[BugFix] Fix use of per-request seed with pipeline parallel (#6698)

This commit is contained in:
Nick Hill
2024-07-30 10:40:08 -07:00
committed by GitHub
parent f058403683
commit 5cf9254a9c
21 changed files with 222 additions and 137 deletions
+1 -2
View File
@@ -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,