From d9e7ca5b48538a854cafb6fdb8a9a7ceb9039b04 Mon Sep 17 00:00:00 2001 From: vfbd Date: Thu, 7 Jul 2022 17:48:00 -0400 Subject: [PATCH] Upload map file for BLOOM --- maps/bloom.json | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+) create mode 100644 maps/bloom.json diff --git a/maps/bloom.json b/maps/bloom.json new file mode 100644 index 00000000..e3f5feb9 --- /dev/null +++ b/maps/bloom.json @@ -0,0 +1,30 @@ +{ + "mtj_compat": "bloom", + "mtj_pe": "alibi", + "mtj_config_map": { + "d_model": "n_embed", + "n_heads": "num_attention_heads", + "layers": "n_layer" + }, + "static_weights": { + "word_embeddings.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}}, + "word_embeddings_layernorm.weight": {"mtj": {"module": "embedding_shard/~/replicated_layer_norm", "param": "scale"}}, + "word_embeddings_layernorm.bias": {"mtj": {"module": "embedding_shard/~/replicated_layer_norm", "param": "offset"}}, + "ln_f.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}}, + "ln_f.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}} + }, + "layer_weights": { + "h.{layer}.self_attention.query_key_value.weight": {"mtj": {"module": "layer_{layer}/~/combined_qkv", "param": "w"}}, + "h.{layer}.self_attention.query_key_value.bias": {"mtj": {"module": "layer_{layer}/~/combined_qkv", "param": "b"}}, + "h.{layer}.self_attention.dense.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}}, + "h.{layer}.self_attention.dense.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}}, + "h.{layer}.mlp.dense_h_to_4h.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}}, + "h.{layer}.mlp.dense_h_to_4h.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}}, + "h.{layer}.mlp.dense_4h_to_h.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}}, + "h.{layer}.mlp.dense_4h_to_h.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}}, + "h.{layer}.input_layernorm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}}, + "h.{layer}.input_layernorm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}}, + "h.{layer}.post_attention_layernorm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}}, + "h.{layer}.post_attention_layernorm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "offset"}} + } +}