minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Revert "[AMD]fix: use CUDA event for targeted draft-to-verify sync in… (#25917)

Browse Source
This commit is contained in:
Hanming Lu
2026-05-20 18:49:01 -07:00
committed by GitHub
parent cf1fd26d16
commit ddf3817924
2 changed files with 0 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/speculative/eagle_worker_v2.py
-13
View File
@@ -807,13 +807,6 @@ class EAGLEWorkerV2(BaseSpecWorker):
):
verify_input: EagleVerifyInput = self.draft_worker.draft(batch)
assert verify_input.is_verify_input()
# Record a CUDA event after draft() GPU work is dispatched.
# This event will be waited on by plan_stream in verify()
# to ensure draft CUDA graph kernels finish before plan_stream
# begins metadata preparation.
if self.plan_stream:
self._draft_done_event = torch.get_device_module(self.device).Event()
self._draft_done_event.record()
batch.spec_info = verify_input
batch_output = self.verify(batch)
with (
@@ -979,12 +972,6 @@ class EAGLEWorkerV2(BaseSpecWorker):
# Batch 1: Target verify
# Prepare for target verify in a separate stream
with self.plan_stream_ctx:
# Wait for the draft CUDA graph to finish before plan_stream
# begins its work. Using an event is more targeted than
# wait_stream(main_stream) — it only waits for draft GPU
# work, not all queued main_stream operations.
if self.plan_stream and hasattr(self, "_draft_done_event"):
self.plan_stream.wait_event(self._draft_done_event)
verify_forward_batch, can_run_cuda_graph = (
verify_input.prepare_for_v2_verify(
self.req_to_token_pool,
python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py
-8
View File
@@ -704,10 +704,6 @@ class MultiLayerEagleWorkerV2(BaseSpecWorker):
)
verify_input: EagleVerifyInput = self.draft_worker.draft(batch)
assert verify_input.is_verify_input()
# Record a CUDA event after draft() GPU work is dispatched.
if self.plan_stream:
self._draft_done_event = torch.get_device_module(self.device).Event()
self._draft_done_event.record()
batch.spec_info = verify_input
batch_output = self.verify(batch)
self.draft_worker._draft_extend_for_decode(batch, batch_output)
@@ -726,10 +722,6 @@ class MultiLayerEagleWorkerV2(BaseSpecWorker):
# Batch 1: Target verify
# Prepare for target verify in a separate stream
with self.plan_stream_ctx:
# Wait for the draft CUDA graph to finish before plan_stream
# begins its work.
if self.plan_stream and hasattr(self, "_draft_done_event"):
self.plan_stream.wait_event(self._draft_done_event)
verify_forward_batch, can_run_cuda_graph = (
verify_input.prepare_for_v2_verify(
self.req_to_token_pool,