[DSV4][Kernel] Fuse shared experts into MegaMoE (#53040)
Signed-off-by: Canlin Guo <canlinguosdu@gmail.com>
This commit is contained in:
@@ -93,7 +93,10 @@ from vllm.models.common.ops.sequence_parallel import (
|
||||
sp_reduce_scatter,
|
||||
sp_shard,
|
||||
)
|
||||
from vllm.models.deepseek_v4.nvidia.model import DeepseekV4MegaMoEExperts
|
||||
from vllm.models.deepseek_v4.nvidia.model import (
|
||||
DeepseekV4MegaMoEExperts,
|
||||
DeepseekV4MLP,
|
||||
)
|
||||
from vllm.models.deepseek_v4.nvidia.ops.prepare_megamoe import prepare_megamoe_inputs
|
||||
from vllm.models.kimi_k3.nvidia.kda import KimiK3DeltaAttention
|
||||
from vllm.models.kimi_k3.nvidia.latent_moe_runner import (
|
||||
@@ -354,7 +357,7 @@ class KimiK3MegaMoEExperts(DeepseekV4MegaMoEExperts):
|
||||
torch.distributed.barrier(group=ep_group.cpu_group)
|
||||
self._synchronized_ep_groups.add(key)
|
||||
|
||||
def finalize_weights(self) -> None:
|
||||
def finalize_weights(self, shared_experts: DeepseekV4MLP | None = None) -> None:
|
||||
if self._transformed_l1_weights is not None:
|
||||
return
|
||||
|
||||
|
||||
Reference in New Issue
Block a user