[Misc] Enable test_silu_mul_fp8_quant_deep_gemm on XPU (#49444)

Signed-off-by: pmanczak <pawel.manczak@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
pmanczak
2026-08-11 23:19:14 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent 4988df2eb8
commit 5b5eae2394
2 changed files with 19 additions and 17 deletions
@@ -211,16 +211,13 @@ def persistent_masked_m_silu_mul_quant(
DeepGemmQuantScaleFMT.UE8M0,
]
device_capability = current_platform.get_device_capability()
assert device_capability is not None
cuda_arch = device_capability.to_int()
if current_platform.is_cuda() and cuda_arch >= 80:
# The C++ kernel requires sm_80+; ROCm and XPU take the Triton path below.
if current_platform.is_cuda() and current_platform.has_device_capability(80):
torch.ops._C.persistent_masked_m_silu_mul_quant(
y, tokens_per_expert, y_q, y_s, ceil_ue8m0
)
else:
# Triton fallback for ROCm -- the C++ kernel is guarded by
# Triton fallback for ROCm and XPU -- the C++ kernel is guarded by
# #ifndef USE_ROCM in activation_kernels.cu.
# https://github.com/ROCm/aiter/issues/2420
stride_cnt_e = tokens_per_expert.stride()[0]