[Feature] SPMD for SGLang + Verl (#3852)
This commit is contained in:
@@ -121,7 +121,7 @@ class DataParallelController:
|
||||
args=(server_args, tmp_port_args, base_gpu_id, dp_rank),
|
||||
)
|
||||
threads.append(thread)
|
||||
base_gpu_id += server_args.tp_size
|
||||
base_gpu_id += server_args.tp_size * server_args.gpu_id_step
|
||||
|
||||
# Free all sockets before starting the threads to launch TP workers
|
||||
for sock in sockets:
|
||||
@@ -177,7 +177,11 @@ class DataParallelController:
|
||||
rank_port_args.nccl_port = port_args.nccl_port
|
||||
|
||||
reader, writer = mp.Pipe(duplex=False)
|
||||
gpu_id = server_args.base_gpu_id + base_gpu_id + tp_rank % tp_size_per_node
|
||||
gpu_id = (
|
||||
server_args.base_gpu_id
|
||||
+ base_gpu_id
|
||||
+ (tp_rank % tp_size_per_node) * server_args.gpu_id_step
|
||||
)
|
||||
proc = mp.Process(
|
||||
target=run_scheduler_process,
|
||||
args=(server_args, rank_port_args, gpu_id, tp_rank, dp_rank, writer),
|
||||
|
||||
Reference in New Issue
Block a user