[DSV4.1][PD][5/N] Decode node support DP attention (#40177)

This commit is contained in:
Siyuan Chen
2026-10-08 14:42:16 -07:00
committed by GitHub
parent d20bdc1443
commit b0e06359ac
2 changed files with 14 additions and 7 deletions
@@ -17,7 +17,7 @@ from sglang.srt.arg_groups.overrides import (
)
from sglang.srt.environ import envs
from sglang.srt.model_executor.cuda_graph_config import Backend, Phase, with_phase
from sglang.srt.runtime_context import attn_dp_enabled_of, get_platform, num_dp_ranks_of
from sglang.srt.runtime_context import attn_dp_enabled_of, get_platform
from sglang.srt.utils.common import is_gfx95_supported, is_npu
if TYPE_CHECKING:
@@ -292,15 +292,14 @@ def validate_deepseek_v41_features(server_args: ServerArgs) -> None:
if (
read_ragged_verify_mode() is not RaggedVerifyMode.STATIC
or cfg.disaggregation_transfer_backend != "mooncake"
or num_dp_ranks_of(cfg) != 1
or attn_dp_enabled_of(cfg)
or cfg.enable_prefill_cp
or cfg.attn_cp_size != 1
or cfg.dcp_size != 1
):
raise ValueError(
"DeepSeek-V4.1 DSpark PD requires static verify, Mooncake, "
"DP=1 and CP=1. Both servers must enable DSpark with the same "
"block size and TP size."
"and CP=1 on both servers. DP attention is supported when "
"both servers use the same block size and target/draft KV layout."
)
prefill_graph = cfg.cuda_graph_config.prefill
@@ -1199,9 +1199,17 @@ class CommonKVManager(BaseKVManager):
"enable DSpark with the same block size and target/draft KV "
"layout. Upgrade both servers together."
)
if info.attn_tp_size != self.attn_tp_size:
if info.attn_cp_size != 1 or self.attn_cp_size != 1:
raise RuntimeError(
"DeepSeek-V4.1 DSpark PD requires the same TP size on both servers"
"DeepSeek-V4.1 DP-only DSpark PD requires CP=1 on both servers"
)
non_cp_mla_layout = info.attn_cp_size == self.attn_cp_size == 1 and (
self.is_mla_backend or self.is_hybrid_mla_backend
)
if info.attn_tp_size != self.attn_tp_size and not non_cp_mla_layout:
raise RuntimeError(
"DeepSeek-V4.1 DSpark PD requires matching attention TP "
"unless both servers use CP=1 with an MLA KV layout"
)
if self.dcp_size > 1: