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

6
vllm/model_executor/models/mamba2.py
View File

@@ -15,6 +15,8 @@ from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncCalculator,
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
)
@@ -228,6 +230,10 @@ class Mamba2ForCausalLM(
conv_kernel=hf_config.conv_kernel,
)
@classmethod
def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]:
return MambaStateCopyFuncCalculator.mamba2_state_copy_func()
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
config = vllm_config.model_config.hf_config