[MoE Refactor] EPLB refactoring for FusedMoE (#41055)
Signed-off-by: Bill Nell <bnell@redhat.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
@@ -10,6 +10,7 @@ import torch
|
||||
|
||||
from tests.kernels.moe.utils import make_test_quant_config
|
||||
from vllm.config import VllmConfig, set_current_vllm_config
|
||||
from vllm.distributed.eplb.eplb_state import EplbLayerState
|
||||
from vllm.distributed.eplb.rebalance_execute import rearrange_expert_weights_inplace
|
||||
from vllm.distributed.parallel_state import (
|
||||
ensure_model_parallel_initialized,
|
||||
@@ -201,7 +202,7 @@ def _test_eplb_fml(env, world_size: int, test_config: TestConfig):
|
||||
dtype=torch.int32,
|
||||
device=device,
|
||||
)
|
||||
fml.enable_eplb = True
|
||||
fml.eplb_state = EplbLayerState()
|
||||
fml.set_eplb_state(
|
||||
lidx,
|
||||
torch.zeros(
|
||||
|
||||
@@ -1004,7 +1004,6 @@ def make_fake_moe_layer(
|
||||
activation: str = "silu",
|
||||
indices_type: torch.dtype | None = None,
|
||||
expert_map: torch.Tensor | None = None,
|
||||
enable_eplb: bool = False,
|
||||
expert_load_view: torch.Tensor | None = None,
|
||||
logical_to_physical_map: torch.Tensor | None = None,
|
||||
logical_replica_count: torch.Tensor | None = None,
|
||||
@@ -1022,7 +1021,6 @@ def make_fake_moe_layer(
|
||||
router = create_fused_moe_router(
|
||||
top_k=top_k,
|
||||
global_num_experts=global_num_experts,
|
||||
# eplb_state=None, # TODO
|
||||
renormalize=renormalize,
|
||||
use_grouped_topk=use_grouped_topk,
|
||||
num_expert_group=num_expert_group,
|
||||
@@ -1032,7 +1030,6 @@ def make_fake_moe_layer(
|
||||
routed_scaling_factor=routed_scaling_factor,
|
||||
e_score_correction_bias=e_score_correction_bias,
|
||||
num_fused_shared_experts=0, # TODO
|
||||
enable_eplb=enable_eplb,
|
||||
# TODO(bnell): once we can construct the MK at init time, we
|
||||
# can make this a value.
|
||||
indices_type_getter=lambda: indices_type,
|
||||
|
||||
@@ -36,9 +36,11 @@ TOP_KS = [2, 4, 6]
|
||||
NUM_EXPERTS = [8, 16, 64]
|
||||
|
||||
|
||||
def setup_eplb_state(enable_eplb: bool, global_num_experts: int) -> EplbLayerState:
|
||||
def setup_eplb_state(
|
||||
enable_eplb: bool, global_num_experts: int
|
||||
) -> EplbLayerState | None:
|
||||
if not enable_eplb:
|
||||
return EplbLayerState()
|
||||
return None
|
||||
|
||||
# Initialize EPLB state with proper tensors for testing
|
||||
# For testing purposes, we use a simple 1:1 mapping (no redundant experts)
|
||||
@@ -349,7 +351,6 @@ def test_fused_topk(
|
||||
top_k=top_k,
|
||||
global_num_experts=global_num_experts,
|
||||
renormalize=renormalize,
|
||||
enable_eplb=enable_eplb,
|
||||
eplb_state=eplb_state,
|
||||
)
|
||||
|
||||
@@ -400,7 +401,6 @@ def test_fused_topk_bias(
|
||||
top_k=top_k,
|
||||
global_num_experts=global_num_experts,
|
||||
renormalize=renormalize,
|
||||
enable_eplb=enable_eplb,
|
||||
eplb_state=eplb_state,
|
||||
)
|
||||
|
||||
@@ -469,7 +469,6 @@ def test_grouped_topk(
|
||||
top_k=top_k,
|
||||
global_num_experts=global_num_experts,
|
||||
renormalize=renormalize,
|
||||
enable_eplb=enable_eplb,
|
||||
eplb_state=eplb_state,
|
||||
)
|
||||
|
||||
@@ -540,7 +539,6 @@ def test_custom(
|
||||
global_num_experts=global_num_experts,
|
||||
custom_routing_function=custom_routing_function,
|
||||
renormalize=renormalize,
|
||||
enable_eplb=enable_eplb,
|
||||
eplb_state=eplb_state,
|
||||
)
|
||||
|
||||
@@ -580,7 +578,6 @@ def test_custom(
|
||||
# router = create_fused_moe_router(
|
||||
# top_k=top_k,
|
||||
# global_num_experts=global_num_experts,
|
||||
# enable_eplb=enable_eplb,
|
||||
# eplb_state=eplb_state,
|
||||
# )
|
||||
|
||||
|
||||
Reference in New Issue
Block a user