[Feature] Support sequence parallel without the need for DP, 1.9%~5.0% E2E Throughput Improvement (#47070)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Co-authored-by: Canlin Guo <canlinguosdu@gmail.com>
This commit is contained in:
Wentao Ye
2026-07-05 05:41:30 -07:00
committed by GitHub
co-authored by Canlin Guo
parent fa4321de3d
commit b6cc46ec3b
7 changed files with 33 additions and 22 deletions
+15 -3
View File
@@ -83,7 +83,10 @@ class DPMetadata:
num_tokens_across_dp_cpu: torch.Tensor,
) -> "DPMetadata":
assert num_tokens_across_dp_cpu is not None
assert parallel_config.data_parallel_size > 1
assert (
parallel_config.data_parallel_size > 1
or parallel_config.use_sequence_parallel_moe
)
assert parallel_config.is_moe_model is not False
dp_rank = parallel_config.data_parallel_rank
batchsize = num_tokens
@@ -277,14 +280,20 @@ def set_forward_context(
dp_metadata: DPMetadata | None = None
if (
vllm_config.parallel_config.data_parallel_size > 1
(
vllm_config.parallel_config.data_parallel_size > 1
or vllm_config.parallel_config.use_sequence_parallel_moe
)
and vllm_config.parallel_config.is_moe_model is not False
and (attn_metadata is not None or num_tokens is not None)
):
# If num_tokens_across_dp hasn't already been initialized, then
# initialize it here. Both DP padding and Microbatching will be
# disabled.
if num_tokens_across_dp is None:
if (
num_tokens_across_dp is None
and vllm_config.parallel_config.data_parallel_size > 1
):
assert ubatch_slices is None
assert num_tokens is not None
_, num_tokens_across_dp, _ = coordinate_batch_across_dp(
@@ -293,6 +302,9 @@ def set_forward_context(
allow_microbatching=False,
)
assert num_tokens_across_dp is not None
elif num_tokens_across_dp is None:
assert num_tokens is not None
num_tokens_across_dp = torch.tensor([num_tokens], dtype=torch.int32)
dp_metadata = DPMetadata.make(
vllm_config.parallel_config, num_tokens or 0, num_tokens_across_dp
)