[gRPC] Fix scheduler startup broken by context parallel refactor (#18933)

Co-authored-by: Chang Su <chang.s.su@oracle.com>
This commit is contained in:
Simo Lin
2026-02-17 08:52:11 -08:00
committed by GitHub
parent 504b2c58cf
commit bf08d3f43c

View File

@@ -112,8 +112,26 @@ def launch_scheduler_process_only(
+ (tp_rank % tp_size_per_node) * server_args.gpu_id_step
)
# Calculate MoE expert parallel rank
moe_ep_rank = tp_rank // (server_args.tp_size // server_args.ep_size)
# Calculate parallelism ranks (matching engine.py logic)
attn_dp_size = (
server_args.dp_size if server_args.enable_dp_attention else 1
)
attn_tp_size = (
server_args.tp_size // attn_dp_size // server_args.attn_cp_size
)
attn_cp_rank = (tp_rank // attn_tp_size) % server_args.attn_cp_size
moe_dp_rank = tp_rank // (
server_args.tp_size // server_args.moe_dp_size
)
moe_ep_rank = (
tp_rank
% (server_args.tp_size // server_args.moe_dp_size)
// (
server_args.tp_size
// server_args.moe_dp_size
// server_args.ep_size
)
)
# Create scheduler process
proc = mp.Process(
@@ -123,6 +141,8 @@ def launch_scheduler_process_only(
port_args,
gpu_id,
tp_rank,
attn_cp_rank,
moe_dp_rank,
moe_ep_rank,
pp_rank,
None, # dp_rank