Software Documentation (Version 1.7.0)

Software Release Notes
Documentation Updates

Cerebras Basics

How Cerebras Works
Types of CS Systems Installations
Cerebras Execution Modes
ML Workflow on Cerebras
Glossary

Getting Started with Cerebras Wafer-Scale Cluster

Software Requirements and Dependencies
Pytorch: Getting Started
- Running Large Models (Weight Streaming Execution)
- Running Small to Medium Models (Pipelined Execution)
TensorFlow: Getting Started
- Running Large Models (Weight Streaming Execution)
- Running Small to Medium Models (Pipelined Execution)

Getting Started with Original Cerebras Installation

Software Requirements and Dependencies
PyTorch: Getting Started
- Running Small to Medium Models (Pipelined Execution)
TensorFlow: Getting Started
- Running Small to Medium Models (Pipelined Execution)

Scripts and Templates (Original Cerebras Installation Only)

The run.py Template
The csrun_cpu Script
The csrun_wse Script

Reference Samples

Cerebras Model Zoo
GPU Requirements for Running Cerebras ModelZoo

Cerebras Advanced

DataLoader Overview
- PyTorch DataLoader
- Basics of TFRecords
Data Formats
Dynamic Loss Scaling
Performance Optimization Practices
Multi-Replica Data Parallel Training
Sparsity
Checkpoint Formats

Develop with TensorFlow (Pipelined Mode Only)

Workflow for TensorFlow on CS
Port TensorFlow to Cerebras
Prepare Input
TensorFlow Dynamic Loss Scaling
Compile on CPU
Train, Eval, and Predict
Early Stopping
Multiple Models
- Multi-model Inference
TensorFlow Variable Sequence Length
Using TensorBoard
Best Practices
Supported TensorFlow Layers
- tf package

Develop with PyTorch (Pipelined Mode Only)

Workflow for PyTorch on CS
Porting PyTorch Model to CS
Taking checkpoints off a Cerebras System
PyTorch Runners
PyTorch Variable Tensor Shape
Limitations of PyTorch on Cerebras
Cerebras PyTorch Layer API

Compiler Reports (Pipelined Mode Only)

Input Function Report
Compile Report
Incremental Compile

Extensions (Original Cerebras Installation Only)

Adding Custom Packages To cbcore Container

Theme by the Executable Book Project

Prepare Input

Prepare Input¶

Multi-Worker Input Pipeline
Sharding For the CS system
The CS_AUTOTUNE
- Using CS_AUTOTUNE
Optimize Input Function
- Automatic execution of the analyzer

previous

Limitations of the CerebrasEstimator

next

Multi-Worker Input Pipeline

© Copyright 2022, Cerebras Systems.
Last updated on Apr 11, 2023, 10:23:47 PM.