Skip to content

PGOLF-619: add public CUDA BF16 train-visible execution posture #263

Description

@AtlantisPleb

Parent: #188, #183

Why this exists:

  • the repo has BF16 policy and accounting posture, but the public CUDA runtime still does not honestly expose broad BF16 train-visible tensor, gradient, and optimizer execution
  • the single-H100 bring-up keeps refusing this blocker explicitly

Scope:

  • widen the public CUDA/runtime surface needed for BF16-forward or FP32-master Parameter Golf training
  • keep storage/compute/accumulator/master-weight posture machine-readable and explicit
  • add targeted tests or evidence for the bounded BF16 train path that actually lands

Exit criteria:

  • the public Rust-owned Parameter Golf lane has a real BF16-or-master-weight CUDA execution path instead of only policy receipts
  • the cuda_bf16_train_precision_contract blocker is retired or materially narrowed with explicit evidence

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