[Bugfix][ROCm] Preserve MoE weight padding for unquantized Triton path (#46381)
Signed-off-by: Uros Markovic <umarkovi@amd.com>
This commit is contained in:
@@ -301,6 +301,13 @@ def convert_to_unquantized_kernel_format(
|
|||||||
is_gated_act_gemm=is_act_and_mul,
|
is_gated_act_gemm=is_act_and_mul,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
if (
|
||||||
|
unquantized_backend == UnquantizedMoeBackend.TRITON
|
||||||
|
and current_platform.is_rocm()
|
||||||
|
and envs.VLLM_ROCM_MOE_PADDING
|
||||||
|
):
|
||||||
|
# Skip .contiguous(): it would undo the ROCm MoE weight padding.
|
||||||
|
return w13_weight, w2_weight
|
||||||
return w13_weight.contiguous(), w2_weight.contiguous()
|
return w13_weight.contiguous(), w2_weight.contiguous()
|
||||||
|
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user