gptj.json

{
  "mtj_compat": "j",
  "mtj_pe": "rotary",
  "mtj_config_map": {
    "pe_rotary_dims": ["rotary_dim", 64],
    "d_model": "n_embd",
    "n_heads": "n_head",
    "layers": "n_layer"
  },
  "static_weights": {
    "transformer.wte.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
    "transformer.wte.bias": {"mtj": {"module": "embedding_shard/~/linear", "param": "b", "transforms": ["vocab_pad"]}},
    "transformer.ln_f.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
    "transformer.ln_f.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}},
    "lm_head.weight": {"mtj": {"module": "projection_shard/~/linear", "param": "w", "transforms": ["vocab_pad"]}},
    "lm_head.bias": {"mtj": {"module": "projection_shard/~/linear", "param": "b", "transforms": ["vocab_pad"]}}
  },
  "layer_weights": {
    "transformer.h.{layer}.attn.bias": {},
    "transformer.h.{layer}.attn.masked_bias": {},
    "transformer.h.{layer}.attn.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
    "transformer.h.{layer}.attn.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
    "transformer.h.{layer}.attn.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
    "transformer.h.{layer}.attn.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
    "transformer.h.{layer}.mlp.fc_in.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
    "transformer.h.{layer}.mlp.fc_in.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
    "transformer.h.{layer}.mlp.fc_out.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
    "transformer.h.{layer}.mlp.fc_out.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
    "transformer.h.{layer}.ln_1.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
    "transformer.h.{layer}.ln_1.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}}
  }
}