[Kimi K3 Refactor] Add UnfinalizedMoEOutput proto following up for #53152 (#53310)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-08-24 11:21:11 -04:00
committed by GitHub
parent 4c56e62c85
commit 29c9af5211
6 changed files with 21 additions and 9 deletions
@@ -14,6 +14,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.moe_output import UnfinalizedMoEOutput
from vllm.model_executor.layers.fused_moe.oracle.nvfp4 import (
NvFp4MoeBackend,
convert_to_nvfp4_moe_kernel_format,
@@ -278,7 +279,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(
@@ -24,6 +24,7 @@ from vllm.model_executor.layers.fused_moe import (
RoutedExperts,
SharedExperts,
)
from vllm.model_executor.layers.fused_moe.moe_output import UnfinalizedMoEOutput
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
Fp8MoeBackend,
convert_to_fp8_moe_kernel_format,
@@ -1624,7 +1625,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(
@@ -10,6 +10,7 @@ from vllm.model_executor.layers.fused_moe import (
RoutedExperts,
SharedExperts,
)
from vllm.model_executor.layers.fused_moe.moe_output import UnfinalizedMoEOutput
from vllm.model_executor.model_loader.reload.layerwise import (
initialize_online_processing,
)
@@ -139,7 +140,7 @@ class OnlineMoEMethodBase(FusedMoEMethodBase):
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(
@@ -26,6 +26,7 @@ from vllm.model_executor.layers.fused_moe.config import (
mxfp4_w4a16_moe_quant_config,
ocp_mx_moe_quant_config,
)
from vllm.model_executor.layers.fused_moe.moe_output import UnfinalizedMoEOutput
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
convert_to_fp8_moe_kernel_format,
make_fp8_moe_kernel,
@@ -1444,7 +1445,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(