LogoLogo
Continuum WebsiteContinuum ApplicationsContinuum KnowledgeAxolotl Platform
Continuum Knowledge
Continuum Knowledge
  • Continuum
  • Data
    • Datasets
      • Pre Training Data
      • Types of Fine Tuning
      • Self Instruct Paper
      • Self-Alignment with Instruction Backtranslation
      • Systematic Evaluation of Instruction-Tuned Large Language Models on Open Datasets
      • Instruction Tuning
      • Instruction Fine Tuning - Alpagasus
      • Less is More For Alignment
      • Enhanced Supervised Fine Tuning
      • Visualising Data using t-SNE
      • UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction
      • Training and Evaluation Datasets
      • What is perplexity?
  • MODELS
    • Foundation Models
      • The leaderboard
      • Foundation Models
      • LLama 2 - Analysis
      • Analysis of Llama 3
      • Llama 3.1 series
      • Google Gemini 1.5
      • Platypus: Quick, Cheap, and Powerful Refinement of LLMs
      • Mixtral of Experts
      • Mixture-of-Agents (MoA)
      • Phi 1.5
        • Refining the Art of AI Training: A Deep Dive into Phi 1.5's Innovative Approach
      • Phi 2.0
      • Phi-3 Technical Report
  • Training
    • The Fine Tuning Process
      • Why fine tune?
        • Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?
        • Explanations in Fine Tuning
      • Tokenization
        • Tokenization Is More Than Compression
        • Tokenization - SentencePiece
        • Tokenization explore
        • Tokenizer Choice For LLM Training: Negligible or Crucial?
        • Getting the most out of your tokenizer for pre-training and domain adaptation
        • TokenMonster
      • Parameter Efficient Fine Tuning
        • P-Tuning
          • The Power of Scale for Parameter-Efficient Prompt Tuning
        • Prefix-Tuning: Optimizing Continuous Prompts for Generation
        • Harnessing the Power of PEFT: A Smarter Approach to Fine-tuning Pre-trained Models
        • What is Low-Rank Adaptation (LoRA) - explained by the inventor
        • Low Rank Adaptation (Lora)
        • Practical Tips for Fine-tuning LMs Using LoRA (Low-Rank Adaptation)
        • QLORA: Efficient Finetuning of Quantized LLMs
        • Bits and Bytes
        • The Magic behind Qlora
        • Practical Guide to LoRA: Tips and Tricks for Effective Model Adaptation
        • The quantization constant
        • QLORA: Efficient Finetuning of Quantized Language Models
        • QLORA and Fine-Tuning of Quantized Language Models (LMs)
        • ReLoRA: High-Rank Training Through Low-Rank Updates
        • SLoRA: Federated Parameter Efficient Fine-Tuning of Language Models
        • GaLora: Memory-Efficient LLM Training by Gradient Low-Rank Projection
      • Hyperparameters
        • Batch Size
        • Padding Tokens
        • Mixed precision training
        • FP8 Formats for Deep Learning
        • Floating Point Numbers
        • Batch Size and Model loss
        • Batch Normalisation
        • Rethinking Learning Rate Tuning in the Era of Language Models
        • Sample Packing
        • Gradient accumulation
        • A process for choosing the learning rate
        • Learning Rate Scheduler
        • Checkpoints
        • A Survey on Efficient Training of Transformers
        • Sequence Length Warmup
        • Understanding Training vs. Evaluation Data Splits
        • Cross-entropy loss
        • Weight Decay
        • Optimiser
        • Caching
      • Training Processes
        • Extending the context window
        • PyTorch Fully Sharded Data Parallel (FSDP)
        • Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
        • YaRN: Efficient Context Window Extension of Large Language Models
        • Sliding Window Attention
        • LongRoPE
        • Reinforcement Learning
        • An introduction to reinforcement learning
        • Reinforcement Learning from Human Feedback (RLHF)
        • Direct Preference Optimization: Your Language Model is Secretly a Reward Model
  • INFERENCE
    • Why is inference important?
      • Grouped Query Attention
      • Key Value Cache
      • Flash Attention
      • Flash Attention 2
      • StreamingLLM
      • Paged Attention and vLLM
      • TensorRT-LLM
      • Torchscript
      • NVIDIA L40S GPU
      • Triton Inference Server - Introduction
      • Triton Inference Server
      • FiDO: Fusion-in-Decoder optimised for stronger performance and faster inference
      • Is PUE a useful measure of data centre performance?
      • SLORA
  • KNOWLEDGE
    • Vector Databases
      • A Comprehensive Survey on Vector Databases
      • Vector database management systems: Fundamental concepts, use-cases, and current challenges
      • Using the Output Embedding to Improve Language Models
      • Decoding Sentence-BERT
      • ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT
      • SimCSE: Simple Contrastive Learning of Sentence Embeddings
      • Questions Are All You Need to Train a Dense Passage Retriever
      • Improving Text Embeddings with Large Language Models
      • Massive Text Embedding Benchmark
      • RocketQAv2: A Joint Training Method for Dense Passage Retrieval and Passage Re-ranking
      • LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders
      • Embedding and Fine-Tuning in Neural Language Models
      • Embedding Model Construction
      • Demystifying Embedding Spaces using Large Language Models
      • Fine-Tuning Llama for Multi-Stage Text Retrieval
      • Large Language Model Based Text Augmentation Enhanced Personality Detection Model
      • One Embedder, Any Task: Instruction-Finetuned Text Embeddings
      • Vector Databases are not the only solution
      • Knowledge Graphs
        • Harnessing Knowledge Graphs to Elevate AI: A Technical Exploration
        • Unifying Large Language Models and Knowledge Graphs: A Roadmap
      • Approximate Nearest Neighbor (ANN)
      • High Dimensional Data
      • Principal Component Analysis (PCA)
      • Vector Similarity Search - HNSW
      • FAISS (Facebook AI Similarity Search)
      • Unsupervised Dense Retrievers
    • Retrieval Augmented Generation
      • Retrieval-Augmented Generation for Large Language Models: A Survey
      • Fine-Tuning or Retrieval?
      • Revolutionising Information Retrieval: The Power of RAG in Language Models
      • A Survey on Retrieval-Augmented Text Generation
      • REALM: Retrieval-Augmented Language Model Pre-Training
      • Retrieve Anything To Augment Large Language Models
      • Generate Rather Than Retrieve: Large Language Models Are Strong Context Generators
      • Active Retrieval Augmented Generation
      • DSPy: LM Assertions: Enhancing Language Model Pipelines with Computational Constraints
      • DSPy: Compiling Declarative Language Model Calls
      • DSPy: In-Context Learning for Extreme Multi-Label Classification
      • Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs
      • HYDE: Revolutionising Search with Hypothetical Document Embeddings
      • Enhancing Recommender Systems with Large Language Model Reasoning Graphs
      • Retrieval Augmented Generation (RAG) versus fine tuning
      • RAFT: Adapting Language Model to Domain Specific RAG
      • Summarisation Methods and RAG
      • Lessons Learned on LLM RAG Solutions
      • Stanford: Retrieval Augmented Language Models
      • Overview of RAG Approaches with Vector Databases
      • Mastering Chunking in Retrieval-Augmented Generation (RAG) Systems
    • Semantic Routing
    • Resource Description Framework (RDF)
  • AGENTS
    • What is agency?
      • Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves
      • Types of Agents
      • The risk of AI agency
      • Understanding Personality in Large Language Models: A New Frontier in AI Psychology
      • AI Agents - Reasoning, Planning, and Tool Calling
      • Personality and Brand
      • Agent Interaction via APIs
      • Bridging Minds and Machines: The Legacy of Newell, Shaw, and Simon
      • A Survey on Language Model based Autonomous Agents
      • Large Language Models as Agents
      • AI Reasoning: A Deep Dive into Chain-of-Thought Prompting
      • Enhancing AI Reasoning with Self-Taught Reasoner (STaR)
      • Exploring the Frontier of AI: The "Tree of Thoughts" Framework
      • Toolformer: Revolutionising Language Models with API Integration - An Analysis
      • TaskMatrix.AI: Bridging Foundational AI Models with Specialised Systems for Enhanced Task Completion
      • Unleashing the Power of LLMs in API Integration: The Rise of Gorilla
      • Andrew Ng's presentation on AI agents
      • Making AI accessible with Andrej Karpathy and Stephanie Zhan
  • Regulation and Ethics
    • Regulation and Ethics
      • Privacy
      • Detecting AI Generated content
      • Navigating the IP Maze in AI: The Convergence of Blockchain, Web 3.0, and LLMs
      • Adverse Reactions to generative AI
      • Navigating the Ethical Minefield: The Challenge of Security in Large Language Models
      • Navigating the Uncharted Waters: The Risks of Autonomous AI in Military Decision-Making
  • DISRUPTION
    • Data Architecture
      • What is a data pipeline?
      • What is Reverse ETL?
      • Unstructured Data and Generatve AI
      • Resource Description Framework (RDF)
      • Integrating generative AI with the Semantic Web
    • Search
      • BM25 - Search Engine Ranking Function
      • BERT as a reranking engine
      • BERT and Google
      • Generative Engine Optimisation (GEO)
      • Billion-scale similarity search with GPUs
      • FOLLOWIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions
      • Neural Collaborative Filtering
      • Federated Neural Collaborative Filtering
      • Latent Space versus Embedding Space
      • Improving Text Embeddings with Large Language Models
    • Recommendation Engines
      • On Interpretation and Measurement of Soft Attributes for Recommendation
      • A Survey on Large Language Models for Recommendation
      • Model driven recommendation systems
      • Recommender AI Agent: Integrating Large Language Models for Interactive Recommendations
      • Foundation Models for Recommender Systems
      • Exploring the Impact of Large Language Models on Recommender Systems: An Extensive Review
      • AI driven recommendations - harming autonomy?
    • Logging
      • A Taxonomy of Anomalies in Log Data
      • Deeplog
      • LogBERT: Log Anomaly Detection via BERT
      • Experience Report: Deep Learning-based System Log Analysis for Anomaly Detection
      • Log-based Anomaly Detection with Deep Learning: How Far Are We?
      • Deep Learning for Anomaly Detection in Log Data: A Survey
      • LogGPT
      • Adaptive Semantic Gate Networks (ASGNet) for log-based anomaly diagnosis
  • Infrastructure
    • The modern data centre
      • Enhancing Data Centre Efficiency: Strategies to Improve PUE
      • TCO of NVIDIA GPUs and falling barriers to entry
      • Maximising GPU Utilisation with Kubernetes and NVIDIA GPU Operator
      • Data Centres
      • Liquid Cooling
    • Servers and Chips
      • The NVIDIA H100 GPU
      • NVIDIA H100 NVL
      • Lambda Hyperplane 8-H100
      • NVIDIA DGX Servers
      • NVIDIA DGX-2
      • NVIDIA DGX H-100 System
      • NVLink Switch
      • Tensor Cores
      • NVIDIA Grace Hopper Superchip
      • NVIDIA Grace CPU Superchip
      • NVIDIA GB200 NVL72
      • Hopper versus Blackwell
      • HGX: High-Performance GPU Platforms
      • ARM Chips
      • ARM versus x86
      • RISC versus CISC
      • Introduction to RISC-V
    • Networking and Connectivity
      • Infiniband versus Ethernet
      • NVIDIA Quantum InfiniBand
      • PCIe (Peripheral Component Interconnect Express)
      • NVIDIA ConnectX InfiniBand adapters
      • NVMe (Non-Volatile Memory Express)
      • NVMe over Fabrics (NVMe-oF)
      • NVIDIA Spectrum-X
      • NVIDIA GPUDirect
      • Evaluating Modern GPU Interconnect
      • Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)
      • Next-generation networking in AI environments
      • NVIDIA Collective Communications Library (NCCL)
    • Data and Memory
      • NVIDIA BlueField Data Processing Units (DPUs)
      • Remote Direct Memory Access (RDMA)
      • High Bandwidth Memory (HBM3)
      • Flash Memory
      • Model Requirements
      • Calculating GPU memory for serving LLMs
      • Transformer training costs
      • GPU Performance Optimisation
    • Libraries and Complements
      • NVIDIA Base Command
      • NVIDIA AI Enterprise
      • CUDA - NVIDIA GTC 2024 presentation
      • RAPIDs
      • RAFT
    • Vast Data Platform
      • Vast Datastore
      • Vast Database
      • Vast Data Engine
      • DASE (Disaggregated and Shared Everything)
      • Dremio and VAST Data
    • Storage
      • WEKA: A High-Performance Storage Solution for AI Workloads
      • Introduction to NVIDIA GPUDirect Storage (GDS)
        • GDS cuFile API
      • NVIDIA Magnum IO GPUDirect Storage (GDS)
      • Vectors in Memory
Powered by GitBook
LogoLogo

Continuum - Accelerated Artificial Intelligence

  • Continuum Website
  • Axolotl Platform

Copyright Continuum Labs - 2023

On this page
  • Key components and techniques of FSDP
  • Model Replication
  • Model Partitioning
  • Model Sharding
  • A metaphor
  • Sharding

Was this helpful?

  1. Training
  2. The Fine Tuning Process
  3. Training Processes

PyTorch Fully Sharded Data Parallel (FSDP)

PreviousExtending the context windowNextTrain Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation

Last updated 1 year ago

Was this helpful?

This September 2023 paper introduces PyTorch Fully Sharded Data Parallel (FSDP), an industry-grade solution for large model training that enables sharding model parameters across multiple devices.

FSDP is designed to provide a user-friendly experience, handle hardware heterogeneity, optimise resource utilisation, and efficiently manage memory.

Key components and techniques of FSDP

Parameter Sharding

  • FSDP divides a model into smaller units and shards the parameters within each unit.

  • Sharded parameters are communicated and recovered on-demand before computations and discarded afterwards.

  • This approach minimizes peak memory consumption by only materializing parameters from one unit at a time.

User Experience

  • FSDP introduces deferred initialization, allowing users to create a model instance on a dummy device and record initialization operations.

  • The model is then initialized and sharded unit by unit by replaying the recorded operations on a real GPU device.

  • This technique provides a user experience similar to local training while effectively scaling large models.

Hardware Heterogeneity

  • FSDP offers configurable sharding strategies that can be customized to match the physical interconnect topology of the cluster.

  • This allows FSDP to handle hardware heterogeneity, such as high-bandwidth islands within machines and low-bandwidth mesh across machines.

Resource Utilisation

  • FSDP minimizes downtime caused by non-computational operations to ensure GPU devices remain fully utilised during distributed training.

  • It employs techniques like operation reordering and parameter prefetching to overlap communication with computation and eliminate execution bubbles.

Memory Planning

  • FSDP optimizes memory usage by prudently restricting the amount of blocks allocated for in-flight unsharded parameters.

  • It can suspend CPU execution if necessary to prevent triggering frequent memory defragmentations, which can slow down training.

Evaluation and Results

  • FSDP was evaluated on various models, including language models and recommendation system models, using up to 512 80GB A100 GPUs.

  • Experiments showed that FSDP achieves similar performance to Distributed Data Parallel (DDP) on small models.

  • FSDP enables training significantly larger models with near-linear scalability in terms of TFLOPS.

FSDP is a beta feature as of PyTorch 2.0 release and has been battle-tested in both industrial and research applications.

It tackles challenges in large model training by providing a user-friendly experience, handling hardware heterogeneity, optimizing resource utilization, and efficiently managing memory.

Model Replication

  • DistributedDataParallel (DDP) is PyTorch's first end-to-end distributed training feature that follows the model replication approach.

  • DDP maintains a model replica on each device and synchronises gradients through AllReduce operations during the backward pass.

  • It overlaps gradient communication with backward computation to accelerate training.

  • However, DDP requires all model parameters, gradients, and optimiser states to fit in the memory of a single GPU device, limiting its ability to support large models.

Model Partitioning

  • As model sizes grow, partitioning the model into smaller components and distributing them across multiple devices becomes a viable solution.

  • Pipeline parallelism breaks a sequence of layers into stages and feeds inputs to different stages in a pipelined manner to optimize resource utilisation.

  • Tensor RPC provides a lower-level toolkit for executing arbitrary computations on remote devices.

  • While these techniques can scale large models across multiple devices, they either limit the model to a sequence of stages or require modifications to the model authoring code, which can be a significant obstacle to adoption.

Model Sharding

  • Sharding model parameters can help reduce memory footprint and support models with sizes beyond the memory capacity of a single GPU device.

  • After sharding, each rank holds only a shard of the model parameters, preventing it from performing the same computations as local training.

  • To ensure correctness, the training process can employ one or both of the following techniques:

a. Perform computations with parameter shards and communicate activations accordingly, avoiding the need to fully materialize any parameter. However, communication appears in the critical path and cannot easily overlap with computations.

b. Perform the same computation as local training by communicating parameters on-demand before computations. This allows parameter communications to overlap with preceding computations, but requires the on-demand communicated parameters to be fully materialised and fit in the memory of a single GPU device.

FSDP falls into the second category of communicating parameters on-demand. Based on observations and experiments, this approach is sufficient to support the majority of large model applications today and in the near future. If the requirement of fully materialising each parameter unit on GPU becomes a blocker, the two techniques can be combined to support such use cases.

A metaphor

Imagine FSDP as a master chef tasked with preparing a grand feast for a large gathering.

The feast represents the training of a large model, and the dishes symbolise the model's parameters. The chef's kitchen is a single GPU device, which has limited space to work with.

To tackle this challenge, the master chef (FSDP) employs a clever strategy. Instead of trying to prepare all the dishes (parameters) at once, which would overwhelm the kitchen (GPU), the chef divides the menu (model) into smaller, manageable units. Each unit consists of a subset of dishes (layers) that can be prepared independently.

During the preparation process (forward and backward computation), the chef focuses on one unit at a time.

The chef gathers all the necessary ingredients (unsharded parameters) for the dishes in the current unit from the kitchen's storage and other sous-chefs (peer ranks). With everything in place, the chef prepares the dishes (performs local computations) for that unit. Once the unit is complete, the chef cleans up the kitchen by putting away the extra ingredients (freeing peer shards) to make space for the next unit.

Throughout the feast preparation, the chef keeps the pantry (optimizer states) organized by storing ingredients (parameters) in smaller, individual portions (shards). This allows the chef to efficiently manage the limited kitchen space.

The kitchen's capacity (memory requirements) depends on the size of the portioned ingredients (sharded model) and the space needed for the largest unit's preparation (fully-materialized FSDP unit).

In this metaphorical representation, FSDP is portrayed as a master chef who efficiently manages the preparation of a grand feast (training a large model) by dividing the menu (model) into smaller units, focusing on one unit at a time, and optimizing kitchen space usage (memory) and cooking processes (computations and communications) to create a successful event.

Sharding

The sharding strategy in FSDP is a crucial element that determines memory footprint and communication overhead.

FSDP supports various sharding strategies, ranging from fully replicated to fully sharded, which are generalised using the sharding factor (𝐹).

The sharding factor represents the number of ranks over which parameters are sharded.

Full Sharding (𝐹 = 𝑊)

  • Lowest memory footprint but highest communication overhead.

  • FSDP organizes parameters into a FlatParameter, which coalesces communications and evenly shards them across ranks.

  • The flatten-concat-chunk algorithm minimizes padding and enables efficient AllGather and ReduceScatter operations.

  • Fine-grained FlatParameter construction reduces peak memory but may decrease throughput due to more collectives.

Hybrid Sharding (1 < 𝐹 < 𝑊)

  • Combines sharding and replication, providing a trade-off between memory saving and throughput degradation.

  • Exploits datacentre locality by mapping the device mesh into the datacentre layout.

  • Reduces cross-host traffic and improves performance of AllReduce collectives.

  • Suitable for medium-sized models that are too large for full replication but not large enough to fully utilize memory with full sharding.

Autograd Integration

  • FlatParameter interoperates with PyTorch's autograd engine for correct gradient propagation and timely gradient reduction.

  • Original parameters are set as views into the unsharded FlatParameter before forward computation.

  • A gradient hook is registered to perform post-backward logic, including gradient reduction.

FSDP incorporates various communication optimization techniques

  1. Overlapping communication and computation using separate CUDA streams.

  2. Backward prefetching to issue the next AllGather before the current ReduceScatter.

  3. Forward prefetching to issue the next AllGather before forward computation of the current FSDP unit.

  4. Gradient accumulation with and without communication, trading off memory usage and communication.

FSDP also addresses memory management challenges related to PyTorch's CUDA caching allocator.

The caching allocator's behavior can lead to over-allocation and blocking cudaFrees and cudaMallocs when the CPU thread runs ahead of GPU execution.

FSDP introduces a rate limiter that intentionally blocks the CPU thread to ensure proper caching allocator block reuse, allowing at most two in-flight AllGathers to achieve communication and computation overlap while minimising memory fragmentation.

LogoPyTorch FSDP: Experiences on Scaling Fully Sharded Data ParallelarXiv.org
PyTorch Fully Sharded Data Parallel (FSDP)
FSDP Algorithm Overview
FSDP is a type of data-parallel training, but unlike traditional data-parallel, which maintains a per-GPU copy of a model’s parameters, gradients and optimizer states, it shards all of these states across data-parallel workers and can optionally offload the sharded model parameters to CPUs.
Page cover image