Skip to content

Fix gated FlashAttention benchmark bindings - #137

Merged
rsuderman merged 4 commits into
AMD-Ecosystem:hrx-graph-develop-v2from
rsuderman:main
Oct 7, 2026
Merged

rsuderman merged 4 commits into
AMD-Ecosystem:hrx-graph-develop-v2from
rsuderman:main

Conversation

@rsuderman

Copy link
Copy Markdown
Collaborator

No description provided.

Materialize the distinct production gate buffer when generating standalone FlashAttention cases, validate its alignment and strided capacity, and cover the contract with focused generator tests.

Assisted-by: Codex
Select a ten-way reduction unroll only for same-format Q8_0 gate/up weights at token count one, preserving the existing schedule for Q4 and generic low-token routes.

Assisted-by: Codex
Integrate reusable Q1_0 decode-state naming and paired-row scale reuse before full model correctness and performance validation.

Assisted-by: Codex
Route contiguous F32 RMSNorm multiply cases with one to four tokens and hidden widths 2048 through 5120 through a striped workgroup schedule. Preserve the existing publication, strided, add, and unsupported-shape routes as controls.

Assisted-by: Codex
@rsuderman
rsuderman merged commit 8108322 into AMD-Ecosystem:hrx-graph-develop-v2 Oct 7, 2026
1 of 2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant