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

convert: dbrx: fix remove wrong ATTN_OUT_NORM tensor, add output layer mapping

Browse source
This commit is contained in:
Pierrick HYMBERT 2024-04-06 20:30:30 +02:00
parent c8e6f903e0
commit 916b91852b
2 changed files with 9 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

1
gguf-py/gguf/constants.py
View file

@ -648,7 +648,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_NORM_2,
MODEL_TENSOR.ATTN_OUT_NORM,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,

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

@ -52,6 +52,7 @@ class TensorNameMap:
"output", # llama-pth bloom internlm2
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
"transformer.wte.weight", # dbrx
),
# Output norm
@ -68,6 +69,7 @@ class TensorNameMap:
"model.norm_f", # mamba-qbert
"backbone.norm_f", # mamba
"transformer.rms_norm", # Grok
"transformer.norm_f.weight", # dbrx
),
# Rope frequencies
@ -176,10 +178,9 @@ class TensorNameMap:
# Attention output norm
MODEL_TENSOR.ATTN_OUT_NORM: (
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"encoder.layers.{bid}.norm1", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
"transformer.blocks.{bid}.norm_attn_norm.attn.out_proj.weight", # dbrx
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"encoder.layers.{bid}.norm1", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
),
# Rotary embeddings
@ -307,9 +308,10 @@ class TensorNameMap:
),
MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
"transformer.blocks.{bid}.norm_attn_norm.attn.out_proj.weight", # dbrx
),
MODEL_TENSOR.SSM_IN: (