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

[Bugfix][Minor] Fix potential NameError in mamba backend selector and misc typos (#35886)

Browse Source 
Signed-off-by: Li <chuali@amd.com>
This commit is contained in:
Chuan (Richard) Li
2026-03-26 08:59:24 -07:00
committed by GitHub
parent e054f152fa
commit cb2263218e
4 changed files with 8 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/v1/attention/backends/flex_attention.py
View File

@@ -749,7 +749,10 @@ class FlexAttentionMetadataBuilder(AttentionMetadataBuilder[FlexAttentionMetadat
prefix_kv_lens = None
suffix_kv_lens = None
if use_cascade:
raise NotImplementedError("Not yet my friend")
raise NotImplementedError(
"Cascade prefix attention is not yet implemented "
"for FlexAttention backend"
)
block_size = self.kv_cache_spec.block_size
max_possible_seq_len = self.model_config.max_model_len

2
vllm/v1/attention/backends/utils.py
View File

@@ -253,7 +253,7 @@ def make_local_attention_virtual_batches(
# seqlens_q_local = [2, 2, 1, 4, 4, 1, 4, 1]
#
# First Get batched arange. (E.g., [2, 4, 2] -> [0, 1, 0, 1, 2, 3, 0, 1])
# (TODO: max a utility to share this code with _prepare_inputs)
# (TODO: make a utility to share this code with _prepare_inputs)
# arange step 1. [2, 4, 2] -> [2, 6, 8]
cu_num_blocks = np.cumsum(local_blocks)
virtual_batches = cu_num_blocks[-1]

4
vllm/v1/attention/selector.py
View File

@@ -149,8 +149,8 @@ def _cached_get_mamba_attn_backend(
selected_backend = MambaAttentionBackendEnum[backend_name]
except KeyError as e:
raise ValueError(
f"Invalid mamba attention backend type: '{backend_name}'. Valid "
f"backends are: {list(MambaAttentionBackendEnum.__members__.keys())}"
f"Invalid mamba attention backend type: '{mamba_type}'. Valid "
f"types are: {list(MAMBA_TYPE_TO_BACKEND_MAP.keys())}"
) from e
mamba_attn_backend = selected_backend.get_class()