Prev Next

Python / PyTorch Fundamentals Interview Questions

What is mixed precision training in PyTorch and how do you implement it with torch.cuda.amp?

Mixed precision training runs most operations in FP16 (or BF16) for speed while keeping a master copy of weights in FP32 for numerical stability. Modern GPUs (Volta and later) have dedicated hardware (Tensor Cores) that make FP16 matrix multiplication significantly faster than FP32.

import torch
import torch.nn as nn
from torch.cuda.amp import autocast, GradScaler

model     = nn.Linear(1024, 512).cuda()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3)
scaler    = GradScaler()   # manages loss scaling to prevent FP16 underflow

x = torch.randn(256, 1024).cuda()
y = torch.randn(256, 512).cuda()

for step in range(100):
    optimizer.zero_grad()

    # autocast: automatically runs eligible ops in FP16/BF16
    with autocast(device_type="cuda", dtype=torch.float16):
        y_hat = model(x)                  # matmul runs in FP16 — faster!
        loss  = nn.MSELoss()(y_hat, y)

    # Loss scaling: inflate loss before backward to prevent small
    # gradients from underflowing to zero in FP16's limited range
    scaler.scale(loss).backward()
    scaler.unscale_(optimizer)            # restore original gradient magnitudes
    nn.utils.clip_grad_norm_(model.parameters(), 1.0)
    scaler.step(optimizer)                # skips the step if grads are inf/NaN
    scaler.update()                       # adjusts scale factor for next iteration

# BFloat16: no GradScaler needed (same exponent range as FP32)
with autocast(device_type="cuda", dtype=torch.bfloat16):
    y_hat = model(x)   # no underflow risk — scaling unnecessary

What problem does GradScaler solve in FP16 mixed precision training?
Why doesn't BFloat16 require a GradScaler while Float16 does?

More Related questions...

What is PyTorch and what are its key advantages over other deep learning frameworks? What is a PyTorch tensor and how does it differ from a NumPy array? What are the most important tensor operations in PyTorch? What are tensor data types (dtypes) in PyTorch and why do they matter? How does broadcasting work in PyTorch and what are the rules? What is autograd in PyTorch and how does it compute gradients? What is the computation graph in PyTorch and how does the dynamic graph differ from a static graph? How do torch.no_grad() and tensor.detach() differ, and when do you use each? What is nn.Module and how do you build a custom neural network in PyTorch? What are nn.Sequential and other container modules in PyTorch? What built-in layers does PyTorch's nn module provide and how do you use the most common ones? What are activation functions in PyTorch and how do you apply them? What are the most important loss functions in PyTorch and when do you use each? What optimizers does PyTorch provide and how do you configure them? What are learning rate schedulers in PyTorch and how do you use them? What are the most common built-in layers in torch.nn and what do they do? How do you initialise weights in a PyTorch model? What loss functions does PyTorch provide and when do you use each? What optimizers does PyTorch provide and how do you choose between them? What are learning rate schedulers in PyTorch and how do you use them? What activation functions are commonly used in PyTorch and how do you choose between them? What loss functions does PyTorch provide and how do you choose the right one? What optimizers does PyTorch provide and what is the difference between SGD, Adam, and AdamW? What is the standard PyTorch training loop and what does each step do? What are Dataset and DataLoader in PyTorch and how do they work together? How do you move tensors and models between CPU and GPU in PyTorch? What is the difference between model.parameters() and model.state_dict() in PyTorch? How do you save and load PyTorch models correctly, including full training checkpoints? What is overfitting and what regularization techniques does PyTorch support to address it? What is the vanishing/exploding gradient problem and how do you detect and fix it in PyTorch? What is weight initialization in PyTorch and why does it matter? What is the difference between nn.Parameter and a regular tensor attribute in nn.Module? How do you implement and use learning rate schedulers in PyTorch? How do you debug a PyTorch training loop where the loss is not decreasing or is NaN? What is the difference between torch.tensor() and torch.Tensor() (capital T) for creating tensors? How does gradient accumulation work in PyTorch and when would you use it? What is mixed precision training in PyTorch and how do you implement it with torch.cuda.amp? What is torch.compile() and how does it speed up PyTorch model execution? What is the difference between batch size, epoch, and iteration in PyTorch training? How do you compute and track evaluation metrics like accuracy during PyTorch training? What is the purpose of torch.manual_seed() and how do you ensure reproducibility in PyTorch? How does PyTorch handle multi-dimensional indexing and slicing of tensors? What is the difference between.view(),.reshape(), and.contiguous() in PyTorch, and why does it matter? How do you freeze layers and perform transfer learning / fine-tuning in PyTorch? What is the purpose of torch.utils.data.random_split() and how do you create train/validation/test splits in PyTorch? What is Batch Normalization in PyTorch and how does it differ from Layer Normalization? How do you implement and use a custom loss function in PyTorch? What is torch.compile() vs TorchScript and how do you export a PyTorch model for production deployment?
Show more question and Answers...

Uvicorn Fundamentals Interview Questions

Comments & Discussions

Interviews Questions

About Javapedia.net

Javapedia.net is for Java and J2EE developers, technologist and college students who prepare of interview. Also this site includes many practical examples.

This site is developed using J2EE technologies by Steve Antony, a senior Developer/lead at one of the logistics based company.

Follow Javapedia updates

contact: javatutorials2016[at]gmail[dot]com

Kindly consider donating for maintaining this website. Thanks.

Copyright © 2026, javapedia.net, all rights reserved. · .

DMCA.com Protection Status