Skip to content

ci: sync the HIP VGPR-spill ignore list with upstream - #341

Merged
TheTom merged 1 commit into
TheTom:feature/turboquant-kv-cachefrom
jasstrong:pr/hip-vgpr-ignore-sync
Sep 3, 2026
Merged

ci: sync the HIP VGPR-spill ignore list with upstream#341
TheTom merged 1 commit into
TheTom:feature/turboquant-kv-cachefrom
jasstrong:pr/hip-vgpr-ignore-sync

Conversation

@jasstrong

Copy link
Copy Markdown

Overview

ubuntu-22-hip-quality-check has been red on every recent PR against this branch (#334, #335, #337) with the same four kernels over the 256-VGPR line, none of which those PRs touch:

kernel VGPRs
rwkv_wkv_f32<128> 387 (64 + 323 spilled)
mul_mat_q<Q2_K, 64, fallback> 317 (128 + 189 spilled)
flash_attn_ext_vec<128, 2, turbo3, q8_0, false> 259 (255 + 4 spilled)
flash_attn_ext_vec<128, 2, turbo3, q8_0, true> 259 (255 + 4 spilled)

The first two are on upstream's current ignore list in scripts/hip/gcn-cdna-vgpr-check.py. This branch carries an older 85-entry copy of that list which predates those entries and no longer matches any kernel that is over the limit. This PR replaces it with upstream's current list and adds the two TurboQuant vector flash-attention instances, which sit three registers over the threshold.

Trimming that 4-VGPR spill is a separate change; this only makes the check report new regressions instead of the baseline.

Additional information

Script-only change, no code. The ignore list is matched by mangled name, so the two added entries are the exact symbols from the failing job's metrics.log.

Requirements

  • I have read and agree with the contributing guidelines
  • AI usage disclosure: YES — Claude Code was used to compare the failing jobs across PRs and to draft this change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NxP6x5bmUDYFvmouceN2mR

…urbo3 vector FA kernel

The ubuntu-22-hip-quality-check job has been red on every recent PR with the
same four kernels, none of which the PRs touch:

  rwkv_wkv_f32<128>                                        387 VGPRs (64 + 323 spilled)
  mul_mat_q<Q2_K, 64, fallback>                            317 VGPRs (128 + 189 spilled)
  flash_attn_ext_vec<128, 2, turbo3, q8_0, {false,true}>   259 VGPRs (255 + 4 spilled)

The first two are on upstream's current ignore list; this branch still carried
an older, longer list that predates those entries and no longer matches any
kernel that is over the limit. Replace it with upstream's list and add the two
TurboQuant vector flash-attention instances, which sit three registers over
the threshold. Trimming that spill is a separate change; this only makes the
check report new regressions instead of the baseline.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NxP6x5bmUDYFvmouceN2mR
@TheTom
TheTom merged commit 4ff4bd3 into TheTom:feature/turboquant-kv-cache Sep 3, 2026
2 of 4 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.

2 participants