Skip to content

[TRTLLM-12478][fix] Synchronize prepare inputs H2D copy #14128

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lfr-0531 merged 2 commits into from
May 15, 2026
Merged

[TRTLLM-12478][fix] Synchronize prepare inputs H2D copy #14128

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1b48ac0
[None][fix] Synchronize prepare inputs H2D copy
jiaganc May 14, 2026
58f6398
[TRTLLM-12478][fix] Synchronize input copy before scheduling
jiaganc May 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions tensorrt_llm/_torch/pyexecutor/model_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -561,6 +561,8 @@ def __init__(

self.kv_cache_dtype_byte_size = self.get_kv_cache_dtype_byte_size()

self._prepare_inputs_event: Optional[torch.cuda.Event] = None

def register_forward_pass_callable(self, callable: Callable):
self.forward_pass_callable = callable

Expand Down Expand Up @@ -4003,6 +4005,8 @@ def forward(self,
new_tensors_device, cache_indirection_buffer,
num_accepted_tokens_device, req_id_to_old_request,
resource_manager, can_run_graph)
self._prepare_inputs_event = torch.cuda.Event()
self._prepare_inputs_event.record()

with with_shared_pool(self.cuda_graph_runner.get_graph_pool()):
if not can_run_graph:
Expand Down Expand Up @@ -4252,3 +4256,11 @@ def _execute_logit_post_processors(self,
lp(request.py_request_id, logits_row, token_ids, None, None)

logits_tensor[idx] = logits_row.view(-1)

def wait_for_input_copy(self):
"""
Wait for input preparation and H2D copy of previous iteration before modifying host input,
otherwise the input of previous iteration will be overwritten.
"""
if self._prepare_inputs_event is not None:
self._prepare_inputs_event.synchronize()
4 changes: 4 additions & 0 deletions tensorrt_llm/_torch/pyexecutor/py_executor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2718,6 +2718,10 @@ def _executor_loop_overlap(self):

self._handle_disagg_cache_errors_synced()

# Need to wait for the copy of previous iteration before modifying any host memory copied to GPU,
# and for scheduler V2, it will modify the host page table, so wait before scheduling.
# This wait is also needed for legacy scheduler, but it can be pushed later, e.g. before model_engine._prepare_inputs().
self.model_engine.wait_for_input_copy()
scheduled_batch, iter_stats = self._prepare_and_schedule_batch()
self._handle_control_request()

Expand Down
Loading