vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update tensor mappings for Phi models (Phi-1, Phi-1.5, Phi-2)

Browse source 
Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
This commit is contained in:
teleprint-me 2023-12-20 15:18:52 -05:00
parent e53d44c0bb
commit e96f40bf99
No known key found for this signature in database
GPG key ID: B0D11345E65C4D48
1 changed files with 8 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

16
gguf-py/gguf/tensor_mapping.py
View file

@ -17,7 +17,7 @@ class TensorNameMap:
"tok_embeddings", # llama-pth "tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert "embeddings.word_embeddings", # bert
"language_model.embedding.word_embeddings", # persimmon "language_model.embedding.word_embeddings", # persimmon
"transformer.embd.wte", # phi2 "transformer.embd.wte", # phi1 phi1_5 phi2
), ),
# Token type embeddings # Token type embeddings
@ -42,7 +42,7 @@ class TensorNameMap:
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen
"output", # llama-pth bloom "output", # llama-pth bloom
"word_embeddings_for_head", # persimmon "word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2 "lm_head.linear", # phi1 phi1_5 phi2
), ),
# Output norm # Output norm
@ -55,7 +55,7 @@ class TensorNameMap:
"transformer.norm_f", # mpt "transformer.norm_f", # mpt
"ln_f", # refact bloom qwen "ln_f", # refact bloom qwen
"language_model.encoder.final_layernorm", # persimmon "language_model.encoder.final_layernorm", # persimmon
"lm_head.ln", # phi2 "lm_head.ln", # phi1 phi1_5 phi2
), ),
# Rope frequencies # Rope frequencies
@ -78,7 +78,7 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.output.LayerNorm", # bert "encoder.layer.{bid}.attention.output.LayerNorm", # bert
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon "language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi "model.layers.{bid}.ln1", # yi
"transformer.h.{bid}.ln", # phi2 "transformer.h.{bid}.ln", # phi1 phi1_5 phi2
), ),
# Attention norm 2 # Attention norm 2
@ -94,7 +94,7 @@ class TensorNameMap:
"transformer.h.{bid}.self_attention.query_key_value", # falcon "transformer.h.{bid}.self_attention.query_key_value", # falcon
"h.{bid}.self_attention.query_key_value", # bloom "h.{bid}.self_attention.query_key_value", # bloom
"language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
"transformer.h.{bid}.mixer.Wqkv", # phi2 "transformer.h.{bid}.mixer.Wqkv", # phi1 phi1_5 phi2
), ),
# Attention query # Attention query
@ -133,7 +133,7 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.output.dense", # bert "encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j "transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon "language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
"transformer.h.{bid}.mixer.out_proj", # phi2 "transformer.h.{bid}.mixer.out_proj", # phi1 phi1_5 phi2
), ),
# Rotary embeddings # Rotary embeddings
@ -173,7 +173,7 @@ class TensorNameMap:
"transformer.h.{bid}.mlp.fc_in", # gpt-j "transformer.h.{bid}.mlp.fc_in", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
"transformer.h.{bid}.mlp.w1", # qwen "transformer.h.{bid}.mlp.w1", # qwen
"transformer.h.{bid}.mlp.fc1", # phi2 "transformer.h.{bid}.mlp.fc1", # phi1 phi1_5 phi2
), ),
MODEL_TENSOR.FFN_UP_EXP: ( MODEL_TENSOR.FFN_UP_EXP: (
@ -205,7 +205,7 @@ class TensorNameMap:
"encoder.layer.{bid}.output.dense", # bert "encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j "transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"transformer.h.{bid}.mlp.fc2", # phi2 "transformer.h.{bid}.mlp.fc2", # phi1 phi1_5 phi2
), ),
MODEL_TENSOR.FFN_DOWN_EXP: ( MODEL_TENSOR.FFN_DOWN_EXP: (