Support dispatching logical to physical experts (#6385)
This commit is contained in:
@@ -6,6 +6,7 @@ from torch.nn import Module
|
||||
|
||||
from sglang.srt.layers.quantization.deep_gemm import _ENABLE_JIT_DEEPGEMM
|
||||
from sglang.srt.managers.expert_location import get_global_expert_location_metadata
|
||||
from sglang.srt.managers.expert_location_dispatch import ExpertLocationDispatchInfo
|
||||
from sglang.srt.managers.schedule_batch import global_server_args_dict
|
||||
|
||||
try:
|
||||
@@ -237,6 +238,9 @@ class EPMoE(torch.nn.Module):
|
||||
correction_bias=self.correction_bias,
|
||||
custom_routing_function=self.custom_routing_function,
|
||||
routed_scaling_factor=self.routed_scaling_factor,
|
||||
expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
|
||||
layer_id=self.layer_id,
|
||||
),
|
||||
)
|
||||
|
||||
reorder_topk_ids, src2dst, seg_indptr = run_moe_ep_preproess(
|
||||
|
||||
Reference in New Issue
Block a user