xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-31 21:57:43 +08:00
Code Issues Packages Projects Releases Wiki Activity

turn yields on for pplx

Browse Source 
Signed-off-by: Sage Moore <sage@neuralmagic.com>
This commit is contained in:
Sage Moore 2025-07-02 22:34:02 +00:00
parent 18f7bfb501
commit ce3ef95c11
1 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py
View File

@ -134,14 +134,14 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
do_recv=not send, do_recv=not send,
) )
# yield_and_switch_from_compute_to_comm_impl(schedule="default") yield_and_switch_from_compute_to_comm_impl(schedule="default")
dispatch(True) # Send dispatch(True) # Send
# torch.cuda.synchronize() # torch.cuda.synchronize()
# print(f"{ubatch_id} AFTER SEND SYNC", flush=True) # print(f"{ubatch_id} AFTER SEND SYNC", flush=True)
dispatch(False) # Recv dispatch(False) # Recv
# torch.cuda.synchronize() # torch.cuda.synchronize()
# print(f"{ubatch_id} AFTER RECV SYNC", flush=True) # print(f"{ubatch_id} AFTER RECV SYNC", flush=True)
# yield_and_switch_from_comm_to_compute_impl(schedule="default") yield_and_switch_from_comm_to_compute_impl(schedule="default")
# torch.cuda.synchronize() # torch.cuda.synchronize()
if expert_x_scale is not None: if expert_x_scale is not None:
expert_x_scale = expert_x_scale[:, :, 0:1] expert_x_scale = expert_x_scale[:, :, 0:1]
@ -185,11 +185,11 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
do_recv=not send, do_recv=not send,
) )
# yield_and_switch_from_compute_to_comm_impl(schedule="default") yield_and_switch_from_compute_to_comm_impl(schedule="default")
combine(True) combine(True)
# torch.cuda.synchronize() # torch.cuda.synchronize()
# print(f"{ubatch_id} AFTER COMBINE SEND SYNC", flush=True) # print(f"{ubatch_id} AFTER COMBINE SEND SYNC", flush=True)
combine(False) combine(False)
# print(f"{ubatch_id} AFTER COMBINE RECV SYNC", flush=True) # print(f"{ubatch_id} AFTER COMBINE RECV SYNC", flush=True)
# yield_and_switch_from_comm_to_compute_impl(schedule="default") yield_and_switch_from_comm_to_compute_impl(schedule="default")
# torch.cuda.synchronize() # torch.cuda.synchronize()