Lecture 6: Infrastructure & Tooling (Full Stack Deep Learning - Spring 2021)

Full Stack Deep Learning - UC Berkeley Spring 2021 - Sergey Karayev, Josh Tobin, Pieter Abbeel
Infrastructure & Tooling

Full Stack Deep Learning - UC Berkeley Spring 2021
Monitor predictions and close data
flywheel loop
Deploy model
Run experiments, review results
Provision compute
Write and debug model code
2
Provide data
Get optimal prediction system 
as scalable API or mobile app
Dream Reality
Aggregate, process, clean, label, 
and version data
Infrastructure & Tooling - Overview

Full Stack Deep Learning - UC Berkeley Spring 2021 3
Goal: add data, see model improve
Andrej Karpathy at PyTorch Devcon 2019 - https://www.youtube.com/watch?v=oBklltKXtDE

• Provide some labeled input-output pairs

• Press a button

• A prediction system that gets 100% accuracy on your data distribution is
live (as an infinitely-scalable API, or running embedded or mobile)
4
SE4ML: Software Engineering for Machine Learning (NIPS 2014

“All-in-one”
Training/Evaluation
6
Deployment
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Frameworks &
Distributed Training
Hyperparameter Tuning
Processing
Compute
Labeling
Software Engineering
Edge
Resource Management
Experiment Management
Data Lake / Warehouse
Feature
Store

“All-in-one”
Training/Evaluation Deployment
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &

Questions?
8

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - Software Engineering

Programming Language
• Python, because of the libraries

• Clear winner in scientific and data computing
10

Editors
11
VS Code
Vim
PyCharm
Emacs
Jupyter

Visual Studio Code
12
• VS Code makes for a very nice Python
experience
Cvjmu.jo!hju!tubhjoh!boe!ejggjoh
Pqfo!xipmf!qspkfdut!sfnpufmz!
Qffl!epdvnfoubujpo
Mjou!dpef!bt!zpvs!xsjuf

Visual Studio Code
13
Cvjmu.jo!hju!tubhjoh!boe!ejggjoh
Pqfo!qspkfdut!sfnpufmz!
Vtf!uif!ufsnjobm
Opufcppl!qpsu!gpsxbsejoh

Linters and Type Hints
• Whatever code style rules can be codified,
should be

• Static analysis can catch some bugs

• Static type checking both documents code
and catches bugs

• Will see in Lab 7
14

Jupyter Notebooks
• Notebooks have become fundamental to
data science

• Great as the "first draft" of a project

• Jeremy Howard from fast.ai good to  
learn from (course.fast.ai videos)

• Diﬃcult to make scalable, reproducible,
well-tested
15

Problems with notebooks
• Hard to version

• Notebook "IDE" is primitive

• Very hard to test

• Out-of-order execution artifacts

• Hard to run long or distributed tasks
16
https://towardsdatascience.com/5-reasons-why-jupyter-notebooks-suck-4dc201e27086

Jupyter Notebooks
• Counter-points:

• Netflix bases all ML workflows
on them
17
https://medium.com/netflix-techblog/notebook-innovation-591ee3221233

NBDev
• Counter-points:

• Jeremy Howard from fast.ai
uses them for everything, with
nbdev
18
https://github.com/fastai/nbdev

Streamlit
• New, but great at fulfilling a common ML need: interactive applets

• Decorate normal Python code

• Smart data caching, quick re-rendering

• In the works: sharing as easy as pushing a web app to Heroku
19
https://streamlit.io

Setting up environment
20
https://github.com/full-stack-deep-learning/conda-piptools
(How we do it in lab)

Questions?
21

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - Compute

23
Development Training/Evaluation
• Function
• Writing code

• Debugging models

• Looking at results

• Desiderata
• Quickly compile models and run training

• Nice-to-have: use GUI

• Solutions
• Desktop with 1-4 GPUs

• Cloud instance with 1-4 GPUs
• Function
• Model architecture / hyperparam search

• Training large models

• Desiderata
• Easy to launch experiments and review
results

• Solutions
• Desktop with 4 GPUs

• Private cluster of GPU machines

• Cloud cluster of GPU instances
or
or
Compute needs

Why compute matters
24
https://openai.com/blog/ai-and-compute/

25

So, or ?
• GPU Basics

• Cloud Options

• On-prem Options

• Analysis and Recommendations
26

GPU Basics
• NVIDIA has been the only game in town

• Google TPUs are the fastest (on GCP only)
27

28
GPU Comparison Table
https://www.microway.com/knowledge-center-articles/comparison-of-nvidia-geforce-gpus-and-nvidia-tesla-gpus/
Card Release Arch Use-case RAM (Gb) 32bit TFlops Tensor TFlops 16bit Cost Cloud
K80 2014H2 Kepler Server 24 5 N/A No - AWS, GCP, MS
P100 2016H1 Pascal Server 16 10 N/A Yes - GCP, MS
1080 Ti 2017H1 Pascal Consumer 11 13 N/A No $

used
V100 2017H1 Volta Server 16 14 120 Yes $$$$
AWS, GCP, MS
2080 Ti 2018H2 Turing Consumer 11 13 107 Yes $$
Titan RTX 2018H2 Turing Enthusiast 24 16 130 Yes $$
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
Ampere Server 40 19.5 312 Yes
$$$$ AWS, GCP

29
• New NVIDIA architecture every year

• Kepler —> Pascal —> Volta -> Turing -> Ampere

• Server version first, then “enthusiast”, then consumer.

• For business, only supposed to use server cards.

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

30
• RAM: should fit meaningful batches of your model
http://timdettmers.com/2018/11/05/which-gpu-for-deep-learning/

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

31

• 32bit vs Tensor TFlops
• Tensor Cores are specifically for deep learning operations (mixed precision)

• Good for convolutional/transformer models

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

32

• 32bit vs Tensor TFlops
• Tensor Cores are specifically for deep learning operations (mixed precision)

• Good for convolutional/transformer models

• Great speedups and bigger batches from 16bit mixed-precision

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

33
Kepler/Maxwell
• 2-4x slower than Pascal/Volta

• Hardware: don’t buy, too old

• Cloud: K80's are cheap (providers are stuck with what they bought)

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

34
Pascal
• Hardware: 1080 Ti still good if buying used, especially for recurrent

• Cloud: P100 is a mid-range option

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

35
Volta/Turing
• Preferred choice right now due to 16bit mixed precision support

• Hardware:

• 2080 Ti is ~1.3x as fast as 1080 Ti in 32bit, but ~2x faster in 16bit

• Titan RTX is 10-20% faster yet. Titan V is just as good (but less RAM), if find used.

• Cloud: V100 is the ultimate for speed
http://timdettmers.com/2018/11/05/which-gpu-for-deep-learning/ https://lambdalabs.com/blog/titan-rtx-tensorflow-benchmarks/

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

36
Ampere
• Latest hardware, with the most Tensor cores.

• At least 30% speedup over Turing
http://timdettmers.com/2018/11/05/which-gpu-for-deep-learning/ https://lambdalabs.com/blog/titan-rtx-tensorflow-benchmarks/

used
AWS, GCP, MS
3090 2021H1
Ampere
Enthusiast 24 ?
285 Yes
$$
A100

00A10000

2020H1
$$$$ AWS, GCP

37
https://timdettmers.com/2019/04/03/which-gpu-for-deep-learning/

A100 vs V100
38
https://lambdalabs.com/deep-learning/servers/hyperplane-a100
https://lambdalabs.com/blog/nvidia-a100-vs-v100-benchmarks/

Great resource
39
https://timdettmers.com/2020/09/07/which-gpu-for-deep-learning/#GPU_Recommendations

Cloud Providers
• Amazon Web Services, Google Cloud Platform, Microsoft Azure are the
heavyweights.

• Heavyweights are largely similar in function and price.

• AWS most expensive

• GCP just about as expensive, and has TPUs

• Azure reportedly has bad user experience

• Startups are Coreweave, Lambda Labs, and more
40

41
Amazon Web Services
Name GPU GPUs GPU RAM vCPU RAM On-demand
p2.16xlarge K80 8 12 64 732 $14.40
p3.16xlarge V100 8 16 64 488 $24.48
p3dn.24xlarge V100 8 32 96 768 $31.22
p4dn.16xlarge A100 8 40 96 1152 $32.78
• Three generations of GPUs

• Prices haven't moved in years

42
Google Cloud Platform
• Same lineup: K80, V100 (also P100), and A100

• In general, a little cheaper than AWS.

• Also has “Tensor Processing Units” (TPUs), the fastest option today.

43
Microsoft Azure
https://azure.microsoft.com/en-us/pricing/details/virtual-machines/linux/
• Same lineup: K80, P100s, V100s (no A100s yet)

• Similar pricing

Lambda Labs Cloud
44
https://lambdalabs.com/service/gpu-cloud

Coreweave
45
https://www.coreweave.com/pricing

On-prem Options
• Build your own

• Up to 4 Turing or 2 Ampere GPUs is easy

• Buy pre-built

• Lambda Labs, NVIDIA, and builders like Supermicro, Cirrascale, etc.
46

Building your own
• Quiet PC with 128GB RAM and 2x
RTX 3090’s: $8000

• (If you can find them)

• One day to build and set up

• Going beyond 4 2000-series or 2
3000-series is painful

• All you need to know: http://
timdettmers.com/2018/12/16/
deep-learning-hardware-guide/
47

Pre-built: Lambda Labs
48
　31&!npsf!fyqfotjwf!
uibo!cvjmejoh!zpvstfmg

Pre-built: Lambda Labs
49

Cost Analysis
• Let’s first compare on-prem and equivalent cloud machines

• Then let’s also consider spot instances for experiment scaling
50

Quad PC vs Quad Cloud
Verdict: not worth it. PC pays for itself in 5-10 weeks.
51
GPU Arch RAM Build Price Cloud Price
Hours =
Build
24/7 Weeks =
Build
16/5 Weeks =
Build
4x RTX
2080 Ti
Volta 12 $10000.00 0
4x V100 Volta 16 $12.00 833 5 10
Gvmm.ujnf!xpslmpbe
Xpsl!xffl!mpbe

Quad PC vs Quad Cloud
52
https://l7.curtisnorthcutt.com/build-pro-deep-learning-workstation

Quad PC vs. Spot Instances
How to think about it: cloud enables quicker experiments.
53
ipvst
ipvst
Length of trial in experiment (hours) 6
Number of trials in experiment 16
Total GPU hours for experiment 96
Cost of 4x RTX 2080 Ti machine $10,000.00
Time to run experiment on 4x machine 24
Time to run experiment on V100 spot instances 6
Cost of provisioning enough pre-emptible V100s $96.00
Number of experiments that equal cost of 4x
machine
104

Quad PC vs. Spot Instances
But at a pretty steep price.
54
ipvst
ipvst
Length of trial in experiment (hours) 6
Number of trials in experiment 16
Total GPU hours for experiment 96
Cost of 4x RTX 2080 Ti machine $10,000.00
Time to run experiment on 4x machine 24
Time to run experiment on V100 spot instances 6
Cost of provisioning enough pre-emptible V100s $96.00
Number of experiments that equal cost of 4x
machine
104

In Practice
• Even though cloud is expensive, it's hard to make on-prem scale past a
certain point

• Dev-ops (declarative infra, repeatable processes) definitely easier in the
cloud

• Maintenance is also a big factor
55

Recommendation for hobbyist
• Development

• Build a 4x Turing or 2x Ampere PC

• Training/Evaluation

• Use the same PC, just always keep it running

• To scale out, use Lambda or Coreweave cloud instances.
56

Recommendation for startup
• Development

• Buy a 4x Turing or 2x Ampere PC per ML Scientist

• To scale out, buy shared server machines or use cloud instances.
57

Recommendation for larger company
• Development

• Buy 8x Turing or Ampere server rack per ML scientist

• Or, go straight to the cloud with V100 instances

• To scale out, use cloud instances with proper provisioning and handling
of failures
58

Questions?
59

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - Resource Management

61
Resource Management
• Function
• Multiple people…

• using multiple GPUs/machines…

• running diﬀerent environments

• Goal
• Easy to launch a batch of experiments, with proper dependencies and resource allocations

• Solutions
• Python scripts

• SLURM

• Docker + Kubernetes

• Software specialized for ML use cases

• Problem we’re solving: allocate free
resources to programs

• Can be scripted pretty easily

• Even better, use old-school cluster job
scheduler

• Job defines necessary resources, gets
queued
Scripts or

• Docker is a way to package up an entire
dependency stack in a lighter-than-a-VM
package

• We will talk more about Docker in the
Deployment lecture, and use it in lab
Docker + Kubernetes

• Kubernetes is a way to run many docker
containers on top of a cluster
Docker + Kubernetes

• Open source project from Google

• Spawn and manage Jupyter
notebooks

• Manage multi-step ML workflows

• Plug-ins for hyperparameter tuning,
model deployment
65
https://github.com/Langhalsdino/Kubernetes-GPU-Guide

This is an active area
• All-in-one solutions like SageMaker and Paperspace Gradient do this

• And some recently-announced startups have it as their goal...
66

Anyscale (makers of Ray, from Berkeley)
67

Grid.ai
68

Questions?
69

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - Frameworks

71
Deep Learning Frameworks
• Unless you have a good reason not to, use
Tensorflow/Keras or PyTorch

• Both have converged to the same point:

• easy development via define-by-run

• multi-platform optimized execution graph

• Today, most new projects use PyTorch, because
of its more Python dev-friendly experience

• fast.ai library builds on PyTorch with best
practices

• PyTorch-Lightning adds a powerful training loop
Good
for
production
Good for development
Tensorflow 2.0
+ eager execution
+ TorchScript
PyTorch 1.0
3126
3126
3128
3129
312:

PyTorch dominates new development
72
https://thegradient.pub/state-of-ml-frameworks-2019-pytorch-dominates-research-tensorflow-dominates-industry/

PyTorch and Tensorflow
73

Why do we need frameworks?
• Deep Learning is not lot of code with a
matrix math library (Numpy)

• BUT: auto-diﬀerentiation and CUDA are a
lot of work

• ...Also all the layer types, optimizers, data
interfaces, etc.
74

• Recent project from Google that's gaining steam

• Numpy + auto-diﬀerentiation and compilation to GPU/TPU code

• Not just for deep learning!
75
https://github.com/google/jax#quickstart-colab-in-the-cloud

Hugging face
• Tons of NLP-focused model architectures (and pre-trained weights) for
both PyTorch and Tensorflow
76

• Using multiple GPUs and/or machines to train a single model.

• More complex than simply running diﬀerent experiments on diﬀerent
GPUs

• A must-do on big datasets and large models
77
Infrastructure & Tooling - Distributed Training

Data Parallelism
• If iteration time is too long, try training in data
parallel regime

• "For convolution, expect 1.9x/3.5x speedup
for 2/4 GPUs. 
78
http://www.cs.cmu.edu/~pengtaox/papers/petuum_15.pdf
https://lambdalabs.com/blog/titan-v-deep-learning-benchmarks/
Uijt!jt!bwfsbhf!pwfs!
ejggfsfou!dpowofut

Model Parallelism
79
http://www.cs.cmu.edu/~pengtaox/papers/petuum_15.pdf
• Model parallelism is necessary when
model does not fit on a single GPU

• Introduces a lot of complexity and is
usually not worth it. (But this is
changing.)

• Better to buy the largest GPU you can,
and/or use gradient checkpointing

Data-parallel PyTorch
• Can be quite easy:
80

Data-parallel PyTorch-Lightning
• Single node: 
python training/
run_experiment.py --gpus=8 --
accelerator=ddp

• And using SLURM to run on
multiple nodes: 
python training/
run_experiment.py --gpus=8 --
accelerator=ddp --nodes=4
81

Horovod
• Distributed training framework for Tensorflow, Keras, and PyTorch

• Uses MPI (standard multi-process communication framework) instead of
Tensorflow parameter servers

• Could be an easier experience for multi-node training
82

Ray (from Anyscale)
• Ray is open-source project
for eﬀortless, stateful
distributed computing in
Python

• From Berkeley!
83
https://ray.readthedocs.io/en/latest/walkthrough.html

Questions?
84

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - Experiment Management

• Even running one experiment at a time, can lose track of which code,
parameters, and dataset generated which trained model.

• When running multiple experiments, problem is much worse.
86
https://towardsdatascience.com/tracking-ml-experiments-using-mlflow-7910197091bb

87
Tensorboard
• A fine solution for single
experiments

• Gets unwieldy to manage many
experiments, and to properly
store past work

88
MLFlow tracking
https://towardsdatascience.com/tracking-ml-experiments-using-mlflow-7910197091bb
• Self-hosted solution
from DataBricks

89
Comet.ml

90
• What we use in lab

91
• Publish reports with
embedded charts,
figures, etc

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - Hyperparameter Tuning

Hyperparameter Optimization
• Useful to have software that helps you search over hyper parameter
settings.

• Could be as simple as being able to provide `—lr=(0.0001, 0.1) --
num_layers=[128, 256, 512]` to training script

• Would be even better if settings were selected intelligently, and
underperforming runs stopped early.
93

Ray Tune
• "Choose among scalable SOTA
algorithms such as Population
Based Training (PBT), Vizier’s
Median Stopping Rule,
HyperBand/ASHA."

• These redirect compute resources
toward promising areas of search
space
95

• What we use in Lab

Questions?
97

“All-in-one”
or or
Data
Web
Versioning
Monitoring
CI / Testing
Exploration
Sources
Processing
Compute
Labeling
Edge
Resource Management
Feature
Store
Frameworks &
Infrastructure & Tooling - All-in-one

All-in-one Solutions
• Single system for everything

• development (hosted notebook)

• scaling experiments to many machines (sometimes even provisioning)

• tracking experiments and versioning models

• deploying models

• monitoring performance
99

https://www.slideshare.net/AmazonWebServices/build-train-and-deploy-machine-learning-models-at-scale

• Open source!

• In-house distributed training module

• HyperBand based hyperparameter
tuning

• Smart GPU scheduling, including spot
instances

• Experiment tracking

Domino Data Lab
107
Provision compute
Track experiments

108
Domino Data Lab
Deploy REST API
Monitor predictions
Publish applets

109
Domino Data Lab
Monitor spend
All projects in one place

Natural place to go for most MLOps companies
110

111
W&B
Paperspace
Gradient
Floyd Determined.ai Domino Data Lab
Amazon
SageMaker
GC ML
Engine
Hardware N/A Paperspace GCP Agnostic Agnostic AWS GCP
Resource Management No Yes Yes Yes Yes Yes Yes
Hyperparam Optimization Yes No No Yes Yes Yes Yes
Storing Models Yes Yes Yes Yes Yes Yes Yes
Reviewing Experiments Yes Yes Yes Yes Yes Yes Yes
Deploying Models as REST API No No Yes No Yes Yes Yes
Monitoring No No No No Yes Yes Yes

Questions?
112

Tooling Tuesdays
113
https://twitter.com/full_stack_dl

Thank you!
114

Lecture 6: Infrastructure & Tooling (Full Stack Deep Learning - Spring 2021)

Recommended

More Related Content

What's hot (20)

Similar to Lecture 6: Infrastructure & Tooling (Full Stack Deep Learning - Spring 2021) (20)

More from Sergey Karayev (20)

Recently uploaded (20)

Lecture 6: Infrastructure & Tooling (Full Stack Deep Learning - Spring 2021)