Fix MoE backend selection for LoRA (unquantized MoE) (#40273)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
@@ -163,6 +163,11 @@ def select_unquantized_moe_backend(
|
||||
if current_platform.is_out_of_tree():
|
||||
return UnquantizedMoeBackend.OOT, None
|
||||
|
||||
if moe_config.is_lora_enabled:
|
||||
return UnquantizedMoeBackend.TRITON, backend_to_kernel_cls(
|
||||
UnquantizedMoeBackend.TRITON
|
||||
)
|
||||
|
||||
# NOTE: the kernels are selected in the following order.
|
||||
AVAILABLE_BACKENDS = _get_priority_backends(moe_config)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user