Fix MoE backend selection for LoRA (unquantized MoE) (#40273)

Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
danisereb
2026-04-19 17:18:40 +00:00
committed by GitHub
parent 982beae809
commit d1135a5087
2 changed files with 90 additions and 0 deletions
@@ -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)