[V1][Spec Decode] Ngram Spec Decode (#12193)
Signed-off-by: LiuXiaoxuanPKU <lilyliupku@gmail.com>
This commit is contained in:
@@ -124,9 +124,8 @@ class CudaPlatformBase(Platform):
|
||||
"vllm.worker.multi_step_worker.MultiStepWorker"
|
||||
elif vllm_config.speculative_config:
|
||||
if envs.VLLM_USE_V1:
|
||||
raise NotImplementedError(
|
||||
"Speculative decoding is not yet supported on VLLM V1."
|
||||
)
|
||||
parallel_config.worker_cls = \
|
||||
"vllm.v1.worker.gpu_worker.Worker"
|
||||
else:
|
||||
parallel_config.worker_cls = \
|
||||
"vllm.spec_decode.spec_decode_worker.create_spec_worker"
|
||||
|
||||
Reference in New Issue
Block a user