Skip to content

[FSDP] Investigate test_fsdp_pure_fp16.py inaccuracy #90784

@awgu

Description

@awgu

test_fsdp_pure_fp16.py has been mostly flaky or broken. For now, we are disabling the parameter check after the training step. However, this means that the backward pass could be incorrect for pure FP16 models using FSDP.

cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @rohan-varma @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501

Metadata

Metadata

Assignees

No one assigned

    Labels

    module: fsdponcall: distributedAdd this issue/PR to distributed oncall triage queuetriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions