[Feature] Expert Parallelism Load Balancer (EPLB) (#18343)

Signed-off-by: Bowen Wang <abmfy@icloud.com>
This commit is contained in:
Bowen Wang
2025-06-26 15:30:21 -07:00
committed by GitHub
parent 07b8fae219
commit e9fd658a73
24 changed files with 2446 additions and 54 deletions
@@ -482,7 +482,15 @@ class AWQMoEMethod(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `AWQMoEMethod` yet.")
assert activation == "silu", "Only SiLU activation is supported."
if apply_router_weight_on_input:
@@ -331,7 +331,15 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for "
"`CompressedTensorsW8A8Fp8MoEMethod` yet.")
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
@@ -593,7 +601,15 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for "
"`CompressedTensorsW8A8Fp8MoECutlassMethod` yet.")
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
@@ -722,7 +738,16 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for "
"`CompressedTensorsW8A8Int8MoEMethod` yet.")
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids = FusedMoE.select_experts(
@@ -1012,7 +1037,16 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for "
"`CompressedTensorsWNA16MarlinMoEMethod` yet.")
assert activation == "silu", (
f"{activation} not supported for Marlin MoE.")
assert not apply_router_weight_on_input, (
@@ -1228,7 +1262,15 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError("EPLB not supported for "
"`CompressedTensorsWNA16MoEMethod` yet.")
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids = FusedMoE.select_experts(
@@ -117,7 +117,15 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ExpertsInt8MoEMethod` yet.")
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids = FusedMoE.select_experts(
@@ -825,7 +825,16 @@ class Fp8MoEMethod(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
assert expert_load_view is not None
assert logical_to_physical_map is not None
assert logical_replica_count is not None
assert isinstance(layer, FusedMoE)
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
@@ -839,6 +848,11 @@ class Fp8MoEMethod(FusedMoEMethodBase):
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
enable_eplb=enable_eplb,
expert_map=expert_map,
expert_load_view=expert_load_view,
logical_to_physical_map=logical_to_physical_map,
logical_replica_count=logical_replica_count,
)
if self.rocm_aiter_moe_enabled:
@@ -520,7 +520,15 @@ class GGUFMoEMethod(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
):
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `GGUFMoEMethod` yet.")
assert activation == "silu", "Only SiLU activation is supported."
if apply_router_weight_on_input:
raise NotImplementedError(
@@ -635,7 +635,15 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `GPTQMarlinMoEMethod` yet.")
assert activation == "silu", "Only SiLU activation is supported."
if apply_router_weight_on_input:
raise NotImplementedError(
@@ -664,7 +664,15 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
):
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ModelOptNvFp4FusedMoE` yet.")
if self.use_marlin:
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
@@ -297,7 +297,15 @@ class MoeWNA16Method(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `MoeWNA16Method` yet.")
from vllm.model_executor.layers.fused_moe import fused_experts
assert activation == "silu", "Only SiLU activation is supported."
topk_weights, topk_ids = FusedMoE.select_experts(
@@ -205,7 +205,15 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `QuarkW8A8Fp8MoEMethod` yet.")
from vllm.model_executor.layers.fused_moe import fused_experts
topk_weights, topk_ids = FusedMoE.select_experts(