[Feature] Hybrid EP and TP (#8590)
This commit is contained in:
@@ -719,6 +719,7 @@ def _launch_subprocesses(
|
||||
+ ((pp_rank % pp_size_per_node) * tp_size_per_node)
|
||||
+ (tp_rank % tp_size_per_node) * server_args.gpu_id_step
|
||||
)
|
||||
moe_ep_rank = tp_rank // (server_args.tp_size // server_args.ep_size)
|
||||
proc = mp.Process(
|
||||
target=run_scheduler_process,
|
||||
args=(
|
||||
@@ -726,6 +727,7 @@ def _launch_subprocesses(
|
||||
port_args,
|
||||
gpu_id,
|
||||
tp_rank,
|
||||
moe_ep_rank,
|
||||
pp_rank,
|
||||
None,
|
||||
writer,
|
||||
|
||||
Reference in New Issue
Block a user