Skip to content

feat(pflash): improve AMD defaults and enable rocWMMA - #669

Draft
Graffioh wants to merge 3 commits into
Luce-Org:mainfrom
Graffioh:codex/pflash-amd-adaptive-ubatch
Draft

feat(pflash): improve AMD defaults and enable rocWMMA#669
Graffioh wants to merge 3 commits into
Luce-Org:mainfrom
Graffioh:codex/pflash-amd-adaptive-ubatch

Conversation

@Graffioh

@Graffioh Graffioh commented Aug 27, 2026

Copy link
Copy Markdown
Contributor

Summary

  • use a HIP-specific sparse-alpha default of 0.95 while preserving user overrides and logging the effective value
  • seed fresh PFlash sessions from the resolved keep ratio, including clamping and sticky-session behavior
  • auto-enable rocWMMA Phase 2 kernels for fresh HIP builds when headers are available, while preserving explicit OFF and warning on q8 fallback

Validation

Tested on the local Strix Halo gfx1151 GPU pinned with HIP_VISIBLE_DEVICES=1.

  • fresh HIP configure: rocWMMA detected and enabled by default
  • explicit -DDFLASH27B_P2_USE_ROCWMMA=OFF: fallback remains disabled as requested
  • affected server and flashprefill targets compile and link with -j$(nproc)
  • fresh-session configured-ratio test passed
  • clamped/sticky session-seed test passed
  • HIP flashprefill kernel test passed on gfx1151
  • git diff --check passed

A full all-target HIP build reaches the unrelated existing test_rms_norm_hip.cpp:50 compatibility failure (cudaErrorNoDevice is not defined for this HIP build). The affected targets compile and link successfully.

@Graffioh
Graffioh force-pushed the codex/pflash-amd-adaptive-ubatch branch from 557f148 to c594b87 Compare August 27, 2026 06:56
@Graffioh Graffioh changed the title feat(pflash): improve AMD defaults and adaptive prefill batching feat(pflash): improve AMD defaults and enable rocWMMA Aug 27, 2026
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