KoboldAI-Client/maps/gpt_neo.json

26 lines
2.2 KiB
JSON

{
"static_weights": {
"transformer.wte.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "axis": 1, "transforms": ["transpose", "vocab_pad"]}},
"transformer.wpe.weight": {"mtj": {"module": "embedding_shard/~/pos_embs", "param": "w", "axis": 2, "transforms": ["transpose"]}},
"transformer.ln_f.weight": {"mtj": {"module": "projection_shard/~/layer_norm", "param": "scale"}},
"transformer.ln_f.bias": {"mtj": {"module": "projection_shard/~/layer_norm", "param": "offset"}}
},
"layer_weights": {
"transformer.h.{layer}.attn.attention.bias": {},
"transformer.h.{layer}.attn.attention.masked_bias": {},
"transformer.h.{layer}.attn.attention.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w", "axis": 2}},
"transformer.h.{layer}.attn.attention.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w", "axis": 2}},
"transformer.h.{layer}.attn.attention.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w", "axis": 2}},
"transformer.h.{layer}.attn.attention.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w", "axis": 1}},
"transformer.h.{layer}.attn.attention.out_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}},
"transformer.h.{layer}.mlp.c_fc.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w", "axis": 2}},
"transformer.h.{layer}.mlp.c_fc.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b", "axis": 1}},
"transformer.h.{layer}.mlp.c_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w", "axis": 1}},
"transformer.h.{layer}.mlp.c_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
"transformer.h.{layer}.ln_1.weight": {"mtj": {"module": "layer_{layer}/~/layer_norm", "param": "scale"}},
"transformer.h.{layer}.ln_1.bias": {"mtj": {"module": "layer_{layer}/~/layer_norm", "param": "offset"}},
"transformer.h.{layer}.ln_2.weight": {"mtj": {"module": "layer_{layer}/~/layer_norm_1", "param": "scale"}},
"transformer.h.{layer}.ln_2.bias": {"mtj": {"module": "layer_{layer}/~/layer_norm_1", "param": "offset"}}
}
}