[ROCm][CI] Fix moe layer tests for fp8 dtype compatibility (#56106)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Divakar Verma
2026-09-09 17:48:14 +00:00
committed by GitHub
co-authored by Andreas Karatzas
parent a207d7ce1d
commit 3fb676bfad
2 changed files with 11 additions and 8 deletions
+4 -1
View File
@@ -84,7 +84,7 @@ def on_gfx950() -> bool:
return False
fp8_dtype = torch.float8_e4m3fn # current_platform.fp8_dtype
fp8_dtype = current_platform.fp8_dtype()
SHAPE_COMBOS = [
(1, 128, 256),
@@ -191,6 +191,9 @@ def mock_normalize_e4m3fn_to_e4m3fnuz(
# NOTE: Not able to use monkeypatch because of the spawned parallel workers.
def override_normalize_e4m3fn_to_e4m3fnuz():
vllm.model_executor.layers.quantization.utils.w8a8_utils.normalize_e4m3fn_to_e4m3fnuz = mock_normalize_e4m3fn_to_e4m3fnuz # noqa: E501
vllm.model_executor.layers.quantization.fp8.normalize_e4m3fn_to_e4m3fnuz = (
mock_normalize_e4m3fn_to_e4m3fnuz # noqa: E501
)
def sp_wrapper(