17. Automatic Mixed Precision (torch.cuda.amp.autocast & GradScaler) for 2x Faster GPU Training • Deep Learning & PyTorch Neural Networks Architecture • Rishabh Ki Kaksha

Deep Learning & PyTorch Neural Networks Architecture › 17. Automatic Mixed Precision (torch.cuda.amp.autocast & GradScaler) for 2x Faster GPU Training

45 mins

← Prev Next →

Course Syllabus

Module 1: PyTorch Tensors, Autograd & Computational Graphs

01. PyTorch Tensor Architecture: Strides, Storage Buffers, Memory Layout (Contiguous vs Non-Contiguous) 02. Dynamic Computational Graphs & The PyTorch Autograd Engine (grad_fn, requires_grad, backward) 03. GPU Acceleration with CUDA: Tensor Device Placement (to("cuda")), CUDA Streams & Pinned Memory

Module 2: Multi-Layer Perceptrons (MLPs) & Custom Training Loops

04. Building Custom nn.Module Classes, Custom Parameters & Activation Functions (ReLU, GELU, Swish) 05. Custom Datasets & High-Performance DataLoaders (num_workers, collate_fn, pin_memory) 06. The Enterprise Training Loop: Forward Pass, Loss Computation, Zero Grad, Backward Pass & Optimizer Step 07. Neural Regularization Techniques: Dropout, Batch Normalization (BatchNorm1d) & Layer Normalization (LayerNorm)

Module 3: Convolutional Neural Networks (CNNs) & Computer Vision Architecture

08. Convolutional Operations: Kernels, Filters, Strides, Padding & Feature Map Dilation 09. Modern CNN Architectures: Residual Connections (ResNet), Squeeze-and-Excitation & ConvNeXt 10. Transfer Learning & Vision Fine-Tuning with Torchvision Pretrained Models

Module 4: Recurrent Architectures: RNNs, LSTMs & GRUs for Sequential Data

11. Sequential Modeling: Vanilla RNNs & The Vanishing / Exploding Gradient Problem 12. Long Short-Term Memory (LSTM) Networks: Forget Gate, Input Gate, Candidate State & Output Gate Math 13. Gated Recurrent Units (GRUs) & Bidirectional Sequence Processing in PyTorch

Module 5: Modern Transformer Architecture from Scratch in PyTorch

14. Scaled Dot-Product Attention: Query, Key, Value Matrices & Attention Softmax Calculation 15. Multi-Head Attention (MHA) & Causal Masking Implementation from Scratch in Pure PyTorch 16. Building a Complete GPT-Style Decoder-Only Transformer Language Model from Scratch

Module 6: PyTorch Distributed (DDP), Mixed Precision (AMP) & Model Checkpointing

17. Automatic Mixed Precision (torch.cuda.amp.autocast & GradScaler) for 2x Faster GPU Training 18. Distributed Data Parallel (DDP) Architecture: Multi-GPU Synchronization with torch.distributed 19. Robust Checkpointing: Saving State Dicts, Resuming Training & Early Stopping Callbacks 20. Model Export & Production Serialization: TorchScript (JIT Tracing), ONNX & TensorRT Compilation

Interactive Lesson Studio

# 17. Automatic Mixed Precision (torch.cuda.amp.autocast & GradScaler) for 2x Faster GPU Training

> Deep Learning & PyTorch Notes by Rishabh - Topic 17

---

### 1. FP32 vs FP16 vs BF16 Numerical Formats ⚡

*   **FP32 (Single Precision):** 1 sign, 8 exponent, 23 mantissa bits (High memory, slow Tensor Cores).
*   **FP16 (Half Precision):** 1 sign, 5 exponent, 10 mantissa (Underflow risk if gradients $< 6 \times 10^{-5}$).
*   **BF16 (Brain Floating Point - Modern Standard 🧠):** 1 sign, **8 exponent bits (Same dynamic range as FP32!)**, 7 mantissa. Zero gradient underflow!

---

### 2. PyTorch Mixed Precision Training Loop

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

model = MyTransformerModel().cuda()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
scaler = GradScaler() # Scales gradients to prevent FP16 underflow

for inputs, targets in dataloader:
    optimizer.zero_grad()
    inputs, targets = inputs.cuda(), targets.cuda()

# Forward pass in FP16 / BF16:
    with autocast(dtype=torch.bfloat16):
        outputs = model(inputs)
        loss = criterion(outputs, targets)

# Scaled Backward pass:
    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()
```

---

### 3. Master Comparison Matrix

| Format | Exponent Bits | Mantissa Bits | Dynamic Range | Needs GradScaler? |
| :--- | :--- | :--- | :--- | :--- |
| **FP32** | 8 | 23 | $10^{\pm 38}$ | ❌ No |
| **FP16** | 5 | 10 | $10^{\pm 4}$ (Small!) | ⚠️ **Yes (Mandatory)** |
| **BF16** | **8** | 7 | **$10^{\pm 38}$ (Massive!)** | 👑 **No (Native Stability)** |

---

### 4. Interview Ke 3 Sawal Is Topic Se

1. **NVIDIA Tensor Cores Mixed Precision par $4\times$ throughput kaise deliver karte hain?** -> FP16/BF16 matrix multiplication perform karke accumulation FP32 me karte hain.
2. **GradScaler gradient underflow kaise fix karta hai?** -> Loss ko large constant $2^{16}$ se multiply karta hai taaki small gradients FP16 range me aa jayein, aur optimizer step se pehle inverse divide karta hai.
3. **BF16 hardware support kab start hua?** -> NVIDIA Ampere architecture (A100, RTX 3090, H100) se.

---
**Next Topic:** 18. Distributed Data Parallel (DDP) Architecture: Multi-GPU Synchronization with torch.distributed

Loading lesson content...

Python 3.12

Full IDE

Live Execution Terminal

Click "Execute Live" to run code.
                

Lesson Mastery AI Quiz

AI Doubt Solver

CERTIFICATE OF COMPLETION

Rishabh Gehlot

PHP 8+ & Laravel 11 Master Track