Skip to content

update dlblas with deepep #4126

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
wants to merge 4 commits into
base: main
Choose a base branch
from
+23 −0
Open

update dlblas with deepep #4126

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
febfa9e
update_dlblas_with_deepep
SHshenhao Nov 14, 2025
c8ce820
update_dlblasanddeepep
SHshenhao Nov 18, 2025
6cc9e6f
Merge remote-tracking branch 'origin/main' into sh/update_dlblas
SHshenhao Nov 18, 2025
6a7c734
update_dlblas_deepep
SHshenhao Nov 18, 2025
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
16 changes: 16 additions & 0 deletions lmdeploy/pytorch/backends/cuda/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -474,6 +474,14 @@ def __init__(self,
except ImportError:
self.use_deep_gemm = False
logger.warning('For higher performance, please install DeepGEMM https://github.com/deepseek-ai/DeepGEMM')
try:
from dlblas.layers.moe.token_dispatcher import DeepEPBuffer, DeepEPMode, use_deepep
self.use_deepep = use_deepep
self.deepep_buffer = DeepEPBuffer
self.deepep_mode = DeepEPMode
except ImportError:
self.use_deepep = False
logger.warning('For higher performance, please install DeepEP https://github.com/deepseek-ai/DeepEP')

# pre-allocate buffer
self.fusedmoe_build(True)
Expand Down Expand Up @@ -592,6 +600,14 @@ def _patched_fusedmoe_forward(*args, **kwargs):

return deepep_moe

def update_dispatch_mode(self):
if self.use_deepep:
deepep_mode_type = self.deepep_mode.NORMAL
step_ctx = get_step_ctx_manager().current_context()
if step_ctx.is_decoding:
deepep_mode_type = self.deepep_mode.LOW_LATENCY
self.deepep_buffer.set_deepep_mode(deepep_mode_type)


class TritonFusedMoEBlockedF8Builder(FusedMoEBlockedF8Builder):
"""Triton fused moe blocked f8 builder."""
Expand Down
7 changes: 7 additions & 0 deletions lmdeploy/pytorch/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1158,6 +1158,11 @@ def get_input_embeddings(self):
"""Get input embeddings."""
return self.model.get_input_embeddings()

def _update_dispatch_mode(self):
if isinstance(self.model.layers[0].mlp, DeepseekV2MoE):
if hasattr(self.model.layers[0].mlp.experts.impl, 'update_dispatch_mode'):
self.model.layers[0].mlp.experts.impl.update_dispatch_mode()

def prepare_inputs_for_generation(
self,
past_key_values: List[List[torch.Tensor]],
Expand All @@ -1169,6 +1174,8 @@ def prepare_inputs_for_generation(
position_ids = context.position_ids
attn_metadata = context.attn_metadata

self._update_dispatch_mode()

return dict(
input_ids=input_ids,
position_ids=position_ids,
Expand Down