xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-25 18:28:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix neuron performance issue (#13589)

Browse Source
This commit is contained in:
ajayvohra2005 2025-02-20 13:59:36 -05:00 committed by GitHub
parent d3ea50113c
commit 6a417b8600
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/worker/neuron_worker.py
View File

@ -76,7 +76,7 @@ class NeuronWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
# Set the number of GPU blocks to be the same as the maximum number of
# sequences that can be processed in a single batch. This is equivalent
# to schedule without PagedAttention.
num_gpu_blocks = self.scheduler_config.max_num_seqs
num_gpu_blocks = self.scheduler_config.max_num_seqs + 1
# Swap not yet supported with Neuron backend.
num_cpu_blocks = 0
@ -90,7 +90,7 @@ class NeuronWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
# Different values are not tested.
assert num_cpu_blocks == 0
assert num_gpu_blocks == self.scheduler_config.max_num_seqs
assert num_gpu_blocks == self.scheduler_config.max_num_seqs + 1
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks