xglm.json

{
  "mtj_compat": "fairseq_lm",
  "mtj_pe": "fairseq_sinusoidal",
  "mtj_config_map": {
    "d_model": "d_model",
    "n_heads": "attention_heads",
    "layers": "num_layers"
  },
  "static_weights": {
    "model.embed_tokens.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
    "model.layer_norm.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
    "model.layer_norm.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}}
  },
  "layer_weights": {
    "model.layers.{layer}.self_attn.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
    "model.layers.{layer}.self_attn.q_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear", "param": "b"}},
    "model.layers.{layer}.self_attn.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
    "model.layers.{layer}.self_attn.v_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "b"}},
    "model.layers.{layer}.self_attn.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
    "model.layers.{layer}.self_attn.k_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "b"}},
    "model.layers.{layer}.self_attn.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
    "model.layers.{layer}.self_attn.out_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}},
    "model.layers.{layer}.fc1.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
    "model.layers.{layer}.fc1.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
    "model.layers.{layer}.fc2.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
    "model.layers.{layer}.fc2.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
    "model.layers.{layer}.self_attn_layer_norm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
    "model.layers.{layer}.self_attn_layer_norm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}},
    "model.layers.{layer}.final_layer_norm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}},
    "model.layers.{layer}.final_layer_norm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "offset"}}
  }
}