A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 12:15 UTC
[Attention][MLA] Re-enable FA4 as default MLA prefill backend (#38819)
Matthew Bonanni committed
Apr 6, 2026 at 17:51 UTC
9c81f35b1ae6c70681661e11c461ddbb7e417aff
1 file changed
+1
-1
vllm/config/attention.py
+1
-1
@@ -30,7 +30,7 @@ class AttentionConfig:
30
use_cudnn_prefill: bool = False
31
"""Whether to use cudnn prefill."""
32
33
- use_trtllm_ragged_deepseek_prefill: bool = True
33
+ use_trtllm_ragged_deepseek_prefill: bool = False
34
"""Whether to use TRTLLM ragged deepseek prefill."""
35
36
use_trtllm_attention: bool | None = None