[BugFix] Fix use of per-request seed with pipeline parallel (#6698)
This commit is contained in:
@@ -7,10 +7,9 @@ from vllm.sequence import ExecuteModelRequest, SamplerOutput
|
||||
from vllm.spec_decode.interfaces import SpeculativeProposals
|
||||
from vllm.spec_decode.proposer_worker_base import NonLLMProposerWorkerBase
|
||||
from vllm.spec_decode.top1_proposer import Top1Proposer
|
||||
from vllm.worker.worker_base import LoraNotSupportedWorkerBase
|
||||
|
||||
|
||||
class NGramWorker(NonLLMProposerWorkerBase, LoraNotSupportedWorkerBase):
|
||||
class NGramWorker(NonLLMProposerWorkerBase):
|
||||
"""NGramWorker provides a light drafter without need for model.
|
||||
|
||||
Current NGramWorker only implements prompt lookup decoding,
|
||||
|
||||
Reference in New Issue
Block a user