Parallel Training Methods

Sam Foreman

Parallel Training Methods

Author

Affiliation

Sam Foreman

ALCF

Published

November 5, 2024

👀 Overview

📊 Slides @ samforeman.me/talks/ai-for-science-2024/slides
- 📄 HTML version: samforeman.me/talks/ai-for-science-2024
Course:
- 📂 ALCF: Intro to AI-driven Science on Supercomputers
  - argonne-lcf/ai-science-training-series

📑 Outline

Scaling: Overview
Data Parallel Training
1. Communication
2. Why Distributed Training?
Beyond Data Parallelism
1. Additional Parallelism Strategies
Large Language Models
Hands On

🚀 Scaling: Overview

✅ Goal:
- Minimize: Cost (i.e. amount of time spent training)
- Maximize: Performance
See 🤗 Performance and Scalability for more details

Single GPU

See 🤗 Methods and tools for efficient training on a single GPU

flowchart LR
    subgraph G0["`GPU0`"]
        subgraph N0["`Network`"]
        end
        L0("`Loss`")
    end
    subgraph D["`Data`"]
        x("`x₀`")
        x1("`x₁`")
        x2("`x₂`")
    end
    x --> N0
    N0 --> L0
    L0 --> N0
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef grey fill:#cccccc,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class x,L0 red
class x1, green
class x2, blue
class x3, grey
class N0,D,G0,n0 block

Figure 1: SLOW !! model size limited by GPU memory

Data Parallel Training

flowchart LR
    subgraph D["`Data`"]
        direction TB
        x("`x₀`")
        x1("`x₁`")
        x2("`x₂`")
    end
    direction LR
    subgraph G0["`GPU0`"]
        direction LR
        subgraph N0["`NN`"]
        end
        %%y0("`y₀`")
        L0["`Loss`"]
    end
    subgraph G1["`GPU1`"]
        direction LR
        subgraph N1["`NN`"]
        end
        L1["`Loss`"]
    end
    subgraph G2["`GPU2`"]
        direction LR
        subgraph N2["`NN`"]
        end
        L2["`Loss`"]
    end
    x --> G0
    x1 --> G1
    x2 --> G2
    N0 --> L0
    N1 --> L1
    N2 --> L2
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
classDef grey fill:#cccccc,stroke:#333,stroke-width:1px,color:#000
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
class x,y0,L0 red
class x1,L1 green
class x2,L2 blue
class x3,ar grey
class D,N0,N1,N2,G0,G1,G2,GU block
class AR block
class bc text

Figure 2: Each GPU receives unique data at each step

Data Parallel Training: Forward Pass

flowchart LR
    subgraph D["`Data`"]
        direction TB
        %%xp("`xₙ₊₁`")
        x("`x₀`")
        x1("`x₁`")
        x2("`x₂`")
    end
    direction LR
    subgraph G0["`GPU0`"]
        direction LR
        subgraph N0["`NN`"]
        end
        %%y0("`y₀`")
        L0["`Loss`"]
    end
    subgraph G1["`GPU1`"]
        direction LR
        subgraph N1["`NN`"]
        end
        L1["`Loss`"]
    end
    subgraph G2["`GPU2`"]
        direction LR
        subgraph N2["`NN`"]
        end
        L2["`Loss`"]
    end
    subgraph AR["`Average Grads`"]
        direction TB
        ar("`(1/n) ∑ gₙ`")
    end
    x --> G0
    x1 --> G1
    x2 --> G2
    N0 --> L0
    N1 --> L1
    N2 --> L2
    G0 -.-> AR
    G1 -.-> AR
    G2 -.-> AR
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef grey fill:#cccccc,stroke:#333,stroke-width:1px,color:#000
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class x,y0,L0 red
class x1,L1 green
class x2,L2 blue
class x3,ar grey
class D,N0,N1,N2,G0,G1,G2,GU block
class AR block
class bc text

Figure 3: Average gradients across all GPUs

Data Parallel Training: Backward Pass

flowchart RL
    subgraph D["`Data`"]
        direction TB
        x("`x₀`")
        x1("`x₁`")
        x2("`x₂`")
    end
    subgraph G0["`GPU0`"]
        direction RL
        subgraph N0["`NN`"]
        end
        L0["`Loss`"]
    end
    subgraph G1["`GPU1`"]
        direction RL
        subgraph N1["`NN`"]
        end
        L1["`Loss`"]
    end
    subgraph G2["`GPU2`"]
        direction RL
        subgraph N2["`NN`"]
        end
        L2["`Loss`"]
    end
    subgraph BC["`Send Updates`"]
        direction TB
    end
    BC -.-> G0
    BC -.-> G1
    BC -.-> G2
    L0 ~~~ N0
    L1 ~~~ N1
    L2 ~~~ N2
    G0 ~~~ x
    G1 ~~~ x1
    G2 ~~~ x2
classDef grey fill:#cccccc,stroke:#333,stroke-width:1px,color:#000
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class x,y0,L0 red
class x1,L1 green
class x2,L2 blue
class x3,ar grey
class D,N0,N1,N2,G0,G1,G2,GU block
class BC block
class bc text

Figure 4: Send global updates back to each GPU

Data Parallel Training

flowchart LR
    subgraph D["`Data`"]
        direction TB
        x("`x₀`")
        x1("`x₁`")
        x2("`x₂`")
    end
    direction LR
    subgraph G0["`GPU0`"]
        direction LR
        subgraph N0["`NN`"]
        end
        L0["`L0`"]
    end
    subgraph G1["`GPU1`"]
        direction LR
        subgraph N1["`NN`"]
        end
        L1["`L1`"]
    end
    subgraph G2["`GPU2`"]
        direction LR
        subgraph N2["`NN`"]
        end
        L2["`L2`"]
    end
    subgraph AR["`Average Grads`"]
        direction TB
        ar("`(1/n) ∑ gₙ`")
        bc("`Update Weights`")
        ar --> bc
    end
    x --> G0
    x1 --> G1
    x2 --> G2
    N0 --> L0
    N1 --> L1
    N2 --> L2
    G0 <-.-> AR
    G1 <-.-> AR
    G2 <-.-> AR
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef grey fill:#cccccc,stroke:#333,stroke-width:1px,color:#000
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class x,y0,L0 red
class x1,L1 green
class x2,L2 blue
class x3,ar grey
class D,N0,N1,N2,G0,G1,G2,GU block
class AR block
class bc text

Figure 5: See: PyTorch / Distributed Data Parallel

Data Parallel Training

Each GPU:
- has identical copy of model
- works on a unique subset of data
Easy to get started (minor modifications to code):
- saforem2/ezpz
- 🔥 PyTorch / DDP
- 🤗 HF / Accelerate
- Microsoft / DeepSpeed
Requires global communication
- every rank must participate (collective communication) !!

flowchart TD
    subgraph D["`Data`"]
        direction LR
        x("`x₀`")
        x1("`x₁`")
        x2("`x₂`")
    end
    subgraph G0["`GPU0`"]
        direction TB
        subgraph N0["`NN`"]
        end
        L0["`L₀`"]
    end
    subgraph G1["`GPU1`"]
        direction TB
        subgraph N1["`NN`"]
        end
        L1["`L₁`"]
    end
    subgraph G2["`GPU2`"]
        direction TB
        subgraph N2["`NN`"]
        end
        L2["`L₂`"]
    end
    subgraph AR["`Average Grads`"]
        direction TB
        ar("`(1/n) ∑ gₙ`")
        bc("`Update Weights`")
        ar --> bc
    end
    x --> G0
    x1 --> G1
    x2 --> G2
    N0 --> L0
    N1 --> L1
    N2 --> L2
    G0 <-.-> AR
    G1 <-.-> AR
    G2 <-.-> AR
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef grey fill:#cccccc,stroke:#333,stroke-width:1px,color:#000
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class x,y0,L0 red
class x1,L1 green
class x2,L2 blue
class x3,ar grey
class D,N0,N1,N2,G0,G1,G2,GU block
class AR block
class bc text

Figure 6

🗣️ Communication

Need mechanism(s) for communicating across GPUs:
- mpi4py
- torch.distributed
Collective Communication:
- Nvidia Collective Communications Library (NCCL)
- Intel oneAPI Collective Communications Library (oneCCL)
⌛ Timeouts
- Collective operations have to be called for each rank to form a complete collective operation.
  
  Failure to do so will result in other ranks waiting indefinitely

AllReduce

Perform reductions on data (e.g. sum, min, max) across ranks, send result back to everyone.

flowchart TD
  subgraph R0["`Rank 0`"]
    x0("`x0`")
  end
  subgraph R1["`Rank 1`"]
    x1("`x1`")
  end
  subgraph R2["`Rank 2`"]
    x2("`x2`")
  end
  subgraph R3["`Rank 3`"]
    x3("`x3`")
  end
  subgraph AR["`Allreduce`"]
    xp["`x' = ∑ xₙ `"]
  end
  subgraph AR3["`Rank 3`"]
    xp3("`x'`")
  end
  subgraph AR2["`Rank 2`"]
    xp2("`x'`")
  end
  subgraph AR1["`Rank 1`"]
    xp1("`x'`")
  end
  subgraph AR0["`Rank 0`"]
    xp0("`x'`")
  end
  x0 --> AR
  x1 --> AR
  x2 --> AR
  x3 --> AR
  AR --> xp0
  AR --> xp1
  AR --> xp2
  AR --> xp3
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef pink fill:#E599F7,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
class R0,R1,R2,R3,AR,AR0,AR1,AR2,AR3 block
class xp,xp0,xp1,xp2,xp3, purple
class x0, red
class x1, green
class x2, blue
class x3, yellow

Figure 7: All-Reduce operation: each rank receives the reduction of input values across ranks.

Reduce

Perform a reduction on data across ranks, send to individual

flowchart TD
  subgraph R0["`Rank 0`"]
    x0("`x0`")
  end
  subgraph R1["`Rank 1`"]
    x1("`x1`")
  end
  subgraph R2["`Rank 2`"]
    x2("`x2`")
  end
  subgraph R3["`Rank 3`"]
    x3("`x3`")
  end
  subgraph AR["`Reduce`"]
    xp["`x'=reduce(x, 2, SUM)`"]
  end
  subgraph AR3["`Rank 3`"]
  end
  subgraph AR2["`Rank 2`"]
    xp2("`x'`")
  end
  subgraph AR1["`Rank 1`"]
  end
  subgraph AR0["`Rank 0`"]
  end
  x0 --> AR
  x1 --> AR
  x2 --> AR
  x3 --> AR
  AR --> AR3
  AR --> xp2
  AR --> AR1
  AR --> AR0
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef pink fill:#E599F7,stroke:#333,stroke-width:1px,color:#000
class R0,R1,R2,R3,AR,AR0,AR1,AR2,AR3, block
class xp,xp2 purple
class x0, red
class x1, green
class x2, blue
class x3, yellow

Figure 8: Reduce operation: one rank receives the reduction of input values across ranks

Broadcast

flowchart TD
  subgraph R3["`Rank 3`"]
  end
  subgraph R2["`Rank 2`"]
    x2("`x2`")
  end
  subgraph R1["`Rank 1`"]
  end
  subgraph R0["`Rank 0`"]
  end
  subgraph AR["` `"]
    xp["`broadcast(x2, 2)`"]
  end
  subgraph AR3["`Rank 3`"]
    xp3("`x2`")
  end
  subgraph AR2["`Rank 2`"]
    xp2("`x2`")
  end
  subgraph AR1["`Rank 1`"]
    xp1("`x2`")
  end
  subgraph AR0["`Rank 0`"]
    xp0("`x2`")
  end
  x2 --> AR
  AR --> AR3
  AR --> AR2
  AR --> AR1
  AR --> AR0
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383,font-weight:500
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,font-weight:500,color:#838383
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
class R0,R1,R2,R3,AR0,AR1,AR2,AR3,AR, block
class x2,xp0,xp1,xp2,xp3 blue
class xp, text

Figure 9: broadcast (send) a tensor x from one rank to all ranks

AllGather

flowchart LR
  subgraph R0["`Rank 0`"]
    x0("`x0`")
  end
  subgraph R1["`Rank 1`"]
    x1("`x1`")
  end
  subgraph R2["`Rank 2`"]
    x2("`x2`")
  end
  subgraph R3["`Rank 3`"]
    x3("`x3`")
  end
  subgraph AG["`Allgather`"]
    %%xp0["`z=[empty_like(x) for _ in range(4)]`"]
    %%xp1["`dist.all_gather(z, x)`"]
  end
  subgraph AG3["`Rank 3`"]
    direction TB
    xp03("`x0`")
    xp13("`x1`")
    xp23("`x2`")
    xp33("`x3`")
  end
  subgraph AG2["`Rank 2`"]
    direction TB
    xp02("`x0`")
    xp12("`x1`")
    xp22("`x2`")
    xp32("`x3`")
  end
  subgraph AG1["`Rank 1`"]
    direction TB
    xp01("`x0`")
    xp11("`x1`")
    xp21("`x2`")
    xp31("`x3`")
  end
  subgraph AG0["`Rank 0`"]
    direction TB
    xp00("`x0`")
    xp10("`x1`")
    xp20("`x2`")
    xp30("`x3`")
  end
  x0 --> AG
  x1 --> AG
  x2 --> AG
  x3 --> AG
  AG --> AG0
  AG --> AG1
  AG --> AG2
  AG --> AG3
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,font-weight:500,color:#838383
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class xp0,xp1, text
class AG0,AG1,AG2,AG3,AG,R0,R1,R2,R3, block
class xp00,xp01,xp02,xp03, red
class xp10,xp11,xp12,xp13, green
class xp20,xp21,xp22,xp23, blue
class xp30,xp31,xp32,xp33, yellow
class x0, red
class x1, green
class x2, blue
class x3, yellow

Figure 10: Gathers tensors from the whole group in a list.

Scatter

flowchart TD
  subgraph R3["`Rank 3`"]
  end
  subgraph R2["`Rank 2`"]
  end
  subgraph R1["`Rank 1`"]
    direction TB
    xp03("`x0`")
    xp13("`x1`")
    xp23("`x2`")
    xp33("`x3`")
  end
  subgraph R0["`Rank 0`"]
  end
  subgraph S["`Scatter`"]
  end
  subgraph AG3["`Rank 3`"]
    x3("`x3`")
  end
  subgraph AG2["`Rank 2`"]
    x2("`x2`")
  end
  subgraph AG1["`Rank 1`"]
    x1("`x1`")
  end
  subgraph AG0["`Rank 0`"]
    x0("`x0`")
  end
  %%R0 --> S
  R1 --> S
  %%R2 --> S
  %%R3 --> S
  S --> AG0
  S --> AG1
  S --> AG2
  S --> AG3
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,font-weight:500,color:#838383
class AG0,AG1,AG2,AG3,S,R0,R1,R2,R3, block
class xp00,xp01,xp02,xp03, red
class xp10,xp11,xp12,xp13, orange
class xp20,xp21,xp22,xp23, yellow
class xp30,xp31,xp32,xp33, blue
class x0, red
class x1, green
class x2, blue
class x3, yellow

Figure 11: Scatters a list of tensors to the whole group

Why Distributed Training?

N workers each processing unique batch¹ of data:
- [micro_batch_size = 1] \times [N GPUs] \rightarrow [global_batch_size = N]
Improved gradient estimators
- Smooth loss landscape
- Less iterations needed for same number of epochs
  - common to scale learning rate lr *= sqrt(N)
See: Large Batch Training of Convolutional Networks

Why Distributed Training? Speedup!

Table 1: Recent progress

Year	Author	GPU	Batch Size	# GPU	TIME (s)	ACC
2016	He	P100	256	8	104,400	75.30%
2019	Yamazaki	V100	81,920	2048	72	75.08%

Dealing with Data

At each training step, we want to ensure that each worker receives unique data
This can be done in one of two ways:
1. Manually partition data (ahead of time)
  - Assign unique subsets to each worker
  - Each worker can only see their local portion of the data
  - Most common approach
2. From each worker, randomly select a mini-batch
  - Each worker can see the full dataset
  - ⚠️ When randomly selecting, it is important that each worker uses different seeds to ensure they receive unique data

Broadcast Initial State

At the start of training (or when loading from a checkpoint), we want all of our workers to be initialized consistently
- Broadcast the model and optimizer states from rank() == 0 worker

flowchart TD
0["GPU0"] --> 1["GPU 1"]
CKPT --> 0
0 --> 2["GPU 2"]
0 --Model + Optim. State-->3["GPU 3"]
0 --> X["`...`"]
0 --> N["GPU N"]
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class 0,1,2,3,N,X,CKPT block

Figure 12: To ensure all workers have the same copies, we load on RANK==0 and broadcast

Best Practices

⏰ Keeping things in Sync

Computation stalls during communication !!

Keeping the communication to computation ratio small is important for effective scaling.

Use parallel IO whenever possible
- Feed each rank from different files
- Use MPI IO to have each rank read its own batch from a file
- Use several ranks to read data, MPI to scatter to remaining ranks
  - Most practical in big at-scale training

Take advantage of data storage
- Use striping on lustre
Use the right optimizations for Aurora, Polaris, etc.
Preload data when possible
- Offloading to a GPU frees CPU cycles for loading the next batch of data
  - minimize IO latency this way

Going Beyond Data Parallelism

✅ Useful when model fits on single GPU:
- ultimately limited by GPU memory
- model performance limited by size
❌ When model does not fit on a single GPU:
- Offloading:
  - DeepSpeed + ZeRO
  - 🔥 PyTorch + FSDP
- Or, resort to additional parallelism strategies…

Going beyond Data Parallelism: DeepSpeed + `ZeRO`

Depending on the ZeRO stage (1, 2, 3), we can offload:
1. Stage 1: optimizer states \left(P_{\mathrm{os}}\right)
2. Stage 2: gradients + opt. states \left(P_{\mathrm{os}+\mathrm{g}}\right)
3. Stage 3: model params + grads + opt. states \left(P_{\mathrm{os}+\mathrm{g}+\mathrm{p}}\right)

Fully Sharded Data Parallel: 🔥 PyTorch + `FSDP`

Instead of maintaining per-GPU copy of {params, grads, opt_states}, FSDP shards (distributes) these across data-parallel workers
- can optionally offload the sharded model params to CPU
Introducing PyTorch Fully Sharded Data Parallel (FSDP) API | PyTorch

Additional Parallelism Strategies

Tensor (/ Model) Parallelism (TP):
- 🤗 Tensor Parallelism
- 🔥 Large Scale Transformer model training with Tensor Parallel (TP)
Pipeline Parallelism (PP):
- 🔥 PyTorch, DeepSpeed
Sequence Parallelism (SP):
argonne-lcf/Megatron-DeepSpeed
- Supports 4D Parallelism (DP + TP + PP + SP)

Pipeline Parallelism (PP)

Model is split up vertically (layer-level) across multiple GPUs
Each GPU:
- has a portion of the full model
- processes in parallel different stages of the pipeline (on a small chunk of the batch)
See:
- 🔥 PyTorch / Pipeline Parallelism
- DeepSpeed / Pipeline Parallelism

flowchart TB
    subgraph G0["`GPU 0`"]
        direction LR
        a0("`Layer 0`")
        b0("`Layer 1`")
    end
    subgraph G1["`GPU 1`"]
        direction LR
        a1("`Layer 2`")
        b1("`Layer 3`")
    end
    a0 -.-> b0
    b0 --> a1
    a1 -.-> b1
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
class G0,G1, block
class a0, red
class b0, green
class a1, blue
class b1, yellow

Figure 15: Pipeline Parallelism

Tensor Parallel (TP)

Each tensor is split up into multiple chunks
Each shard of the tensor resides on its designated GPU
During processing each shard gets processed separately (and in parallel) on different GPUs
- synced at the end of the step
See: 🤗 Model Parallelism for additional details

flowchart LR
   subgraph G0["`GPU0`"]
    direction TB
    a0("`Layer 0`")
    b0("`Layer 1`")
    c0("`Layer 2`")
    d0("`Layer 3`")
   end
   subgraph G1["`GPU1`"]
    direction TB
    a1("`Layer 0`")
    b1("`Layer 1`")
    c1("`Layer 2`")
    d1("`Layer 3`")
   end
   a0 <-.-> a1
   b0 <-.-> b1
   c0 <-.-> c1
   d0 <-.-> d1
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
class G0,G1, block
class a0,a1 red
class b0,b1 green
class c0,c1 blue
class d0,d1 yellow

Figure 16: Tensor Parallel Training

Tensor Parallel (TP)

Suitable when the model is too large to fit onto a single device (CPU / GPU)
Typically more complicated to implement than data parallel training
- This is what one may call horizontal parallelism
- Communication whenever dataflow between two subsets
argonne-lcf/Megatron-DeepSpeed
🤗 huggingface/nanotron

flowchart LR
   subgraph G0["`GPU0`"]
    direction TB
    a0("`Layer 0`")
    b0("`Layer 1`")
    c0("`Layer 2`")
    d0("`Layer 3`")
   end
   subgraph G1["`GPU1`"]
    direction TB
    a1("`Layer 0`")
    b1("`Layer 1`")
    c1("`Layer 2`")
    d1("`Layer 3`")
   end
   a0 <-.-> a1
   b0 <-.-> b1
   c0 <-.-> c1
   d0 <-.-> d1
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
class G0,G1, block
class a0,a1 red
class b0,b1 green
class c0,c1 blue
class d0,d1 yellow

Figure 17: Tensor Parallel Training

Split up network over multiple workers
- Each receives disjoint subset
- All communication associated with subsets are distributed
Communication whenever dataflow between two subsets
Typically more complicated to implement than data parallel training
Suitable when the model is too large to fit onto a single device (CPU / GPU)
argonne-lcf/Megatron-DeepSpeed
🤗 huggingface/nanotron

See: 🤗 Model Parallelism for additional details

Tensor (/ Model) Parallel Training: Example

Want to compute: y = \sum_{i} x_{i} W_{i} = x_0 W_0 + x_1 W_1 + x_2 W_2 where each GPU only has only its portion of the full weights as shown below

Compute: x_{0} W_{0}\rightarrow GPU1
Compute: x_{0} W_{0} + x_{1} W_{1}\rightarrow GPU2
Compute: y = \sum_{i} x_{i} W_{i} ✅

flowchart TD
  subgraph X2["`GPU2`"]
    direction LR
    c("`W₂`")
  end
  subgraph X1["`GPU1`"]
    direction TB
    b("`W₁`")
  end
  subgraph X0["`GPU0`"]
    a("`W₀`")
  end
  X0 <-.-> X1
  X1 <-.-> X2
  t0("`x₀`") --> X0
  t1("`x₁`") --> X1
  t2("`x₂`") --> X2
classDef redText fill:#CCCCCC02,stroke:#FF8181,stroke-width:2px,color:#838383,font-weight:500
classDef orangeText fill:#CCCCCC02,stroke:#FFC47F,stroke-width:2px,color:#838383
classDef yellowText fill:#CCCCCC02,stroke:#FFFF7F,stroke-width:2px,color:#838383
classDef blueText fill:#CCCCCC02,stroke:#7DCAff,stroke-width:2px,color:#838383
classDef greenText fill:#CCCCCC02,stroke:#98E6A5,stroke-width:2px,color:#838383
classDef red fill:#ff8181,stroke:#333,stroke-width:1px,color:#000
classDef orange fill:#FFC47F,stroke:#333,stroke-width:1px,color:#000
classDef yellow fill:#FFFF7F,stroke:#333,stroke-width:1px,color:#000
classDef green fill:#98E6A5,stroke:#333,stroke-width:1px,color:#000
classDef blue fill:#7DCAFF,stroke:#333,stroke-width:1px,color:#000
classDef purple fill:#FFCBE6,stroke:#333,stroke-width:1px,color:#000
classDef block fill:#CCCCCC02,stroke:#838383,stroke-width:1px,color:#838383
classDef text fill:#CCCCCC02,stroke:#838383,stroke-width:0px,color:#838383
class a, red
class b, green
class c, blue
class X0,X1,X2, block
%%class t0, redText
%%class t1, greenText
%%class t2, blueText
class a0,b0,c0, text

Tensor (Model) Parallelism²

In Tensor Paralleism each GPU processes only a slice of a tensor and only aggregates the full tensor for operations that require the whole thing.
- The main building block of any transformer is a fully connected nn.Linear followed by a nonlinear activation GeLU.
  - Y = GeLU(XA), where X and Y are the input and output vectors, and A is the weight matrix.
- If we look at the computation in matrix form, it’s easy to see how the matrix multiplication can be split between multiple GPUs:

Tensor Parallelism

Figure 18: Tensor Parallel GEMM. This information is based on (the much more in-depth) TP Overview by @anton-l

3D Parallelism

DP + TP + PP (3D) Parallelism

Deciding on a Parallelism Strategy

Model fits onto a single GPU:
- Normal use
Model DOES NOT fit on a single GPU:
- ZeRO + Offload CPU (or, optionally, NVMe)
Largest layer DOES NOT fit on a single GPU:
- ZeRO + Enable Memory Centric Tiling (MCT)
  - MCT Allows running of arbitrarily large layers by automatically splitting them and executing them sequentially.

Model fits onto a single GPU
- DDP
- ZeRO

Model DOES NOT fit onto a single GPU

With sufficiently fast connectivity between nodes, these three strategies should be comparable.
- Otherwise, PP > ZeRO \simeq TP.

When you have fast inter-node connectivity:
- ZeRO (virtually NO modifications)
- PP + ZeRO + TP + DP (less communication, at the cost of MAJOR modifications)
  - when you have slow inter-node connectivity and still low on GPU memory:
```
DP + PP + TP + ZeRO-1
```
- NOTE: TP is almost always used within a single node, e.g.
  TP <= GPUS_PER_NODE

Large Language Models

Emergent Abilities

Training LLMs

Figure 22: Visualization from Yang et al. (2023)

Figure 23: It’s hungry! Wei et al. (2022)

♻️ Life-Cycle of the LLM

Data collection + preprocessing
Pre-training
- Architecture decisions, model size, etc.
Supervised Fine-Tuning
- Instruction Tuning
- Alignment
Deploy (+ monitor, re-evaluate, etc.)

Figure 24: **Pre-training**: Virtually *all of the compute* used during pre-training⁴.

🎀 Life-Cycle of the LLM

Data collection + preprocessing
Pre-training
- Architecture decisions, model size, etc.
Supervised Fine-Tuning
- Instruction Tuning
- Alignment
Deploy (+ monitor, re-evaluate, etc.)

Figure 25: **Fine-tuning**: Fine-tuning actually updates the model’s weights to make the model better at a certain task⁵.

Forward Pass

Video

Figure 26: Language Model trained for causal language modeling⁶.

Generating Text

Video

Figure 27: Language Model trained for causal language modeling⁷.

👋 Hands On

ai-science-training-series / 06_parallel_training

🧑‍💻 Hands On: Getting Started

🌱 Clone Repo(s):

saforem2/wordplay

git clone https://github.com/saforem2/wordplay
cd wordplay

saforem2/ezpz

git clone https://github.com/saforem2/ezpz deps/ezpz

🐍 Setup Python:

export PBS_O_WORKDIR=$(pwd) && source deps/ezpz/src/ezpz/bin/utils.sh
ezpz_setup_python
ezpz_setup_job

📦 Install {`ezpz`, `wordplay`}

Install Python packages:

saforem2/ezpz:

python3 -m pip install -e "./deps/ezpz" --require-virtualenv

saforem2/wordplay:

# from inside `wordplay/`
python3 -m pip install -e . --require-virtualenv

Test distributed setup:

mpirun -n "${NGPUS}" python3 -m ezpz.test_dist

See: 🍋 ezpz/test_dist.py

`ezpz`: Example [video]

Figure 28: Example: using 🍋 ezpz.test_dist to train a small model using DDP

Prepare Data

python3 wordplay/data/shakespeare_char/prepare.py
# Using HF_DATASETS_CACHE=/home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/data/shakespeare_char/.cache/huggingface
# length of dataset in characters: 1,115,394
# all the unique characters:
#  !$&\',-.3:;?ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz
# vocab size: 65
# train has 1,003,854 tokens
# val has 111,540 tokens

Launch Training (DDP)

launch python3 -m wordplay \
    train.backend=DDP \
    train.eval_interval=100 \
    data=shakespeare \
    train.dtype=bf16 \
    model.batch_size=64 \
    model.block_size=1024 \
    train.max_iters=1000 \
    train.log_interval=10 \
    train.compile=false \
    | tee wordplay-gpt2-DDP.log

Training: Example Output


```bash
[2024-07-17 07:42:11.746540][INFO][__init__:156] - Setting logging level to 'INFO' on 'RANK == 0'
[2024-07-17 07:42:11.748763][INFO][__init__:157] - Setting logging level to 'CRITICAL' on all others 'RANK != 0'
[2024-07-17 07:42:11.749453][INFO][__init__:160] - To disable this behavior, and log from ALL ranks (not recommended), set: 'export LOG_FROM_ALL_RANKS=1'  in your environment, and re-run.
[2024-07-17 07:42:11.772718][INFO][configs:81] - Setting HF_DATASETS_CACHE to /home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/.cache/huggingface/datasets
[2024-07-17 07:42:15.341532][INFO][dist:358] - [device='cuda'][rank=2/3][local_rank=2/3][node=0/0]
[2024-07-17 07:42:15.342381][INFO][dist:358] - [device='cuda'][rank=1/3][local_rank=1/3][node=0/0]
[2024-07-17 07:42:15.342430][INFO][dist:358] - [device='cuda'][rank=3/3][local_rank=3/3][node=0/0]
[2024-07-17 07:42:15.348657][INFO][dist:95] -

[dist_info]:
  • DEVICE=cuda
  • DEVICE_ID=cuda:0
  • DISTRIBUTED_BACKEND=nccl
  • GPUS_PER_NODE=4
  • HOSTS=['x3101c0s13b0n0.hsn.cm.polaris.alcf.anl.gov']
  • HOSTFILE=/var/spool/pbs/aux/2024084.polaris-pbs-01.hsn.cm.polaris.alcf.anl.gov
  • HOSTNAME=x3101c0s13b0n0.hsn.cm.polaris.alcf.anl.gov
  • LOCAL_RANK=0
  • MACHINE=Polaris
  • NUM_NODES=1
  • NGPUS=4
  • NGPUS_AVAILABLE=4
  • NODE_ID=0
  • RANK=0
  • SCHEDULER=PBS
  • WORLD_SIZE_TOTAL=4
  • WORLD_SIZE_IN_USE=4
  • LAUNCH_CMD=mpiexec --verbose --envall -n 4 -ppn 4 --hostfile /var/spool/pbs/aux/2024084.polaris-pbs-01.hsn.cm.polaris.alcf.anl.gov --cpu-bind depth -d 16

[2024-07-17 07:42:15.351446][INFO][dist:725] - [0/4] Using device='cuda' with backend='DDP' + 'nccl' for distributed training.
[2024-07-17 07:42:15.356169][INFO][dist:358] - [device='cuda'][rank=0/3][local_rank=0/3][node=0/0]
[2024-07-17 07:42:15.356692][WARNING][dist:364] - Using [4 / 4] available "cuda" devices !!
[2024-07-17 07:42:15.359571][INFO][configs:317] - Loading val from /home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/data/shakespeare_char/val.bin
[2024-07-17 07:42:15.360138][INFO][configs:317] - Loading train from /home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/data/shakespeare_char/train.bin
[2024-07-17 07:42:15.361154][INFO][configs:442] - Tokens per iteration: 262,144
[2024-07-17 07:42:15.361574][INFO][configs:465] - Using self.ptdtype=torch.float16 on self.device_type='cuda'
[2024-07-17 07:42:15.362002][INFO][configs:471] - Initializing a new model from scratch
[2024-07-17 07:42:15.362529][INFO][dist:874] - Setting up wandb from rank: 0
[2024-07-17 07:42:15.362896][INFO][dist:875] - Using: WB PROJECT: WordPlay
[2024-07-17 07:42:16.451786][INFO][dist:905] - W&B RUN: [still-frog-17](https://wandb.ai/aurora_gpt/WordPlay/runs/6by9vpcj)
[2024-07-17 07:42:16.464106][INFO][dist:312] - Updating wandb.run: still-frog-17 config with "DIST_INFO"
[2024-07-17 07:42:16.469424][INFO][dist:938] - Running on machine='Polaris'
[2024-07-17 07:42:16.471151][WARNING][__main__:89] - {
    "train": {
        "framework": "pytorch",
        "backend": "DDP",
        "device": null,
        "seed": null,
        "port": null,
        "ds_config_path": null,
        "precision": null,
        "ngpus": null,
        "use_wandb": true,
        "eval_interval": 100,
        "log_interval": 10,
        "eval_iters": 200,
        "eval_only": false,
        "always_save_checkpoint": false,
        "init_from": "scratch",
        "wandb_project": "WordPlay",
        "max_iters": 1000,
        "warmup_iters": 100,
        "dtype": "bf16",
        "compile": false
    },
    "model": {
        "n_layer": 12,
        "n_head": 12,
        "n_embd": 768,
        "batch_size": 64,
        "block_size": 1024,
        "activation": "gelu",
        "dropout": 0.0,
        "bias": false,
        "vocab_size": 65
    },
    "data": {
        "dataset": "shakespeare_char",
        "out_dir": "out-shakespeare-char",
        "root_path": null
    },
    "optimizer": {
        "gas": 1,
        "name": "AdamW",
        "learning_rate": 0.0006,
        "weight_decay": 0.1,
        "beta1": 0.9,
        "beta2": 0.95,
        "grad_clip": 1.0,
        "decay_lr": true,
        "lr_decay_iters": 600000,
        "min_lr": 6e-05
    }
}
[2024-07-17 07:42:16.474305][WARNING][__main__:90] - Output dir: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13
[2024-07-17 07:42:16.474922][INFO][trainer:246] - Initializing a new model from scratch
[2024-07-17 07:42:17.258904][INFO][model:255] - number of parameters: 85.00M
[2024-07-17 07:42:17.290004][INFO][trainer:264] - Model size: num_params=85003776
[2024-07-17 07:42:17.292626][INFO][model:445] - num decayed parameter tensors: 50, with 85,771,008 parameters
[2024-07-17 07:42:17.293296][INFO][model:449] - num non-decayed parameter tensors: 25, with 19,200 parameters
[2024-07-17 07:42:17.515324][CRITICAL][trainer:316] - "devid='cuda:1'"
[2024-07-17 07:42:17.515340][CRITICAL][trainer:316] - "devid='cuda:2'"
[2024-07-17 07:42:17.515465][CRITICAL][trainer:316] - "devid='cuda:3'"
[2024-07-17 07:42:18.431814][INFO][model:465] - using fused AdamW: True
[2024-07-17 07:42:18.432620][CRITICAL][trainer:316] - "devid='cuda:0'"
[2024-07-17 07:42:19.951020][INFO][trainer:356] - • self.model=GPT(
  (transformer): ModuleDict(
    (wte): Embedding(65, 768)
    (wpe): Embedding(1024, 768)
    (drop): Dropout(p=0.0, inplace=False)
    (h): ModuleList(
      (0-11): 12 x Block(
        (ln_1): LayerNorm()
        (attn): CausalSelfAttention(
          (c_attn): Linear(in_features=768, out_features=2304, bias=False)
          (c_proj): Linear(in_features=768, out_features=768, bias=False)
          (attn_dropout): Dropout(p=0.0, inplace=False)
          (resid_dropout): Dropout(p=0.0, inplace=False)
        )
        (ln_2): LayerNorm()
        (mlp): MLP(
          (c_fc): Linear(in_features=768, out_features=3072, bias=False)
          (act_fn): GELU(approximate='none')
          (c_proj): Linear(in_features=3072, out_features=768, bias=False)
          (dropout): Dropout(p=0.0, inplace=False)
        )
      )
    )
    (ln_f): LayerNorm()
  )
  (lm_head): Linear(in_features=768, out_features=65, bias=False)
)
[2024-07-17 07:42:19.955340][INFO][trainer:357] - • self.grad_scaler=<torch.cuda.amp.grad_scaler.GradScaler object at 0x145a38f0f090>
[2024-07-17 07:42:19.956897][INFO][trainer:358] - • self.model_engine=DistributedDataParallel(
  (module): GPT(
    (transformer): ModuleDict(
      (wte): Embedding(65, 768)
      (wpe): Embedding(1024, 768)
      (drop): Dropout(p=0.0, inplace=False)
      (h): ModuleList(
        (0-11): 12 x Block(
          (ln_1): LayerNorm()
          (attn): CausalSelfAttention(
            (c_attn): Linear(in_features=768, out_features=2304, bias=False)
            (c_proj): Linear(in_features=768, out_features=768, bias=False)
            (attn_dropout): Dropout(p=0.0, inplace=False)
            (resid_dropout): Dropout(p=0.0, inplace=False)
          )
          (ln_2): LayerNorm()
          (mlp): MLP(
            (c_fc): Linear(in_features=768, out_features=3072, bias=False)
            (act_fn): GELU(approximate='none')
            (c_proj): Linear(in_features=3072, out_features=768, bias=False)
            (dropout): Dropout(p=0.0, inplace=False)
          )
        )
      )
      (ln_f): LayerNorm()
    )
    (lm_head): Linear(in_features=768, out_features=65, bias=False)
  )
)
[2024-07-17 07:42:19.961066][INFO][trainer:359] - • self.optimizer=AdamW (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.95)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: True
    lr: 0.0006
    maximize: False
    weight_decay: 0.1

Parameter Group 1
    amsgrad: False
    betas: (0.9, 0.95)
    capturable: False
    differentiable: False
    eps: 1e-08
    foreach: None
    fused: True
    lr: 0.0006
    maximize: False
    weight_decay: 0.0
)
[2024-07-17 07:42:19.988827][INFO][trainer:802] - Startup time: 6.7125
                Training Legend
┏━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃    abbr     ┃ desc                           ┃
┡━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│    step     │ Current training iteration     │
│    loss     │ Loss value                     │
│     dt      │ Elapsed time per training step │
│     dtf     │ Elapsed time per forward step  │
│     dtb     │ Elapsed time per backward step │
│     sps     │ Samples per second             │
│ sps_per_gpu │ Samples per second (per GPU)   │
│     tps     │ Tokens per second              │
│ tps_per_gpu │ Tokens per second (per GPU)    │
│     mfu     │ Model flops utilization        │
│ train_loss  │ Training loss value            │
│  val_loss   │ Validation loss value          │
└─────────────┴────────────────────────────────┘
[2024-07-17 07:42:21.451865][INFO][trainer:820] - ['prompt']: 'What is an LLM?'
[2024-07-17 07:42:21.452667][INFO][trainer:824] - ['response']:
What is an LLM?eelEl\'$nltPwBSWal,;PWw bbu\'HiyP\'FWwF &AhW:ygrn kk-\'\'KFlMwnlEfflkc,elpWaWtgml$Pgglhllw lglhFllzczPAFHpeAAPPSltgkrWPPhlEMgcrN ggPWt-WPSSzHSkkrzzk.FFrtSSkgMll&gFXr,hghaueaVPW-pHFF-gg,,,FF,,kbApgg gg\'aWWzzkk\'a\'CggHl$bGeA,FFk,,SF;UF,,aZ ;gglee$,k.US&kg:S,,zVzzc
[2024-07-17 07:43:01.573073][INFO][trainer:885] - step=10 loss=3.154310 dt=0.282833 dtf=0.005247 dtb=0.011417 sps=14.142633 sps_per_gpu=3.535658 tps=926851.609409 tps_per_gpu=231712.902352 mfu=46.288281 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:04.402750][INFO][trainer:885] - step=20 loss=2.660851 dt=0.306263 dtf=0.005233 dtb=0.011419 sps=13.060678 sps_per_gpu=3.265170 tps=855944.613638 tps_per_gpu=213986.153409 mfu=45.934162 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:07.237507][INFO][trainer:885] - step=30 loss=2.543283 dt=0.283021 dtf=0.005238 dtb=0.011245 sps=14.133211 sps_per_gpu=3.533303 tps=926234.088226 tps_per_gpu=231558.522057 mfu=45.966490 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:10.077248][INFO][trainer:885] - step=40 loss=2.503963 dt=0.285001 dtf=0.005213 dtb=0.011471 sps=14.035061 sps_per_gpu=3.508765 tps=919801.749941 tps_per_gpu=229950.437485 mfu=45.963461 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:12.917039][INFO][trainer:885] - step=50 loss=2.477469 dt=0.283532 dtf=0.005166 dtb=0.011294 sps=14.107763 sps_per_gpu=3.526941 tps=924566.380009 tps_per_gpu=231141.595002 mfu=45.984530 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:15.760749][INFO][trainer:885] - step=60 loss=2.471083 dt=0.284630 dtf=0.005140 dtb=0.011224 sps=14.053326 sps_per_gpu=3.513332 tps=920998.786204 tps_per_gpu=230249.696551 mfu=45.985675 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:18.602785][INFO][trainer:885] - step=70 loss=2.458894 dt=0.283926 dtf=0.005219 dtb=0.010383 sps=14.088155 sps_per_gpu=3.522039 tps=923281.352698 tps_per_gpu=230820.338174 mfu=45.998106 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:21.451433][INFO][trainer:885] - step=80 loss=2.489088 dt=0.285537 dtf=0.005183 dtb=0.011373 sps=14.008683 sps_per_gpu=3.502171 tps=918073.060430 tps_per_gpu=229518.265108 mfu=45.983282 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:24.302241][INFO][trainer:885] - step=90 loss=2.471990 dt=0.300767 dtf=0.005445 dtb=0.010290 sps=13.299337 sps_per_gpu=3.324834 tps=871585.359388 tps_per_gpu=217896.339847 mfu=45.737774 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:27.153275][INFO][trainer:885] - step=100 loss=2.445556 dt=0.285869 dtf=0.005182 dtb=0.011251 sps=13.992403 sps_per_gpu=3.498101 tps=917006.151328 tps_per_gpu=229251.537832 mfu=45.743655 train_loss=4.125778 val_loss=4.128809
[2024-07-17 07:43:28.182553][INFO][trainer:820] - ['prompt']: 'What is an LLM?'
[2024-07-17 07:43:28.183179][INFO][trainer:824] - ['response']:

What is an LLM?

Goupay my winghimithell bls ger t bon sinthard ht omind be,
And lereind h py balithand frd oforondof wimon me hageas thinero mand,
Thacanes,
An frift ghik med d herthecke ntore thack couthen ale, t thit ang d m t h chy me fache ag, wit my hathan glat ng
[2024-07-17 07:44:06.025837][INFO][trainer:760] - Saving checkpoint to: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13
[2024-07-17 07:44:06.026607][INFO][trainer:761] - Saving model to: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13/model.pth
[2024-07-17 07:44:07.682968][INFO][configs:141] - Appending /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13 to /home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/src/ckpts/checkpoints.log
[2024-07-17 07:44:10.519506][INFO][trainer:885] - step=110 loss=2.433923 dt=0.285038 dtf=0.005757 dtb=0.011762 sps=14.033209 sps_per_gpu=3.508302 tps=919680.367894 tps_per_gpu=229920.091974 mfu=45.762304 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:13.362148][INFO][trainer:885] - step=120 loss=2.429014 dt=0.284445 dtf=0.005222 dtb=0.011486 sps=14.062460 sps_per_gpu=3.515615 tps=921597.361532 tps_per_gpu=230399.340383 mfu=45.788661 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:16.210694][INFO][trainer:885] - step=130 loss=2.402059 dt=0.285559 dtf=0.005199 dtb=0.011765 sps=14.007633 sps_per_gpu=3.501908 tps=918004.211586 tps_per_gpu=229501.052897 mfu=45.794438 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:19.061546][INFO][trainer:885] - step=140 loss=2.374062 dt=0.285476 dtf=0.005239 dtb=0.011453 sps=14.011662 sps_per_gpu=3.502916 tps=918268.297093 tps_per_gpu=229567.074273 mfu=45.800956 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:21.917283][INFO][trainer:885] - step=150 loss=2.365385 dt=0.285846 dtf=0.005125 dtb=0.011320 sps=13.993568 sps_per_gpu=3.498392 tps=917082.475791 tps_per_gpu=229270.618948 mfu=45.800900 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:24.771924][INFO][trainer:885] - step=160 loss=2.317337 dt=0.280788 dtf=0.005173 dtb=0.011249 sps=14.245602 sps_per_gpu=3.561401 tps=933599.792506 tps_per_gpu=233399.948127 mfu=45.883340 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:27.626812][INFO][trainer:885] - step=170 loss=2.256231 dt=0.284973 dtf=0.005141 dtb=0.011299 sps=14.036416 sps_per_gpu=3.509104 tps=919890.544506 tps_per_gpu=229972.636126 mfu=45.889069 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:30.480952][INFO][trainer:885] - step=180 loss=2.216419 dt=0.286555 dtf=0.005180 dtb=0.011402 sps=13.958906 sps_per_gpu=3.489726 tps=914810.852170 tps_per_gpu=228702.713043 mfu=45.868857 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:33.337342][INFO][trainer:885] - step=190 loss=2.145123 dt=0.291456 dtf=0.005409 dtb=0.019347 sps=13.724205 sps_per_gpu=3.431051 tps=899429.467247 tps_per_gpu=224857.366812 mfu=45.773849 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:36.194584][INFO][trainer:885] - step=200 loss=2.068149 dt=0.285703 dtf=0.005153 dtb=0.011286 sps=14.000555 sps_per_gpu=3.500139 tps=917540.393411 tps_per_gpu=229385.098353 mfu=45.778791 train_loss=2.439494 val_loss=2.478951
[2024-07-17 07:44:37.224149][INFO][trainer:820] - ['prompt']: 'What is an LLM?'
[2024-07-17 07:44:37.224745][INFO][trainer:824] - ['response']:

What is an LLM?

LORTESS LA:
No, sighappat selace? don downd sourciceans note cancen up sof liond
This and my man, werame, of re thee
Thise not will I on land brond sul me a fingore?

FLER:
Tisint your not nare lame o igen,-to brorst.

SamERS:
Sin:
I\'l hell she lor hen w
[2024-07-17 07:45:14.409129][INFO][trainer:760] - Saving checkpoint to: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13
[2024-07-17 07:45:14.409820][INFO][trainer:761] - Saving model to: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13/model.pth
[2024-07-17 07:45:16.366935][INFO][configs:141] - Appending /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13 to /home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/src/ckpts/checkpoints.log
[2024-07-17 07:45:19.245061][INFO][trainer:885] - step=210 loss=1.982169 dt=0.283305 dtf=0.005223 dtb=0.011284 sps=14.119042 sps_per_gpu=3.529760 tps=925305.515083 tps_per_gpu=231326.378771 mfu=45.822019 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:22.092430][INFO][trainer:885] - step=220 loss=1.897731 dt=0.284759 dtf=0.005217 dtb=0.011187 sps=14.046945 sps_per_gpu=3.511736 tps=920580.608106 tps_per_gpu=230145.152026 mfu=45.837327 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:24.942639][INFO][trainer:885] - step=230 loss=1.817213 dt=0.285266 dtf=0.005208 dtb=0.011446 sps=14.022003 sps_per_gpu=3.505501 tps=918945.985503 tps_per_gpu=229736.496376 mfu=45.842940 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:27.797910][INFO][trainer:885] - step=240 loss=1.779287 dt=0.285465 dtf=0.005189 dtb=0.011220 sps=14.012250 sps_per_gpu=3.503062 tps=918306.793546 tps_per_gpu=229576.698387 mfu=45.844800 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:30.653597][INFO][trainer:885] - step=250 loss=1.704220 dt=0.289284 dtf=0.005471 dtb=0.010346 sps=13.827253 sps_per_gpu=3.456813 tps=906182.836379 tps_per_gpu=226545.709095 mfu=45.785926 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:33.512769][INFO][trainer:885] - step=260 loss=1.671318 dt=0.287679 dtf=0.005125 dtb=0.011250 sps=13.904380 sps_per_gpu=3.476095 tps=911237.442617 tps_per_gpu=227809.360654 mfu=45.758182 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:36.373461][INFO][trainer:885] - step=270 loss=1.650952 dt=0.298661 dtf=0.005118 dtb=0.011520 sps=13.393107 sps_per_gpu=3.348277 tps=877730.651421 tps_per_gpu=219432.662855 mfu=45.565875 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:39.236930][INFO][trainer:885] - step=280 loss=1.573242 dt=0.285970 dtf=0.005171 dtb=0.011290 sps=13.987477 sps_per_gpu=3.496869 tps=916683.279847 tps_per_gpu=229170.819962 mfu=45.587333 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:42.100605][INFO][trainer:885] - step=290 loss=1.533265 dt=0.286487 dtf=0.005432 dtb=0.011288 sps=13.962259 sps_per_gpu=3.490565 tps=915030.617828 tps_per_gpu=228757.654457 mfu=45.598392 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:44.964424][INFO][trainer:885] - step=300 loss=1.492064 dt=0.288480 dtf=0.005355 dtb=0.011480 sps=13.865774 sps_per_gpu=3.466443 tps=908707.340870 tps_per_gpu=227176.835218 mfu=45.576766 train_loss=2.045786 val_loss=2.148510
[2024-07-17 07:45:45.995833][INFO][trainer:820] - ['prompt']: 'What is an LLM?'
[2024-07-17 07:45:45.996497][INFO][trainer:824] - ['response']:

What is an LLM?

RICHMORD:
Char stire? how in those are name the range hone.

GLOUCESTER:
Nay, in lond's time the palt are worder more
That wilt in the purpose be a pey
And thou thine onter hands, and the which broth.

ELBOWINCA:
At lie my lord with the me an arms be a s
[2024-07-17 07:46:23.549987][INFO][trainer:760] - Saving checkpoint to: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13
[2024-07-17 07:46:23.550696][INFO][trainer:761] - Saving model to: /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13/model.pth
[2024-07-17 07:46:25.496559][INFO][configs:141] - Appending /home/foremans/tmp/polaris-talk/outputs/runs/pytorch/DDP/2024-07-17/07-42-13 to /home/foremans/tmp/polaris-talk/2024-07-17-073327/wordplay/src/ckpts/checkpoints.log
[2024-07-17 07:46:28.374854][INFO][trainer:885] - step=310 loss=1.444200 dt=0.299907 dtf=0.005333 dtb=0.010637 sps=13.337481 sps_per_gpu=3.334370 tps=874085.133345 tps_per_gpu=218521.283336 mfu=45.384395 train_loss=1.495372 val_loss=1.713714
[2024-07-17 07:46:31.223079][INFO][trainer:885] - step=320 loss=1.429350 dt=0.285238 dtf=0.005245 dtb=0.011485 sps=14.023353 sps_per_gpu=3.505838 tps=919034.479880 tps_per_gpu=229758.619970 mfu=45.435743 train_loss=1.495372 val_loss=1.713714
[2024-07-17 07:46:34.074957][INFO][trainer:885] - step=330 loss=1.362220 dt=0.285027 dtf=0.005165 dtb=0.011407 sps=14.033736 sps_per_gpu=3.508434 tps=919714.904826 tps_per_gpu=229928.726207 mfu=45.485355 train_loss=1.495372 val_loss=1.713714
[2024-07-17 07:46:36.929464][INFO][trainer:885] - step=340 loss=1.350888 dt=0.284436 dtf=0.005199 dtb=0.011287 sps=14.062893 sps_per_gpu=3.515723 tps=921625.744709 tps_per_gpu=230406.436177 mfu=45.539549 train_loss=1.495372 val_loss=1.713714
```

`wordplay`: Example [video]

Figure 30: Training a LLM to talk like Shakespeare using saforem2/wordplay 🎮💬

❤️ Thank you!

Organizers
Feel free to reach out!

Acknowledgements

This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC02-06CH11357

📓 References

Title slide (Tetris animation) from: https://emilhvitfeldt.github.io/quarto-iframe-examples/tetris/index.html

References

Wei, Jason, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, et al. 2022. “Emergent Abilities of Large Language Models.” https://arxiv.org/abs/2206.07682.

Yang, Jingfeng, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023. “Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond.” https://arxiv.org/abs/2304.13712.

Yao, Shunyu, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. “Tree of Thoughts: Deliberate Problem Solving with Large Language Models.” https://arxiv.org/abs/2305.10601.

Footnotes

micro_batch_size = batch_size per GPU↩︎
Efficient Large-Scale Language Model Training on GPU Clusters ↩︎
Source: Hannibal046/Awesome-LLM↩︎
Figure from The Illustrated Transformer ↩︎
Figure from The Illustrated Transformer ↩︎
Video from: 🤗 Generation with LLMs ↩︎
Video from: 🤗 Generation with LLMs ↩︎

Citation

BibTeX citation:

@unpublished{foreman2024,
  author = {Foreman, Sam},
  title = {Parallel {Training} {Methods}},
  date = {2024-11-05},
  url = {https://samforeman.me/talks/ai-for-science-2024/slides},
  langid = {en}
}

For attribution, please cite this work as:

Foreman, Sam. 2024. “Parallel Training Methods.” November 5. https://samforeman.me/talks/ai-for-science-2024/slides.

👀 Overview

📑 Outline

🚀 Scaling: Overview

Single GPU

Data Parallel Training

Data Parallel Training: Forward Pass

Data Parallel Training: Backward Pass

Data Parallel Training

Data Parallel Training

🗣️ Communication

AllReduce

Reduce

Broadcast

AllGather

Scatter

Why Distributed Training?

Why Distributed Training? Speedup!

Dealing with Data

Broadcast Initial State

Best Practices

Going Beyond Data Parallelism

Going beyond Data Parallelism: DeepSpeed + ZeRO

Fully Sharded Data Parallel: 🔥 PyTorch + FSDP

Additional Parallelism Strategies

Pipeline Parallelism (PP)

Tensor Parallel (TP)

Tensor Parallel (TP)

Tensor (/ Model) Parallel Training: Example

Tensor (Model) Parallelism2

Tensor Parallelism

3D Parallelism

Deciding on a Parallelism Strategy

Large Language Models

Emergent Abilities

Training LLMs

♻️ Life-Cycle of the LLM

🎀 Life-Cycle of the LLM

Forward Pass

Generating Text

👋 Hands On

🧑‍💻 Hands On: Getting Started

📦 Install {ezpz, wordplay}

ezpz: Example [video]

Install wordplay 🎮💬

Prepare Data

Launch Training (DDP)

Training: Example Output

wordplay: Example [video]

❤️ Thank you!

📓 References

References

Footnotes

Citation

Going beyond Data Parallelism: DeepSpeed + `ZeRO`

Fully Sharded Data Parallel: 🔥 PyTorch + `FSDP`

Tensor (Model) Parallelism²

📦 Install {`ezpz`, `wordplay`}

`ezpz`: Example [video]

Install `wordplay` 🎮💬

`wordplay`: Example [video]