{ "attention_backend": "sdpa", "compute_dtype": "torch.bfloat16", "n_embd": 512, "n_head": 4, "n_kv_head": 4, "n_layer": 8, "sequence_len": 2048, "use_activation_checkpointing": true, "vocab_size": 16384, "window_pattern": "SSSL" }