Skip to content

fix(dlinfer): build the MoE expert-id map on the active device #5039

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
li-lizhe wants to merge 1 commit into
base: main
Choose a base branch
from
+15 −5
Open

fix(dlinfer): build the MoE expert-id map on the active device #5039

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 15 additions & 5 deletions lmdeploy/pytorch/backends/dlinfer/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,12 +52,22 @@ def __init__(self,
self.ep_group = ep_group
self.expert_ids_per_ep_rank = None
self.chunked_moe_layout = None
if self.ep_size > 1:

def _get_expert_ids_per_ep_rank(self, device: torch.device):
"""Build the expert-id map of the current rank on the given device.

The map is created from the input device instead of a CUDA device index, so the non-CUDA accelerators served by
this backend (Ascend NPU, MACA, Cambricon) do not have to be CUDA at construction time.
"""
if self.ep_size <= 1:
return None
if self.expert_ids_per_ep_rank is None or self.expert_ids_per_ep_rank.device != device:
self.expert_ids_per_ep_rank = torch.tensor(
[i % (self.num_experts // self.ep_size) for i in range(num_experts)],
[i % (self.num_experts // self.ep_size) for i in range(self.num_experts)],
dtype=torch.int32,
device=torch.cuda.current_device(),
device=device,
)
return self.expert_ids_per_ep_rank

def ep_expert_list(self, world_size: int, rank: int):
"""Experts list of current rank."""
Expand Down Expand Up @@ -110,7 +120,7 @@ def forward(self,
step_context = get_step_ctx_manager().current_context()
moe_metadata = getattr(step_context, 'moe_metadata', None)
if moe_metadata is not None:
moe_metadata.expert_ids_per_ep_rank = self.expert_ids_per_ep_rank
moe_metadata.expert_ids_per_ep_rank = self._get_expert_ids_per_ep_rank(hidden_states.device)
return fused_moe(hidden_states, gate_up_weights, down_weights, topk_weights, topk_ids, self.top_k,
self.renormalize, moe_metadata, self.chunked_moe_layout)

Expand Down Expand Up @@ -169,7 +179,7 @@ def forward(self,
moe_metadata = getattr(step_context, 'moe_metadata', None)
if moe_metadata is None:
raise RuntimeError('Dlinfer W8A8 MoE requires moe_metadata in the current step context.')
moe_metadata.expert_ids_per_ep_rank = self.expert_ids_per_ep_rank
moe_metadata.expert_ids_per_ep_rank = self._get_expert_ids_per_ep_rank(hidden_states.device)
return fused_moe_w8a8(hidden_states, gate_up_weights, gate_up_scale, down_weights, down_scale, topk_weights,
topk_ids, self.top_k, self.renormalize, moe_metadata)

Expand Down
Loading