---

name: ml-deep-learning description: > Use this skill when asked about PyTorch, TensorFlow, Keras, neural network, CNN, RNN, LSTM, transformer, GAN, autoencoder, training loop, backpropagation, gradient descent, CUDA, distributed training, or mixed precision. This skill enforces: PyTorch training loop with nn.Module and DataLoader, TensorFlow Keras Sequential/Functional APIs, CNN architectures (ResNet, EfficientNet), RNN/LSTM/GRU for sequences, transformer implementations (self-attention, positional encoding), distributed training (DDP, FSDP), mixed precision (AMP), and CUDA optimization. Do NOT use for: classical ML models, feature engineering, or MLOps pipeline configuration. version: "1.0.0" author: "j4flmao" license: "MIT" compatibility: claude-code: true cursor: true codex: true windsurf: true tags: [ml, deep-learning, neural, phase-11]

ML Deep Learning

Purpose

Design and train deep neural networks with PyTorch and TensorFlow. Build CNN, RNN, and transformer architectures. Optimize training with distributed strategies, mixed precision, and proper GPU utilization.

Agent Protocol

Trigger

Exact user phrases: "PyTorch", "TensorFlow", "Keras", "neural network", "CNN", "RNN", "LSTM", "transformer", "GAN", "autoencoder", "training loop", "backpropagation", "gradient descent", "CUDA", "distributed training", "mixed precision", "DDP", "FSDP", "AMP", "CUDA", "ResNet", "EfficientNet".

Input Context

Before activating, verify:

• Framework (PyTorch, TensorFlow, JAX)
• Problem type (image classification, NLP, time-series, generative)
• Dataset size (images, sequences, total samples)
• Hardware (single GPU, multi-GPU, TPU, CPU)
• GPU type and memory (A100, V100, RTX, etc.)
• Training constraints (time budget, memory budget)
• Prior experiments or baseline results

Output Artifact

Model architecture definition, training loop, distributed strategy config, and optimization settings as Python.

Response Format

# Model architecture (nn.Module or Keras)
# Training loop with AMP and logging
# Distributed training launcher

# Training hyperparameters
# Hardware configuration
# Mixed precision settings

No preamble. No postamble. No explanations. No filler/hedging/transitions. Compress output — why use many token when few do trick.

Completion Criteria

• Framework selected (PyTorch preferred for research, TF for production)
• Model architecture defined (CNN, RNN, transformer)
• Data loading configured (DataLoader, tf.data with prefetch/augmentation)
• Training loop with gradient descent, logging, checkpointing
• Distributed training configured (DDP for multi-GPU, FSDP for large models)
• Mixed precision (AMP) configured for throughput
• Hyperparameter optimization strategy defined
• Evaluation on held-out test set with proper metrics

Max Response Length

300 lines of code and configuration.

Workflow

Step 1: Framework Selection

PyTorch: imperative/debuggable, research community standard, HuggingFace ecosystem, dynamic computation graphs, easier custom architectures. TensorFlow/Keras: production deployment (TF Serving, TF Lite, TF.js), static graph optimization, TFX pipeline integration, larger industrial ecosystem. JAX: functional, XLA-compiled, high-performance research, growing but smaller ecosystem. Recommendation: PyTorch for research and custom models, TensorFlow for production deployment pipelines.

# PyTorch: explicit training loop
import torch.nn as nn
import torch.optim as optim

class ImageClassifier(nn.Module):
    def __init__(self, num_classes=10):
        super().__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(2),
            nn.Conv2d(64, 128, 3, padding=1),
            nn.BatchNorm2d(128),
            nn.ReLU(inplace=True),
            nn.AdaptiveAvgPool2d(1),
        )
        self.classifier = nn.Linear(128, num_classes)

    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), -1)
        return self.classifier(x)

Step 2: Data Loading

PyTorch: Dataset class + DataLoader with num_workers, prefetch_factor, pin_memory. Augmentation: torchvision.transforms for images, custom transforms for audio/text. Multiprocessing: num_workers = 4-8 per GPU, prefetch_factor=2. TensorFlow: tf.data.Dataset with interleave, map (with num_parallel_calls), cache, prefetch. Data pipeline must never be the bottleneck.

from torch.utils.data import Dataset, DataLoader
from torchvision import transforms

class ImageDataset(Dataset):
    def __init__(self, file_paths, labels, transform=None):
        self.file_paths = file_paths
        self.labels = labels
        self.transform = transform or transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ColorJitter(0.2, 0.2, 0.2, 0.1),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

    def __len__(self): return len(self.file_paths)

    def __getitem__(self, idx):
        image = Image.open(self.file_paths[idx]).convert("RGB")
        if self.transform:
            image = self.transform(image)
        return image, self.labels[idx]

train_loader = DataLoader(
    train_dataset, batch_size=64, shuffle=True,
    num_workers=8, pin_memory=True, prefetch_factor=2,
    persistent_workers=True,
)

Step 3: Training Loop

PyTorch: model.train(), optimizer.zero_grad(), loss.backward(), optimizer.step(). Gradient clipping for stability. Learning rate scheduling: cosine annealing, ReduceLROnPlateau, OneCycleLR. Checkpointing: save model state, optimizer state, epoch, best metric. Early stopping: patience=5-10 epochs on validation metric. Logging: loss, learning rate, metrics per epoch.

def train_epoch(model, loader, optimizer, criterion, device, scaler=None):
    model.train()
    total_loss = 0
    for batch_idx, (inputs, targets) in enumerate(loader):
        inputs, targets = inputs.to(device), targets.to(device)

        optimizer.zero_grad()

        if scaler:
            with torch.amp.autocast(device_type="cuda"):
                outputs = model(inputs)
                loss = criterion(outputs, targets)
            scaler.scale(loss).backward()
            scaler.unscale_(optimizer)
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
            scaler.step(optimizer)
            scaler.update()
        else:
            outputs = model(inputs)
            loss = criterion(outputs, targets)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
            optimizer.step()

        total_loss += loss.item()

    return total_loss / len(loader)


def validate(model, loader, criterion, device):
    model.eval()
    total_loss = 0
    correct = 0
    total = 0
    with torch.no_grad():
        for inputs, targets in loader:
            inputs, targets = inputs.to(device), targets.to(device)
            outputs = model(inputs)
            loss = criterion(outputs, targets)
            total_loss += loss.item()
            _, predicted = outputs.max(1)
            total += targets.size(0)
            correct += predicted.eq(targets).sum().item()
    return total_loss / len(loader), 100.0 * correct / total

Step 4: CNN Architectures

ResNet: residual connections for deep networks (18, 34, 50, 101, 152). ResNet50 is the best accuracy/compute trade-off. EfficientNet: compound scaling (depth, width, resolution). EfficientNet-B0 to B7. MobileNet: depthwise separable convolutions for mobile/edge. ConvNeXt: modernized ResNet with Swish, LayerNorm. Transfer learning: freeze pretrained backbone, train new classifier head. Fine-tune whole network at low learning rate (1e-5 to 1e-4).

import torchvision.models as models

model = models.resnet50(weights="DEFAULT")
for param in model.parameters():
    param.requires_grad = False

num_features = model.fc.in_features
model.fc = nn.Sequential(
    nn.Dropout(0.2),
    nn.Linear(num_features, 256),
    nn.ReLU(),
    nn.Linear(256, num_classes),
)

optimizer = optim.Adam(model.fc.parameters(), lr=1e-3)

for param in model.parameters():
    param.requires_grad = True
optimizer = optim.Adam(model.parameters(), lr=1e-5)

Step 5: RNN and Transformer

LSTM: for sequence modeling (time-series, text). Bidirectional LSTM: context from both directions. GRU: lighter than LSTM, similar performance. Transformer: self-attention for parallel processing, positional encoding for sequence order. Multi-head attention: multiple attention heads capture different relationships.

class TransformerClassifier(nn.Module):
    def __init__(self, vocab_size, d_model=256, nhead=8, num_layers=4, num_classes=2):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, d_model)
        self.pos_encoder = PositionalEncoding(d_model, dropout=0.1)
        encoder_layer = nn.TransformerEncoderLayer(d_model, nhead, dim_feedforward=1024, dropout=0.1)
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers)
        self.classifier = nn.Linear(d_model, num_classes)

    def forward(self, x):
        x = self.embedding(x) * math.sqrt(x.size(-1))
        x = self.pos_encoder(x)
        x = self.transformer(x)
        x = x.mean(dim=0)
        return self.classifier(x)


class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super().__init__()
        self.dropout = nn.Dropout(dropout)
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0).transpose(0, 1)
        self.register_buffer("pe", pe)

    def forward(self, x):
        x = x + self.pe[:x.size(0), :]
        return self.dropout(x)

Step 6: Distributed Training

DDP (DistributedDataParallel): each GPU has model copy, gradient sync at each step. Best for models fitting in single GPU memory. FSDP (Fully Sharded Data Parallel): shard model parameters, gradients, optimizer states across GPUs. Best for large models (1B+ params). Launch: torchrun for DDP/FSDP. Gradient checkpointing: trade compute for memory (train larger models on fewer GPUs).

import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data.distributed import DistributedSampler

def setup_ddp(rank, world_size):
    dist.init_process_group("nccl", rank=rank, world_size=world_size)
    torch.cuda.set_device(rank)

def cleanup_ddp():
    dist.destroy_process_group()

def train_ddp(rank, world_size):
    setup_ddp(rank, world_size)
    model = ImageClassifier().to(rank)
    model = DistributedDataParallel(model, device_ids=[rank])
    train_sampler = DistributedSampler(train_dataset, num_replicas=world_size, rank=rank)
    train_loader = DataLoader(train_dataset, batch_size=64, sampler=train_sampler, num_workers=4)
    optimizer = optim.AdamW(model.parameters(), lr=1e-4)

    for epoch in range(10):
        train_sampler.set_epoch(epoch)
        train_epoch(model, train_loader, optimizer, criterion, rank)
    cleanup_ddp()

if __name__ == "__main__":
    world_size = torch.cuda.device_count()
    torch.multiprocessing.spawn(train_ddp, args=(world_size,), nprocs=world_size)

Step 7: Training Visualization and Logging

Use TensorBoard, WandB, or MLflow for tracking. Log loss, learning rate, gradients, and activations histograms. Monitor GPU utilization with nvidia-smi. Detect vanishing/exploding gradients by tracking gradient norms.

from torch.utils.tensorboard import SummaryWriter

writer = SummaryWriter("runs/experiment_001")

for epoch in range(num_epochs):
    train_loss = train_epoch(...)
    val_loss, val_acc = validate(...)

    writer.add_scalar("Loss/train", train_loss, epoch)
    writer.add_scalar("Loss/val", val_loss, epoch)
    writer.add_scalar("Acc/val", val_acc, epoch)
    writer.add_scalar("LR", optimizer.param_groups[0]["lr"], epoch)

    for name, param in model.named_parameters():
        if param.grad is not None:
            writer.add_histogram(f"grad/{name}", param.grad, epoch)

    scheduler.step()

Step 8: Hyperparameter Optimization

Use Optuna for Bayesian optimization of hyperparameters. Define search space for learning rate, batch size, optimizer, architecture parameters. Use pruning (MedianPruner, HyperbandPruner) to terminate unpromising trials early.

import optuna

def objective(trial):
    lr = trial.suggest_float("lr", 1e-5, 1e-2, log=True)
    batch_size = trial.suggest_categorical("batch_size", [32, 64, 128, 256])
    dropout = trial.suggest_float("dropout", 0.1, 0.5)
    optimizer_name = trial.suggest_categorical("optimizer", ["Adam", "AdamW", "SGD"])

    model = ModifiedResNet(dropout=dropout)
    optimizer = getattr(optim, optimizer_name)(model.parameters(), lr=lr)
    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=50)

    for epoch in range(50):
        train_loss = train_epoch(model, train_loader, optimizer, criterion, device)
        val_loss, val_acc = validate(model, val_loader, criterion, device)
        scheduler.step()
        trial.report(val_acc, epoch)
        if trial.should_prune():
            raise optuna.TrialPruned()

    return val_acc

study = optuna.create_study(direction="maximize", pruner=optuna.pruners.MedianPruner())
study.optimize(objective, n_trials=50)

Architecture / Decision Trees

Architecture Selection

Input type
  ├── Images
  │   ├── Classification → ResNet, EfficientNet, ViT
  │   ├── Detection → YOLO, DETR, Faster R-CNN
  │   └── Segmentation → U-Net, DeepLab, Mask R-CNN
  ├── Text / Sequences
  │   ├── Short sequences (< 512 tokens) → BERT, RoBERTa
  │   ├── Long sequences (> 512 tokens) → Longformer, BigBird
  │   ├── Generation → GPT, LLaMA, Mistral
  │   └── Time series → LSTM, GRU, TCN, PatchTST
  ├── Audio / Speech
  │   ├── Classification → Wav2Vec2, HuBERT
  │   └── Generation → WaveNet, MusicGen
  └── Multi-modal
      ├── Image + Text → CLIP, BLIP, LLaVA
      └── Any modality → Perceiver IO

Common Pitfalls

1. Data pipeline bottleneck: GPU idle while CPU loads data. Fix: num_workers=4-8, prefetch, pin_memory.
2. Learning rate too high/low: divergence or slow convergence. Fix: learning rate range test, cosine annealing.
3. Overfitting on small datasets: insufficient regularization. Fix: dropout, weight decay, data augmentation.
4. Gradient vanishing/exploding: common in deep networks and RNNs. Fix: residual connections, gradient clipping, proper initialization.
5. Batch size too large for GPU memory: CUDA OOM. Fix: gradient accumulation, gradient checkpointing, mixed precision.
6. No validation during training: no signal for early stopping or overfitting detection.
7. Not shuffling training data: batches have same distribution bias.
8. Saving model without optimizer state: can't resume training after interruption.
9. Mixed precision without loss scaling: underflow for small gradients. AMP handles this automatically in PyTorch 1.6+.
10. DDP without DistributedSampler: each GPU sees full dataset. Fix: use DistributedSampler with set_epoch.
11. Not setting dropout to eval mode: dropout active during validation causes inconsistent metrics.
12. Forgetting model.to(device): tensors on CPU causing device mismatch errors.

Best Practices

• Use PyTorch for research and custom architectures, TF for production pipelines.
• DataLoader with num_workers=4-8 and pin_memory=True always.
• Mixed precision (AMP) for 1.5-3x throughput on Volta+ GPUs.
• Gradient clipping (max_norm=1.0) for RNNs and transformers.
• Learning rate scheduling always (cosine annealing or ReduceLROnPlateau).
• Checkpoint every N epochs, keep best 3 checkpoints.
• DDP for model fitting in single GPU, FSDP for larger models.
• Gradient checkpointing when memory constrained.
• Validate on GPU, no .cpu() in validation loop.
• Set deterministic flags for reproducibility (at cost of speed).
• Use tensor shape comments for readability: # (B, C, H, W).
• Log hyperparameters, metrics, and model graph to experiment tracker.
• Profile with PyTorch Profiler to identify bottlenecks.
• Test with a single batch before full training to catch shape errors.
• Freeze batch norm during fine-tuning with small batches.
• Use label smoothing for classification (helps calibration).
• Warm up learning rate for first 5-10% of training steps.

Compared With

Feature	PyTorch	TensorFlow	JAX
Graph type	Dynamic (Eager)	Static (Graph) + Eager	JIT (XLA)
Debugging	Python-native	tfdbg	print inside jit
Distributed	DDP, FSDP, DeepSpeed	MirroredStrategy, TPUStrategy	pmap, shmap
Production	TorchServe, ONNX	TF Serving, TF Lite, TF.js	Limited
Community	Research dominant	Industry dominant	Growing research
HuggingFace	Primary	Supported	Supported
Deployment	ONNX, mobile	TF Lite, TF.js, mobile	XLA devices
Ecosystem	torchvision, torchaudio	TFX, KerasCV, KerasNLP	Flax, Haiku, Equinox

Deep learning vs classical ML: deep learning requires more data (typically > 100K samples), more compute (GPU), and more hyperparameter tuning. It excels with unstructured data (images, audio, text, video) and learns hierarchical representations automatically. Classical ML (gradient boosting) often beats deep learning on structured/tabular data with < 100K rows.

Performance

• AMP throughput gain: 1.5-3x on Volta+ (V100, A100, H100), minimal on older GPUs.
• DDP scaling: near-linear up to 8 GPUs, 80-90% efficiency at 64 GPUs with proper batch size scaling.
• FSDP memory reduction: 2-8x for large models (1B+ params) compared to DDP.
• Gradient checkpointing: 1.5-2x memory reduction, 15-25% compute overhead.
• DataLoader: 8 workers with prefetch = 2 saturates most NVMe/SSD read speeds.
• Profile with: torch.cuda.profiler.profile(), torch.autograd.profiler.profile().
• Memory optimization: use inplace ops, share buffers, avoid unnecessary tensor creation.
• Inference optimization: torch.compile (2-5x speedup), ONNX (2-3x), TensorRT (3-8x).
• Batch size: larger batches give better GPU utilization but may hurt generalization (use learning rate warmup).
• Memory-efficient attention: FlashAttention (2-4x faster attention, 5-10x memory reduction) for transformers.

Scalability: single GPU for models up to 7B params (quantized). 4-8 GPUs for 7B-70B. FSDP + CPU offloading for 70B+. TPU pods for 100B+ models.

Tooling

Tool	Purpose
PyTorch	Primary deep learning framework
TensorFlow / Keras	Production deep learning framework
JAX + Flax	High-performance research framework
HuggingFace Transformers	Pre-trained models, tokenizers, trainers
DeepSpeed	Large model training, ZeRO optimization
FSDP	Fully sharded data parallel (PyTorch native)
Lightning	Training loop abstraction, multi-GPU/TPU
Optuna	Hyperparameter optimization with pruning
Weights & Biases	Experiment tracking and visualization
TensorBoard	Training metrics visualization
ONNX	Cross-platform model export
TorchServe / TF Serving	Model serving
NVIDIA DALI	GPU-accelerated data loading
FlashAttention	Memory-efficient attention implementation
bitsandbytes	4-bit/8-bit quantization for large models

Rules

• Use PyTorch for research/custom models, TensorFlow for TFX production
• DataLoader with num_workers=4-8 and pin_memory=True always
• Mixed precision (AMP) for 1.5-3x throughput on Volta+ GPUs
• Gradient clipping (max_norm=1.0) for RNNs and transformers
• Learning rate scheduling always (cosine annealing or ReduceLROnPlateau)
• Checkpoint every N epochs, keep best 3 checkpoints
• DDP for model fitting in single GPU, FSDP for larger models
• Gradient checkpointing when memory constrained
• Validate on GPU, no .cpu() in validation loop
• Set deterministic flags for reproducibility (at the cost of speed)
• Profile data pipeline: GPU should never wait for data
• Use gradient accumulation for effective large batch size when GPU memory limited
• Freeze batch norm during fine-tuning with small batches
• Log all metrics, hyperparameters, and model graph
• Test with a single batch before full training
• Handle device placement explicitly — never assume tensors are on correct device
• Use tensor shape comments: # (batch, channels, height, width)

References

• references/architectures.md — Modern Deep Learning Architectures
• references/deep-learning-advanced.md — Deep Learning Advanced Topics
• references/deep-learning-fundamentals.md — Deep Learning Fundamentals
• references/generative-models.md — Generative Models
• references/pytorch-tensorflow.md — PyTorch and TensorFlow Reference
• references/training-optimization.md — Training Optimization Reference
• references/deep-learning-architecture-patterns.md — Deep Learning Architecture Patterns
• references/deep-learning-training-optimization.md — Training Optimization Deep Dive

Handoff

ml-experiment-tracking for logging deep learning experiments ml-feature-engineering for deep learning feature extraction

Architecture Decision Trees

Architecture Selection

Decision Point	Option A	Option B	Decision Criteria
Data type	Sequential → RNN/LSTM/Transformer	Spatial → CNN/ViT	Input structure, domain
Dataset size	Small → Transfer learning (pre-trained)	Large → Train from scratch	Data availability, compute budget
Interpretability	Simple (linear layers, attention)	Complex (deep residual networks)	Regulatory requirements, debugging
Latency requirement	Low → Shallow network, quantization	High throughput → Deep network, ensemble	Deployment constraints

Optimizer Selection

• Small dataset, sparse features → Adam with weight decay
• CV tasks, large batch → SGD with cosine annealing
• NLP/Transformers → AdamW with linear warmup
• GANs → Adam with beta1=0.5, beta2=0.999

Implementation Patterns

Transformer for Time Series

`python import torch import torch.nn as nn

class TimeSeriesTransformer(nn.Module): def init(self, input_dim, d_model=128, nhead=8, num_layers=4): super().init() self.input_proj = nn.Linear(input_dim, d_model) self.pos_encoding = nn.Parameter(torch.randn(1, 1000, d_model)) encoder_layer = nn.TransformerEncoderLayer( d_model=d_model, nhead=nhead, dim_feedforward=512, dropout=0.1, activation='gelu', batch_first=True ) self.transformer = nn.TransformerEncoder(encoder_layer, num_layers) self.output_head = nn.Linear(d_model, 1)

def forward(self, x):
    x = self.input_proj(x) + self.pos_encoding[:, :x.size(1), :]
    x = self.transformer(x)
    return self.output_head(x[:, -1, :])

`

Training Loop with Mixed Precision

`python import torch from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler() optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4, weight_decay=0.01) scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=100)

for epoch in range(num_epochs): for batch in dataloader: optimizer.zero_grad() with autocast(): outputs = model(batch['input']) loss = criterion(outputs, batch['target']) scaler.scale(loss).backward() scaler.unscale_(optimizer) torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0) scaler.step(optimizer) scaler.update() scheduler.step() `

Production Considerations

Scalability

• Distributed training: Use FSDP or DeepSpeed for large models. Enable activation checkpointing for memory efficiency.
• Model parallelism: Split large models across GPUs using pipeline parallelism. Use Tensor Parallelism for very large models.
• Inference serving: Deploy with TorchServe or Triton Inference Server. Use dynamic batching for throughput optimization.

Reproducibility

• Seed everything: Set seeds for all random generators. Log seed with each experiment in MLflow.
• Deterministic algorithms: Use torch.backends.cudnn.deterministic = True. Accept potential 10-20% training slowdown.
• Environment pinning: Pin CUDA, cuDNN, PyTorch minor versions. Use Docker container for consistent environment.

Anti-Patterns

Anti-Pattern	Symptom	Solution
Too big learning rate	Loss diverges to NaN	Use learning rate finder, cosine warmup
No validation during training	No early stopping, overfit	Always hold out validation set, monitor loss
Ignoring gradient clipping	Training instability with transformers	Clip gradients to max_norm=1.0
Single batch norm statistics	Batch norm fails on small batches	Use LayerNorm or GroupNorm for small batches
Overfitting on augmentation	Model fails on clean data	Cross-validate with clean validation set

Performance Optimization

Memory Efficiency

• Gradient checkpointing: Trade compute for memory, reduce memory 3-4x. Use torch.utils.checkpoint.checkpoint.
• Mixed precision training: Train in bfloat16 on Ampere+ GPUs. Maintains model quality while halving memory.
• Activation offloading: Offload activations to CPU during backward pass. Useful for extremely large models.

Throughput

• Data pipeline optimization: Use WebDataset for large-scale training. Pre-shard data to avoid IO bottleneck.
• Compiled models: Use torch.compile with inductor backend. Achieves 1.5-2x training speedup on modern GPUs.
• Flash Attention: Replace standard attention with Flash Attention v2. Achieves 2-4x attention speedup for long sequences.

Security Considerations

Model Security

• Backdoor detection: Scan model for hidden triggers using activation analysis. Validate all training data sources.
• Extraction attacks: Monitor API for suspicious pattern of queries. Rate-limit and add jitter to predictions.
• Inversion attacks: Prevent training data reconstruction from gradients. Use gradient noise or secure aggregation.

Infrastructure Security

• GPU isolation: Ensure GPU memory isolation in multi-tenant environments. Use CUDA MPS with memory limits.
• Model distribution: Sign model artifacts with GPG/containers. Only load models from signed and verified registries.