xPos (arXiv 2212.10554, ``use_xpos`` in the Toto-2.0 release): rotary
position embedding with per-dimension exponential decay
``ΞΆΜ_i = (i/(d/2) + Ξ³)/(1 + Ξ³)``, Ξ³ = 0.4 β queries scaled by ``ΞΆΜ^m`` and
keys by ``ΞΆΜ^{-m}`` over the sequence axis of ``(B, H, T, hd)``. Follows
the official torchscale implementation: the exponent is centered and
divided by ``scale_base`` (512) so ``ΞΆΜ^{Β±m}`` stays representable."""
T = q.shape[-2]
t = torch.arange(T, device=q.device, dtype=inv_freq.dtype)
freqs = torch.outer(t, inv_freq) # (T, hd/2)
cos = freqs.cos().repeat_interleave(2, dim=-1) # (T, hd)
sin = freqs.sin().repeat_interleave(2, dim=-1)
power = ((t - T // 2) / scale_base)[:, None] # (T, 1)
scale = (zeta[None, :] ** power).repeat_interleave(2, dim=-1) # (T, hd)
def rotate(x):
x1 = x[..., 0::2]
x2 = x[..., 1::2]
return torch.stack((-x2, x1), dim=-1).flatten(-2)
return (q * cos + rotate(q) * sin) * scale, (k * cos + rotate(k) * sin) / scale
class _Block(nn.Module):
"""Pre-norm multi-head attention + GELU MLP.
``axis="time"``: causal over the patch axis with rotary positions.
``axis="variate"``: full attention over the variate axis (no positions β
variates are unordered). Both use PerDimScale query scaling with ``1/d_k``
attention scaling (ΞΌP-compatible), biases on attention projections only.
"""
def __init__(self, cfg: Toto2Config, axis: str, block_idx: int = 0):
super().__init__()
self.cfg = cfg
self.axis = axis
inner = cfg.num_heads * cfg.head_dim
# norm_eps = 1e-4, norm_include_weight = false β per the released config.
self.norm1 = nn.LayerNorm(cfg.d_model, eps=1e-4, elementwise_affine=False)
self.qkv = nn.Linear(cfg.d_model, 3 * inner, bias=True)
self.proj = nn.Linear(inner, cfg.d_model, bias=True)
self.norm2 = nn.LayerNorm(cfg.d_model, eps=1e-4, elementwise_affine=False)
hidden = cfg.ffn_hidden
self.mlp = nn.Sequential(
nn.Linear(cfg.d_model, hidden, bias=False),
nn.GELU(),
nn.Linear(hidden, cfg.d_model, bias=False),
)
# PerDimScale: learned per-dimension query scaling; softplus(0) = ln 2
# normalizer so the init is an exact no-op.
self.per_dim_scale = nn.Parameter(torch.zeros(cfg.head_dim))
if axis == "time":
half = cfg.head_dim // 2
idx = torch.arange(half).float() / max(1, half)
self.register_buffer("inv_freq", 1.0 / (10000.0**idx), persistent=False)
self.register_buffer("zeta", (idx + 0.4) / 1.4, persistent=False) # xPos Ξ³=0.4
# u-ΞΌP residual scheme (u-ΞΌP eq. 25β31; Toto 2.0 Β§4.4): stream and
# branch combine as x β bΒ·x + aΒ·branch with aΒ² + bΒ² = 1, keeping the
# residual stream at unit scale. Ξ±_res = residual_mult = 0.75 and
# Ξ±_res-attn-ratio = sqrt(S/log S) with S = context patches β exactly
# the released config's residual_mult / residual_attn_ratio (β5.136
# at S = 128). Branches count attention and MLP separately (L = 2Β·layers).
S = max(2.0, cfg.context_length / cfg.patch_size)
ratio2 = S / math.log(S) # Ξ±_res-attn-ratioΒ²
af2 = 2.0 * cfg.residual_mult**2 / (ratio2 + 1.0)
aa2 = ratio2 * af2
L = 2.0 * cfg.num_layers
i = block_idx
tau2_attn = aa2 / (L / 2.0 + i * aa2 + i * af2)
tau2_mlp = af2 / (L / 2.0 + (i + 1) * aa2 + i * af2)
self.attn_a = math.sqrt(tau2_attn / (tau2_attn + 1.0))
self.attn_b = math.sqrt(1.0 / (tau2_attn + 1.0))
self.mlp_a = math.sqrt(tau2_mlp / (tau2_mlp + 1.0))
self.mlp_b = math.sqrt(1.0 / (tau2_mlp + 1.0))
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, T, _ = x.shape
h = self.norm1(x)
qkv = self.qkv(h).view(B, T, 3, self.cfg.num_heads, self.cfg.head_dim)
q, k, v = qkv.unbind(dim=2)
q, k, v = (t.transpose(1, 2) for t in (q, k, v)) # (B, H, T, hd)
if self.axis == "time":
q, k = _xpos(q, k, self.inv_freq, self.zeta)
q = q * (F.softplus(self.per_dim_scale) / math.log(2.0))
attn = F.scaled_dot_product_attention(
q, k, v, is_causal=(self.axis == "time"), scale=1.0 / self.cfg.head_dim
)
attn = attn.transpose(1, 2).reshape(B, T, self.cfg.num_heads * self.cfg.head_dim)
x = self.attn_b * x + self.attn_a * self.proj(attn)
x = self.mlp_b * x + self.mlp_a * self.mlp(self.norm2(x))
return x
class Toto2Model(nn.Module):
"""Patch transformer with contiguous patch masking and alternating
time/variate attention, predicting the next patch's per-step quantiles.
Each input patch carries a binary mask channel (1 = unobserved entry);
masked entries are zeroed on input, so a masked patch contributes only its
position and mask bits. Training masks random contiguous spans (CPM);
inference appends fully-masked horizon patches and reads every horizon
patch's quantiles from a single forward pass.
"""
def __init__(self, cfg: Toto2Config):
super().__init__()
self.cfg = cfg
# values β mask channel β 2Γpatch_size inputs per patch.
self.patch_embed = nn.Linear(cfg.patch_size * 2, cfg.d_model)
self.embed_mlp = _ResidualMLP(cfg.d_model, cfg.ffn_hidden)
# grouped layers: variate-axis attention closes each group of
# ``layer_group_size`` (Toto-2.0's 3-time-then-1-variate pattern).
self.blocks = nn.ModuleList(
_Block(cfg, axis=cfg.layer_axis(i), block_idx=i)
for i in range(cfg.num_layers)
)
self.norm = nn.LayerNorm(cfg.d_model, eps=1e-4, elementwise_affine=False)
self.out_mlp = _ResidualMLP(cfg.d_model, cfg.ffn_hidden)
# each position predicts the NEXT patch: patch_size steps Γ num_quantiles
self.head = nn.Linear(cfg.d_model, cfg.patch_size * cfg.num_quantiles)
self.apply(self._init_weights)
def _init_weights(self, m: nn.Module) -> None:
# u-ΞΌP-flavoured init: linear weights ~ N(0, 1/fan_in); the operator can
# swap in exact u-ΞΌP multipliers and pin base_arch_digest accordingly.
if isinstance(m, nn.Linear):
fan_in = m.weight.shape[1]
nn.init.normal_(m.weight, mean=0.0, std=1.0 / math.sqrt(fan_in))
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Embedding):
nn.init.normal_(m.weight, mean=0.0, std=0.02)
def forward(self, patches: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
"""``patches``: ``(B, P, patch_size)`` univariate or
``(B, C, P, patch_size)`` multivariate; ``mask``: optional binary
patch-level (``(B, P)`` / ``(B, C, P)``) or per-entry (same + trailing
``patch_size`` axis), 1 = unobserved. Returns predicted quantiles for
each position's *next* patch, shaped like the input with a trailing
``num_q`` axis: ``(B, [C,] P, patch_size, num_q)``."""
squeeze_variates = patches.dim() == 3
if squeeze_variates:
patches = patches[:, None] # (B, 1, P, ps)
if mask is not None:
mask = mask[:, None]
B, C, P, ps = patches.shape
if mask is None:
mask = torch.zeros_like(patches)
else:
if mask.dim() == 3:
mask = mask[..., None].expand(B, C, P, ps)
mask = mask.to(patches.dtype)
x = torch.cat([patches * (1.0 - mask), mask], dim=-1)
x = self.embed_mlp(self.patch_embed(x)) # (B, C, P, d)
for blk in self.blocks:
if blk.axis == "time":
x = blk(x.reshape(B * C, P, -1)).view(B, C, P, -1)
else:
x = (
blk(x.transpose(1, 2).reshape(B * P, C, -1))
.view(B, P, C, -1)
.transpose(1, 2)
)
x = self.out_mlp(self.norm(x))
out = self.head(x) # (B, C, P, ps*num_q)
out = out.view(B, C, P, ps, self.cfg.num_quantiles)
return out[:, 0] if squeeze_variates else out
def pinball_loss(pred_q: torch.Tensor, target: torch.Tensor, levels: tuple[float, ...]) -> torch.Tensor:
"""Mean pinball (quantile) loss. ``pred_q`` ``(..., num_q)``, ``target``
``(...)`` broadcast over the quantile axis."""
q = torch.tensor(levels, device=pred_q.device, dtype=pred_q.dtype)
err = target.unsqueeze(-1) - pred_q
return torch.maximum(q * err, (q - 1.0) * err).mean()
|