[MoE Refactor][5/N] Isolate zero expert to LongCatFlash (#28891)

Signed-off-by: baonudesifeizhai <85092850+baonudesifeizhai@users.noreply.github.com>
Signed-off-by: Dongjie Zou <85092850+baonudesifeizhai@users.noreply.github.com>
Signed-off-by: baonudesifeizhai <baonudesifeizhai@gmail.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
This commit is contained in:
baonudesifeizhai
2025-12-20 18:22:04 +00:00
committed by GitHub
co-authored by Robert Shaw Robert Shaw
parent 560ae9638c
commit 54c8924384
19 changed files with 264 additions and 109 deletions
@@ -764,7 +764,7 @@ class AWQMarlinMoEMethod(FusedMoEMethodBase):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert layer.activation == "silu", "Only SiLU activation is supported."
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -500,7 +500,7 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -574,7 +574,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
e_score_correction_bias=layer.e_score_correction_bias,
)
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -1166,7 +1166,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -1403,7 +1403,7 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -1765,7 +1765,7 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
f"{layer.activation} not supported for Marlin MoE."
)
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -1991,7 +1991,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -2607,7 +2607,7 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
"EPLB not supported for `CompressedTensorsW4A8Fp8MoEMethod` yet."
)
assert self.moe_quant_config is not None
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -142,7 +142,7 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
+2 -10
View File
@@ -1292,13 +1292,11 @@ class Fp8MoEMethod(FusedMoEMethodBase):
apply_router_weight_on_input=layer.apply_router_weight_on_input,
)
select_result = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
topk_weights, topk_ids, zero_expert_result = select_result
if self.rocm_aiter_moe_enabled:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import ( # noqa: E501
rocm_aiter_fused_experts,
@@ -1353,13 +1351,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
apply_router_weight_on_input=layer.apply_router_weight_on_input,
)
if layer.zero_expert_num != 0 and layer.zero_expert_type is not None:
assert not isinstance(result, tuple), (
"Shared + zero experts are mutually exclusive not yet supported"
)
return result, zero_expert_result
else:
return result
return result
class Fp8OnlineMoEMethod(Fp8MoEMethod):
@@ -639,7 +639,7 @@ class GGUFMoEMethod(FusedMoEMethodBase):
"fused GGUF MoE method."
)
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -900,7 +900,7 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert layer.activation == "silu", "Only SiLU activation is supported."
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -796,7 +796,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
)
# Expert selection
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -1599,7 +1599,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
x_routing, _ = x
else:
x_routing = x
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x_routing,
router_logits=router_logits,
)
@@ -370,7 +370,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
from vllm.model_executor.layers.fused_moe import fused_experts
assert layer.activation == "silu", "Only SiLU activation is supported."
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -896,7 +896,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
raise NotImplementedError("EPLB is not supported for mxfp4")
if self.mxfp4_backend == Mxfp4Backend.MARLIN:
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -990,7 +990,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
):
from vllm.utils.flashinfer import flashinfer_cutlass_fused_moe
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -338,7 +338,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -530,7 +530,7 @@ class QuarkW4A8Fp8MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -738,7 +738,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
@@ -359,7 +359,7 @@ class RTNMoEMethod(FusedMoEMethodBase):
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)