frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 23:03:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Remove unnecessary fallback to prefill-decode attention (#19138)

Browse Source 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
This commit is contained in:
vllmellm
2025-06-05 16:08:26 +08:00
committed by GitHub
parent da40380214
commit 18093084be
1 changed files with 1 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@ -171,10 +171,7 @@ class TritonAttentionImpl(AttentionImpl):
# Whenever making a change in this method, please benchmark the
# performance to make sure it does not introduce any overhead.
num_queries_per_kv = query.shape[1] // key.shape[1]
num_q_is_pow2 = (num_queries_per_kv & (num_queries_per_kv - 1)) == 0
use_prefill_decode_attn = (self.force_prefill_decode_attn
or not num_q_is_pow2)
use_prefill_decode_attn = self.force_prefill_decode_attn
num_actual_tokens = attn_metadata.num_actual_tokens
if use_prefill_decode_attn: