File size: 19,648 Bytes
131da64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Console logger utilities.

Copied from https://github.com/HazyResearch/transformers/blob/master/src/utils/utils.py
Copied from https://docs.python.org/3/howto/logging-cookbook.html#using-a-context-manager-for-selective-logging
"""

import math
import os
from pathlib import Path
from typing import List, Optional

import fsspec
import torch
from timm.scheduler import CosineLRScheduler
import omegaconf
import rich
import rich.syntax
import rich.tree

from decoupled_utils import rank_zero_fn, rprint
from decoupled_utils import (get_hostname, get_num_devices, get_tpu_devices, gprint,
                             is_torch_cuda_available, is_torch_xla_available, rprint)


def print_trainable_parameters(model):
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    print(
        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
    )

def fsspec_exists(filename):
    """Check if a file exists using fsspec."""
    fs, _ = fsspec.core.url_to_fs(filename)
    return fs.exists(filename)


def fsspec_listdir(dirname):
    """Listdir in manner compatible with fsspec."""
    fs, _ = fsspec.core.url_to_fs(dirname)
    return fs.ls(dirname)


def fsspec_mkdirs(dirname, exist_ok=True):
    """Mkdirs in manner compatible with fsspec."""
    fs, _ = fsspec.core.url_to_fs(dirname)
    fs.makedirs(dirname, exist_ok=exist_ok)


def print_nans(tensor, name):
    if torch.isnan(tensor).any():
        gprint(f"{name} has nans: {tensor}")


class CosineDecayWarmupLRScheduler(CosineLRScheduler, torch.optim.lr_scheduler._LRScheduler):
    """Wrap timm.scheduler.CosineLRScheduler
    Enables calling scheduler.step() without passing in epoch.
    Supports resuming as well.
    Adapted from:
      https://github.com/HazyResearch/hyena-dna/blob/main/src/utils/optim/schedulers.py
    """

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._last_epoch = -1
        self.step(epoch=0)

    def step(self, epoch=None):
        if epoch is None:
            self._last_epoch += 1
        else:
            self._last_epoch = epoch
        # We call either step or step_update, depending on
        # whether we're using the scheduler every epoch or every
        # step.
        # Otherwise, lightning will always call step (i.e.,
        # meant for each epoch), and if we set scheduler
        # interval to "step", then the learning rate update will
        # be wrong.
        if self.t_in_epochs:
            super().step(epoch=self._last_epoch)
        else:
            super().step_update(num_updates=self._last_epoch)


class Sampler:
    def __init__(self, shape):
        self.shape = shape

    def _sampling_noise(self):
        pass

    def _hard_sample(self, logits):
        pass

    def _soft_sample(self, logits):
        return 0

    def sample(self, logits):
        noise = self._sampling_noise()
        noise = noise[: logits.shape[0], :]
        logits = logits + noise.to(dtype=logits.dtype, device=logits.device)
        hard_sample = self._hard_sample(logits)
        soft_sample = self._soft_sample(logits)
        return soft_sample + (hard_sample - soft_sample).detach()


class TopKSampler(Sampler):
    def __init__(self, k, shape, gamma_tau=1.0):
        super().__init__(shape)
        self.k = k
        self.gamma_tau = gamma_tau
        self.num_betas = 10
        self.sampler = torch.distributions.gamma.Gamma(1 / k * torch.ones(self.num_betas, *self.shape), 1.0)

    def _sampling_noise(self):
        noise = self.sampler.sample()
        beta = self.k / torch.arange(1, self.num_betas + 1, 1, dtype=torch.float32)
        beta = beta[:, None, None]
        assert beta.ndim == noise.ndim
        s = noise / beta
        s = torch.sum(s, axis=0)
        s = s - math.log(10.0)
        s = self.gamma_tau * (s / self.k)
        return s

    def _hard_sample(self, logits):
        assert logits.ndim == 2
        thresholds, _ = torch.sort(logits, dim=-1)
        thresholds = thresholds[:, -self.k][:, None]
        return (logits >= thresholds).type(logits.dtype)

    def _soft_sample(self, logits):
        soft_top_k = logits - torch.mean(logits, dim=-1, keepdim=True)
        return soft_top_k / torch.norm(soft_top_k, dim=-1, keepdim=True)


class DeterministicTopK(TopKSampler):
    def __init__(self, k):
        super().__init__(k, shape=(1, 1))

    def _sampling_noise(self):
        return 0

    def discreize(self, x):
        hard_sample = self._hard_sample(x)
        soft_sample = self._soft_sample(x)
        return soft_sample + (hard_sample - soft_sample).detach()


class GumbelSampler(Sampler):

    def __init__(self, shape, temperature=1.0):
        super().__init__(shape)
        self.temperature = temperature

    def _sampling_noise(self):
        return -(1e-10 - (torch.rand(*self.shape) + 1e-10).log()).log()

    def _hard_sample(self, logits):
        assert logits.ndim == 2
        indices = torch.argmax(logits, dim=-1)
        zeros = logits * 0
        ones = torch.ones_like(logits[:, :, :1])
        return torch.scatter(zeros, -1, indices[:, :, None], ones)

    def _soft_sample(self, logits):
        return torch.nn.functional.softmax(logits / self.temperature, dim=-1)


class BinarySampler(GumbelSampler):

    def sample(self, probs):
        # TODO(subhamsahoo): use the temperature parameter.
        pos_noise = self._sampling_noise().to(dtype=probs.dtype, device=probs.device)
        neg_noise = self._sampling_noise().to(dtype=probs.dtype, device=probs.device)
        del_noise_exp = (neg_noise - pos_noise).exp()
        hard_sample = (probs * (1 + del_noise_exp) > 1).to(probs.dtype)
        soft_sample = probs / (probs + (1 - probs) * del_noise_exp)
        return soft_sample + (hard_sample - soft_sample).detach()


class GaussianSampler:
    def __init__(self):
        self.softplus = torch.nn.Softplus()

    def sample(self, x):
        assert x.ndim == 2
        n = x.shape[-1] // 2
        mu = x[:, :n]
        sigma = self.softplus(x[:, n:]).sqrt()
        return mu + sigma * torch.randn_like(mu)


def is_global_rank_zero():
    """Helper function to determine if the current process is global_rank 0 (the main process)"""
    # Try to get the pytorch RANK env var
    # RANK is set by torch.distributed.launch
    rank = os.environ.get("RANK", None)
    if rank is not None:
        return rank == 0

    # Try to get the SLURM global rank env var
    # SLURM_PROCID is set by SLURM
    slurm_rank = os.environ.get("SLURM_PROCID", None)
    if slurm_rank is not None:
        return slurm_rank == 0

    # Try to get the MPI global rank env var
    mpi_rank = os.environ.get("OMPI_COMM_WORLD_RANK", None)
    if mpi_rank is not None:
        return mpi_rank == 0

    # if neither pytorch, SLURM nor MPI env vars are set
    # check NODE_RANK/GROUP_RANK and LOCAL_RANK env vars
    # assume global_rank is zero if undefined
    node_rank = os.environ.get("NODE_RANK", os.environ.get("GROUP_RANK", 0))
    local_rank = os.environ.get("LOCAL_RANK", 0)
    return node_rank == 0 and local_rank == 0


def get_rank():
    """Helper function that returns torch.distributed.get_rank() if DDP has been initialized otherwise it returns 0."""

    if is_global_rank_zero():
        return 0
    else:
        return torch.distributed.get_rank()


def set_numa_affinity(gpu_index, verbose=False):
    import pynvml as nvml

    nvml.nvmlInit()
    """This util will assign to the current process the cpu cores set that resides on the same NUMA
    node as the GPU. Typically if you have 8 GPUs, then the first 4 are on the first NUMA node and
    the remaining 4 are on the second.

    `gpu_index` is typically the same as `LOCAL_RANK` in the distributed training, but beware that
    `CUDA_VISIBLE_DEVICES` could impact that. e.g. `CUDA_VISIBLE_DEVICES=0,7` won't do the right
    thing - then you will probably want to remap the ids with something like:

    ```
    if "CUDA_VISIBLE_DEVICES" in os.environ:
        ids = list(map(int, os.environ.get("CUDA_VISIBLE_DEVICES", "").split(",")))
        gpu_index = ids[gpu_index] # remap
    ```

    """

    num_elements = math.ceil(os.cpu_count() / 64)
    handle = nvml.nvmlDeviceGetHandleByIndex(gpu_index)
    affinity_string = ""
    for j in nvml.nvmlDeviceGetCpuAffinity(handle, num_elements):
        # assume nvml returns list of 64 bit ints
        affinity_string = f"{j:064b}{affinity_string}"
    affinity_list = [int(x) for x in affinity_string]
    affinity_list.reverse()  # so core 0 is the 0th element
    affinity_to_set = [i for i, e in enumerate(affinity_list) if e != 0]

    if verbose:
        cores = os.sched_getaffinity(0)
        gprint(f"before: {len(cores)} visible cpu cores: {cores}")

    try:
        os.sched_setaffinity(0, affinity_to_set)
    except Exception as e:
        gprint(f"Failed to set affinity: {e}")

    if verbose:
        cores = os.sched_getaffinity(0)
        gprint(f"after: {len(cores)} visible cpu cores: {cores}")


from typing import Dict, Union

import torch
from torch.nn import Module


def grad_norm(module: Module, norm_type: Union[float, int, str], group_separator: str = "/") -> Dict[str, float]:
    """Compute each parameter's gradient's norm and their overall norm.

    The overall norm is computed over all gradients together, as if they
    were concatenated into a single vector.

    Args:
        module: :class:`torch.nn.Module` to inspect.
        norm_type: The type of the used p-norm, cast to float if necessary.
            Can be ``'inf'`` for infinity norm.
        group_separator: The separator string used by the logger to group
            the gradients norms in their own subfolder instead of the logs one.

    Return:
        norms: The dictionary of p-norms of each parameter's gradient and
            a special entry for the total p-norm of the gradients viewed
            as a single vector.

    """
    norm_type = float(norm_type)
    if norm_type <= 0:
        raise ValueError(f"`norm_type` must be a positive number or 'inf' (infinity norm). Got {norm_type}")

    norms = {f"{group_separator}{name}": p.grad.data.norm(norm_type) for name, p in module.named_parameters() if p.grad is not None}
    total_norm = torch.tensor(list(norms.values())).norm(norm_type)
    return norms, total_norm

has_set_omega_conf_resolvers = False

def set_omega_conf_resolvers():
    global has_set_omega_conf_resolvers
    if has_set_omega_conf_resolvers:
        return
    has_set_omega_conf_resolvers = True
    import omegaconf
    from omegaconf import OmegaConf

    def get_dir_name(_root_):
        if str(_root_.mode) == "eval":
            return "eval"
        elif _root_.debug:
            return "debug"
        else:
            return _root_.data.train

    def getpythoncmd(_root_):
        return _root_.python_orig + "--multi_gpu \\\n" if (_root_.trainer.devices * _root_.trainer.num_nodes > 1) else _root_.python_orig

    def custom_batch_size():
        if is_torch_cuda_available() and torch.cuda.get_device_properties(0).total_memory >= 23 * 1024 * 1024 * 1024:
            return 64
        elif is_torch_cuda_available() and torch.cuda.get_device_properties(0).total_memory >= 10 * 1024 * 1024 * 1024:
            return 32
        else:
            return 28

    def get_slurm_name(_root_):
        return _root_.slurm_name if hasattr(_root_, "slurm_name") and _root_.slurm_name is not None else _root_.wandb.project
    
    partition_time_limit_min = {
        "partition_name": 60 * 6,
    }
    
    gpu_constraints = {
        "cluster_name": "gpu_constraints", # e.g. "A5000|A6000"
    }

    partitions = {
        "cluster_name": "partition_name",
    }


    babel_exclude_nodes = set()
    if os.environ.get("BAD_NODES", None) is not None:
        babel_exclude_nodes.update(os.environ.get("BAD_NODES").split(","))

    exclude_nodes = {
        "cluster_name": "nodes_to_exclude",
    }

    def get_hostname_split():
        return get_hostname().split("-")[0].split(".")[0]

    omegaconf.OmegaConf.register_new_resolver("getpythoncmd", getpythoncmd)
    omegaconf.OmegaConf.register_new_resolver("get_dir_name", get_dir_name)
    omegaconf.OmegaConf.register_new_resolver("cwd", os.getcwd)
    omegaconf.OmegaConf.register_new_resolver("device_count", get_num_devices)
    omegaconf.OmegaConf.register_new_resolver("eval", eval)
    omegaconf.OmegaConf.register_new_resolver("div_up", lambda x, y: (x + y - 1) // y)
    omegaconf.OmegaConf.register_new_resolver("find_grad_accum", lambda x, y: round(x / y))
    omegaconf.OmegaConf.register_new_resolver("find_partition", lambda: partitions[get_hostname_split()] if get_hostname_split() in partitions else "all")
    omegaconf.OmegaConf.register_new_resolver("find_constraint", lambda: gpu_constraints[get_hostname_split()] if get_hostname_split() in gpu_constraints else "")
    omegaconf.OmegaConf.register_new_resolver("is_ar", lambda parameterization: parameterization == "ar")
    omegaconf.OmegaConf.register_new_resolver("kv_cache_batch_size", lambda eval_batch_size, cfg: eval_batch_size * 2 if cfg is not None else eval_batch_size)
    omegaconf.OmegaConf.register_new_resolver("exclude_nodes", lambda: exclude_nodes[get_hostname_split()] if get_hostname_split() in exclude_nodes else "")
    omegaconf.OmegaConf.register_new_resolver("get_slurm_name", get_slurm_name)
    

    def adjust_n_blocks(_root_):
        return (
            (_root_.model.base_n_blocks - 1 if _root_.model.base_n_blocks < 24 else _root_.model.base_n_blocks - 2)
            if str(_root_.backbone) == "maskdit"
            else _root_.model.base_n_blocks
        )

    omegaconf.OmegaConf.register_new_resolver("adjust_n_blocks", adjust_n_blocks)
    omegaconf.OmegaConf.register_new_resolver("partition_limit", lambda x: partition_time_limit_min[x] if x in partition_time_limit_min else 60 * 6)
    omegaconf.OmegaConf.register_new_resolver("custom_batch_size", custom_batch_size)
    omegaconf.OmegaConf.register_new_resolver("get_repo_dir", lambda: os.getenv("UNIDISC_DIR", str(Path(__file__).parent)))


@rank_zero_fn
def _print_config(config, resolve: bool = True, save_cfg: bool = True) -> None:
    """Prints content of DictConfig using Rich library and its tree structure.

    Args:
      config (DictConfig): Configuration composed by Hydra.
      resolve (bool): Whether to resolve reference fields of DictConfig.
      save_cfg (bool): Whether to save the configuration tree to a file.
    """

    style = "dim"
    tree = rich.tree.Tree("CONFIG", style=style, guide_style=style)

    fields = config.keys()
    for field in fields:
        branch = tree.add(field, style=style, guide_style=style)

        config_section = config.get(field)
        branch_content = str(config_section)
        if isinstance(config_section, omegaconf.DictConfig):
            branch_content = omegaconf.OmegaConf.to_yaml(config_section, resolve=resolve)

        branch.add(rich.syntax.Syntax(branch_content, "yaml"))

    rich.print(tree)
    if save_cfg:
        with fsspec.open("config_tree.txt", "w") as fp:
            rich.print(tree, file=fp)

def set_torch_defaults(benchmark=True):
    torch.set_float32_matmul_precision("medium")
    if is_torch_cuda_available():
        rprint(f"Setting torch defaults")
        exec("import torch.backends.cuda as cuda")
        exec("import torch.backends.cudnn as cudnn")
        exec("cudnn.enabled = True")
        if benchmark:
            exec("cudnn.benchmark = True")
        else:
            rprint(f"Warning: Not benchmarking")
        exec("cudnn.allow_tf32 = True")
        exec("cuda.matmul.allow_tf32 = True")
        exec("cudnn.deterministic = False")
    else:
        rprint(f"Warning: CUDA not available. Not setting defaults.")

from torch.distributed.elastic.multiprocessing.errors import (ChildFailedError,
                                                              record)
from torch.distributed.elastic.multiprocessing.errors.handlers import \
    get_error_handler

_NOT_AVAILABLE = "<N/A>"
class ErrorHandler:
    def __init__(self, error_handler=None):
        self.error_handler = error_handler or get_error_handler()

    def __enter__(self):
        assert self.error_handler is not None
        self.error_handler.initialize()
        return self

    def __exit__(self, exc_type, exc_value, traceback):
        if exc_type is not None:
            if issubclass(exc_type, SystemExit) and exc_value.code == 0:
                return True  # Prevents SystemExit with code 0 from stopping the program
            elif issubclass(exc_type, ChildFailedError):
                rank, failure = exc_value.get_first_failure()
                if failure.error_file != _NOT_AVAILABLE:
                    self.error_handler.dump_error_file(failure.error_file, failure.exitcode)
                else:
                    rprint(
                        "local_rank %s FAILED with no error file. "
                        "Decorate your entrypoint fn with @record for traceback info. "
                        "See: https://pytorch.org/docs/stable/elastic/errors.html",
                        rank
                    )
                return False  # Re-raises the exception
            self.error_handler.record_exception(exc_value)
        return False  # Any other exceptions will be re-raised


def convert_state_dict_keys(state_dict):
    new_state_dict = {}
    for k, v in state_dict.items():
        if "attn_out" in k:
            new_key = k.replace("attn_out", "attention.attn_out")
        elif "attn_qkv" in k:
            new_key = k.replace("attn_qkv", "attention.attn_qkv")
        else:
            new_key = k
        new_state_dict[new_key] = v
    return new_state_dict

from accelerate.utils import extract_model_from_parallel
def apply_compile(model, **compile_kwargs):
    """
    Apply torch.compile to each TransformerBlock, which makes compilation efficient due to
    repeated structure. Alternatively one can compile the whole model (after applying DP).
    """
    for layer_id, transformer_block in extract_model_from_parallel(model).blocks.named_children():
        transformer_block = torch.compile(transformer_block, **compile_kwargs)
        extract_model_from_parallel(model).blocks.register_module(layer_id, transformer_block)

    output_layer = torch.compile(extract_model_from_parallel(model).output_layer, **compile_kwargs)
    extract_model_from_parallel(model).register_module("output_layer", output_layer)

def compile_model(config, model):
    compile_kwargs = dict()

    if config.backbone == "maskdit":
        compile_kwargs["dynamic"] = True

    compile_kwargs["mode"] = config.trainer.compile_mode
    rprint(f"Using compile mode: {config.trainer.compile_mode}")

    if getattr(config.trainer, "sd3_compile_config", True):
        torch._inductor.config.conv_1x1_as_mm = True
        torch._inductor.config.coordinate_descent_tuning = True
        torch._inductor.config.epilogue_fusion = False
        torch._inductor.config.coordinate_descent_check_all_directions = True
        rprint(f"Using SD3 compile config")

    if config.trainer.compile_fullgraph:
        compile_kwargs["fullgraph"] = True
        rprint(f"Using fullgraph compile")

    if getattr(config.trainer, "compile_per_layer", False):
        apply_compile(model, **compile_kwargs)
    else:
        model = torch.compile(model, **compile_kwargs)

    return model