rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Merge pull request #349 from Zurnaz/llama_config

Browse Source 
feat: llama config and updated mtj requirement
This commit is contained in:
henk717
2023-05-06 16:52:22 +02:00
committed by GitHub
parent 2730879c61 91463a4d97
commit 04592e5086
2 changed files with 36 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
maps/llama.json Normal file
View File

@@ -0,0 +1,35 @@
{
"mtj_compat": "llama",
"mtj_pe": "neox_rotary",
"mtj_config_map": {
"norm": ["norm", "layernorm-nobias"],
"pe_rotary_dims": ["pe_rotary_dims", 128],
"d_model": "hidden_size",
"n_heads": "num_attention_heads",
"n_vocab": "vocab_size",
"layers": "num_hidden_layers",
"seq": "max_position_embeddings",
"tokenizer_class": ["tokenizer_class", "LlamaTokenizer"],
"tokenizer": ["tokenizer", "llama"]
},
"static_weights": {
"model.embed_tokens.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
"model.norm.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
"lm_head.weight": {"mtj": {"module": "projection_shard/~/linear", "param": "w", "transforms": ["vocab_pad"]}}
},
"layer_weights": {
"transformer.h.{layer}.attn.attention.bias": {},
"transformer.h.{layer}.attn.attention.masked_bias": {},
"model.layers.{layer}.self_attn.rotary_emb.inv_freq": {},
"model.layers.{layer}.self_attn.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
"model.layers.{layer}.self_attn.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
"model.layers.{layer}.self_attn.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
"model.layers.{layer}.self_attn.o_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
"model.layers.{layer}.mlp.gate_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
"model.layers.{layer}.mlp.down_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
"model.layers.{layer}.mlp.up_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_6", "param": "w"}},
"model.layers.{layer}.input_layernorm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
"model.layers.{layer}.post_attention_layernorm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}}
}
}

2
requirements_mtj.txt
View File

@@ -9,7 +9,7 @@ transformers == 4.28.0
chex == 0.1.5 chex == 0.1.5
huggingface_hub==0.12.1 huggingface_hub==0.12.1
progressbar2 progressbar2
git+https://github.com/VE-FORBRYDERNE/mesh-transformer-jax@ck git+https://github.com/Zurnaz/mesh-transformer-jax.git@llama_tpu
Flask==2.2.3 Flask==2.2.3
Flask-SocketIO==5.3.2 Flask-SocketIO==5.3.2
python-socketio==5.7.2 python-socketio==5.7.2