SIGN IN SIGN UP

[rocm-libraries] ROCm/rocm-libraries#5438 (commit 7000562)

[CK_TILE] Normalize gpu_target before LDS_SIZE_MAP lookup
 (#5438)

GPU targets passed with feature suffixes (e.g. `gfx950:xnack+`) were
falling through to `DEFAULT_LDS_SIZE` instead of matching their entry in
`LDS_SIZE_MAP`, potentially causing incorrect tile acceptance/rejection.

## Changes

- **`gemm_validation_utils.py`**: Strip everything after `:` from
`gpu_target` before the `LDS_SIZE_MAP` lookup; use the normalized base
arch name in the error message as well.

```python
# Before
hw_lds_size = LDS_SIZE_MAP.get(gpu_target, DEFAULT_LDS_SIZE)

# After
base_gpu_target = gpu_target.split(":")[0] if gpu_target else gpu_target
hw_lds_size = LDS_SIZE_MAP.get(base_gpu_target, DEFAULT_LDS_SIZE)
```
S
Sami Remes committed
b619c374eb7c042bee67411f224d228cc7ab239e
Parent: 8bd8094
Committed by assistant-librarian[bot] <assistant-librarian[bot]@users.noreply.github.com> on 5/29/2026, 4:33:15 PM