biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[BugFix] Add block_size validation for mamba cache align mode (#34445)

Browse Source 
Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
This commit is contained in:
Harry Huang
2026-02-13 10:18:07 +08:00
committed by GitHub
parent d707678dfb
commit 6f019e6e0a
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/config/vllm.py
View File

@@ -1110,6 +1110,15 @@ class VllmConfig:
self.scheduler_config.disable_hybrid_kv_cache_manager = False
if self.cache_config.mamba_cache_mode == "align":
assert (
self.cache_config.block_size
<= self.scheduler_config.max_num_batched_tokens
), (
"In Mamba cache align mode, block_size "
f"({self.cache_config.block_size}) must be <= "
"max_num_batched_tokens "
f"({self.scheduler_config.max_num_batched_tokens})."
)
if self.scheduler_config.long_prefill_token_threshold > 0:
assert (
self.scheduler_config.long_prefill_token_threshold