config.json

{
  "attention_backend": "sdpa",
  "compute_dtype": "torch.bfloat16",
  "expert_dim": 2048,
  "n_embd": 512,
  "n_head": 4,
  "n_kv_head": 4,
  "n_layer": 8,
  "num_experts": 8,
  "num_experts_per_tok": 2,
  "num_shared_experts": 1,
  "sequence_len": 2048,
  "use_activation_checkpointing": true,
  "use_moe": true,
  "vocab_size": 16384,
  "window_pattern": "SSSL"
}