34 lines
2.5 KiB
JSON
34 lines
2.5 KiB
JSON
|
{
|
||
|
"mtj_compat": "neox",
|
||
|
"mtj_pe": "neox_rotary",
|
||
|
"mtj_config_map": {
|
||
|
"pe_rotary_pct": ["rotary_pct", 0.25],
|
||
|
"d_model": "hidden_size",
|
||
|
"n_heads": "num_attention_heads",
|
||
|
"layers": "num_hidden_layers"
|
||
|
},
|
||
|
"static_weights": {
|
||
|
"gpt_neox.embed_in.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
|
||
|
"gpt_neox.final_layer_norm.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
|
||
|
"gpt_neox.final_layer_norm.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}},
|
||
|
"embed_out.weight": {"mtj": {"module": "projection_shard/~/linear", "param": "w", "transforms": ["vocab_pad"]}}
|
||
|
},
|
||
|
"layer_weights": {
|
||
|
"gpt_neox.layers.{layer}.attention.bias": {},
|
||
|
"gpt_neox.layers.{layer}.attention.masked_bias": {},
|
||
|
"gpt_neox.layers.{layer}.attention.rotary_emb.inv_freq": {},
|
||
|
"gpt_neox.layers.{layer}.attention.query_key_value.weight": {"mtj": {"module": "layer_{layer}/~/combined_qkv", "param": "w"}},
|
||
|
"gpt_neox.layers.{layer}.attention.query_key_value.bias": {"mtj": {"module": "layer_{layer}/~/combined_qkv", "param": "b"}},
|
||
|
"gpt_neox.layers.{layer}.attention.dense.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
|
||
|
"gpt_neox.layers.{layer}.attention.dense.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}},
|
||
|
"gpt_neox.layers.{layer}.mlp.dense_h_to_4h.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
|
||
|
"gpt_neox.layers.{layer}.mlp.dense_h_to_4h.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
|
||
|
"gpt_neox.layers.{layer}.mlp.dense_4h_to_h.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
|
||
|
"gpt_neox.layers.{layer}.mlp.dense_4h_to_h.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
|
||
|
"gpt_neox.layers.{layer}.input_layernorm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
|
||
|
"gpt_neox.layers.{layer}.input_layernorm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}},
|
||
|
"gpt_neox.layers.{layer}.post_attention_layernorm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}},
|
||
|
"gpt_neox.layers.{layer}.post_attention_layernorm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "offset"}}
|
||
|
}
|
||
|
}
|