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

Remove V0 attention backends (#25351)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-09-21 16:03:28 -07:00
committed by GitHub
parent af7dfb0d1a
commit bc6e542d9f
28 changed files with 143 additions and 7376 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/kernels/attention/test_attention.py
View File

@@ -18,7 +18,7 @@ if not current_platform.is_rocm():
from xformers import ops as xops
from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
from vllm.attention.backends.xformers import _make_alibi_bias
from tests.kernels.utils import make_alibi_bias
FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
# This will change depending on the compute capability.
@@ -429,8 +429,8 @@ def test_multi_query_kv_attention(
alibi_bias = None
if use_alibi:
alibi_slopes = torch.randn(num_query_heads, dtype=torch.float)
attn_bias = _make_alibi_bias(alibi_slopes, num_kv_heads, dtype,
seq_lens)
attn_bias = make_alibi_bias(alibi_slopes, num_kv_heads, dtype,
seq_lens)
output = torch.empty_like(query)
start = 0
# Dynamic sequence length not supported with custom attn_bias.