# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: OpenMDW-1.1 from __future__ import annotations import collections import collections.abc import functools import json import os import random from contextlib import ContextDecorator, nullcontext from dataclasses import fields from typing import Any, Callable, List, Tuple, TypeVar, Union import numpy as np from loguru import logger as logging try: # pyrefly: ignore # import-error import straggler except ImportError: straggler = None import termcolor import torch from torch.distributed._functional_collectives import AsyncCollectiveTensor from torch.distributed._tensor.api import DTensor from cosmos_framework.utils import distributed, log from cosmos_framework.utils.distributed import all_gather_tensor from cosmos_framework.utils.easy_io import easy_io from cosmos_framework.utils.timer import Timer def requires_grad(model: torch.nn.Module, value: bool = True) -> None: """Set a model to require gradients or not. Args: model (torch.nn.Module): Neural network model. value (bool): Whether the network requires gradients or not. """ for p in model.parameters(): p.requires_grad = value def to( data: Any, device: str | torch.device | None = None, dtype: torch.dtype | None = None, memory_format: torch.memory_format = torch.preserve_format, ) -> Any: """Recursively cast data into the specified device, dtype, and/or memory_format. The input data can be a tensor, a list of tensors, a dict of tensors. See the documentation for torch.Tensor.to() for details. Args: data (Any): Input data. device (str | torch.device): GPU device (default: None). dtype (torch.dtype): data type (default: None). memory_format (torch.memory_format): memory organization format (default: torch.preserve_format). Returns: data (Any): Data cast to the specified device, dtype, and/or memory_format. """ assert device is not None or dtype is not None or memory_format is not None, ( "at least one of device, dtype, memory_format should be specified" ) if isinstance(data, torch.Tensor): if ( memory_format == torch.channels_last and data.dim() != 4 or memory_format == torch.channels_last_3d and data.dim() != 5 ): memory_format = torch.preserve_format # do not change the memory format is_cpu = (isinstance(device, str) and device == "cpu") or ( isinstance(device, torch.device) and device.type == "cpu" ) data = data.to( device=device, dtype=dtype, memory_format=memory_format, non_blocking=(not is_cpu), ) return data elif isinstance(data, collections.abc.Mapping): return type(data)({key: to(data[key], device=device, dtype=dtype, memory_format=memory_format) for key in data}) elif isinstance(data, collections.abc.Sequence) and not isinstance(data, (str, bytes)): return type(data)([to(elem, device=device, dtype=dtype, memory_format=memory_format) for elem in data]) else: return data def serialize(data: Any) -> Any: """Serialize data by hierarchically traversing through iterables. Args: data (Any): Input data. Returns: data (Any): Serialized data. """ if isinstance(data, collections.abc.Mapping): return type(data)({key: serialize(data[key]) for key in data}) elif isinstance(data, collections.abc.Sequence) and not isinstance(data, (str, bytes)): return type(data)([serialize(elem) for elem in data]) else: try: json.dumps(data) except TypeError: data = str(data) return data def print_environ_variables(env_vars: list[str]) -> None: """Print a specific list of environment variables. Args: env_vars (list[str]): List of specified environment variables. """ for env_var in env_vars: if env_var in os.environ: log.info(f"Environment variable {Color.green(env_var)}: {Color.yellow(os.environ[env_var])}") else: log.warning(f"Environment variable {Color.green(env_var)} not set!") def set_random_seed(seed: int, by_rank: bool = False) -> None: """Set random seed. This includes random, numpy, Pytorch. Args: seed (int): Random seed. by_rank (bool): if true, each GPU will use a different random seed. """ if by_rank: seed += distributed.get_rank() log.info(f"Using random seed {seed}.") random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) # sets seed on the current CPU & all GPUs def arch_invariant_rand( shape: List[int] | Tuple[int], dtype: torch.dtype, device: str | torch.device, seed: int | None = None ): """Produce a GPU-architecture-invariant randomized Torch tensor. Args: shape (list or tuple of ints): Output tensor shape. dtype (torch.dtype): Output tensor type. device (torch.device): Device holding the output. seed (int): Optional randomization seed. Returns: tensor (torch.tensor): Randomly-generated tensor. """ # Create a random number generator, optionally seeded rng = np.random.RandomState(seed) # Generate random numbers using the generator random_array = rng.standard_normal(shape).astype(np.float32) # Use standard_normal for normal distribution # Convert to torch tensor and return return torch.from_numpy(random_array).to(dtype=dtype, device=device) def get_data_batch_size(data: dict[str, torch.Tensor] | torch.Tensor) -> int: """Get the batch size from a data batch, a (possibly hierarchical) dictionary of tensors. Args: data (dict[str, torch.Tensor]): Data batch (dictionary of tensors). Returns: batch_size (int): Data batch size. """ def _get_batch_size(input_data: Any) -> Union[int, None]: """ Helper function that recursively finds a tensor in the input data (could be a nested dictionary or list of tensors) and returns its batch size. """ if isinstance(input_data, torch.Tensor): return len(input_data) elif isinstance(input_data, collections.abc.Mapping): for key, value in input_data.items(): batch_size = _get_batch_size(value) if batch_size is not None: return batch_size elif isinstance(input_data, (list, tuple)) and len(input_data) > 0: # Handle list/tuple of tensors (variable-length batches) # The batch size is the length of the list # We are verifying if input_data[0] is indeed a tensor. If so, return the length of the list. if isinstance(input_data[0], torch.Tensor): return len(input_data) # Recurse into first element if it's a nested structure return _get_batch_size(input_data[0]) return None batch_size = _get_batch_size(data) if not isinstance(batch_size, int): raise ValueError(f"Batch size ({batch_size}) obtained from invalid data: {data}") return batch_size def parameters_to_buffer(module: torch.nn.Module, persistent: bool = True): """Convert parameters in a module to buffers. Buffers do not have its own gradients and thus not updated by backpropagation. Args: module (torch.nn.Module): a module to convert parameters persistent (bool): If True, buffers are included in state_dict. """ named_params = dict() for name, param in module.named_parameters(): named_params[name] = param for name, param in named_params.items(): module_hierarchy = name.split(".") submodule_name = ".".join(module_hierarchy[:-1]) submodule = module.get_submodule(submodule_name) subname = module_hierarchy[-1] delattr(submodule, subname) submodule.register_buffer(subname, param, persistent=persistent) return T = TypeVar("T", bound=Callable[..., Any]) class timer(Timer): """Simple CPU timer for timing the execution of code. It can be used as either a context manager or a function decorator. The timing result will be logged upon exit. Example: def func_a(): time.sleep(1) with timer("func_a"): func_a() @timer("func_b) def func_b(): time.sleep(1) func_b() """ def __init__(self, context: str, debug: bool = False): super().__init__( tag=context, measure_cpu=True, measure_cuda=False, unit="s", debug=debug, ) class memory_checker(ContextDecorator): # noqa: N801 """Simple memory checker for a given block of code. It can be used as either a context manager or a function decorator. The memory usage will be logged upon exit. Example: def func_a(): torch.rand([int(1024**2)]).float().cuda() with memory_checker("func_a"): func_a() >>> 0.004GB memory used @memory_checker("func_b") def func_b(): random_var = torch.rand([int(1024**2)]).cuda() func_b() """ def __init__(self, context: str, debug: bool = False): self.context = context self.debug = debug def __enter__(self) -> None: torch.cuda.synchronize() torch.cuda.reset_peak_memory_stats() self.initial_memory = torch.cuda.max_memory_allocated() def __exit__(self, exc_type, exc_value, traceback) -> None: # noqa: ANN001 torch.cuda.synchronize() final_memory = torch.cuda.max_memory_allocated() message = f"Memory used within {self.context}: {(final_memory - self.initial_memory) / 1024**3:.4f} GB" if self.debug: log.debug(message) else: log.info(message) def __call__(self, func: T) -> T: @functools.wraps(func) def wrapper(*args, **kwargs): # noqa: ANN202 torch.cuda.synchronize() torch.cuda.reset_peak_memory_stats() initial_memory = torch.cuda.max_memory_allocated() result = func(*args, **kwargs) torch.cuda.synchronize() final_memory = torch.cuda.max_memory_allocated() message = f"Memory used within {self.context}: {(final_memory - initial_memory) / 1024**3:.4f} GB" if self.debug: log.debug(message) else: log.info(message) return result return wrapper # type: ignore class TrainingTimer: """Timer for timing the execution of code, aggregating over multiple training iterations. It is used as a context manager to measure the execution time of code and store the timing results for each function. The context managers can be nested. Attributes: results (dict): A dictionary to store timing results for various code. Example: timer = Timer() for i in range(100): with timer("func_a"): func_a() avg_time = sum(timer.results["func_a"]) / len(timer.results["func_a"]) print(f"func_a() took {avg_time} seconds.") """ def __init__(self) -> None: self.results = dict() self.average_results = dict() self.timers = [] self.func_stack = [] self.reset() def reset(self) -> None: self.results = {key: [] for key in self.results} def __enter__(self) -> TrainingTimer: timer = Timer(measure_cpu=True, measure_cuda=False, debug=True, unit="s") self.timers.append(timer) timer.start() return self def __exit__(self, exc_type, exc_value, traceback) -> None: # noqa: ANN001 timer = self.timers.pop() timer.end() result = timer.get_cpu_time() key = self.func_stack.pop() self.results.setdefault(key, []) self.results[key].append(result) def __call__(self, func_name: str) -> TrainingTimer: self.func_stack.append(func_name) return self def __getattr__(self, func_name: str) -> TrainingTimer: return self.__call__(func_name) def nested(self, func_name: str) -> TrainingTimer: return self.__call__(func_name) def compute_average_results(self) -> dict[str, float]: results = dict() for key, value_list in self.results.items(): results[key] = sum(value_list) / len(value_list) return results def timeout_handler(timeout_period: float, signum: int, frame: int) -> None: # What to do when the process gets stuck. For now, we simply end the process. error_message = f"Timeout error: more than {timeout_period} seconds passed since the last iteration." if distributed.is_rank0(): import wandb wandb.alert(title="Timeout error!", text=error_message, level=wandb.AlertLevel.ERROR) raise TimeoutError(error_message) class Color: """A convenience class to colorize strings in the console. Example: import print("This is {Color.red('important')}.") """ @staticmethod def red(x: str) -> str: return termcolor.colored(str(x), color="red") @staticmethod def green(x: str) -> str: return termcolor.colored(str(x), color="green") @staticmethod def blue(x: str) -> str: return termcolor.colored(str(x), color="blue") @staticmethod def cyan(x: str) -> str: return termcolor.colored(str(x), color="cyan") @staticmethod def yellow(x: str) -> str: return termcolor.colored(str(x), color="yellow") @staticmethod def magenta(x: str) -> str: return termcolor.colored(str(x), color="magenta") @staticmethod def grey(x: str) -> str: return termcolor.colored(str(x), color="grey") class BufferCnt: """ Buffer counter which keeps track of the condition when called and returns True when the condition in met "thres" amount of times, otherwise returns False. Example usage: buf = BufferCnt(thres=3) for _ in range(5): if buf(random.random() > 0.5): print("We got lucky 3 times out of 5.") Args: thres (int): The amount of times the expression needs to be True before returning True. reset_over_thres (bool): Whether to reset the buffer after returning True. """ def __init__(self, thres=10, reset_over_thres=False): self._cnt = 0 self.thres = thres self.reset_over_thres = reset_over_thres def __call__(self, expre, thres=None): if expre is True: self._cnt += 1 else: self._cnt = 0 if thres is None: thres = self.thres if self._cnt >= thres: if self.reset_over_thres: self.reset() return True return False @property def cnt(self): return self._cnt def reset(self): self._cnt = 0 def dataclass_instance_to_dict(dataclass: Any) -> dict: """Convert a dataclass to a dictionary. Args: dataclass (Any): Dataclass object. Returns: dict: Dictionary representation of the dataclass. """ return {f.name: getattr(dataclass, f.name) for f in fields(dataclass)} def get_local_tensor_if_DTensor(tensor: torch.Tensor | DTensor) -> torch.tensor: if isinstance(tensor, DTensor): local = tensor.to_local() # As per PyTorch documentation, if the communication is not finished yet, we need to wait for it to finish # https://pytorch.org/docs/stable/distributed.tensor.html#torch.distributed.tensor.DTensor.to_local if isinstance(local, AsyncCollectiveTensor): return local.wait() else: return local return tensor def set_torch_compile_options(recompile_limit: int = 8, use_duck_shape: bool = True): """ Set some of the torch compile config options. The default values of arguments are default config values in PyTorch as of 2.10 version. The value recompile_limit=32 is useful for Wan Tokenizer encoding compilation, as the standard value of 8 can easily overflow. The value of use_duck_shape=False is useful for Cosmos3 MoT training to reduce recompilations. Args: recompile_limit (int): Controls the maximum number of cache entries with a guard on same ID_MATCH'd object. use_duck_shape (bool): This flag changes whether we should use the same symbolic variable to represent input sizes that are the same """ try: # PyTorch >= 2.7 torch._dynamo.config.recompile_limit = recompile_limit torch.fx.experimental._config.use_duck_shape = use_duck_shape except AttributeError: try: torch._dynamo.config.cache_size_limit = recompile_limit torch.fx.experimental._config.use_duck_shape = use_duck_shape except AttributeError as e: log.warning("torch.compile is not available due to missing config options.") raise e class NVTXRangeContext: """ Context manager which inserts NVTX range around the current context and optionally calls torch.cuda.synchronize at the start and the end of the context. Args: name (str): Name of the NVTX range. enabled (bool): Whether the context manager is enabled. When disabled, it does nothing. Default: True. synchronize (bool): Whether to call torch.cuda.synchronize() at the start and the end of the context. Default: True. """ def __init__(self, name: str, enabled: bool = True, synchronize: bool = True): self.name = name self.enabled = enabled self.synchronize = synchronize def __enter__(self): if not self.enabled: return if self.synchronize: torch.cuda.synchronize() torch.cuda.nvtx.range_push(self.name) def __exit__(self, exc_type, exc_val, exc_tb): if not self.enabled: return if self.synchronize: torch.cuda.synchronize() torch.cuda.nvtx.range_pop() class StragglerDetectorV2: """StragglerDetectorV2 is a class that allows you to easily integrate the "straggler" tool. This tool detects stragglers using low-level CUPTI tool, which can gather kernel execution time with very low overhead. The execution times are compared across different ranks, as well as to the execution time of the exact same kernels in the past. This tool can be easily integrated, as it's resilient to any synchronizations, since it captures kernels execution time. It means that we can wrap the entire forward or backward passes and the stragglers will be identified regardless of synchronizations happening during the iteration. Args: enabled (bool): Whether the straggler detection is enabled. When disabled, it does nothing. Default: True. report_freq (int): Generate a report each report_freq iterations that analyzes the GPUs performance. Defaults to 100. profile_freq (int): Enable the CUPTI profiling each profile_freq iterations. Since the overhead is very low, the default value is 1. max_diff (float): Defines the maximum relative difference between the fastest and the slowest rank to determine the slowdown. Defaults to 2.0 raise_error (bool): Whether to raise error when stragglers are detected enough times. Defaults to True.""" def __init__( self, enabled: bool = True, report_freq: int = 100, profile_freq: int = 1, max_diff: float = 2.0, raise_error: bool = True, save_s3: bool = False, ): self.enabled = enabled self.report_freq = report_freq self.profile_freq = profile_freq self.name = self.__class__.__name__ self.slowdown_count = BufferCnt(thres=10, reset_over_thres=True) self.max_diff = max_diff self.raise_error = raise_error self.save_s3 = save_s3 def initialize(self): if self.enabled: if not straggler: raise RuntimeError( "Please install the `straggler` package before using StragglerDetectionV2." ) straggler.Detector.initialize( scores_to_compute=["relative_perf_scores", "individual_perf_scores"], gather_on_rank0=False, # all ranks results will be available on rank 0 profiling_interval=self.profile_freq, ) def profile_section(self, name: str, section_enabled: bool, profile_cuda: bool = True): if section_enabled and self.enabled: return straggler.Detector.detection_section(name, profile_cuda=profile_cuda) else: return nullcontext() def _aggregate_section_results(self, local_section_summaries): data = [] for key in local_section_summaries: # straggler reports time in ms data.append(local_section_summaries[key][straggler.Statistic.MAX] / 1000) return distributed.all_gather_tensor(torch.tensor(data).cuda()) def generate_report(self, iteration): if self.enabled and iteration % self.report_freq == 0: report = straggler.Detector.generate_report() gpu_relative_perf_score = report.gpu_relative_perf_scores[distributed.get_rank()] gpu_relative_perf_score_gather_list = distributed.all_gather_tensor( torch.tensor([gpu_relative_perf_score]).cuda() ) local_section_data = self._aggregate_section_results(report.local_section_summaries) if distributed.get_rank() == 0: stragglers = report.identify_stragglers(gpu_rel_threshold=1 / self.max_diff) wandb_info = { f"{self.name}/relative_gpu_perf_{rank}": perf[0].item() for rank, perf in enumerate(gpu_relative_perf_score_gather_list) } for key_id, key in enumerate(report.local_section_summaries): wandb_info.update( {f"{self.name}/{key}_{rank:03d}": v[key_id].item() for rank, v in enumerate(local_section_data)} ) data_tensor = torch.tensor(gpu_relative_perf_score_gather_list) slowest_rank_id = torch.argmin(data_tensor) wandb_info.update( { f"slowest_rank/{self.name}_rank": slowest_rank_id.item(), f"slowest_rank/{self.name}_relative_perf": torch.min(data_tensor).item(), } ) for key_id, key in enumerate(report.local_section_summaries): data_tensor = torch.tensor([v[key_id] for v in local_section_data]) wandb_info.update( { f"slowest_rank/slowest_{key}_rank": torch.argmax(data_tensor).item(), f"slowest_rank/slowest_{key}_time": torch.max(data_tensor).item(), } ) import wandb if wandb.run: wandb.log(wandb_info, step=iteration) import cosmos_framework.utils.launch if cosmos_framework.utils.launch.S3_READY and (iteration % (5 * self.report_freq) == 0) and self.save_s3: easy_io.dump( wandb_info, f"s3://rundir/{self.__class__.__name__}/iter_{iteration:09d}.yaml", ) easy_io.dump( report, f"s3://rundir/{self.__class__.__name__}/report_iter_{iteration:09d}.pkl", ) # Which GPUs are slower than other GPUs, based on the execution time of kernels relative_stragglers = stragglers["straggler_gpus_relative"] # Which GPUs are slower than itself in the past, based on the past execution time of kernels. individual_stragglers = stragglers["straggler_gpus_individual"] is_slowdown = relative_stragglers or individual_stragglers if is_slowdown: hostname = torch.ByteTensor(bytearray(os.uname().nodename, "utf-8")).cuda() whole_hostname = all_gather_tensor(hostname) slowest_hostname = whole_hostname[slowest_rank_id].cpu().numpy().tobytes().decode("utf-8") logging.critical(f"Slowest rank hostname: {slowest_hostname}") if self.slowdown_count(is_slowdown) and self.raise_error: raise RuntimeError( f"Detected GPU {slowest_rank_id} to be too slow compared to other GPUs." f" The relative performance of {slowest_rank_id} rank was {report.gpu_relative_perf_scores[slowest_rank_id]}. Terminating the training." )