Skip to content

Jobs hang with fsdp only (1d) sharding with >=64 tpu chips #363

@jialei777

Description

@jialei777

Training worked locally with 4 chips fsdp=4: export PJRT_DEVICE=TPU; export TORCHPRIME_TPU_TYPE=v6e-4 && python torchprime/torch_xla_models/train.py model=flex-qwen-1b

MFU: 0.21

On a v5p-128 cluster with command tp run --name jialei-0812-qwen-fsdp32tensor2 torchprime/torch_xla_models/train.py model=flex-qwen-1b task.global_batch_size=64 ici_mesh.fsdp=x ici_mesh.tensor=y

  • fsdp64 tp 1: hang >.<
  • fsdp 32 tp 2: finished MFU 0.22
  • fsdp 16 tp 4: finished: MFU 0.19
  • fsdp 8 tp 8: finished, MFU 0.11

Also see cluster log here https://b.corp.google.com/issues/436664633#comment40.
Runs for llama model: https://b.corp.google.com/issues/436664633#comment33

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions