Nvidia GPU Tech Conference - Optimizing, Profiling, and Deploying TensorFlow AI Models in...

HIGH PERFORMANCE TENSORFLOW IN PRODUCTION WITH GPUSNVIDIA GPU TECH CONFERENCEWASHINGTON DC, NOV 2017CHRIS FREGLY, FOUNDER@PIPELINE.AI

INTRODUCTIONS: ME§ Chris Fregly, Founder & Engineer @ PipelineAI

§ Formerly Netflix and Databricks

§ Advanced Spark and TensorFlow MeetupPlease Join Our 50,000+ Members Globally!

Contact Mechris@pipeline.ai

@cfregly

* San Francisco

* Chicago* Austin* Washington DC* Dusseldorf* London

INTRODUCTIONS: YOU§ Software Engineer or Data {Scientist, Engineer, Analyst}

§ Interested in Optimizing + Deploying TF Models to Production

§ Nice to have working knowledge of TensorFlow, but not required

CONTENT BREAKDOWN50% Training Optimizations (GPUs, Pipeline, XLA+JIT)50% Prediction Optimizations (XLA+AOT, TF Serving)

Why Heavy Focus on Model Prediction vs. Model Training?

50 Data Scientists <<< 120 Million App Users

TrainingBoring & Batch

PredictionExciting & Real-Time!!

100% OPEN SOURCE CODE§ https://github.com/PipelineAI/pipeline/

§ Please ! this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu.ml

HANDS-ON EXERCISES§ Combo of Jupyter Notebooks and Command Line§ Command Line through Jupyter Terminal

§ Some Exercises Based on Experimental Features

You May See Errors. Stay Calm. You Will Be OK!!

OTHER TENSORFLOW TALKS AT GTC!

4:30-4:55pm Todayin Hemisphere B.

AGENDA

Part 0: PipelineAI Research

Part 1: TensorFlow Model Training

Part 2: TensorFlow Model Serving

PIPELINE.AI OVERVIEW

AGENDA

Part 0: PipelineAI Research

§ Package, Deploy, and Tune Both Model + Runtime§ Experiment Safely in Production§ Compare Models + Runtimes Both Offline + Online§ Shift Traffic to Winning Model or Cheaper Cloud!

PACKAGE MODEL + RUNTIME AS ONE

§ Package Model + Runtime into Immutable Docker Image

§ Same Package: Local, Dev, and Prod

§ No Dependency Surprises in Production

§ Deploy (and Tune )Model + Runtime as a Single Unit

OPTIMIZE MODEL + RUNTIME AS ONE

§ Tune BOTH Model Params + Runtime Configs Together

§ Generate Native CPU + GPU Code

§ Quantize Model Weights + Activations

§ Swap Runtimes: TensorFlow Serving, Nvidia TensorRT, …

NVIDIA TENSOR-RT RUNTIME

§ Post-Training Model Optimizations§ Similar to TF Graph Transform Tool

§ GPU-Optimized Prediction Runtime§ Alternative to TensorFlow Serving

§ PipelineAI Supports TensorRT!

AGENDA

Part 0: PipelineAI Research

§ Package, Deploy, and Tune Both Model + Runtime§ Experiment Safely in Production§ Compare Models + Runtimes Both Offline + Online§ Shift Traffic to Winning Model or Cheaper Cloud!

EXPERIMENT SAFELY IN PRODUCTION

§ Setup Experiments Directly from Jupyter Notebooks

§ Deploy to 1% Prod Traffic

§ Deploy in Traffic-Shadow Mode

§ Tear-Down or Rollback Experiments Quickly

AGENDA

Part 0: PipelineAI Research

§ Package, Deploy, and Tune Both Model + Runtime§ Experiment Safely in Production§ Compare Models + Runtimes Both Offline + Online§ Shift Traffic to Winning Model or Cheaper Cloud!

COMPARE MODELS OFFLINE + ONLINE

§ Offline Metrics§ Training Accuracy§ Validation Accuracy

§ Online Real-Time Metrics§ Prediction Precision§ Latency + Throughput

PREDICTION PROFILING + TUNING§ Pinpoint Performance Bottlenecks

§ Fine-Grained Prediction Metrics

§ 3 Logic Steps in a Prediction1.transform_request()2.predict()3.transform_response()

VIEW LIVE PREDICTION STREAMS§ Visually Compare Real-Time Predictions

AGENDA

Part 0: PipelineAI Research

§ Package, Deploy, and Tune Both Model + Runtime§ Experiment Safely in Production§ Compare Models + Runtimes Both Offline + Online§ Shift Traffic to Winning Model or Cheaper Cloud!

SHIFT TRAFFIC TO MAXIMIZE REVENUE§ Shift Traffic to Winning Model using AI Bandit Algorithms

SHIFT TRAFFIC TO MINIMIZE COST

§ Real-Time Cost Per Prediction

§ Across Clouds + On-Premise

§ Explore/Exploit Bandits

CONTINUOUS MODEL TRAINING§ Identify and Fix Borderline Predictions (50-50% Confidence)

§ Fix Along Class Boundaries

§ Retrain on New Labeled Data

§ Enables Crowd Sourcing

§ Game-ify Labeling Process

AGENDA

Part 0: PipelineAI Research

Part 1: TensorFlow Model Training

Part 2: TensorFlow Model Serving

AGENDA

Part 1: TensorFlow Model Training

§ GPUs and TensorFlow§ Train, Inspect, and Debug TensorFlow Models§ TensorFlow Distributed Model Training on a Cluster§ Optimize Training with JIT XLA Compiler

EVERYBODY GETS A GPU!

SETUP ENVIRONMENT

§ Step 1: Browse to the following:http://allocator.community.pipeline.ai/allocate

§ Step 2: Browse to the following:http://<ip-address>

§ Step 3: Browse around. I will provide a Jupyter Username/Password soon.

Need Help? Use the Chat!

VERIFY SETUP

http://<ip-address>

Any username,Any password!

LET’S EXPLORE OUR ENVIRONMENT§ Navigate to the following notebook:

01_Explore_Environment

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

PULSE CHECK

BREAK

§ Please ! this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu.ml

Need Help? Use the Chat!

SETTING UP TENSORFLOW WITH GPUS

§ Very Painful!

§ Especially inside Docker§ Use nvidia-docker

§ Especially on Kubernetes!§ Use Kubernetes 1.7+

§ http://pipeline.ai for GitHub + DockerHub Links

GPU HALF-PRECISION SUPPORT§ FP32 is “Full Precision”, FP16 is “Half Precision”§ Supported by Pascal P100 (2016) and Volta V100 (2017)§ Half-Precision is OK for Approximate Deep Learning Use Cases§ Fit Two(2) FP16’s into FP32 GPU Cores for 2x Throughput!

You Can Set TF_FP16_MATMUL_USE_FP32_COMPUTE=0

on GPU w/ Compute Capability(CC) 5.3+

VOLTA V100 (2017) VS. PASCAL P100 (2016)§ 84 Streaming Multiprocessors (SM’s)§ 5,376 GPU Cores§ 672 Tensor Cores (ie. Google TPU)

§ Mixed FP16/FP32 Precision § More Shared Memory§ New L0 Instruction Cache§ Faster L1 Data Cache§ V100 vs. P100 Performance

§ 12x TFLOPS @ Peak Training§ 6x Inference Throughput

V100 AND CUDA 9§ Independent Thread Scheduling - Finally!!

§ Similar to CPU fine-grained thread synchronization semantics§ Allows GPU to yield execution of any thread

§ Still Optimized for SIMT (Same Instruction Multiple Thread)§ SIMT units automatically scheduled together

§ Explicit Synchronization

P100 V100

GPU CUDA PROGRAMMING

§ Barbaric, But Fun Barbaric

§ Must Know Hardware Very Well

§ Hardware Changes are Painful

§ Many Great Debuggers Exist

CUDA STREAMS

§ Asynchronous I/O Transfer

§ Overlap Compute and I/O

§ Keeps GPUs Saturated

§ Fundamental to Queue Framework in TensorFlow

LET’S SEE WHAT THIS THING CAN DO!§ Navigate to the following notebook:

01a_Explore_GPU01b_Explore_Numba

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

AGENDA

Part 1: TensorFlow Model Training

§ GPUs and TensorFlow§ Train, Inspect, and Debug TensorFlow Models§ TensorFlow Distributed Model Training on a Cluster§ Optimize Training with JIT XLA Compiler

TRAINING TERMINOLOGY§ Tensors: N-Dimensional Arrays§ ie. Scalar, Vector, Matrix

§ Operations: MatMul, Add, SummaryLog,…§ Graph: Graph of Operations (DAG)§ Session: Contains Graph(s)§ Feeds: Feed Inputs into Placeholder§ Fetches: Fetch Output from Operation§ Variables: What We Learn Through Training§ aka “weights”, “parameters”

§ Devices: Hardware device on which we train

-TensorFlow-Trains

Variables

-User-FetchesOutputs

-User-FeedsInputs

-TensorFlow-Performs

Operations

-TensorFlow-Flows

Tensors

with tf.device(“/gpu:0,/gpu:1”):

TENSORFLOW MODEL§ MetaGraph

§ Combines GraphDef and Metadata§ GraphDef

§ Architecture of your model (nodes, edges)

§ Metadata§ Asset: Accompanying assets to your model§ SignatureDef: Maps external : internal tensors

§ Variables§ Stored separately during training (checkpoint)§ Allows training to continue from any checkpoint§ Variables are “frozen” into Constants when deployed for inference

GraphDef

x

W

mul add

b

MetaGraphMetadata

AssetsSignatureDef

TagsVersion

Variables:“W” : 0.328“b” : -1.407

TENSORFLOW SESSION

Session

graph: GraphDef

Variables:“W” : 0.328“b” : -1.407

Variables arePeriodically

Checkpointed

GraphDefis Static

EXTEND EXISTING DATA PIPELINES§ Data Processing

§ HDFS/Hadoop§ Spark

§ Containers§ Docker

§ Schedulers§ Kubernetes§ Mesos

<dependency> <groupId>org.tensorflow</groupId> <artifactId>tensorflow-hadoop</artifactId>

</dependency>

https://github.com/tensorflow/ecosystem

DON’T USE FEED_DICT!!§ Not Optimized for Production Pipelines§ Single-Threaded, Synchronous, SLOW!§ feed_dict Requires Python <-> C++ Serialization§ Retrieves Next Batch After Current Batch is Done§ CPUs/GPUs Not Fully Utilized!§ Use Queue or Dataset API

sess.run(train_step, feed_dict={…}

QUEUES

§ More than traditional Queue§ Uses CUDA Streams§ Perform I/O, pre-processing, cropping, shuffling, …§ Pull from HDFS, S3, Google Storage, Kafka, ...§ Combine many small files into large TFRecord files§ Use CPUs to free GPUs for compute§ Helps saturate CPUs and GPUs

QUEUE CAPACITY PLANNING§ batch_size

§ # examples / batch (ie. 64 jpg)§ Limited by GPU RAM

§ num_processing_threads§ CPU threads pull and pre-process batches of data§ Limited by CPU Cores

§ queue_capacity§ Limited by CPU RAM (ie. 5 * batch_size)

DETECT UNDERUTILIZED CPUS, GPUS

§ Instrument training code to generate “timelines”

§ Analyze with Google Web Tracing Framework (WTF)

§ Monitor CPU with `top`, GPU with `nvidia-smi`

http://google.github.io/tracing-framework/

from tensorflow.python.client import timeline

trace = timeline.Timeline(step_stats=run_metadata.step_stats)

with open('timeline.json', 'w') as trace_file:trace_file.write(trace.generate_chrome_trace_format(show_memory=True))

LET’S FEED DATA WITH A QUEUE§ Navigate to the following notebook:

02_Feed_Queue_HDFS

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

PULSE CHECK

BREAK

§ Please ! this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu.ml

Need Help? Use the Chat!

LET’S TRAIN A MODEL (CPU)§ Navigate to the following notebook:

03_Train_Model_CPU

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

LET’S TRAIN A MODEL (GPU)§ Navigate to the following notebook:

03a_Train_Model_GPU

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

TENSORFLOW DEBUGGER§ Step through Operations§ Inspect Inputs and Outputs§ Wrap Session in Debug Session

sess = tf.Session(config=config)sess =

tf_debug.LocalCLIDebugWrapperSession(sess)

LET’S DEBUG A MODEL§ Navigate to the following notebook:

04_Debug_Model

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

AGENDA

Part 1: TensorFlow Model Training

§ GPUs and TensorFlow§ Train, Inspect, and Debug TensorFlow Models§ TensorFlow Distributed Model Training on a Cluster§ Optimize Training with JIT XLA Compiler

SINGLE NODE, MULTI-GPU TRAINING§ cpu:0

§ By default, all CPUs§ Requires extra config to target a CPU

§ gpu:0..n§ Each GPU has a unique id§ TF usually prefers a single GPU

§ xla_cpu:0, xla_gpu:0..n§ “JIT Compiler Device”§ Hints TensorFlow to attempt JIT Compile

with tf.device(“/cpu:0”):

with tf.device(“/gpu:0”):

with tf.device(“/gpu:1”):

GPU 0 GPU 1

DISTRIBUTED, MULTI-NODE TRAINING§ TensorFlow Automatically Inserts Send and Receive Ops into Graph§ Parameter Server Synchronously Aggregates Updates to Variables§ Nodes with Multiple GPUs will Pre-Aggregate Before Sending to PS

Worker0 Worker0

Worker1

Worker0 Worker1 Worker2

gpu0 gpu1

gpu2 gpu3

gpu0 gpu1

gpu2 gpu3

gpu0 gpu1

gpu2 gpu3

gpu0

gpu1

gpu0

gpu0

DATA PARALLEL VS MODEL PARALLEL

§ Data Parallel (“Between-Graph Replication”)§ Send exact same model to each device§ Each device operates on its partition of data

§ ie. Spark sends same function to many workers§ Each worker operates on their partition of data

§ Model Parallel (“In-Graph Replication”)§ Send different partition of model to each device§ Each device operates on all data

Very Difficult!!

Required for Large Models.(GPU RAM Limitation)

SYNCHRONOUS VS. ASYNCHRONOUS§ Synchronous

§ Nodes compute gradients§ Nodes update Parameter Server (PS)§ Nodes sync on PS for latest gradients

§ Asynchronous§ Some nodes delay in computing gradients§ Nodes don’t update PS§ Nodes get stale gradients from PS§ May not converge due to stale reads!

CHIEF WORKER§ Worker Task 0 is Usually the Chief

§ Task 0 is guaranteed to exist

§ Performs Maintenance Tasks§ Writes log summaries§ Instructs PS to checkpoint vars§ Performs PS health checks§ (Re-)Initialize variables at (re-)start of training

NODE AND PROCESS FAILURES

§ Checkpoint to Persistent Storage (HDFS, S3)§ Use MonitoredTrainingSession and Hooks§ Use a Good Cluster Orchestrator (ie. Kubernetes,Mesos)§ Understand Failure Modes and Recovery States

Stateless, Not Bad: Training Continues Stateful, Bad: Training Must Stop Dios Mio! Long Night Ahead…

LET’S TRAIN DISTRIBUTED TENSORFLOW§ Navigate to the following notebook:

05_Train_Model_Distributed_CPUor 05a_Train_Model_Distributed_GPU

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

USE EXPERIMENT AND ESTIMATOR API§ Unified API: Local + Distributed§ Clean Hooks for Data Retrieval§ Config: TF_CONFIG + RunConfig§ Hyper-Parameter Tuning with HParams§ run() or tune() using learn_runner API§ Works Well with Google Cloud ML (Surprised?!)

https://github.com/GoogleCloudPlatform/cloudml-samples/blob/master/census/customestimator/trainer/{model.py, task.py}

BATCH NORMALIZATION

§ Each Mini-Batch May Have Wildly Different Distributions§ Normalize per Batch (and Layer)§ Faster Training§ Weights are Learned Quicker§ Final Model is More Accurate§ Final mean + variance Are Folded Into Our Graph Later

-- (Almost)Always Use Batch Normalization! --

z = tf.matmul(a_prev, W)a = tf.nn.relu(z)

a_mean, a_var = tf.nn.moments(a, [0])

scale = tf.Variable(tf.ones([depth/channels]))beta = tf.Variable(tf.zeros ([depth/channels]))

bn = tf.nn.batch_normalizaton(a, a_mean, a_var, beta, scale, 0.001)

OPTIMIZE GRAPH EXECUTION ORDER

§ https://github.com/yaroslavvb/stuff

Linearize to minimize graphmemory usage

SEPARATE TRAINING + VALIDATION

§ Separate Training and Validation Clusters

§ Validate Upon Checkpoint

§ Avoid Resource Contention

§ Let Training Continue in Parallel with Validation

TrainingCluster

ValidationCluster

Parameter ServerCluster

PULSE CHECK

BREAK

§ Please ! this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu.ml

Need Help? Use the Chat!

AGENDA

Part 1: TensorFlow Model Training

§ GPUs and TensorFlow§ Train, Inspect, and Debug TensorFlow Models§ TensorFlow Distributed Model Training on a Cluster§ Optimize Training with JIT XLA Compiler

XLA FRAMEWORK§ Accelerated Linear Algebra (XLA)§ Goals:

§ Reduce reliance on custom operators§ Improve execution speed§ Improve memory usage§ Reduce mobile footprint§ Improve portability

§ Helps TF Stay Flexible and Performant

XLA HIGH LEVEL OPTIMIZER (HLO)

§ Compiler Intermediate Representation (IR)§ Independent of source and target language§ Define Graphs using HLO Language§ XLA Step 1 Emits Target-Independent HLO § XLA Step 2 Emits Target-Dependent LLVM§ LLVM Emits Native Code Specific to Target § Supports x86-64, ARM64 (CPU), and NVPTX (GPU)

JIT COMPILER§ Just-In-Time Compiler§ Built on XLA Framework§ Goals:

§ Reduce memory movement – especially useful on GPUs§ Reduce overhead of multiple function calls

§ Similar to Spark Operator Fusing in Spark 2.0§ Unroll Loops, Fuse Operators, Fold Constants, …§ Scope to session, device, or `with jit_scope():`

VISUALIZING JIT COMPILER IN ACTION

Before After

Google Web Tracing Framework:http://google.github.io/tracing-framework/

from tensorflow.python.client import timelinetrace = timeline.Timeline(step_stats=run_metadata.step_stats)with open('timeline.json', 'w') as trace_file:trace_file.write(

trace.generate_chrome_trace_format(show_memory=True))

VISUALIZING FUSING OPERATORS

pip install graphviz

dot -Tpng \/tmp/hlo_graph_1.w5LcGs.dot \-o hlo_graph_1.png

GraphViz:http://www.graphviz.org

hlo_*.dot files generated by XLA

LET’S TRAIN WITH XLA CPU§ Navigate to the following notebook:

06_Train_Model_XLA_CPU

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

LET’S TRAIN WITH XLA GPU§ Navigate to the following notebook:

06a_Train_Model_XLA_GPU

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

AGENDA

Part 0: PipelineAI Research

Part 1: TensorFlow Model Training

Part 2: TensorFlow Model Serving

AGENDA

Part 2: TensorFlow Model Serving

§ AOT XLA Compiler and Graph Transform Tool§ Key Components of TensorFlow Serving§ Deploy Optimized TensorFlow Model§ Optimize TensorFlow Serving Runtime

AOT COMPILER§ Standalone, Ahead-Of-Time (AOT) Compiler§ Built on XLA framework§ tfcompile§ Creates executable with minimal TensorFlow Runtime needed

§ Includes only dependencies needed by subgraph computation§ Creates functions with feeds (inputs) and fetches (outputs)

§ Packaged as cc_libary header and object files to link into your app§ Commonly used for mobile device inference graph

§ Currently, only CPU x86-64 and ARM are supported - no GPU

GRAPH TRANSFORM TOOL (GTT)

§ Post-Training Optimization to Prepare for Inference§ Remove Training-only Ops (checkpoint, drop out, logs)§ Remove Unreachable Nodes between Given feed -> fetch§ Fuse Adjacent Operators to Improve Memory Bandwidth§ Fold Final Batch Norm mean and variance into Variables§ Round Weights/Variables to improve compression (ie. 70%)§ Quantize (FP32 -> INT8) to Speed Up Math Operations

BEFORE OPTIMIZATIONS

GRAPH TRANSFORM TOOLtransform_graph \--in_graph=tensorflow_inception_graph.pb \ ß Original Graph--out_graph=optimized_inception_graph.pb \ ß Transformed Graph--inputs='Mul' \ ß Feed (Input)--outputs='softmax' \ ß Fetch (Output) --transforms=' ß List of Transforms strip_unused_nodesremove_nodes(op=Identity, op=CheckNumerics) fold_constants(ignore_errors=true) fold_batch_normsfold_old_batch_normsquantize_weightsquantize_nodes'

AFTER STRIPPING UNUSED NODES

§ Optimizations§ strip_unused_nodes

§ Results§ Graph much simpler§ File size much smaller

AFTER REMOVING UNUSED NODES

§ Optimizations§ strip_unused_nodes§ remove_nodes

§ Results§ Pesky nodes removed§ File size a bit smaller

AFTER FOLDING CONSTANTS

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants

§ Results§ Placeholders (feeds) -> Variables*

(*Why Variables and not Constants?)

AFTER FOLDING BATCH NORMS

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms

§ Results§ Graph remains the same§ File size approximately the same

AFTER QUANTIZING WEIGHTS

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms§ quantize_weights

§ Results§ Graph is same, file size is smaller, compute is faster

WEIGHT QUANTIZATION

§ FP16 and INT8 Are Smaller and Computationally Simpler§ Weights/Variables are Constants§ Easy to Linearly Quantize

LET’S OPTIMIZE FOR INFERENCE§ Navigate to the following notebook:

07_Optimize_Model**Why just CPU version? Why not GPU?

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

BUT WAIT, THERE’S MORE!

ACTIVATION QUANTIZATION§ Activations Not Known Ahead of Time

§ Depends on input, not easy to quantize§ Requires Additional Calibration Step

§ Use a “representative” dataset§ Per Neural Network Layer…

§ Collect histogram of activation values§ Generate many quantized distributions with different saturation thresholds§ Choose threshold to minimize…

KL_divergence(ref_distribution, quant_distribution)

§ Not Much Time or Data is Required (Minutes on Commodity Hardware)

AFTER ACTIVATION QUANTIZATION

§ Optimizations§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms§ quantize_weights§ quantize_nodes (activations)

§ Results§ Larger graph, needs calibration!

Requires additional freeze_requantization_ranges

LET’S OPTIMIZE FOR INFERENCE§ Navigate to the following notebook:

08_Optimize_Model_Activations

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

FREEZING MODEL FOR DEPLOYMENT§ Optimizations

§ strip_unused_nodes§ remove_nodes§ fold_constants§ fold_batch_norms§ quantize_weights§ quantize_nodes§ freeze_graph

§ Results§ Variables -> Constants

Finally!We’re Ready to Deploy!!

AGENDA

Part 2: TensorFlow Model Serving

§ AOT XLA Compiler and Graph Transform Tool§ Key Components of TensorFlow Serving§ Deploy Optimized TensorFlow Model§ Optimize TensorFlow Serving Runtime

MODEL SERVING TERMINOLOGY§ Inference

§ Only Forward Propagation through Network§ Predict, Classify, Regress, …

§ Bundle§ GraphDef, Variables, Metadata, …

§ Assets§ ie. Map of ClassificationID -> String§ {9283: “penguin”, 9284: “bridge”}

§ Version§ Every Model Has a Version Number (Integer)

§ Version Policy§ ie. Serve Only Latest (Highest), Serve Both Latest and Previous, …

TENSORFLOW SERVING FEATURES§ Supports Auto-Scaling§ Custom Loaders beyond File-based§ Tune for Low-latency or High-throughput§ Serve Diff Models/Versions in Same Process§ Customize Models Types beyond HashMap and TensorFlow§ Customize Version Policies for A/B and Bandit Tests§ Support Request Draining for Graceful Model Updates§ Enable Request Batching for Diff Use Cases and HW§ Supports Optimized Transport with GRPC and Protocol Buffers

PREDICTION SERVICE§ Predict (Original, Generic)

§ Input: List of Tensor§ Output: List of Tensor

§ Classify§ Input: List of tf.Example (key, value) pairs§ Output: List of (class_label: String, score: float)

§ Regress§ Input: List of tf.Example (key, value) pairs§ Output: List of (label: String, score: float)

PREDICTION INPUTS + OUTPUTS§ SignatureDef

§ Defines inputs and outputs§ Maps external (logical) to internal (physical) tensor names§ Allows internal (physical) tensor names to change

from tensorflow.python.saved_model import utilsfrom tensorflow.python.saved_model import signature_constantsfrom tensorflow.python.saved_model import signature_def_utils

graph = tf.get_default_graph()x_observed = graph.get_tensor_by_name('x_observed:0') y_pred = graph.get_tensor_by_name('add:0') inputs_map = {'inputs': x_observed} outputs_map = {'outputs': y_pred} predict_signature = signature_def_utils.predict_signature_def(inputs=inputs_map,

outputs=outputs_map)

MULTI-HEADED INFERENCE

§ Multiple “Heads” of Model§ Return class and scores to be fed into another model§ Inputs Propagated Forward Only Once§ Optimizes Bandwidth, CPU, Latency, Memory, Coolness

BUILD YOUR OWN MODEL SERVER§ Adapt GRPC(Google) <-> HTTP (REST of the World)§ Perform Batch Inference vs. Request/Response§ Handle Requests Asynchronously§ Support Mobile, Embedded Inference§ Customize Request Batching§ Add Circuit Breakers, Fallbacks§ Control Latency Requirements§ Reduce Number of Moving Parts

#include “tensorflow_serving/model_servers/server_core.h”

class MyTensorFlowModelServer {ServerCore::Options options;

// set options (model name, path, etc)std::unique_ptr<ServerCore> core;

TF_CHECK_OK(ServerCore::Create(std::move(options), &core)

);}

Compile and Link withlibtensorflow.so

NVIDIA TENSOR-RT RUNTIME

§ Post-Training Model Optimizations§ Similar to TF Graph Transform Tool

§ GPU-Optimized Prediction Runtime§ Alternative to TensorFlow Serving

§ PipelineAI Supports TensorRT!

AGENDA

Part 2: TensorFlow Model Serving

§ AOT XLA Compiler and Graph Transform Tool§ Key Components of TensorFlow Serving§ Deploy Optimized TensorFlow Model§ Optimize TensorFlow Serving Runtime

SAVED MODEL FORMAT§ Navigate to the following notebook:

09_Deploy_Optimized_Model

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

AGENDA

Part 2: TensorFlow Model Serving

§ AOT XLA Compiler and Graph Transform Tool§ Key Components of TensorFlow Serving§ Deploy Optimized TensorFlow Model§ Optimize TensorFlow Serving Runtime

REQUEST BATCH TUNING§ max_batch_size

§ Enables throughput/latency tradeoff§ Bounded by RAM

§ batch_timeout_micros§ Defines batch time window, latency upper-bound§ Bounded by RAM

§ num_batch_threads§ Defines parallelism§ Bounded by CPU cores

§ max_enqueued_batches§ Defines queue upper bound, throttling§ Bounded by RAM

Reaching either thresholdwill trigger a batch

BATCH SCHEDULER STRATEGIES§ BasicBatchScheduler

§ Best for homogeneous request types (ie. always classify or always regress)§ Async callback upon max_batch_size or batch_timeout_micros§ BatchTask encapsulates unit of work to be batched

§ SharedBatchScheduler§ Best for heterogeneous request types, multi-step inference, ensembles, …§ Groups BatchTasks into separate queues to form homogenous batches§ Processes batches fairly through interleaving

§ StreamingBatchScheduler§ Mixed CPU/GPU/IO-bound workloads§ Provides fine-grained control for complex, multi-phase inference logic

Experiment to Find the Best Batching Strategy for You!!

Co-locate Similar Prediction Workloads

LET’S DEPLOY OPTIMIZED MODEL§ Navigate to the following notebook:

10_Optimize_Model_Server

§ https://github.com/PipelineAI/pipeline/tree/master/gpu.ml/notebooks

AGENDA

Part 0: PipelineAI Research

Part 1: TensorFlow Model Training

Part 2: TensorFlow Model Serving

THANK YOU!! QUESTIONS?§ https://github.com/PipelineAI/pipeline/

§ Please ! this GitHub Repo!

§ All slides, code, notebooks, and Docker images here:https://github.com/PipelineAI/pipeline/tree/master/gpu.ml

Contact Mechris@pipeline.ai

@cfregly