EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

reset default block_size from 16 to 128 (#84)

Browse Source 
### What this PR does / why we need it?

Changed default block_size in platform.py from 16 to 128, as Ascend
Devices have a better affinity for block size 128.

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
CI passed

Signed-off-by: hzji210@gmail.com <hzji210@gmail.com>
This commit is contained in:
Huazhong Ji
2025-02-18 14:19:38 +08:00
committed by GitHub
parent 7606977739
commit 8ea8523744
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/platform.py
View File

@@ -96,7 +96,7 @@ class NPUPlatform(Platform):
parallel_config.worker_cls = "vllm_ascend.worker.NPUWorker" parallel_config.worker_cls = "vllm_ascend.worker.NPUWorker"
cache_config = vllm_config.cache_config cache_config = vllm_config.cache_config
if cache_config and cache_config.block_size is None: if cache_config and cache_config.block_size is None:
cache_config.block_size = 16 cache_config.block_size = 128
@classmethod @classmethod
def get_attn_backend_cls(cls, selected_backend, head_size, dtype, def get_attn_backend_cls(cls, selected_backend, head_size, dtype,