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

[V1][Hybrid] Mamba Prefix Caching with align mode (#30877)

Browse Source 
Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
Harry Huang
2026-01-24 01:56:48 +08:00
committed by GitHub
parent fec9da0af4
commit 5206e5e28c
42 changed files with 1774 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/config/vllm.py
View File

@@ -999,6 +999,17 @@ class VllmConfig:
# Default to enable HMA if not explicitly disabled by user or logic above.
self.scheduler_config.disable_hybrid_kv_cache_manager = False
if self.cache_config.mamba_cache_mode == "align":
if self.scheduler_config.long_prefill_token_threshold > 0:
assert (
self.scheduler_config.long_prefill_token_threshold
>= self.cache_config.block_size
)
assert not self.scheduler_config.disable_chunked_mm_input, (
"Chunked MM input is required because we need the flexibility to "
"schedule a multiple of block_size tokens even if they are in the "
"middle of a mm input"
)
if self.compilation_config.debug_dump_path:
self.compilation_config.debug_dump_path = (
self.compilation_config.debug_dump_path.absolute().expanduser()