Skip to content

rocm: add hipBLASLt 7.14.0 dat.zlib overlay support - #14

Open
dannybaths wants to merge 1 commit into
amd:mainfrom
dannybaths:hipblaslt-7140-dat-zlib-overlay
Open

rocm: add hipBLASLt 7.14.0 dat.zlib overlay support#14
dannybaths wants to merge 1 commit into
amd:mainfrom
dannybaths:hipblaslt-7140-dat-zlib-overlay

Conversation

@dannybaths

@dannybaths dannybaths commented Sep 5, 2026

Copy link
Copy Markdown

Summary

This PR extends rocm/scripts/hipblaslt_patch_setup.sh to support ROCm 7.14.0.

ROCm 7.14.0 stores the relevant Tensile catalogue leaves as *.dat.zlib under
gfx942/, so the patcher now detects compressed leaves, decompresses them for
msgpack editing, and recompresses them when writing the overlay.

The 7.14.0 case patches only the fp16 forward leaf:

TensileLibrary_HH_HH_HA_Bias_SAV_UA_Type_HH_HPA_Contraction_l_Alik_Bljk_Cijk_Dijk_gfx942.dat.zlib

The 120 small-N fp16 shapes are routed to existing in-pool solution 243113.
No kernel binary is built or changed. fp32 and bf16 pass stock on 7.14.0, and
the CPX WorkGroupMappingXCC rewrite is skipped for 7.14.0 because that defect is
already fixed upstream for this release.

Why

On MI300A, stock ROCm 7.14.0 misses the fp16 forward small-N family and returns
zero solutions for the 120-shape sweep. The existing Dock patcher handled older
plain .dat layouts, but not the 7.14.0 .dat.zlib layout.

This makes the proven overlay reproducible through the normal Dock packaging
path instead of requiring a manual SDK edit.

Validation

  • The Dock 7.14.0 case was extracted and run against the 7.14.0 SDK without
    sudo. It produced one patched .dat.zlib leaf, byte-identical to the
    empirically proven overlay leaf, md5 3caa7a436470c3aa7a75822504ed4c0d.
  • AAC7 job 10521: fp16, fp32, and bf16 all passed with the generated overlay
    active.
  • AAC7 job 10534: standalone rerun reproduced the all-dtype pass with the same
    overlay active.
  • AAC7 job 10539: focused in-harness Phase C run proved the row-level overlay
    environment reached the compute node and Pytorch_HIPBLASLt_Regression_Check
    passed.
  • AAC6 focused check passed in Slurm job 19668 on PPAC_MI300A_SPX with the PR rocm: add hipBLASLt 7.14.0 dat.zlib overlay support #14 scratch overlay active. Pytorch_HIPBLASLt_Regression_Check passed for fp16, fp32, and bf16; no sudo, live modulefile, SDK, or /opt edits were made.

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