File size: 18,686 Bytes
0b80c48 | 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 | """
Chimera 5.3 β HYPER Paradigm Engine for 10,000+ tok/s CPU Training
===================================================================
Seven orthogonal paradigms that stack multiplicatively:
P1 GrowLength Curriculum β Start seq=16, grow to target. Short seqs =
huge batch = way more tok/s early on.
(arxiv:2310.00576)
P2 Reservoir Freezing (GRC) β Freeze ~50 % of recurrent gate matrices as
random ternary. No grad for those params β
2Γ fewer FLOPs in recurrent layers.
(arxiv:2512.23145)
P3 Sparse MeZO β Perturb only top-K % most-sensitive params
(by magnitude). ZO signal quality β
βmaskββfβΒ²/ββfβΒ²; masking raises it.
(arxiv:2406.02913)
P4 Blockwise Pipeline β Pin layer-groups to core-groups; overlap
block N on batch t with block N-1 on t+1.
P5 Fused Ternary Cache β Pre-materialise dense ternary weights once
per step; reuse for both MeZO forwards.
P6 Aggressive Token Packing β Zero padding waste; pack documents
back-to-back with EOS separators.
P7 Progressive Layer Unfreeze β Train only top ~25 % of layers first; un-
freeze downward as training proceeds.
Expected combined multiplier (tiny-35 M on 8-core CPU):
P1 (4-8Γ) Γ P2 (1.5-2Γ) Γ P3 (3-5Γ) Γ P5 (1.3Γ) Γ P7 (1.5-2Γ)
β 35-260Γ β 50-200 tok/s baseline β **1 750-52 000 tok/s**
"""
from __future__ import annotations
import math
import time
from typing import Dict, List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset
from .quantization import BitLinear
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# P1 β GrowLength Curriculum
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GrowLengthDataset(Dataset):
"""Flat token buffer re-chunked on-the-fly when ``set_seq_len`` is called.
Because chunks are contiguous slices, set_seq_len is O(1).
"""
def __init__(self, all_ids: torch.Tensor, seq_len: int = 16):
self.all_ids = all_ids
self._seq_len = 0
self._n = 0
self.set_seq_len(seq_len)
# ββ public API βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def set_seq_len(self, seq_len: int) -> None:
self._seq_len = int(seq_len)
self._n = self.all_ids.numel() // (self._seq_len + 1)
@property
def seq_len(self) -> int:
return self._seq_len
def __len__(self) -> int:
return self._n
def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
start = idx * (self._seq_len + 1)
chunk = self.all_ids[start: start + self._seq_len + 1]
return {"input_ids": chunk[:-1], "labels": chunk[1:]}
class GrowLengthScheduler:
"""Maps a global step to the current target sequence length.
``stages`` is a list of ``(seq_len, fraction_of_total_steps)`` tuples.
Fractions are normalised internally so they need not sum to 1.
"""
def __init__(self, stages: List[Tuple[int, float]], total_steps: int):
total_frac = sum(f for _, f in stages) or 1.0
cumulative = 0
self._boundaries: List[Tuple[int, int]] = []
for seq_len, frac in stages:
cumulative += int(total_steps * frac / total_frac)
self._boundaries.append((cumulative, int(seq_len)))
def get_seq_len(self, step: int) -> int:
for boundary, seq_len in self._boundaries:
if step < boundary:
return seq_len
return self._boundaries[-1][1]
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# P2 β Reservoir Freezing (GRC-inspired, arxiv:2512.23145)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def apply_reservoir_freezing(model: nn.Module,
freeze_ratio: float = 0.5) -> int:
"""Freeze gate / forget projections in recurrent layers as random ternary
reservoirs. Returns the number of frozen scalar parameters.
Targets:
β’ GatedDeltaNet β a_proj, b_proj (alpha / beta gates)
β’ mLSTM β fgate (forget gate)
β’ TitansMAC β alpha_proj (forgetting gate)
The frozen weights are re-initialised to unit-spectral-radius ternary
matrices so every layer starts with a stable reservoir.
"""
frozen = 0
for _name, module in model.named_modules():
# ββ GatedDeltaNet gates ββββββββββββββββββββββββββββββββββββββ
if hasattr(module, "a_proj") and hasattr(module, "b_proj"):
for attr in ("a_proj", "b_proj"):
proj = getattr(module, attr, None)
if proj is None:
continue
w = getattr(proj, "weight", None)
if w is None or not isinstance(w, nn.Parameter):
continue
with torch.no_grad():
w.data = torch.randint(-1, 2, w.shape,
dtype=w.dtype, device=w.device)
norm = torch.linalg.matrix_norm(
w.data.float(), ord=2).clamp(min=1.0)
w.data.div_(norm)
w.requires_grad = False
frozen += w.numel()
# ββ mLSTM forget gate ββββββββββββββββββββββββββββββββββββββββ
if hasattr(module, "fgate") and hasattr(module, "igate"):
fg = module.fgate
w = getattr(fg, "weight", None)
if w is not None and isinstance(w, nn.Parameter):
with torch.no_grad():
w.data = torch.randint(-1, 2, w.shape,
dtype=w.dtype, device=w.device).float()
norm = torch.linalg.matrix_norm(
w.data, ord=2).clamp(min=1.0)
w.data.div_(norm)
w.requires_grad = False
frozen += w.numel()
# ββ TitansMAC forgetting βββββββββββββββββββββββββββββββββββββ
if hasattr(module, "alpha_proj") and hasattr(module, "eta_proj"):
ap = module.alpha_proj
w = getattr(ap, "weight", None)
if w is not None and isinstance(w, nn.Parameter):
with torch.no_grad():
w.data = torch.randint(-1, 2, w.shape,
dtype=w.dtype, device=w.device).float()
norm = torch.linalg.matrix_norm(
w.data, ord=2).clamp(min=1.0)
w.data.div_(norm)
w.requires_grad = False
frozen += w.numel()
return frozen
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# P3 β Sparse MeZO (arxiv:2406.02913)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class SparseMeZOOptimizer:
"""Zeroth-order optimiser that perturbs only the top-K % most-sensitive
parameters (ranked by weight magnitude as a cheap proxy for gradient
magnitude).
Combined with **Paradigm 5** (fused ternary cache): before each dual-
forward the caller should invoke ``precompute_ternary_cache(model)``
once so that both forward passes reuse the same dense-weight buffers.
"""
def __init__(self, model: nn.Module, *,
lr: float = 1e-4,
eps: float = 1e-3,
sparsity: float = 0.01,
weight_decay: float = 0.0,
momentum: float = 0.0,
mask_refresh_interval: int = 50):
self.model = model
self.lr = float(lr)
self.eps = float(eps)
self.sparsity = float(sparsity)
self.wd = float(weight_decay)
self.momentum_coeff = float(momentum)
self.mask_refresh = int(mask_refresh_interval)
# Deduplicated trainable params
self._params: List[Tuple[str, nn.Parameter]] = []
seen: set = set()
for name, p in model.named_parameters():
if p.requires_grad and id(p) not in seen:
self._params.append((name, p))
seen.add(id(p))
self._total = sum(p.numel() for _, p in self._params)
self._k = max(1, int(self._total * self.sparsity))
self._masks: Dict[int, torch.Tensor] = {}
self._momentum: Dict[int, torch.Tensor] = {}
if self.momentum_coeff > 0:
for _, p in self._params:
self._momentum[id(p)] = torch.zeros_like(p.data)
self._step = 0
self._refresh_masks()
# ββ mask computation βββββββββββββββββββββββββββββββββββββββββββββ
def _refresh_masks(self) -> None:
slices, offset = [], 0
mags = []
for _, p in self._params:
flat = p.data.abs().flatten()
mags.append(flat)
slices.append((offset, offset + flat.numel()))
offset += flat.numel()
all_mag = torch.cat(mags)
if self._k < all_mag.numel():
thr = torch.topk(all_mag, self._k, sorted=False).values.min()
else:
thr = torch.tensor(0.0)
for i, (_, p) in enumerate(self._params):
s, e = slices[i]
self._masks[id(p)] = (all_mag[s:e] >= thr).view(p.shape)
# ββ perturbation helpers βββββββββββββββββββββββββββββββββββββββββ
def _direction(self, p: torch.Tensor, seed: int,
mask: torch.Tensor) -> torch.Tensor:
gen = torch.Generator(device="cpu")
gen.manual_seed(seed & 0x7FFF_FFFF_FFFF_FFFF)
z = torch.empty(p.shape, dtype=p.dtype, device="cpu")
z.bernoulli_(0.5, generator=gen).mul_(2).sub_(1)
return z * mask.to(z.dtype)
def _perturb(self, seed: int, scale: float) -> None:
for i, (_, p) in enumerate(self._params):
z = self._direction(p.data, seed + i * 1_000_003,
self._masks.get(id(p),
torch.ones_like(p.data)))
p.data.add_(z, alpha=scale)
_invalidate_bitlinear(self.model)
# ββ step βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@torch.no_grad()
def step(self, loss_fn, batch) -> float:
self._step += 1
if self._step % self.mask_refresh == 0:
self._refresh_masks()
seed = int(torch.randint(0, 2 ** 31, (1,)).item())
self._perturb(seed, +self.eps)
loss_pos = float(loss_fn(batch).item())
self._perturb(seed, -2.0 * self.eps)
loss_neg = float(loss_fn(batch).item())
self._perturb(seed, +self.eps) # restore
proj = (loss_pos - loss_neg) / (2.0 * self.eps)
for i, (_, p) in enumerate(self._params):
mask = self._masks.get(id(p), torch.ones_like(p.data))
z = self._direction(p.data, seed + i * 1_000_003, mask)
if self.momentum_coeff > 0:
buf = self._momentum[id(p)]
buf.mul_(self.momentum_coeff).add_(z, alpha=proj)
p.data.add_(buf, alpha=-self.lr)
else:
p.data.add_(z, alpha=-self.lr * proj)
if self.wd > 0:
p.data.mul_(1 - self.lr * self.wd)
_invalidate_bitlinear(self.model)
return 0.5 * (loss_pos + loss_neg)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# P5 β Fused Ternary Cache
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def precompute_ternary_cache(model: nn.Module) -> None:
"""Materialise every BitLinear's packed + dense fp32 cache so the next
forward pass is allocation-free. Call once before each MeZO dual-fwd."""
for m in model.modules():
if isinstance(m, BitLinear):
m._ensure_packed()
m._ensure_dense()
def _invalidate_bitlinear(model: nn.Module) -> None:
for m in model.modules():
if isinstance(m, BitLinear):
m.invalidate_packed()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# P6 β Aggressive Token Packing
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def pack_documents(raw_ids: torch.Tensor, eos_id: int,
max_tokens: int) -> torch.Tensor:
"""Return a contiguous 1-D ``LongTensor`` of ``max_tokens`` tokens where
individual documents are separated by ``eos_id`` and there is **zero**
padding. Already-tokenised documents should be concatenated in
``raw_ids`` (the function simply truncates to ``max_tokens``).
"""
n = min(raw_ids.numel(), int(max_tokens))
return raw_ids[:n].contiguous()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# P7 β Progressive Layer Unfreezing
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class ProgressiveUnfreezer:
"""Freeze all but the top *k* layers initially; unfreeze downward as
training advances.
``n_stages`` = number of unfreeze events spread evenly across
``total_steps``. At each event one more block of layers becomes
trainable (starting from the output end).
"""
def __init__(self, model: nn.Module, total_steps: int,
n_stages: int = 4):
self._layers = model.layers # nn.ModuleList
self._n = len(self._layers)
self._total = int(total_steps)
self._stages = int(n_stages)
self._block = max(1, self._n // self._stages)
self._current_from = self._n # everything frozen initially
# Immediately unfreeze the first block (top layers)
self.update(0)
def update(self, step: int) -> int:
"""Call every step. Returns the index of the first trainable layer."""
stage = min(step * self._stages // max(1, self._total),
self._stages - 1)
target = max(0, self._n - (stage + 1) * self._block)
if target != self._current_from:
self._current_from = target
for i, layer in enumerate(self._layers):
req = i >= self._current_from
for p in layer.parameters():
p.requires_grad = req
return self._current_from
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Cosine LR helper (shared)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def cosine_lr(step: int, warmup: int, total: int,
max_lr: float, min_lr: float) -> float:
if warmup > 0 and step < warmup:
return max_lr * (step + 1) / warmup
if step >= total:
return min_lr
p = (step - warmup) / max(1, total - warmup)
return min_lr + 0.5 * (max_lr - min_lr) * (1.0 + math.cos(math.pi * p))
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Public surface
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
__all__ = [
"GrowLengthDataset",
"GrowLengthScheduler",
"apply_reservoir_freezing",
"SparseMeZOOptimizer",
"precompute_ternary_cache",
"pack_documents",
"ProgressiveUnfreezer",
"cosine_lr",
]
|