Add flag that enables NCCL mlp sync batch for overlap scheduler (#17288)
This commit is contained in:
@@ -246,6 +246,7 @@ class Envs:
|
||||
SGLANG_PREFILL_DELAYER_TOKEN_USAGE_LOW_WATERMARK = EnvFloat(None)
|
||||
SGLANG_DATA_PARALLEL_BUDGET_INTERVAL = EnvInt(1)
|
||||
SGLANG_QUEUED_TIMEOUT_MS = EnvInt(-1)
|
||||
SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH = EnvBool(False)
|
||||
|
||||
# Test: pd-disaggregation
|
||||
SGLANG_TEST_PD_DISAGG_BACKEND = EnvStr("mooncake")
|
||||
|
||||
@@ -170,7 +170,10 @@ def prepare_mlp_sync_batch_raw(
|
||||
local_batch.is_extend_in_batch = is_extend_in_batch
|
||||
|
||||
tbo_preparer = TboDPAttentionPreparer()
|
||||
if len(offload_tags) == 0 and disable_overlap_schedule:
|
||||
if len(offload_tags) == 0 and (
|
||||
disable_overlap_schedule
|
||||
or envs.SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH.get()
|
||||
):
|
||||
group = tp_group.device_group
|
||||
device = tp_group.device
|
||||
else:
|
||||
|
||||
Reference in New Issue
Block a user