[Perf] Fix moe reduce_scatter perf regression by removing additional comm, 5% E2E throughput gain back. (#48763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-07-25 16:36:19 +00:00
committed by GitHub
parent 2e0da24150
commit 7fe6d3c76b
2 changed files with 13 additions and 37 deletions
+3 -25
View File
@@ -44,25 +44,6 @@ from .utils import (
)
def _restore_full_token_layout_if_needed(
hidden_states: torch.Tensor,
residual: torch.Tensor,
num_tokens: int,
is_sequence_parallel: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Restore full token rows for the MTP proposer after SP MoE layers."""
if not is_sequence_parallel and hidden_states.shape[0] == num_tokens:
return hidden_states, residual
combined_states = torch.cat([hidden_states, residual], dim=-1)
combined_states = tensor_model_parallel_all_gather(combined_states, 0)
combined_states = combined_states[:num_tokens]
hidden_states, residual = combined_states.split(
[hidden_states.shape[-1], residual.shape[-1]], dim=-1
)
return hidden_states, residual
class SharedHead(nn.Module):
def __init__(
self,
@@ -142,13 +123,10 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
hidden_states=hidden_states,
residual=None,
)
hidden_states, residual = _restore_full_token_layout_if_needed(
hidden_states,
residual,
positions.shape[0],
is_sequence_parallel=self.mtp_block.use_sequence_parallel_moe,
)
hidden_states = residual + hidden_states # pre-final-norm (logits hidden)
if self.mtp_block.use_sequence_parallel_moe:
hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
hidden_states = hidden_states[: positions.shape[0]]
# Recycle the post-final-norm hidden into the next draft step.
# compute_logits applies shared_head (== final norm) to the pre-norm
# element, so logits and the recycle each get exactly one final-norm.
+10 -12
View File
@@ -8,7 +8,10 @@ import torch.nn as nn
from vllm._aiter_ops import rocm_aiter_ops
from vllm.config import VllmConfig
from vllm.distributed import tensor_model_parallel_all_reduce
from vllm.distributed import (
tensor_model_parallel_all_gather,
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.fused_moe import (
fused_moe_make_expert_params_mapping,
)
@@ -24,10 +27,7 @@ from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.deepseek_mtp import (
SharedHead,
_restore_full_token_layout_if_needed,
)
from vllm.model_executor.models.deepseek_mtp import SharedHead
from vllm.model_executor.models.deepseek_v2 import (
DeepseekV2MixtureOfExperts,
DeepseekV2MoE,
@@ -95,13 +95,8 @@ class DeepseekV32MultiTokenPredictorLayer(nn.Module):
hidden_states, residual = self.mtp_block(
positions=positions, hidden_states=hidden_states, residual=None
)
hidden_states, residual = _restore_full_token_layout_if_needed(
hidden_states,
residual,
positions.shape[0],
is_sequence_parallel=self.mtp_block.use_sequence_parallel_moe,
)
if not self.mtp_block.use_sequence_parallel_moe:
is_sequence_parallel = self.mtp_block.use_sequence_parallel_moe
if not is_sequence_parallel:
# Without sequence parallelism, the MoE output is left un-reduced.
hidden_states = tensor_model_parallel_all_reduce(hidden_states)
# Recycle the POST-final-norm hidden into the next draft step. The
@@ -115,6 +110,9 @@ class DeepseekV32MultiTokenPredictorLayer(nn.Module):
# the legacy proposer (model_returns_tuple is True for the
# DeepSeekMTPModel architecture).
hidden_states, _ = self.shared_head.norm(hidden_states, residual)
if is_sequence_parallel:
hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
hidden_states = hidden_states[: positions.shape[0]]
return hidden_states, hidden_states