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

[v1][attention] Support Hybrid Allocator + FlashInfer (#21412)

Browse Source 
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
Chen Zhang
2025-07-29 18:45:29 -07:00
committed by GitHub
parent 0e36abf993
commit 555e7225bc
16 changed files with 85 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/v1/attention/backends/mamba_attn.py
View File

@@ -87,8 +87,8 @@ class Mamba2AttentionMetadata:
class Mamba2AttentionMetadataBuilder(
AttentionMetadataBuilder[Mamba2AttentionMetadata]):
def __init__(self, kv_cache_spec: AttentionSpec, vllm_config: VllmConfig,
device: torch.device):
def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
vllm_config: VllmConfig, device: torch.device):
assert isinstance(kv_cache_spec, MambaSpec)
self.kv_cache_spec = kv_cache_spec
self.chunk_size = vllm_config.model_config.get_mamba_chunk_size()