File size: 31,922 Bytes
f1f9265
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0

import math
from typing import Callable, Optional, Tuple

import numpy as np
import torch
from came_pytorch import CAME
from mmcv import Config
from mmcv.runner import OPTIMIZER_BUILDERS, OPTIMIZERS, DefaultOptimizerConstructor
from mmcv.runner import build_optimizer as mm_build_optimizer
from mmcv.utils import _BatchNorm, _InstanceNorm
from torch.nn import GroupNorm, LayerNorm
from torch.optim.optimizer import Optimizer

from .logger import get_root_logger


def auto_scale_lr(effective_bs, optimizer_cfg, rule="linear", base_batch_size=256):
    assert rule in ["linear", "sqrt"]
    logger = get_root_logger()
    # scale by world size
    if rule == "sqrt":
        scale_ratio = math.sqrt(effective_bs / base_batch_size)
    elif rule == "linear":
        scale_ratio = effective_bs / base_batch_size
    optimizer_cfg["lr"] *= scale_ratio
    logger.info(f'Automatically adapt lr to {optimizer_cfg["lr"]:.5f} (using {rule} scaling rule).')
    return scale_ratio


@OPTIMIZER_BUILDERS.register_module()
class MyOptimizerConstructor(DefaultOptimizerConstructor):
    def add_params(self, params, module, prefix="", is_dcn_module=None):
        """Add all parameters of module to the params list.

        The parameters of the given module will be added to the list of param
        groups, with specific rules defined by paramwise_cfg.

        Args:
            params (list[dict]): A list of param groups, it will be modified
                in place.
            module (nn.Module): The module to be added.
            prefix (str): The prefix of the module

        """
        # get param-wise options
        custom_keys = self.paramwise_cfg.get("custom_keys", {})
        # first sort with alphabet order and then sort with reversed len of str
        # sorted_keys = sorted(sorted(custom_keys.keys()), key=len, reverse=True)

        bias_lr_mult = self.paramwise_cfg.get("bias_lr_mult", 1.0)
        bias_decay_mult = self.paramwise_cfg.get("bias_decay_mult", 1.0)
        norm_decay_mult = self.paramwise_cfg.get("norm_decay_mult", 1.0)
        bypass_duplicate = self.paramwise_cfg.get("bypass_duplicate", False)

        # special rules for norm layers and depth-wise conv layers
        is_norm = isinstance(module, (_BatchNorm, _InstanceNorm, GroupNorm, LayerNorm))

        for name, param in module.named_parameters(recurse=False):
            base_lr = self.base_lr
            if name == "bias" and not (is_norm or is_dcn_module):
                base_lr *= bias_lr_mult

            # apply weight decay policies
            base_wd = self.base_wd
            if self.base_wd is not None:
                # norm decay
                if is_norm:
                    base_wd *= norm_decay_mult
                # bias lr and decay
                elif name == "bias" and not is_dcn_module:
                    # TODO: current bias_decay_mult will have affect on DCN
                    base_wd *= bias_decay_mult

            param_group = {"params": [param]}
            if not param.requires_grad:
                param_group["requires_grad"] = False
                params.append(param_group)
                continue
            if bypass_duplicate and self._is_in(param_group, params):
                logger = get_root_logger()
                logger.warn(f"{prefix} is duplicate. It is skipped since " f"bypass_duplicate={bypass_duplicate}")
                continue
            # if the parameter match one of the custom keys, ignore other rules
            is_custom = False
            for key in custom_keys:
                if isinstance(key, tuple):
                    scope, key_name = key
                else:
                    scope, key_name = None, key
                if scope is not None and scope not in f"{prefix}":
                    continue
                if key_name in f"{prefix}.{name}":
                    is_custom = True
                    if "lr_mult" in custom_keys[key]:
                        # if 'base_classes' in f'{prefix}.{name}' or 'attn_base' in f'{prefix}.{name}':
                        #     param_group['lr'] = self.base_lr
                        # else:
                        param_group["lr"] = self.base_lr * custom_keys[key]["lr_mult"]
                    elif "lr" not in param_group:
                        param_group["lr"] = base_lr
                    if self.base_wd is not None:
                        if "decay_mult" in custom_keys[key]:
                            param_group["weight_decay"] = self.base_wd * custom_keys[key]["decay_mult"]
                        elif "weight_decay" not in param_group:
                            param_group["weight_decay"] = base_wd

            if not is_custom:
                # bias_lr_mult affects all bias parameters
                # except for norm.bias dcn.conv_offset.bias
                if base_lr != self.base_lr:
                    param_group["lr"] = base_lr
                if base_wd != self.base_wd:
                    param_group["weight_decay"] = base_wd
            params.append(param_group)

        for child_name, child_mod in module.named_children():
            child_prefix = f"{prefix}.{child_name}" if prefix else child_name
            self.add_params(params, child_mod, prefix=child_prefix, is_dcn_module=is_dcn_module)


def build_optimizer(model, optimizer_cfg):
    # default parameter-wise config
    logger = get_root_logger()

    if hasattr(model, "module"):
        model = model.module
    # set optimizer constructor
    optimizer_cfg.setdefault("constructor", "MyOptimizerConstructor")
    # parameter-wise setting: cancel weight decay for some specific modules
    custom_keys = dict()
    for name, module in model.named_modules():
        if hasattr(module, "zero_weight_decay"):
            custom_keys.update({(name, key): dict(decay_mult=0) for key in module.zero_weight_decay})

    paramwise_cfg = Config(dict(cfg=dict(custom_keys=custom_keys)))
    given_cfg = optimizer_cfg.get("paramwise_cfg")
    if given_cfg:
        paramwise_cfg.merge_from_dict(dict(cfg=given_cfg))
    optimizer_cfg["paramwise_cfg"] = paramwise_cfg.cfg
    # build optimizer
    optimizer = mm_build_optimizer(model, optimizer_cfg)

    weight_decay_groups = dict()
    lr_groups = dict()
    for group in optimizer.param_groups:
        if not group.get("requires_grad", True):
            continue
        lr_groups.setdefault(group["lr"], []).append(group)
        weight_decay_groups.setdefault(group["weight_decay"], []).append(group)

    learnable_count, fix_count = 0, 0
    for p in model.parameters():
        if p.requires_grad:
            learnable_count += 1
        else:
            fix_count += 1
    fix_info = f"{learnable_count} are learnable, {fix_count} are fix"
    lr_info = "Lr group: " + ", ".join([f"{len(group)} params with lr {lr:.5f}" for lr, group in lr_groups.items()])
    wd_info = "Weight decay group: " + ", ".join(
        [f"{len(group)} params with weight decay {wd}" for wd, group in weight_decay_groups.items()]
    )
    opt_info = f"{optimizer.__class__.__name__} Optimizer: total {len(optimizer.param_groups)} param groups, {fix_info}. {lr_info}; {wd_info}."
    logger.info(opt_info)

    return optimizer


@OPTIMIZERS.register_module()
class Lion(Optimizer):
    def __init__(
        self,
        params,
        lr: float = 1e-4,
        betas: Tuple[float, float] = (0.9, 0.99),
        weight_decay: float = 0.0,
    ):
        assert lr > 0.0
        assert all([0.0 <= beta <= 1.0 for beta in betas])

        defaults = dict(lr=lr, betas=betas, weight_decay=weight_decay)

        super().__init__(params, defaults)

    @staticmethod
    def update_fn(p, grad, exp_avg, lr, wd, beta1, beta2):
        # stepweight decay
        p.data.mul_(1 - lr * wd)

        # weight update
        update = exp_avg.clone().lerp_(grad, 1 - beta1).sign_()
        p.add_(update, alpha=-lr)

        # decay the momentum running average coefficient
        exp_avg.lerp_(grad, 1 - beta2)

    @staticmethod
    def exists(val):
        return val is not None

    @torch.no_grad()
    def step(self, closure: Optional[Callable] = None):

        loss = None
        if self.exists(closure):
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            for p in filter(lambda p: self.exists(p.grad), group["params"]):

                grad, lr, wd, beta1, beta2, state = (
                    p.grad,
                    group["lr"],
                    group["weight_decay"],
                    *group["betas"],
                    self.state[p],
                )

                # init state - exponential moving average of gradient values
                if len(state) == 0:
                    state["exp_avg"] = torch.zeros_like(p)

                exp_avg = state["exp_avg"]

                self.update_fn(p, grad, exp_avg, lr, wd, beta1, beta2)

        return loss


@OPTIMIZERS.register_module()
class CAMEWrapper(torch.optim.Optimizer):
    """Implements CAME algorithm.
    This implementation is based on:
    `CAME: Confidence-guided Adaptive Memory Efficient Optimization`
    Args:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float, optional): external learning rate (default: None)
        eps (tuple[float, float]): regularization constants for square gradient
            and instability respectively (default: (1e-30, 1e-16))
        clip_threshold (float): threshold of root-mean-square of
            final gradient update (default: 1.0)
        betas (tuple[float, float, float]): coefficient used for computing running averages of
        update, square gradient and instability (default: (0.9, 0.999, 0.9999)))
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
    """

    def __init__(
        self,
        params,
        lr=None,
        eps=(1e-30, 1e-16),
        clip_threshold=1.0,
        betas=(0.9, 0.999, 0.9999),
        weight_decay=0.0,
    ):
        assert lr > 0.0
        assert all([0.0 <= beta <= 1.0 for beta in betas])

        defaults = dict(
            lr=lr,
            eps=eps,
            clip_threshold=clip_threshold,
            betas=betas,
            weight_decay=weight_decay,
        )
        super().__init__(params, defaults)

    @property
    def supports_memory_efficient_fp16(self):
        return True

    @property
    def supports_flat_params(self):
        return False

    def _get_options(self, param_shape):
        if len(param_shape) == 4:  # Conv layer
            if param_shape[2] == 1 and param_shape[3] == 1:  # 1x1 conv
                return True, "1x1_conv"
            else:  # 3x3 conv or others
                return False, "conv"
        elif len(param_shape) == 2:  # Linear layer, exactly 2D
            return True, "linear"
        return False, "other"

    def _rms(self, tensor):
        return tensor.norm(2) / (tensor.numel() ** 0.5)

    def _approx_sq_grad(self, exp_avg_sq_row, exp_avg_sq_col):
        r_factor = (exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True)).rsqrt_().unsqueeze(-1)
        c_factor = exp_avg_sq_col.unsqueeze(-2).rsqrt()
        return torch.mul(r_factor, c_factor)

    def step(self, closure=None):
        """Performs a single optimization step.
        Args:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:
            for p in group["params"]:
                if p.grad is None:
                    continue
                grad = p.grad.data
                if grad.dtype in {torch.float16, torch.bfloat16}:
                    grad = grad.float()
                if grad.is_sparse:
                    raise RuntimeError("CAME does not support sparse gradients.")

                state = self.state[p]
                grad_shape = grad.shape

                # factored = self._get_options(grad_shape)
                factored, layer_type = self._get_options(grad_shape)
                # State Initialization
                if len(state) == 0:
                    state["step"] = 0

                    state["exp_avg"] = torch.zeros_like(grad)
                    if factored:
                        if layer_type == "1x1_conv" or layer_type == "linear":
                            # 1x1 conv and linear layers can be handled the same way
                            state["exp_avg_sq_row"] = torch.zeros(grad_shape[0]).type_as(grad)
                            state["exp_avg_sq_col"] = torch.zeros(grad_shape[1]).type_as(grad)
                            state["exp_avg_res_row"] = torch.zeros(grad_shape[0]).type_as(grad)
                            state["exp_avg_res_col"] = torch.zeros(grad_shape[1]).type_as(grad)
                        else:
                            state["exp_avg_sq"] = torch.zeros_like(grad)

                    else:
                        state["exp_avg_sq"] = torch.zeros_like(grad)

                    state["RMS"] = 0

                state["step"] += 1
                state["RMS"] = self._rms(p.data)

                update = (grad**2) + group["eps"][0]
                if factored:
                    exp_avg_sq_row = state["exp_avg_sq_row"]
                    exp_avg_sq_col = state["exp_avg_sq_col"]

                    if layer_type == "1x1_conv" or layer_type == "linear":
                        # Handle dimensions
                        if len(grad_shape) == 4:  # 1x1 conv
                            update_reshaped = update.squeeze(-1).squeeze(-1)  # Remove last two dimensions
                        else:
                            update_reshaped = update

                        exp_avg_sq_row.mul_(group["betas"][1]).add_(
                            update_reshaped.mean(dim=1), alpha=1.0 - group["betas"][1]
                        )
                        exp_avg_sq_col.mul_(group["betas"][1]).add_(
                            update_reshaped.mean(dim=0), alpha=1.0 - group["betas"][1]
                        )

                    # Approximate calculation
                    update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col)
                    if layer_type == "1x1_conv":
                        # Need to reshape back to 4D
                        update = update.view(grad_shape[0], grad_shape[1], 1, 1)
                    update.mul_(grad)
                else:
                    # 3x3 conv or other cases: use standard AdamW approach
                    exp_avg_sq = state["exp_avg_sq"]
                    exp_avg_sq.mul_(group["betas"][1]).add_(update, alpha=1.0 - group["betas"][1])
                    update = exp_avg_sq.rsqrt().mul_(grad)

                update.div_((self._rms(update) / group["clip_threshold"]).clamp_(min=1.0))

                exp_avg = state["exp_avg"]
                exp_avg.mul_(group["betas"][0]).add_(update, alpha=1 - group["betas"][0])

                # Confidence-guided strategy
                # Calculation of instability
                res = (update - exp_avg) ** 2 + group["eps"][1]

                if factored:
                    exp_avg_res_row = state["exp_avg_res_row"]
                    exp_avg_res_col = state["exp_avg_res_col"]

                    if layer_type == "1x1_conv" or layer_type == "linear":
                        # Handle dimensions
                        if len(grad_shape) == 4:  # 1x1 conv
                            res_reshaped = res.squeeze(-1).squeeze(-1)  # Remove last two dimensions
                        else:
                            res_reshaped = res

                        # Update residual statistics
                        exp_avg_res_row.mul_(group["betas"][2]).add_(
                            res_reshaped.mean(dim=1), alpha=1.0 - group["betas"][2]
                        )
                        exp_avg_res_col.mul_(group["betas"][2]).add_(
                            res_reshaped.mean(dim=0), alpha=1.0 - group["betas"][2]
                        )

                    # Approximate calculation
                    res_approx = self._approx_sq_grad(exp_avg_res_row, exp_avg_res_col)
                    if layer_type == "1x1_conv":
                        # 需要reshape回4D
                        res_approx = res_approx.view(grad_shape[0], grad_shape[1], 1, 1)
                    update = res_approx.mul_(exp_avg)
                else:
                    update = exp_avg.clone()

                if group["weight_decay"] != 0:
                    p.data.add_(p.data, alpha=-group["weight_decay"] * group["lr"])

                update.mul_(group["lr"])
                p.data.add_(-update)

        return loss


@OPTIMIZERS.register_module()
class CAME8BitWrapper(torch.optim.Optimizer):
    """Implements 8bit-CAME algorithm.

    Args:
        params (iterable): parameters to optimize or dicts defining parameter groups
        lr (float, optional): external learning rate (default: None)
        eps (tuple[float, float]): regularization constants for square gradient
            and instability respectively (default: (1e-30, 1e-16))
        clip_threshold (float): threshold of root-mean-square of
            final gradient update (default: 1.0)
        betas (tuple[float, float, float]): coefficient used for computing running averages of
            update, square gradient and instability (default: (0.9, 0.999, 0.9999)))
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
        block_size (int): quantization block size, larger memory efficiency, but may reduce accuracy
        min_8bit_size (int): minimum parameter size for using 8bit quantization, only layers larger than this value will be quantized

    Note:
        1. Only use 8bit quantization for large Linear layers and 1x1 Conv layers
        2. Keep all statistics (exp_avg_sq_row, etc.) in 32bit to ensure stability
        3. Use simple min-max quantization strategy, quantize each block separately
    """

    def __init__(
        self,
        params,
        lr=None,
        eps=(1e-30, 1e-16),
        clip_threshold=1.0,
        betas=(0.9, 0.999, 0.9999),
        weight_decay=0.0,
        block_size=2048,
        min_8bit_size=16384,
    ):
        assert lr > 0.0
        assert all([0.0 <= beta <= 1.0 for beta in betas])

        logger = get_root_logger()
        logger.info(f"Initializing CAME8bit with block_size={block_size}, min_8bit_size={min_8bit_size}")

        defaults = dict(
            lr=lr,
            eps=eps,
            clip_threshold=clip_threshold,
            betas=betas,
            weight_decay=weight_decay,
            block_size=block_size,
            min_8bit_size=min_8bit_size,
        )
        super().__init__(params, defaults)

    def print_layer_info(self, param_shape, use_8bit):
        """Print layer information, including parameter size and whether 8bit quantization is used

        Args:
            param_shape (tuple): parameter shape
            use_8bit (bool): whether 8bit quantization is used
        """
        size = np.prod(param_shape)
        layer_type = "unknown"
        if len(param_shape) == 1:
            layer_type = "1D Layer"
        elif len(param_shape) == 2:
            layer_type = "Linear"
        elif len(param_shape) == 4:
            if param_shape[2] == 1 and param_shape[3] == 1:
                layer_type = "1x1 Conv"
            else:
                layer_type = "Conv"

        status = "8bit" if use_8bit else "32bit"
        print(f"{layer_type} layer with shape {param_shape}: {size:,} params -> using {status}")

    def _should_use_8bit(self, param_shape):
        """Determine if a parameter should be quantized to 8bit

        Rules:
        1. linear layers: parameter size > min_8bit_size
        2. 1x1 conv layers: parameter size > min_8bit_size
        3. other layers: use 32bit
        """
        if len(param_shape) == 2:  # linear layer
            return param_shape[0] * param_shape[1] > self.defaults["min_8bit_size"]
        elif len(param_shape) == 4 and param_shape[2] == 1 and param_shape[3] == 1:
            return param_shape[0] * param_shape[1] > self.defaults["min_8bit_size"]
        return False  # other layers are not quantized

    def _quantize_state(self, state_tensor, block_size=2048):
        """Quantize a state tensor to 8bit

        Args:
            state_tensor: tensor to be quantized
            block_size: quantization block size

        Returns:
            list of quantized data blocks, each block contains:
            - data: uint8 data
            - scale: quantization scale
            - min: minimum value
        """
        if state_tensor.numel() <= 1:
            return state_tensor

        quantized_chunks = []
        for chunk in state_tensor.split(block_size):
            # Calculate quantization parameters
            chunk_min = chunk.min()
            chunk_max = chunk.max()
            scale = (chunk_max - chunk_min) / 255

            # Quantize to 0-255 range
            quantized_chunk = ((chunk - chunk_min) / scale).round().byte()
            quantized_chunks.append({"data": quantized_chunk, "scale": scale, "min": chunk_min})
        return quantized_chunks

    def _dequantize_state(self, quantized_chunks):
        """Dequantize 8bit quantized data to 32bit float

        Args:
            quantized_chunks: list of quantized data blocks

        Returns:
            dequantized 32bit float tensor
        """
        if not isinstance(quantized_chunks, list):
            return quantized_chunks

        chunks = []
        for chunk_dict in quantized_chunks:
            # Dequantize: value = data * scale + min
            chunk = chunk_dict["data"].float() * chunk_dict["scale"] + chunk_dict["min"]
            chunks.append(chunk)
        return torch.cat(chunks)

    def _dequantize_state_first_step(self, quantized_chunks):
        """Efficient dequantization for the first step"""
        if not isinstance(quantized_chunks, list):
            return quantized_chunks

        # 1. Dequantize all chunks to CPU
        dequantized_chunks = []
        for chunk_dict in quantized_chunks:
            chunk = chunk_dict["data"].float() * chunk_dict["scale"] + chunk_dict["min"]
            dequantized_chunks.append(chunk)
            del chunk_dict["data"]
            torch.cuda.empty_cache()

        # 2. Concatenate all chunks
        result = torch.cat(dequantized_chunks)

        del dequantized_chunks
        torch.cuda.empty_cache()

        return result

    def _get_options(self, param_shape):
        if len(param_shape) == 4:
            if param_shape[2] == 1 and param_shape[3] == 1:
                return True, "1x1_conv"
            else:
                return False, "conv"
        elif len(param_shape) == 2:
            return True, "linear"
        return False, "other"

    def _rms(self, tensor):
        return tensor.norm(2) / (tensor.numel() ** 0.5)

    def _approx_sq_grad(self, exp_avg_sq_row, exp_avg_sq_col):
        r_factor = (exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True)).rsqrt_().unsqueeze(-1)
        c_factor = exp_avg_sq_col.unsqueeze(-2).rsqrt()
        return torch.mul(r_factor, c_factor)

    def step(self, closure=None):
        """Perform a single optimization step

        Main steps:
        1. Determine if 8bit quantization is needed
        2. Update first and second moment estimates
        3. Compute update step
        4. Apply confidence-guided strategy
        """
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:
            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad.data
                if grad.dtype in {torch.float16, torch.bfloat16}:
                    grad = grad.float()
                if grad.is_sparse:
                    raise RuntimeError("CAME8bit does not support sparse gradients.")

                state = self.state[p]
                grad_shape = grad.shape
                factored, layer_type = self._get_options(grad_shape)

                # Determine if 8bit quantization is used
                use_8bit = self._should_use_8bit(grad_shape)

                # State Initialization
                if len(state) == 0:
                    self.print_layer_info(grad_shape, use_8bit)

                    state["step"] = 0
                    # Only use 8bit quantization for large matrices
                    if use_8bit:
                        state["exp_avg"] = self._quantize_state(torch.zeros_like(grad), group["block_size"])
                    else:
                        state["exp_avg"] = torch.zeros_like(grad)

                    if factored:
                        if layer_type == "1x1_conv" or layer_type == "linear":
                            # Keep row and column statistics in 32bit
                            state["exp_avg_sq_row"] = torch.zeros(grad_shape[0]).type_as(grad)
                            state["exp_avg_sq_col"] = torch.zeros(grad_shape[1]).type_as(grad)
                            state["exp_avg_res_row"] = torch.zeros(grad_shape[0]).type_as(grad)
                            state["exp_avg_res_col"] = torch.zeros(grad_shape[1]).type_as(grad)
                        else:
                            if use_8bit:
                                state["exp_avg_sq"] = self._quantize_state(torch.zeros_like(grad), group["block_size"])
                            else:
                                state["exp_avg_sq"] = torch.zeros_like(grad)
                    else:
                        if use_8bit:
                            state["exp_avg_sq"] = self._quantize_state(torch.zeros_like(grad), group["block_size"])
                        else:
                            state["exp_avg_sq"] = torch.zeros_like(grad)
                    state["RMS"] = 0

                state["step"] += 1
                state["RMS"] = self._rms(p.data)

                exp_avg = self._dequantize_state(state["exp_avg"]) if use_8bit else state["exp_avg"]

                update = (grad**2) + group["eps"][0]
                if factored:
                    exp_avg_sq_row = state["exp_avg_sq_row"]  # 32bit
                    exp_avg_sq_col = state["exp_avg_sq_col"]  # 32bit

                    if layer_type == "1x1_conv" or layer_type == "linear":
                        if len(grad_shape) == 4:
                            update_reshaped = update.squeeze(-1).squeeze(-1)
                        else:
                            update_reshaped = update

                        # Update row and column statistics
                        exp_avg_sq_row.mul_(group["betas"][1]).add_(
                            update_reshaped.mean(dim=1), alpha=1.0 - group["betas"][1]
                        )
                        exp_avg_sq_col.mul_(group["betas"][1]).add_(
                            update_reshaped.mean(dim=0), alpha=1.0 - group["betas"][1]
                        )

                    update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col)
                    if layer_type == "1x1_conv":
                        update = update.view(grad_shape[0], grad_shape[1], 1, 1)
                    update.mul_(grad)
                else:
                    exp_avg_sq = self._dequantize_state(state["exp_avg_sq"]) if use_8bit else state["exp_avg_sq"]
                    exp_avg_sq.mul_(group["betas"][1]).add_(update, alpha=1.0 - group["betas"][1])
                    if use_8bit:
                        state["exp_avg_sq"] = self._quantize_state(exp_avg_sq, group["block_size"])
                    else:
                        state["exp_avg_sq"] = exp_avg_sq
                    update = exp_avg_sq.rsqrt().mul_(grad)

                # Gradient clipping
                update.div_((self._rms(update) / group["clip_threshold"]).clamp_(min=1.0))

                # Update first moment
                exp_avg.mul_(group["betas"][0]).add_(update, alpha=1 - group["betas"][0])

                # Re-quantize (if needed)
                if use_8bit:
                    state["exp_avg"] = self._quantize_state(exp_avg, group["block_size"])
                else:
                    state["exp_avg"] = exp_avg

                # Confidence-guided strategy
                res = (update - exp_avg) ** 2 + group["eps"][1]

                if factored:
                    exp_avg_res_row = state["exp_avg_res_row"]  # 32bit
                    exp_avg_res_col = state["exp_avg_res_col"]  # 32bit

                    if layer_type == "1x1_conv" or layer_type == "linear":
                        if len(grad_shape) == 4:
                            res_reshaped = res.squeeze(-1).squeeze(-1)
                        else:
                            res_reshaped = res

                        # Update residual statistics
                        exp_avg_res_row.mul_(group["betas"][2]).add_(
                            res_reshaped.mean(dim=1), alpha=1.0 - group["betas"][2]
                        )
                        exp_avg_res_col.mul_(group["betas"][2]).add_(
                            res_reshaped.mean(dim=0), alpha=1.0 - group["betas"][2]
                        )

                    res_approx = self._approx_sq_grad(exp_avg_res_row, exp_avg_res_col)
                    if layer_type == "1x1_conv":
                        res_approx = res_approx.view(grad_shape[0], grad_shape[1], 1, 1)
                    update = res_approx.mul_(exp_avg)
                else:
                    update = exp_avg.clone()

                # Weight decay
                if group["weight_decay"] != 0:
                    p.data.add_(p.data, alpha=-group["weight_decay"] * group["lr"])

                # Apply update
                update.mul_(group["lr"])
                p.data.add_(-update)

        return loss

    def load_state_dict(self, state_dict):
        """Load state dict and convert relevant states to 8bit"""
        super().load_state_dict(state_dict)

        for state in self.state.values():
            for key in [
                "exp_avg",
                "exp_avg_sq",
                "exp_avg_sq_row",
                "exp_avg_sq_col",
                "exp_avg_res_row",
                "exp_avg_res_col",
            ]:
                if key in state:
                    if isinstance(state[key], list):
                        state[key] = [
                            {
                                "data": exp["data"].byte(),  # Convert data to 8bit directly
                                "scale": exp["scale"],  # Keep scale unchanged
                                "min": exp["min"],  # Keep min unchanged
                            }
                            for exp in state[key]
                        ]
                    elif isinstance(state[key], torch.Tensor):
                        # If tensor, keep as 32bit
                        state[key] = state[key].float()  # Ensure 32bit

        del state_dict
        torch.cuda.empty_cache()