Adjust overlap event loop (#11507)
This commit is contained in:
@@ -168,6 +168,7 @@ class TpModelWorker:
|
||||
)[0]
|
||||
set_random_seed(self.random_seed)
|
||||
|
||||
self.enable_overlap = not server_args.disable_overlap_schedule
|
||||
self.hicache_layer_transfer_counter = None
|
||||
|
||||
def register_hicache_layer_transfer_counter(self, counter: LayerDoneCounter):
|
||||
@@ -266,9 +267,18 @@ class TpModelWorker:
|
||||
# Skip sampling and return logits for target forward
|
||||
return batch_result
|
||||
|
||||
if model_worker_batch.delay_sample_launch:
|
||||
batch_result.delay_sample_launch = True
|
||||
batch_result.forward_batch = forward_batch
|
||||
if (
|
||||
self.enable_overlap
|
||||
and model_worker_batch.sampling_info.grammars is not None
|
||||
):
|
||||
|
||||
def sample_batch_func():
|
||||
batch_result.next_token_ids = self.model_runner.sample(
|
||||
logits_output, forward_batch
|
||||
)
|
||||
return batch_result
|
||||
|
||||
batch_result.delay_sample_func = sample_batch_func
|
||||
return batch_result
|
||||
|
||||
if model_worker_batch.is_prefill_only:
|
||||
|
||||
Reference in New Issue
Block a user