fix: default attn to sdpa (no flash_attn dependency at inference)

config.json

@@ -2,7 +2,7 @@
   "architectures": [
     "MossTTSNanoForCausalLM"
   ],
-  "attn_implementation": null,
+  "attn_implementation": "sdpa",
   "audio_assistant_slot_token_id": 9,
   "audio_codebook_sizes": [
     1024,
@@ -70,7 +70,7 @@
   "im_end_token_id": 5,
   "im_start_token_id": 4,
   "initializer_range": 0.02,
-  "local_transformer_attn_implementation": "flash_attention_2",
+  "local_transformer_attn_implementation": "sdpa",
   "local_transformer_layers": 1,
   "max_position_embeddings": 32768,
   "model_architecture": "global_local_transformer",
@@ -81,4 +81,4 @@
   "tokenizer_use_fast": false,
   "transformers_version": "4.57.1",
   "vocab_size": 16384
-}
+}