[MoE Refactor] Make SharedExperts class for use with DefaultMoERunner (#35153)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-04-01 09:44:08 -04:00
committed by GitHub
parent 5e30e9b9a9
commit 7cf56a59a2
34 changed files with 556 additions and 397 deletions
@@ -444,7 +444,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
if self.rocm_aiter_moe_enabled:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,
@@ -634,7 +634,7 @@ class QuarkW4A8Fp8MoEMethod(QuarkMoEMethod):
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,
)
@@ -1027,7 +1027,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
if not self.emulate:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,