modify default attention backend to flash_attn

config.json

@@ -13,8 +13,8 @@
   "causal_transformer_context_duration": 10.0,
   "number_channels": 2,
   "enable_channel_interleave": true,
-  "attention_implementation": "sdpa",
-  "compute_dtype": "fp32",
+  "attention_implementation": "flash_attention_2",
+  "compute_dtype": "bf16",
   "dtype": "float32",
   "code_dim": 768,
   "encoder_kwargs": [