Repository navigation
Expand file tree
/
Copy pathinitialization.py
More file actions
754 lines (685 loc) · 37.6 KB
/
Copy pathinitialization.py
File metadata and controls
754 lines (685 loc) · 37.6 KB
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
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
# -*- coding: utf-8 -*-
"""DAMI-specific initialization and thin Hugging Face training integration.
FLA supplies every model, recurrence and CUDA kernel. Transformers supplies the
training loop. This module contains DAMI's targets, calibration, architecture
adapters, memory-only export, and the unified entry point for 15 conditions.
"""
from __future__ import annotations
import argparse
import hashlib
import importlib
import inspect
import json
import math
from pathlib import Path
import textwrap
import torch
from torch import nn
from torch.nn import functional as F
ROOT = Path(__file__).resolve().parent
ARCHITECTURES = ("gdn", "kda", "gdn2", "mamba2", "rwkv7")
INITIALIZATIONS = ('random', 'post', 'dami')
FAMILIES = {
"gdn": ("gated_deltanet", "GatedDeltaNet"),
"kda": ("kda", "KDA"),
"gdn2": ("gdn2", "GDN2"),
"mamba2": ("mamba2", "Mamba2"),
"rwkv7": ("rwkv7", "RWKV7"),
}
def targets(layers: int, heads: int, tau_min=1.0, tau_max=4096.0):
"""Head 0 has the longest e-folding time. GDN uses a sine depth profile."""
if layers < 1 or heads < 1 or not 0 < tau_min <= tau_max:
raise ValueError("Positive dimensions and 0 < tau_min <= tau_max required")
tau = torch.logspace(math.log10(tau_max), math.log10(tau_min), heads,
dtype=torch.float64)
depth = torch.linspace(0, 1, layers, dtype=torch.float64)
sensitivity = .25 + .5 * torch.sin(math.pi * depth)
return tau, sensitivity
def inverse_softplus(x):
if not torch.isfinite(x).all() or (x <= 0).any():
raise ValueError("inverse_softplus requires finite positive values")
return x + torch.log(-torch.expm1(-x))
def solve_bias(z, target, iterations=64):
"""Solve mean_tokens sigmoid(z+b) = target, independently per column."""
z = z.double()
target = torch.as_tensor(target, device=z.device, dtype=z.dtype)
if not torch.isfinite(z).all() or not ((target > 0) & (target < 1)).all():
raise ValueError("Finite activations and sensitivity in (0, 1) required")
# Data-dependent brackets work even when the logits lie outside [-60, 60].
logit = torch.logit(target)
lo = logit - z.amax(0) - 32
hi = logit - z.amin(0) + 32
for _ in range(iterations):
mid = (lo + hi) / 2
below = torch.sigmoid(z + mid).mean(0) < target
lo = torch.where(below, mid, lo)
hi = torch.where(below, hi, mid)
return (lo + hi) / 2
def centered_channel_noise(layer, head, channels, std=.03):
"""The deterministic channel perturbation used for paper KDA/GDN2."""
if channels == 1:
return torch.zeros(1)
gen = torch.Generator(device="cpu").manual_seed(42000 + layer * 997 + head * 37)
noise = torch.randn(channels, generator=gen)
noise -= noise.mean()
noise *= std / noise.std(unbiased=False)
return noise - noise.mean()
def _layers(model):
backbone = getattr(model, 'model', None)
if backbone is None:
backbone = getattr(model, 'backbone', None)
if backbone is None or not hasattr(backbone, 'layers'):
raise ValueError('Unsupported FLA model backbone')
return list(backbone.layers)
def _normalize_save_metadata(model):
# Transformers 5 uses a mapping for tied-weight metadata. These paper
# configurations have untied embeddings; old FLA list metadata is empty
# semantically and can be normalized without changing any tensor.
for module in model.modules():
if isinstance(getattr(module, '_tied_weights_keys', None), list):
if getattr(model.config, 'tie_word_embeddings', False):
raise ValueError('This integration supports the paper untied-embedding configurations')
module._tied_weights_keys = {}
# The upstream L2Wrap returns a view; Trainer scales loss in place. Cloning
# this scalar preserves its backward while satisfying the autograd contract.
if getattr(model.config, 'use_l2warp', False):
_fix_l2warp_dtype()
model.register_forward_hook(_clone_wrapped_loss)
return model
def _clone_wrapped_loss(_module, _inputs, output):
if getattr(output, 'loss', None) is not None:
output.loss = output.loss.clone()
return output
def _fix_l2warp_dtype():
"""Keep the native L2Wrap gradient, fixing its AMP scatter dtype contract."""
module = importlib.import_module('fla.modules.l2warp')
cls = module.L2Wrap
if getattr(cls, '_dami_dtype_fix', False):
return
source = textwrap.dedent(inspect.getsource(cls.backward))
anchor = 'glogits.scatter_(-1, ids, maxx * grad_output)'
if source.count(anchor) != 1:
raise RuntimeError('Unsupported FLA L2Wrap API; install the pinned dependency revision')
source = source.replace('@staticmethod\n', '').replace(
anchor, 'glogits.scatter_(-1, ids, (maxx * grad_output).to(glogits.dtype))')
namespace = dict(vars(module))
exec(compile('from __future__ import annotations\n'+source,
'<FLA L2Wrap AMP compatibility>', 'exec'), namespace)
cls.backward = staticmethod(namespace['backward'])
cls._dami_dtype_fix = True
def _is_memory(key):
return key.endswith((".A_log", ".dt_bias", ".decay_log_scale",
".w_lora.lora.2.bias"))
def memory_parameters(model):
return {k: v.detach().cpu().contiguous().clone()
for k, v in model.state_dict().items() if _is_memory(k)}
def paper_parameters(architecture, scheme='dami'):
"""Load any of the fifteen 100M memory initializations without FLA."""
if architecture not in ARCHITECTURES or scheme not in INITIALIZATIONS:
raise ValueError('Unknown architecture or initialization')
name = '100m.json' if scheme == 'dami' else 'controls_100m.json'
source = json.loads((ROOT / 'reference' / name).read_text(encoding='utf-8'))
key = architecture if scheme == 'dami' else architecture + '_' + scheme
return {k: torch.tensor(v["values"], dtype=getattr(torch, v["dtype"].split('.')[-1]))
.reshape(v["shape"])
for k, v in source['models'][key]['tensors'].items()}
def load_memory(model, state):
"""Load memory tensors strictly, while leaving non-memory weights untouched."""
destination = {k: v for k, v in model.state_dict().items() if _is_memory(k)}
if destination.keys() != state.keys():
raise ValueError(f"Memory keys differ: missing={destination.keys()-state.keys()}, "
f"unexpected={state.keys()-destination.keys()}")
for key, value in state.items():
if destination[key].shape != value.shape:
raise ValueError(f"Shape mismatch for {key}: {value.shape} vs {destination[key].shape}")
if not torch.isfinite(value).all():
raise ValueError(f"Non-finite parameter: {key}")
with torch.no_grad():
for key, value in state.items():
destination[key].copy_(value.to(destination[key]))
def _rwkv_logits(self, x):
# Preserve the paper implementation's fp32 master bias with bf16 projections.
seq = self.w_lora.lora
hidden = seq[1](seq[0](x))
out = F.linear(hidden, seq[2].weight, None)
return (out.float() + seq[2].bias.float()).to(out.dtype)
def install_rwkv_adapter():
"""Replace RWKV's native constant with a frozen per-head scale.
The dependency's forward implementation is retained. Only its decay-line
expression changes; no recurrence or CUDA kernel is copied into this repo.
"""
module = importlib.import_module("fla.layers.rwkv7")
cls = module.RWKV7Attention
if getattr(cls, "_dami_adapter", False):
return
source = textwrap.dedent(inspect.getsource(cls.forward))
anchor = "w = -0.6065306597126334 * self.w_lora(xw).sigmoid()"
if source.count(anchor) != 1:
raise RuntimeError("Unsupported FLA RWKV-7 API. Install the pinned FLA revision in requirements.txt.")
source = source.replace(anchor,
"w = -self.decay_log_scale.float().exp().repeat_interleave(self.head_dim) "
"* self._dami_logits(xw).sigmoid()")
# Match the paper training backend's 32-token chunk schedule.
if source.count('chunk_size=64,') != 1:
raise RuntimeError('Unsupported RWKV-7 chunk scheduler in the FLA dependency')
source = source.replace('chunk_size=64,', 'chunk_size=32,')
namespace = dict(vars(module))
exec(compile("from __future__ import annotations\n" + source,
"<DAMI RWKV-7 decay adapter>", "exec"), namespace)
class DAMIRWKV7Attention(cls):
_dami_adapter = True
forward = namespace["forward"]
_dami_logits = _rwkv_logits
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.decay_log_scale = nn.Parameter(
torch.full((self.num_heads,), -.5, dtype=torch.float32), requires_grad=False)
self.w_lora.lora[2].bias.data = self.w_lora.lora[2].bias.data.float()
def _apply(self, fn, recurse=True):
super()._apply(fn, recurse=recurse)
self.decay_log_scale.data = self.decay_log_scale.data.float()
self.w_lora.lora[2].bias.data = self.w_lora.lora[2].bias.data.float()
return self
module.RWKV7Attention = DAMIRWKV7Attention
importlib.import_module("fla.models.rwkv7.modeling_rwkv7").RWKV7Attention = DAMIRWKV7Attention
def build_model(architecture, config=None, init_seed=42):
"""Build a real FLA model. A size label alone does not define its geometry."""
if architecture not in FAMILIES:
raise ValueError(f"architecture must be one of {ARCHITECTURES}")
import fla # noqa: F401 -- registers the FLA model families
if architecture == "rwkv7":
install_rwkv_adapter()
cfg = config or json.loads((ROOT / f"configs/{architecture}_100m.json").read_text(encoding="utf-8"))
cfg = dict(cfg)
torch.manual_seed(init_seed)
if architecture == 'gdn2':
# The pinned public FLA commit has the GDN2 layer/kernels, but no HF
# language-model wrapper. Its norm/MLP/residual stack matches KDA's.
# Reuse that stack, constructing genuine GDN2 attention in each block.
from fla.models.kda import KDAConfig, KDAForCausalLM
from fla.layers.gdn2 import GatedDeltaNet2
class GDN2AttentionAdapter(GatedDeltaNet2):
# KDA's HF initialization checks this flag; GDN2 uses its unbounded
# Softplus gate. This flag does not change the GDN2 forward operator.
safe_gate = False
modeling = importlib.import_module('fla.models.kda.modeling_kda')
cfg.pop('model_type', None)
cfg['dami_architecture'] = 'gdn2'
original = modeling.KimiDeltaAttention
modeling.KimiDeltaAttention = GDN2AttentionAdapter
try:
model = KDAForCausalLM(KDAConfig(**cfg))
finally:
modeling.KimiDeltaAttention = original
if not all(isinstance(layer.attn, GatedDeltaNet2) for layer in _layers(model)):
raise RuntimeError('GDN2 adapter did not replace every attention block')
return _normalize_save_metadata(model)
submodule, prefix = FAMILIES[architecture]
module = importlib.import_module(f"fla.models.{submodule}")
config_cls = getattr(module, prefix + "Config")
model_cls = getattr(module, prefix + "ForCausalLM")
return _normalize_save_metadata(model_cls(config_cls(**cfg)))
@torch.no_grad()
def calibrate_gdn(model, batches, tau=None, sensitivity=None):
"""LSUV-style sequential calibration on actual a_proj outputs.
batches: iterable of packed [batch, length] token tensors (no padding), or
dictionaries with input_ids and an all-one attention_mask. Outputs from the
calibrated layer are recomputed before propagating to the next layer.
"""
if batches is None:
raise ValueError('GDN calibration requires real packed token batches')
layers = _layers(model)
h = layers[0].attn.A_log.numel()
default_tau, default_s = targets(len(layers), h)
tau = torch.as_tensor(default_tau if tau is None else tau, dtype=torch.float64)
sensitivity = torch.as_tensor(default_s if sensitivity is None else sensitivity)
if tau.shape != (h,) or sensitivity.shape != (len(layers),) or (tau <= 0).any():
raise ValueError("tau must have one positive value per head; sensitivity one per layer")
if not ((sensitivity > 0) & (sensitivity < 1)).all():
raise ValueError("Sensitivity targets must lie in (0, 1)")
device = next(model.parameters()).device
was_training = model.training
model.eval()
hidden_batches = []
try:
for batch in batches:
if isinstance(batch, dict):
mask = batch.get("attention_mask")
if mask is not None and not torch.as_tensor(mask).bool().all():
raise ValueError("Use packed, unpadded calibration sequences")
batch = batch["input_ids"]
ids = torch.as_tensor(batch, dtype=torch.long, device=device)
if ids.ndim != 2:
raise ValueError("Calibration input_ids must have shape [batch, length]")
hidden_batches.append(model.get_input_embeddings()(ids).cpu())
if not hidden_batches:
raise ValueError("Calibration data is empty")
report = []
for index, layer in enumerate(layers):
attn = layer.attn
if attn.dt_bias.numel() != h:
raise ValueError("calibrate_gdn requires scalar, head-wise GDN dt_bias")
captured = []
handle = attn.a_proj.register_forward_hook(
lambda _m, _i, out: captured.append(out.detach().float().cpu().reshape(-1, h)))
try:
for hidden in hidden_batches:
layer(hidden.to(device), use_cache=False)
finally:
handle.remove()
z = torch.cat(captured)
bias = solve_bias(z, sensitivity[index])
mean_dt = F.softplus(z.double() + bias).mean(0)
a_log = -tau.log() - mean_dt.log()
attn.dt_bias.copy_(bias.to(attn.dt_bias))
attn.A_log.copy_(a_log.to(attn.A_log))
hidden_batches = [layer(hidden.to(device), use_cache=False)[0].cpu()
for hidden in hidden_batches]
# Report realized values AFTER casting to the model's storage dtype.
b = attn.dt_bias.detach().double().cpu()
a = attn.A_log.detach().double().cpu()
measured_s = torch.sigmoid(z.double() + b).mean(0)
measured_tau = 1 / (a.exp() * F.softplus(z.double() + b).mean(0))
report.append({"layer":index, "target_sensitivity":float(sensitivity[index]),
"realized_sensitivity":measured_s.tolist(), "target_tau":tau.tolist(),
"realized_tau":measured_tau.tolist()})
return report
finally:
model.train(was_training)
@torch.no_grad()
def transfer_gdn_to_channels(model, gdn_state, channel_std=.03):
"""Paper KDA/GDN2: transfer head means, add centered channel perturbations."""
for li, layer in enumerate(_layers(model)):
a = gdn_state[f"model.layers.{li}.attn.A_log"].float().flatten()
b = gdn_state[f"model.layers.{li}.attn.dt_bias"].float().flatten()
attn = layer.attn
h = attn.A_log.numel()
if a.numel() != h or b.numel() != h:
raise ValueError("GDN reference must match the target layer/head geometry")
d = attn.dt_bias.numel() // h
values = []
for head in range(h):
channel_values = b[head] + centered_channel_noise(li, head, d, channel_std)
values.append(channel_values - channel_values.mean() + b[head])
biases = torch.stack(values)
attn.A_log.copy_(a.reshape_as(attn.A_log).to(attn.A_log))
attn.dt_bias.copy_(biases.reshape_as(attn.dt_bias).to(attn.dt_bias))
@torch.no_grad()
def design_rwkv7(model, tau_min=1., tau_max=4096.):
"""Static channel-mean sigmoid response, NOT sigmoid derivative sensitivity."""
layers = _layers(model)
h = layers[0].attn.num_heads
tau, _ = targets(len(layers), h, tau_min, tau_max)
depth = torch.linspace(0, 1, len(layers), dtype=torch.float64)
parabola = 4 * depth * (1 - depth)
s = .25 + .5 * parabola / parabola.max().clamp_min(1e-12)
for li, layer in enumerate(layers):
attn = layer.attn
first_projection = attn.w_lora.lora[0].weight
if first_projection.detach().abs().max().item() != 0:
raise ValueError("Static RWKV calibration requires the native zero input projection at step zero")
d = attn.head_dim
x = torch.linspace(-1, 1, d, dtype=torch.float64)
fan = 2.5 * x * x.abs()
macro = solve_bias(fan[:, None], s[li])[0]
bias = (fan + macro).repeat(h)
attn.w_lora.lora[2].bias.copy_(bias.to(attn.w_lora.lora[2].bias))
a_log = -tau.log() - s[li].log()
attn.decay_log_scale.copy_(a_log.to(attn.decay_log_scale))
attn.decay_log_scale.requires_grad_(False)
@torch.no_grad()
def design_mamba2(model, gdn_state, tau_max=4096.):
"""Transfer the GDN static Softplus profile; preserve the global bias scale."""
layers = _layers(model)
keys = sorted(k for k in gdn_state if k.endswith('.dt_bias'))
# Numerical layer order is intentional for newly designed models.
keys.sort(key=lambda k: int(k.split('layers.')[1].split('.')[0]))
source = torch.stack([F.softplus(gdn_state[k].double()).mean() for k in keys])
multipliers = F.interpolate(source[None,None], size=len(layers), mode='linear',
align_corners=True).flatten()
multipliers /= multipliers.log().mean().exp()
native = torch.stack([F.softplus(layer.mixer.dt_bias.detach().double().cpu()).mean()
for layer in layers])
multipliers /= (native * multipliers).mean() / native.mean()
for li, layer in enumerate(layers):
mixer = layer.mixer
h = mixer.A_log.numel()
a = torch.linspace(0, math.log(tau_max), h, dtype=torch.float64)
dt = F.softplus(mixer.dt_bias.detach().double()) * multipliers[li].to(mixer.dt_bias.device)
mixer.A_log.copy_(a.reshape_as(mixer.A_log).to(mixer.A_log))
mixer.dt_bias.copy_(inverse_softplus(dt).to(mixer.dt_bias))
def initialize(model, architecture, *, scheme='dami', mode="paper", calibration_batches=None,
gdn_parameters=None, tau_min=1., tau_max=4096., init_seed=42):
"""Initialize a FRESH model with Random, PoST, or DAMI; return memory tensors.
paper: use the included exact-shape 100M parameters for all 15 conditions.
calibrate: DAMI GDN with real packed calibration inputs.
design: fresh-size Random/PoST, or DAMI architecture adaptations.
"""
if architecture not in ARCHITECTURES or scheme not in INITIALIZATIONS:
raise ValueError('Unknown architecture or initialization')
if scheme != 'dami' and mode == 'calibrate':
raise ValueError('Activation calibration belongs to DAMI, not Random/PoST')
if mode == "paper":
expected = json.loads((ROOT/f"configs/{architecture}_100m.json").read_text(encoding='utf-8'))
for name in ['hidden_size','num_hidden_layers','head_dim','num_heads','vocab_size']:
if name in expected and getattr(model.config,name,None) != expected[name]:
raise ValueError(f"Paper parameters require {name}={expected[name]}; use a new design for other sizes")
load_memory(model, paper_parameters(architecture, scheme))
elif mode == 'design' and scheme in ('random', 'post'):
design_control(model, architecture, scheme, gdn_parameters, tau_min, tau_max, init_seed)
elif architecture == "gdn" and mode == "calibrate":
h = _layers(model)[0].attn.A_log.numel()
tau,s = targets(len(_layers(model)),h,tau_min,tau_max)
model.dami_calibration_report = calibrate_gdn(model,calibration_batches,tau,s)
elif mode == "design":
gdn_parameters = paper_parameters('gdn') if gdn_parameters is None else gdn_parameters
if architecture in ('kda','gdn2'):
transfer_gdn_to_channels(model,gdn_parameters)
elif architecture == 'mamba2':
design_mamba2(model,gdn_parameters,tau_max)
elif architecture == 'rwkv7':
design_rwkv7(model,tau_min,tau_max)
else:
raise ValueError("GDN requires real calibration inputs; use mode='calibrate'")
else:
raise ValueError("Unsupported mode/architecture combination")
return memory_parameters(model)
@torch.no_grad()
def design_control(model, architecture, scheme, gdn_parameters=None,
tau_min=1., tau_max=4096., init_seed=42):
"""Paper-style controls on a fresh model, keeping the common non-memory base."""
if architecture in ('kda', 'gdn2'):
if gdn_parameters is None:
# Use a same-geometry GDN Random reference for the channel baselines.
config = model.config.to_dict()
config.pop('model_type', None)
reference = build_model('gdn', config, init_seed)
gdn_parameters = memory_parameters(reference)
del reference
transfer_gdn_to_channels(model, gdn_parameters)
if architecture == 'rwkv7':
for layer in _layers(model):
layer.attn.decay_log_scale.fill_(-.5)
layer.attn.decay_log_scale.requires_grad_(False)
if scheme == 'random':
return
for layer in _layers(model):
if architecture == 'rwkv7':
attn = layer.attn
bias = attn.w_lora.lora[2].bias.detach().double()
sigma = bias.sigmoid().reshape(attn.num_heads, attn.head_dim).mean(-1)
tau, _ = targets(len(_layers(model)), attn.num_heads, tau_min, tau_max)
value = -tau.to(sigma.device).log() - sigma.log()
attn.decay_log_scale.copy_(value.to(attn.decay_log_scale))
else:
mixer = layer.mixer if architecture == 'mamba2' else layer.attn
ladder = torch.linspace(0., math.log(tau_max), mixer.A_log.numel(),
device=mixer.A_log.device, dtype=mixer.A_log.dtype)
mixer.A_log.copy_(ladder.reshape_as(mixer.A_log))
def export_parameters(parameters, output_dir, architecture, metadata=None, *, scheme='dami'):
from safetensors.torch import save_file
output = Path(output_dir)
output.mkdir(parents=True, exist_ok=True)
if scheme not in INITIALIZATIONS:
raise ValueError('Unknown initialization')
file = output/'memory_parameters.safetensors'
state = {k:v.detach().cpu().contiguous() for k,v in parameters.items()}
if any(not _is_memory(k) for k in state):
raise ValueError('Export accepts memory tensors only, not a full model state dict')
if not state or any(not torch.isfinite(v).all() for v in state.values()):
raise ValueError("Empty or non-finite memory parameter export")
label = {'random':'Random', 'post':'PoST', 'dami':'DAMI'}[scheme]
save_file(state,str(file),metadata={"method":label,"architecture":architecture,'initialization':scheme})
manifest = {"method":label, 'initialization':scheme, "architecture":architecture, "format_version":2,
"sha256":hashlib.sha256(file.read_bytes()).hexdigest(),
"tensors":{k:{"shape":list(v.shape),"dtype":str(v.dtype)} for k,v in state.items()},
"metadata":metadata or {}}
(output/'initialization.json').write_text(json.dumps(manifest,indent=2,default=str)+'\n',encoding='utf-8')
return file
def _token_batches(path, batch_size):
rows = []
with Path(path).open(encoding='utf-8') as f:
for line in f:
ids = json.loads(line)['input_ids']
rows.append(torch.tensor(ids,dtype=torch.long))
if not rows or len({len(row) for row in rows}) != 1:
raise ValueError("Calibration JSONL must contain equal-length, nonempty input_ids")
return [torch.stack(rows[i:i+batch_size]) for i in range(0,len(rows),batch_size)]
def _packed_examples(files, tokenizer_path, seq_len):
"""Minimal input adapter; datasets and Transformers own IO/training."""
from datasets import load_dataset
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
dataset = load_dataset('parquet',data_files=list(files),split='train',streaming=True)
buffer = []
for row in dataset:
ids = tokenizer(row['text'],add_special_tokens=False)['input_ids']
# Match the calibration packer: concatenate documents without separators.
buffer.extend(ids)
while len(buffer) >= seq_len:
block,buffer = buffer[:seq_len],buffer[seq_len:]
yield {'input_ids':block,'labels':block.copy()}
def _training_defaults(args):
if args.batch_size is None:
args.batch_size = 8 if args.architecture in ('mamba2','rwkv7') else 16
if args.warmup_steps is None:
args.warmup_steps = 2000 if args.architecture in ('mamba2','rwkv7') else 1000
if args.seed is None:
args.seed = 43 if (args.architecture,args.initialization)==('rwkv7','dami') else 42
return args
def _run_matrix(p, args, argv):
import copy
import os
import subprocess
import sys
if int(os.environ.get('WORLD_SIZE','1')) > 1:
p.error('--all is a single-process orchestrator; use torchrun for individual combinations')
if args.architecture is not None or any(a.startswith('--initialization') for a in argv):
p.error('--all selects all architectures and initializations; do not also select one combination')
if args.config or args.model_size.upper()!='100M' or args.init_mode!='paper' or args.memory_file or args.calibration_jsonl or args.gdn_parameters or args.resume_from_checkpoint:
p.error('--all uses the 15 bundled 100M paper presets; use individual runs for custom initialization/configuration/resume')
common=[]
i=0
while i<len(argv):
value=argv[i]
if value=='--all':i+=1;continue
if value=='--output-dir':i+=2;continue
if value.startswith('--output-dir='):i+=1;continue
common.append(value);i+=1
runs=[]
root=Path(args.output_dir)
visible_device=os.environ.get('CUDA_VISIBLE_DEVICES','0').split(',')[0].strip()
for arch in ARCHITECTURES:
for scheme in INITIALIZATIONS:
child=copy.copy(args)
child.architecture=arch;child.initialization=scheme
_training_defaults(child)
directory=root/arch/scheme
runs.append({'architecture':arch,'initialization':scheme,'seed':child.seed,
'batch_size':child.batch_size,'warmup_steps':child.warmup_steps,
'cuda_visible_devices':visible_device,
'output_dir':str(directory),'status':'pending',
'command':[sys.executable,str(ROOT/'train.py'),*common,
'--architecture',arch,'--initialization',scheme,
'--output-dir',str(directory)]})
if args.dry_run:
print(json.dumps({'count':15,'runs':runs},indent=2));return
if any((Path(run['output_dir'])/'final').exists() for run in runs):
p.error('A planned run already has a final model. Select a new output root or resume individual runs.')
root.mkdir(parents=True,exist_ok=True)
journal=root/'matrix.json'
def save():
journal.write_text(json.dumps({'count':15,'runs':runs},indent=2)+'\n',encoding='utf-8')
save()
for index,run in enumerate(runs,1):
print(f"[{index}/15] {run['architecture']} / {run['initialization']}",flush=True)
run['status']='running';save()
try:
child_env=dict(os.environ,CUDA_VISIBLE_DEVICES=visible_device)
subprocess.run(run['command'],check=True,env=child_env)
except BaseException:
run['status']='failed';save();raise
run['status']='completed';save()
def training_main(argv=None):
"""One entry point for five architectures times three initializations."""
import sys
argv = list(sys.argv[1:] if argv is None else argv)
p = argparse.ArgumentParser(description='Train Random, PoST or DAMI on five architectures')
p.add_argument('--architecture',choices=ARCHITECTURES)
p.add_argument('--initialization',choices=INITIALIZATIONS,default='dami')
p.add_argument('--all',action='store_true',help='Run all 15 paper presets sequentially')
p.add_argument('--train-files',nargs='+',required=True,help='Explicit local parquet paths or glob patterns')
p.add_argument('--tokenizer',required=True,help='HF identifier or local tokenizer directory')
p.add_argument('--output-dir',required=True)
p.add_argument('--config',type=Path)
p.add_argument('--model-size',default='100M',help='Only 100M has a bundled paper configuration')
p.add_argument('--init-mode',choices=['paper','calibrate','design'],default='paper')
p.add_argument('--calibration-jsonl',type=Path)
p.add_argument('--memory-file',type=Path,help='Load an exported memory safetensors file labeled for the selected initialization')
p.add_argument('--gdn-parameters',type=Path)
p.add_argument('--init-seed',type=int,default=42)
p.add_argument('--seed',type=int)
p.add_argument('--steps',type=int,default=9537)
p.add_argument('--seq-len',type=int,default=4096)
p.add_argument('--batch-size',type=int)
p.add_argument('--gradient-accumulation',type=int)
p.add_argument('--lr',type=float,default=1e-3)
p.add_argument('--warmup-steps',type=int)
p.add_argument('--save-steps',type=int,default=1000)
p.add_argument('--resume-from-checkpoint')
p.add_argument('--dry-run',action='store_true',help='Print resolved recipe without constructing a model')
args = p.parse_args(argv)
if args.all:
return _run_matrix(p,args,argv)
if args.architecture is None:
p.error('Choose --architecture, or --all for the complete matrix')
_training_defaults(args)
architecture=args.architecture
if args.initialization!='dami' and args.init_mode=='calibrate':
p.error('Only DAMI uses activation calibration')
import os
import glob
world = int(os.environ.get('WORLD_SIZE','1'))
accum = args.gradient_accumulation
if accum is None:
if 128 % (args.batch_size * world):
p.error('Set --gradient-accumulation explicitly; global batch 128 is not divisible by microbatch * GPUs')
accum = 128 // (args.batch_size * world)
if min(accum,args.steps,args.seq_len,args.batch_size) < 1:
p.error('Batch, accumulation, steps and sequence length must be positive')
if args.model_size.upper() != '100M' and args.config is None:
p.error('A different size requires --config with explicit layer/head dimensions')
recipe = dict(vars(args),gradient_accumulation=accum,
global_batch=args.batch_size*world*accum,
tokens=args.batch_size*world*accum*args.steps*args.seq_len)
if args.dry_run:
print(json.dumps(recipe,default=str,indent=2));return
from safetensors.torch import load_file
from transformers import Trainer, TrainingArguments, AutoTokenizer, default_data_collator, set_seed
from datasets import IterableDataset
if not torch.cuda.is_available():
raise RuntimeError('FLA training requires an NVIDIA GPU and Linux CUDA environment')
if world==1 and torch.cuda.device_count()>1:
p.error('Single-process training needs one visible GPU. Set CUDA_VISIBLE_DEVICES, or use torchrun for distributed training.')
files = sorted({str(Path(f).resolve()) for pattern in args.train_files for f in glob.glob(pattern)})
if not files:
p.error('No training parquet files matched --train-files')
local_rank = int(os.environ.get('LOCAL_RANK','0'))
torch.cuda.set_device(local_rank)
config = json.loads(args.config.read_text(encoding='utf-8')) if args.config else None
model = build_model(architecture,config,args.init_seed).to(f'cuda:{local_rank}')
if architecture in ('mamba2','rwkv7') or args.init_mode == 'calibrate':
model.to(torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer)
if len(tokenizer) != model.config.vocab_size:
raise ValueError(f'Tokenizer vocabulary {len(tokenizer)} differs from model vocabulary {model.config.vocab_size}')
if args.memory_file:
manifest=args.memory_file.parent/'initialization.json'
if manifest.is_file():
data=json.loads(manifest.read_text(encoding='utf-8'))
tag=data.get('initialization',data.get('method','').lower())
if tag and tag!=args.initialization:
raise ValueError(f'Memory file is labeled {tag}, not {args.initialization}')
load_memory(model,load_file(str(args.memory_file)))
else:
calibration = _token_batches(args.calibration_jsonl,8) if args.calibration_jsonl else None
gdn = load_file(str(args.gdn_parameters)) if args.gdn_parameters else None
initialize(model,architecture,scheme=args.initialization,mode=args.init_mode,
calibration_batches=calibration,gdn_parameters=gdn,init_seed=args.init_seed)
# The GDN calibration runs in bf16; the GDN-family trainer keeps fp32 masters.
if architecture in ('gdn','kda','gdn2'):
model.float()
model.config.use_cache = False
# Train-mode RNG is separate from the common step-zero initialization seed.
set_seed(args.seed)
rank = int(os.environ.get('RANK','0'))
output = Path(args.output_dir)
if rank == 0:
export_parameters(memory_parameters(model),output/'initialization',architecture,recipe,scheme=args.initialization)
model.config.save_pretrained(output/'initialization')
(output/'recipe.json').write_text(json.dumps(recipe,default=str,indent=2)+'\n',encoding='utf-8')
data = IterableDataset.from_generator(_packed_examples,gen_kwargs={
'files':files,'tokenizer_path':args.tokenizer,'seq_len':args.seq_len})
# Respect FLA's no-weight-decay markers for A_log and dt_bias.
groups = [[],[]]
for name,value in model.named_parameters():
if value.requires_grad:
no_decay = value.ndim < 2 or getattr(value,'_no_weight_decay',False) or name.endswith('bias')
groups[int(no_decay)].append(value)
optimizer = torch.optim.AdamW([{'params':groups[0],'weight_decay':.1},
{'params':groups[1],'weight_decay':0.}],
lr=args.lr,betas=(.9,.95),eps=1e-15)
def lr_factor(step):
if step < args.warmup_steps:
return step / max(1,args.warmup_steps)
progress = min(1.,(step-args.warmup_steps)/max(1,args.steps-args.warmup_steps))
return .1 + .9 * .5 * (1 + math.cos(math.pi*progress))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer,lr_factor)
options = TrainingArguments(output_dir=str(output),max_steps=args.steps,
per_device_train_batch_size=args.batch_size,gradient_accumulation_steps=accum,
bf16=True,learning_rate=args.lr,max_grad_norm=1.,logging_steps=10,
save_steps=args.save_steps,save_total_limit=2,report_to='none',seed=args.seed,
data_seed=args.seed,dataloader_num_workers=0,remove_unused_columns=False,
ddp_find_unused_parameters=False,torch_compile=False)
trainer = Trainer(model=model,args=options,train_dataset=data,
data_collator=default_data_collator,optimizers=(optimizer,scheduler))
trainer.train(resume_from_checkpoint=args.resume_from_checkpoint)
trainer.save_model(str(output/'final'))
if rank == 0:
tokenizer.save_pretrained(output/'final')
def main():
p = argparse.ArgumentParser(description='Export Random/PoST/DAMI parameters or calibrate a new DAMI model')
p.add_argument('--architecture',choices=ARCHITECTURES,required=True)
p.add_argument('--initialization',choices=INITIALIZATIONS,default='dami')
p.add_argument('--output-dir',required=True)
p.add_argument('--mode',choices=['paper','calibrate','design'],default='paper')
p.add_argument('--model-size',default='100M',help='Other sizes require an explicit --config')
p.add_argument('--config',type=Path)
p.add_argument('--calibration-jsonl',type=Path)
p.add_argument('--gdn-parameters',type=Path)
p.add_argument('--device',default='cuda')
p.add_argument('--dtype',choices=['float32','bfloat16'],default='bfloat16')
p.add_argument('--tau-min',type=float,default=1.)
p.add_argument('--tau-max',type=float,default=4096.)
args = p.parse_args()
if args.model_size.upper() != '100M' and args.config is None:
p.error('A different size requires --config with explicit model dimensions')
if args.mode == 'paper':
if args.config or args.calibration_jsonl or args.tau_min != 1. or args.tau_max != 4096.:
p.error('Paper mode exports the fixed 100M archive; custom configurations require calibrate/design')
params = paper_parameters(args.architecture,args.initialization)
metadata = {'mode':'paper','model_size':'approximately 100M',
'reference':'reference/100m.json' if args.initialization=='dami' else 'reference/controls_100m.json'}
else:
config = json.loads(args.config.read_text(encoding='utf-8')) if args.config else None
model = build_model(args.architecture,config).to(device=args.device,dtype=getattr(torch,args.dtype))
batches = _token_batches(args.calibration_jsonl,8) if args.calibration_jsonl else None
from safetensors.torch import load_file
gdn = load_file(str(args.gdn_parameters)) if args.gdn_parameters else None
params = initialize(model,args.architecture,scheme=args.initialization,mode=args.mode,calibration_batches=batches,
gdn_parameters=gdn,tau_min=args.tau_min,tau_max=args.tau_max)
metadata = {'mode':args.mode,'config':model.config.to_dict(),
'calibration':getattr(model,'dami_calibration_report',None)}
print(export_parameters(params,args.output_dir,args.architecture,metadata,scheme=args.initialization))
if __name__ == '__main__':
main()