File size: 8,866 Bytes
54c5666 |
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 |
"""Real-time training metrics and monitoring"""
import torch
import numpy as np
from typing import Dict, List, Optional, Any
from dataclasses import dataclass, field
import time
import logging
from collections import deque
logger = logging.getLogger(__name__)
@dataclass
class TrainingMetrics:
"""Track comprehensive training metrics"""
# Loss metrics
train_loss: List[float] = field(default_factory=list)
val_loss: List[float] = field(default_factory=list)
# Performance metrics
tokens_per_second: List[float] = field(default_factory=list)
samples_per_second: List[float] = field(default_factory=list)
gpu_memory_allocated: List[float] = field(default_factory=list)
gpu_memory_reserved: List[float] = field(default_factory=list)
gpu_utilization: List[float] = field(default_factory=list)
# Gradient metrics
grad_norm: List[float] = field(default_factory=list)
param_norm: List[float] = field(default_factory=list)
grad_variance: List[float] = field(default_factory=list)
# Learning rate
lr_history: List[float] = field(default_factory=list)
# Time tracking
step_times: List[float] = field(default_factory=list)
epoch_times: List[float] = field(default_factory=list)
# Step counter
global_step: int = 0
def log_step(
self,
loss: float,
lr: float,
model: torch.nn.Module,
batch_size: int = 1,
seq_length: int = 512,
step_time: Optional[float] = None
):
"""Log metrics for current step"""
self.global_step += 1
self.train_loss.append(loss)
self.lr_history.append(lr)
if step_time is not None:
self.step_times.append(step_time)
# Calculate throughput
tokens = batch_size * seq_length
self.tokens_per_second.append(tokens / step_time)
self.samples_per_second.append(batch_size / step_time)
# GPU metrics
if torch.cuda.is_available():
self.gpu_memory_allocated.append(
torch.cuda.memory_allocated() / 1e9 # GB
)
self.gpu_memory_reserved.append(
torch.cuda.memory_reserved() / 1e9
)
try:
import pynvml
pynvml.nvmlInit()
handle = pynvml.nvmlDeviceGetHandleByIndex(0)
util = pynvml.nvmlDeviceGetUtilizationRates(handle)
self.gpu_utilization.append(util.gpu)
except (ImportError, Exception):
pass
# Gradient norms
try:
total_norm = 0.0
param_norm = 0.0
grad_list = []
for p in model.parameters():
if p.grad is not None:
grad_norm = p.grad.data.norm(2).item()
total_norm += grad_norm ** 2
grad_list.append(grad_norm)
if p.data is not None:
param_norm += p.data.norm(2).item() ** 2
self.grad_norm.append(total_norm ** 0.5)
self.param_norm.append(param_norm ** 0.5)
if grad_list:
self.grad_variance.append(np.var(grad_list))
except Exception as e:
logger.debug(f"Failed to compute gradient metrics: {e}")
def log_validation(self, val_loss: float):
"""Log validation metrics"""
self.val_loss.append(val_loss)
def get_summary(self, window: int = 100) -> Dict[str, float]:
"""Get summary statistics over recent window"""
def get_recent_mean(data: List[float], window: int) -> float:
if not data:
return 0.0
recent = data[-window:]
return sum(recent) / len(recent) if recent else 0.0
summary = {
'avg_train_loss': get_recent_mean(self.train_loss, window),
'avg_lr': get_recent_mean(self.lr_history, window),
'avg_tokens_per_sec': get_recent_mean(self.tokens_per_second, window),
'avg_grad_norm': get_recent_mean(self.grad_norm, window),
'avg_gpu_memory_gb': get_recent_mean(self.gpu_memory_allocated, window),
'global_step': self.global_step
}
if self.val_loss:
summary['latest_val_loss'] = self.val_loss[-1]
return summary
def export_tensorboard(self, writer, step: Optional[int] = None):
"""Export to TensorBoard"""
if step is None:
step = self.global_step
if self.train_loss:
writer.add_scalar('Loss/train', self.train_loss[-1], step)
if self.val_loss:
writer.add_scalar('Loss/validation', self.val_loss[-1], step)
if self.lr_history:
writer.add_scalar('Learning_Rate', self.lr_history[-1], step)
if self.tokens_per_second:
writer.add_scalar('Performance/tokens_per_second', self.tokens_per_second[-1], step)
if self.gpu_memory_allocated:
writer.add_scalar('GPU/memory_allocated_gb', self.gpu_memory_allocated[-1], step)
if self.grad_norm:
writer.add_scalar('Gradients/norm', self.grad_norm[-1], step)
if self.param_norm:
writer.add_scalar('Parameters/norm', self.param_norm[-1], step)
def export_mlflow(self, step: Optional[int] = None):
"""Export metrics to MLflow"""
try:
import mlflow
except ImportError:
logger.warning("mlflow not installed, skipping export")
return
if step is None:
step = self.global_step
metrics_dict = {}
if self.train_loss:
metrics_dict['train/loss'] = float(self.train_loss[-1])
if self.val_loss:
metrics_dict['val/loss'] = float(self.val_loss[-1])
if self.lr_history:
metrics_dict['train/learning_rate'] = float(self.lr_history[-1])
if self.tokens_per_second:
metrics_dict['perf/tokens_per_second'] = float(self.tokens_per_second[-1])
if self.gpu_memory_allocated:
metrics_dict['gpu/memory_gb'] = float(self.gpu_memory_allocated[-1])
if self.grad_norm:
metrics_dict['train/grad_norm'] = float(self.grad_norm[-1])
if metrics_dict:
try:
mlflow.log_metrics(metrics_dict, step=int(step))
except Exception as _e:
logger.debug(f"MLflow export skipped: {_e}")
class MetricsLogger:
"""Enhanced metrics logger with rolling statistics"""
def __init__(self, window_size: int = 100, log_freq: int = 10):
self.window_size = window_size
self.log_freq = log_freq
self.metrics = TrainingMetrics()
# Rolling windows for smoothing
self.loss_window = deque(maxlen=window_size)
self.lr_window = deque(maxlen=window_size)
self.step_count = 0
self.last_log_time = time.time()
def log(
self,
loss: float,
lr: float,
model: torch.nn.Module,
batch_size: int = 1,
seq_length: int = 512
):
"""Log metrics and print periodically"""
current_time = time.time()
step_time = current_time - self.last_log_time
self.last_log_time = current_time
# Update metrics
self.metrics.log_step(loss, lr, model, batch_size, seq_length, step_time)
self.loss_window.append(loss)
self.lr_window.append(lr)
self.step_count += 1
# Log periodically
if self.step_count % self.log_freq == 0:
avg_loss = sum(self.loss_window) / len(self.loss_window)
avg_lr = sum(self.lr_window) / len(self.lr_window)
log_msg = (
f"Step {self.step_count} | "
f"Loss: {avg_loss:.4f} | "
f"LR: {avg_lr:.2e} | "
)
if self.metrics.tokens_per_second:
log_msg += f"Tokens/s: {self.metrics.tokens_per_second[-1]:.0f} | "
if self.metrics.gpu_memory_allocated:
log_msg += f"GPU Mem: {self.metrics.gpu_memory_allocated[-1]:.2f}GB"
logger.info(log_msg)
def get_metrics(self) -> TrainingMetrics:
"""Get the underlying metrics object"""
return self.metrics
|