Skip to content

SOTA Open Source TTS

Disable flash attention 2 by default

Lengyue2919eafd19e7
diff --git a/fish_speech/models/text2semantic/llama.py b/fish_speech/models/text2semantic/llama.py
index 9b42386..0b27e6f 100644
--- a/fish_speech/models/text2semantic/llama.py
+++ b/fish_speech/models/text2semantic/llama.py
@@ -38,7 +38,7 @@ class ModelArgs:
codebook_padding_idx: int = 0
# Use flash attention
- use_flash_attention: bool = is_flash_attn_2_available()
+ use_flash_attention: bool = False
# Gradient checkpointing
use_gradient_checkpointing: bool = True