Skip to content

Grouped-GEMM and non-fp32 redesign for GemmArgs #56

Description

@drunkcoding

Follow-up to #48.

When #48 landed (commit f298c6d), GemmArgs was collapsed to single-shape fp32 because:

  • Every callsite (proxy_cli, cpu/gpu workers, benches, tests) hardcoded group_size = 1.
  • The group_size > 1 branch in CalculateTaskSizes was LOG_FATAL.
  • No fp16/bf16/int8 GEMM consumer existed.

The struct now ships with a GemmDtype enum (kFloat32 only) as an explicit extension point and static_assert(sizeof(float) == 4) to make the assumption observable. Reopen / pick this up when a real consumer requires:

  1. Grouped GEMM: variadic per-group shapes (re-introduce the MAX_GEMM_DIM-indexed layout or use a heap-allocated descriptor array).
  2. Non-fp32 operands: fp16 / bf16 / int8 paths. Will need cuBLASXt / MKL backend support and probably a dispatch table keyed on GemmDtype.

Until then, the single-shape fp32 surface is the documented contract.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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