Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions perf_tools/hlo_eval_tools/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -726,6 +726,14 @@ running `inference/1gpu` with command buffers on:
| science | `alphafold3`, `colabfold` |
| large_language_models | `gpt_j_6b`, `flan_t5_large` |

> **Fixed in TheRock ROCm nightlies.** These failures are fixed by CLR commit
> [`a9dad6342cd`](https://github.com/ROCm/rocm-systems/commit/a9dad6342cdf0ad3ec74d8cd9d103a8ee15b4e78)
> ([ROCm/rocm-systems#12274](https://github.com/ROCm/rocm-systems/pull/12274)), available since nightly
> `10.3.0a20261103`. With it and MIOpen
> [ROCm/rocm-libraries#12265](https://github.com/ROCm/rocm-libraries/pull/12265) (since `10.3.0a20261104`),
> every model above runs with command buffers ON on gfx950. These nightlies are development builds of TheRock
> `main` (rocm-systems and rocm-libraries `develop`), not a 10.3 release candidate.

These run fine **with command buffers ON** (the flag is not required, but the
default `CMD_BUFFER=off` is harmless for them):

Expand Down
Loading