Spaces:
Runtime error
Runtime error
File size: 12,471 Bytes
0558aa4 |
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 |
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# 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.
"""
OneLogger callback for NeMo training.
This module provides a callback that integrates OneLogger telemetry with NeMo training.
"""
import os
from typing import Any, Dict
from lightning.pytorch import Trainer
from lightning.pytorch.callbacks.model_checkpoint import ModelCheckpoint
from nv_one_logger.api.config import OneLoggerConfig
from nv_one_logger.training_telemetry.api.callbacks import on_app_start
from nv_one_logger.training_telemetry.api.config import TrainingTelemetryConfig
from nv_one_logger.training_telemetry.api.training_telemetry_provider import TrainingTelemetryProvider
from nv_one_logger.training_telemetry.integration.pytorch_lightning import TimeEventCallback as OneLoggerPTLCallback
from nemo.lightning.base_callback import BaseCallback
# Export all symbols for testing and usage
__all__ = ['OneLoggerNeMoCallback']
def get_one_logger_init_config() -> Dict[str, Any]:
"""Generate minimal configuration for OneLogger initialization.
This function provides the absolute minimal configuration needed for OneLogger initialization.
It only includes the required fields and uses defaults for everything else to avoid
dependencies on exp_manager during early import.
Returns:
Dictionary containing minimal initialization configuration
"""
if "EXP_NAME" in os.environ:
session_tag = os.environ.get("EXP_NAME") # For NeMo v1
else:
session_tag = os.environ.get("SLURM_JOB_NAME", "nemo-run")
world_size = int(os.environ.get('WORLD_SIZE', 1))
# Minimal configuration - required fields only
init_config = {
# Required fields (from OneLoggerConfig) - no defaults
"application_name": "nemo",
"session_tag_or_fn": session_tag,
# Important fields with defaults - provide if available from config
"enable_for_current_rank": _should_enable_for_current_rank(),
"world_size_or_fn": world_size,
}
return init_config
def _get_base_callback_config(
trainer: Any,
global_batch_size: int,
seq_length: int,
) -> Dict[str, Any]:
"""Generate base configuration for OneLogger training telemetry.
This function provides the common configuration needed for both NeMo v1 and v2.
It extracts basic training information from trainer object and uses provided
batch size and sequence length values.
Args:
trainer: PyTorch Lightning trainer instance
global_batch_size: Global batch size (calculated by version-specific function)
seq_length: Sequence length (calculated by version-specific function)
Returns:
Dictionary containing base training callback configuration
"""
# Extract values from trainer
# Get job name from multiple sources in order of reliability
if "EXP_NAME" in os.environ:
job_name = os.environ.get("EXP_NAME") # For NeMo v1
else:
job_name = os.environ.get("SLURM_JOB_NAME", "nemo-run")
world_size = int(os.environ.get('WORLD_SIZE', 1))
max_steps = getattr(trainer, 'max_steps', 1)
log_every_n_steps = getattr(trainer, 'log_every_n_steps', 10)
micro_batch_size = global_batch_size // world_size
# Get PERF_VERSION_TAG from environment
perf_version_tag = os.environ.get('PERF_VERSION_TAG', '0.0.0')
# Calculate performance tag
perf_tag = f"{job_name}_{perf_version_tag}_bf{global_batch_size}_se{seq_length}_ws{world_size}"
# Calculate train samples target
train_samples_target = max_steps * global_batch_size
# Fallback values
is_save_checkpoint_enabled = False
is_validation_iterations_enabled = False
save_checkpoint_strategy = "sync"
checkpoint_callbacks = [cb for cb in trainer.callbacks if isinstance(cb, ModelCheckpoint)]
is_save_checkpoint_enabled = len(checkpoint_callbacks) > 0
val_check_interval = getattr(trainer, 'val_check_interval', -1)
is_validation_iterations_enabled = val_check_interval > 0
# Check for async_save in trainer strategy (handle both dict and object cases)
if hasattr(trainer, 'strategy') and trainer.strategy is not None:
if isinstance(trainer.strategy, dict):
if trainer.strategy.get('async_save', False):
save_checkpoint_strategy = "async"
else:
if hasattr(trainer.strategy, 'async_save') and trainer.strategy.async_save:
save_checkpoint_strategy = "async"
for callback in checkpoint_callbacks:
if hasattr(callback, 'async_save') and callback.async_save:
save_checkpoint_strategy = "async"
break
# Base training telemetry configuration
base_config = {
# Performance tag (REQUIRED in TrainingTelemetryConfig)
"perf_tag_or_fn": perf_tag,
# Batch information (REQUIRED in TrainingTelemetryConfig)
"global_batch_size_or_fn": global_batch_size,
"micro_batch_size_or_fn": micro_batch_size,
"seq_length_or_fn": seq_length,
# Training targets
"train_iterations_target_or_fn": max_steps,
"train_samples_target_or_fn": train_samples_target,
# Logging frequency
"log_every_n_train_iterations": log_every_n_steps,
'is_validation_iterations_enabled_or_fn': is_validation_iterations_enabled,
'is_save_checkpoint_enabled_or_fn': is_save_checkpoint_enabled,
'save_checkpoint_strategy': save_checkpoint_strategy,
}
return base_config
def get_nemo_v1_callback_config(trainer: Any) -> Dict[str, Any]:
"""Generate NeMo v1 specific configuration for OneLogger training callback.
This function provides NeMo v1 specific configuration by extracting values from
the exp_manager_config object and trainer object.
Args:
trainer: PyTorch Lightning trainer instance
Returns:
Dictionary containing NeMo v1 training callback configuration
"""
global_batch_size = 1 # Default fallback
seq_length = 1 # Default fallback
if (
hasattr(trainer, 'lightning_module')
and trainer.lightning_module is not None
and hasattr(trainer.lightning_module, 'cfg')
):
model_cfg = trainer.lightning_module.cfg
if hasattr(model_cfg, 'train_ds'):
train_ds = model_cfg.train_ds
micro_batch_size = getattr(train_ds, 'batch_size', None)
if micro_batch_size is not None:
# Standard fixed-size batching
global_batch_size = int(micro_batch_size) * int(os.environ.get('WORLD_SIZE', 1))
else:
# Try bucketing average first if available
if hasattr(train_ds, 'bucket_batch_size'):
# For ASR with bucketing, use the average batch size
bucket_batch_sizes = train_ds.bucket_batch_size
# Handle both ListConfig and regular list types
if hasattr(bucket_batch_sizes, '__len__') and len(bucket_batch_sizes) > 0:
# Convert to list if it's a ListConfig, otherwise use as is
bucket_list = (
list(bucket_batch_sizes) if hasattr(bucket_batch_sizes, '__iter__') else bucket_batch_sizes
)
avg_batch_size = sum(bucket_list) / len(bucket_list)
global_batch_size = int(avg_batch_size) * int(os.environ.get('WORLD_SIZE', 1))
if hasattr(model_cfg, 'encoder') and hasattr(model_cfg.encoder, 'd_model'):
seq_length = model_cfg.encoder.d_model
# Get base configuration with calculated values
config = _get_base_callback_config(
trainer=trainer,
global_batch_size=global_batch_size,
seq_length=seq_length,
)
return config
def get_nemo_v2_callback_config(
trainer: Any,
data: Any,
) -> Dict[str, Any]:
"""Generate NeMo v2 specific configuration for the OneLogger training callback.
This function extracts the global batch size and sequence length from the provided NeMo v2 data module,
and uses them to construct the configuration dictionary for the OneLogger training callback.
Args:
trainer: PyTorch Lightning trainer instance.
data: NeMo v2 data module (required).
Returns:
Dictionary containing the NeMo v2 training callback configuration.
"""
# NeMo v2: Extract batch size and sequence length from data module (most reliable source)
global_batch_size = 1 # Default fallback
seq_length = 1 # Default fallback
if data is not None:
seq_length = data.seq_length
# Prefer explicit global_batch_size if provided by the data module
if hasattr(data, 'global_batch_size') and getattr(data, 'global_batch_size') is not None:
global_batch_size = int(getattr(data, 'global_batch_size'))
else:
# Fall back to micro_batch_size multiplied by WORLD_SIZE when global_batch_size is unavailable
micro_batch_size = getattr(data, 'micro_batch_size', None)
if micro_batch_size is not None:
world_size = int(os.environ.get('WORLD_SIZE', 1))
global_batch_size = int(micro_batch_size) * world_size
# Get base configuration with calculated values
config = _get_base_callback_config(
trainer=trainer,
global_batch_size=global_batch_size,
seq_length=seq_length,
)
return config
def _should_enable_for_current_rank() -> bool:
"""Determine if OneLogger should be enabled for the current rank.
Uses environment variables instead of torch.distributed to avoid circular imports.
In distributed training, typically only rank 0 (or the last rank) should
enable OneLogger to avoid duplicate telemetry data.
Returns:
True if OneLogger should be enabled for the current rank, False otherwise
"""
rank = int(os.environ.get('RANK', -1))
# Enable for rank 0 or the last rank (common pattern)
return rank == 0
class OneLoggerNeMoCallback(OneLoggerPTLCallback, BaseCallback):
"""Adapter extending OneLogger's PTL callback with init + config update.
__init__ configures the provider from meta info, then calls super().__init__.
update_config computes TrainingTelemetryConfig and applies it.
"""
_instance = None
def __new__(cls, *args, **kwargs):
if cls._instance is None:
cls._instance = super().__new__(cls)
return cls._instance
def __init__(self) -> None:
if getattr(self, '_initialized', False):
return
init_config = get_one_logger_init_config()
one_logger_config = OneLoggerConfig(**init_config)
TrainingTelemetryProvider.instance().with_base_config(
one_logger_config
).with_export_config().configure_provider()
# Initialize underlying OneLogger PTL callback
super().__init__(TrainingTelemetryProvider.instance(), call_on_app_start=False)
# Explicitly signal application start after provider configuration
on_app_start()
def update_config(self, nemo_version: str, trainer: Trainer, **kwargs) -> None:
# Avoid this function being called multiple times
if TrainingTelemetryProvider.instance().config.telemetry_config is not None:
return
if nemo_version == 'v1':
config = get_nemo_v1_callback_config(trainer=trainer)
elif nemo_version == 'v2':
# v2 expects data module in kwargs
data = kwargs.get('data', None)
config = get_nemo_v2_callback_config(trainer=trainer, data=data)
else:
config = get_nemo_v1_callback_config(trainer=trainer)
training_telemetry_config = TrainingTelemetryConfig(**config)
TrainingTelemetryProvider.instance().set_training_telemetry_config(training_telemetry_config)
|