Spaces:
Running on L40S
Running on L40S
File size: 12,685 Bytes
9f818c5 | 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 | # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: OpenMDW-1.1
from typing import Any, Optional
import torch
from cosmos_framework.model._base import ImaginaireModel
from cosmos_framework.utils.callback import Callback
from cosmos_framework.utils.log import logger
from cosmos_framework.utils.misc import get_data_batch_size
from cosmos_framework.utils.training_telemetry.utils import (
get_checkpoint_strategy,
get_telemetry_recorder,
import_training_telemetry,
)
class TelemetryCallback(Callback):
"""Callback for Telemetry"""
def __init__(
self,
) -> None:
super().__init__()
self.training_telemetry = import_training_telemetry()
self.recorder = get_telemetry_recorder()
self.spans: dict[self.training_telemetry.events.EventName, self.training_telemetry.Span] = {}
self.iteration_elapsed = 0.0
self.forward_elapsed = 0.0
self.backward_elapsed = 0.0
self.dataloader_elapsed = 0.0
self.validation_elapsed = 0.0
self.validation_loss = 0.0
self.validation_iter = 0
self.checkpoint_strategy = "async"
self.checkpoint_interval = 0
self.validation_interval = 0
def _start_span(self, span_name: Any, metrics: Any = None, verbosity: Any = None, color: Any = None) -> None:
"""Start a span with the given event name and optional metrics, color, and verbosity"""
if span_name in self.spans:
logger.warning(f"Span {span_name} already started, stopping it but this is unexpected")
self.recorder.stop(self.spans[span_name])
del self.spans[span_name]
if verbosity is None:
verbosity = self.training_telemetry.Verbosity.INFO
self.spans[span_name] = self.recorder.start(
name=span_name,
color=color,
verbosity=verbosity,
metrics=metrics,
)
def _stop_span(self, span_name: Any, metrics: Any = None) -> float:
"""Stop a span and return the elapsed time"""
span = self.spans.get(span_name)
if span is None:
logger.warning(f"Span {span_name} was not started, this is unexpected")
return 0.0
if metrics is not None:
span.add_metrics(metrics)
self.recorder.stop(span)
del self.spans[span_name]
return span.duration.elapsed
def on_train_start(self, model: ImaginaireModel, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.TRAINING_LOOP
self._start_span(span_name)
self.iteration_elapsed = 0
self.forward_elapsed = 0
self.backward_elapsed = 0
self.dataloader_elapsed = 0
try:
self.checkpoint_strategy = get_checkpoint_strategy(self.config.checkpoint)
except Exception as e:
logger.warning(f"Failed to get checkpoint strategy using default {self.checkpoint_strategy}: {e}")
try:
self.checkpoint_interval = self.config.checkpoint.save_iter
self.validation_interval = self.config.trainer.validation_iter
except Exception as e:
logger.warning(
f"Failed to get intervals using default {self.checkpoint_interval} and {self.validation_interval}: {e}"
)
def on_training_step_batch_start(
self, model: ImaginaireModel, data: dict[str, torch.Tensor], iteration: int = 0
) -> None:
span_name = self.training_telemetry.SpanName.ITERATION
self._start_span(
span_name,
color=self.training_telemetry.SpanColor.BLUE,
verbosity=self.training_telemetry.Verbosity.PROFILING,
)
def on_before_forward(self, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.MODEL_FORWARD
self._start_span(
span_name,
color=self.training_telemetry.SpanColor.GREEN,
verbosity=self.training_telemetry.Verbosity.PROFILING,
)
def on_after_forward(self, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.MODEL_FORWARD
self.forward_elapsed += self._stop_span(span_name)
def on_before_backward(self, model: ImaginaireModel, loss: torch.Tensor, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.MODEL_BACKWARD
self._start_span(
span_name,
color=self.training_telemetry.SpanColor.YELLOW,
verbosity=self.training_telemetry.Verbosity.PROFILING,
)
def on_after_backward(self, model: ImaginaireModel, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.MODEL_BACKWARD
self.backward_elapsed += self._stop_span(span_name)
def on_before_dataloading(self, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.DATA_LOADING
self._start_span(
span_name,
color=self.training_telemetry.SpanColor.GREEN,
verbosity=self.training_telemetry.Verbosity.PROFILING,
)
def on_after_dataloading(self, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.DATA_LOADING
self.dataloader_elapsed += self._stop_span(span_name)
def on_optimizer_init_start(self) -> None:
span_name = self.training_telemetry.SpanName.OPTIMIZER_INIT
self._start_span(span_name)
def on_optimizer_init_end(self) -> None:
span_name = self.training_telemetry.SpanName.OPTIMIZER_INIT
self._stop_span(span_name)
def on_before_optimizer_step(
self,
model: ImaginaireModel,
optimizer: torch.optim.Optimizer,
scheduler: torch.optim.lr_scheduler.LRScheduler,
grad_scaler: torch.amp.GradScaler,
iteration: int = 0,
) -> None:
pass
def on_before_zero_grad(
self,
model: ImaginaireModel,
optimizer: torch.optim.Optimizer,
scheduler: torch.optim.lr_scheduler.LRScheduler,
iteration: int = 0,
) -> None:
pass
def on_training_step_batch_end(
self,
model: ImaginaireModel,
data_batch: dict[str, torch.Tensor],
output_batch: dict[str, torch.Tensor],
loss: torch.Tensor,
iteration: int = 0,
) -> None:
span_name = self.training_telemetry.SpanName.ITERATION
self.iteration_elapsed += self._stop_span(span_name)
def on_training_step_end(
self,
model: ImaginaireModel,
data_batch: dict[str, torch.Tensor],
output_batch: dict[str, torch.Tensor],
loss: torch.Tensor,
iteration: int = 0,
) -> None:
"""Unlike on_training_step_batch_end, this function is called when the optimizer is updated, and the iteration incremented."""
if iteration % self.config.trainer.logging_iter == 0:
event_name = self.training_telemetry.events.EventName.TRAINING_ITERATIONS
avg_iteration_time = self.iteration_elapsed / self.config.trainer.logging_iter
avg_forward_time = self.forward_elapsed / self.config.trainer.logging_iter
avg_backward_time = self.backward_elapsed / self.config.trainer.logging_iter
avg_dataloader_time = self.dataloader_elapsed / self.config.trainer.logging_iter
batch_size = get_data_batch_size(data_batch)
# throughput = num_floating_point_operations(batch_size) / (
# avg_iteration_time * 10**12 * distributed.get_world_size)
metrics = self.training_telemetry.IterationMetrics.create(
current_iteration=iteration,
num_iterations=self.config.trainer.logging_iter,
interval=self.config.trainer.logging_iter,
average_iteration_time=avg_iteration_time,
average_forward_time=avg_forward_time,
average_backward_time=avg_backward_time,
average_dataloader_time=avg_dataloader_time,
tflops=0.0,
tokens_per_second=0.0,
loss=loss.item(),
batch_size=batch_size,
)
self.recorder.event(self.training_telemetry.events.Event.create(event_name, metrics))
self.iteration_elapsed = 0
self.forward_elapsed = 0
self.backward_elapsed = 0
self.dataloader_elapsed = 0
def on_validation_start(
self, model: ImaginaireModel, dataloader_val: torch.utils.data.DataLoader, iteration: int = 0
) -> None:
span_name = self.training_telemetry.SpanName.VALIDATION_LOOP
self._start_span(span_name)
self.validation_loss = 0.0
self.validation_iter = 0
self.validation_elapsed = 0.0
def on_validation_step_start(
self, model: ImaginaireModel, data: dict[str, torch.Tensor], iteration: int = 0
) -> None:
span_name = self.training_telemetry.SpanName.ITERATION
self._start_span(
span_name,
color=self.training_telemetry.SpanColor.RED,
verbosity=self.training_telemetry.Verbosity.PROFILING,
)
def on_validation_step_end(
self,
model: ImaginaireModel,
data_batch: dict[str, torch.Tensor],
output_batch: dict[str, torch.Tensor],
loss: torch.Tensor,
iteration: int = 0,
) -> None:
self.validation_loss += loss.item()
self.validation_iter += 1
span_name = self.training_telemetry.SpanName.ITERATION
self.validation_elapsed += self._stop_span(span_name)
def on_validation_end(self, model: ImaginaireModel, iteration: int = 0) -> None:
if self.validation_iter == 0:
return
span_name = self.training_telemetry.SpanName.VALIDATION_LOOP
avg_validation_time = self.validation_elapsed / self.validation_iter
metrics = self.training_telemetry.IterationMetrics.create(
current_iteration=iteration,
num_iterations=self.validation_iter,
interval=self.validation_interval,
average_iteration_time=avg_validation_time,
loss=self.validation_loss / self.validation_iter,
)
self._stop_span(span_name, metrics=metrics)
self.validation_elapsed = 0.0
self.validation_loss = 0.0
self.validation_iter = 0
def on_load_checkpoint_start(self, model: ImaginaireModel) -> None:
span_name = self.training_telemetry.SpanName.CHECKPOINT_LOAD
self._start_span(span_name)
def on_load_checkpoint_end(
self, model: ImaginaireModel, iteration: int = 0, checkpoint_path: Optional[str] = None
) -> None:
span_name = self.training_telemetry.SpanName.CHECKPOINT_LOAD
metrics = self.training_telemetry.CheckpointMetrics.create(
checkpoint_type=self.training_telemetry.CheckPointType.GLOBAL,
current_iteration=iteration,
checkpoint_directory=checkpoint_path,
)
self._stop_span(span_name, metrics=metrics)
def on_load_checkpoint(self, model: ImaginaireModel, state_dict: dict[Any]) -> None:
pass
def on_save_checkpoint_start(self, model: ImaginaireModel, iteration: int = 0) -> None:
span_name = (
self.training_telemetry.SpanName.CHECKPOINT_SAVE_SYNC
if self.checkpoint_strategy == "sync"
else self.training_telemetry.SpanName.CHECKPOINT_SAVE_ASYNC
)
checkpoint_metrics = self.training_telemetry.CheckpointMetrics.create(
checkpoint_type=self.training_telemetry.CheckPointType.GLOBAL,
current_iteration=iteration,
interval=self.checkpoint_interval,
)
self._start_span(span_name, metrics=checkpoint_metrics)
def on_save_checkpoint_end(self, model: ImaginaireModel, iteration: int = 0) -> None:
span_name = (
self.training_telemetry.SpanName.CHECKPOINT_SAVE_SYNC
if self.checkpoint_strategy == "sync"
else self.training_telemetry.SpanName.CHECKPOINT_SAVE_ASYNC
)
self._stop_span(span_name)
def on_save_checkpoint_success(self, iteration: int = 0, elapsed_time: float = 0) -> None:
pass
def on_save_checkpoint(self, model: ImaginaireModel, state_dict: dict[Any]) -> None:
pass
def on_train_end(self, model: ImaginaireModel, iteration: int = 0) -> None:
span_name = self.training_telemetry.SpanName.TRAINING_LOOP
self._stop_span(span_name)
def on_app_end(self) -> None:
pass
|