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

convert : for Mamba, also consider the "MambaLMHeadModel" arch name

Browse source 
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
This commit is contained in:
Francis Couture-Harpin 2024-02-04 09:00:42 -05:00
parent a3f4a1c7dc
commit 9f55809f72
3 changed files with 1 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

2
convert-hf-to-gguf.py
View file

@ -1844,7 +1844,7 @@ class StarCoder2Model(Model):
model_arch = gguf.MODEL_ARCH.STARCODER2 model_arch = gguf.MODEL_ARCH.STARCODER2
@Model.register("MambaForCausalLM") @Model.register("MambaForCausalLM", "MambaLMHeadModel")
class MambaModel(Model): class MambaModel(Model):
model_arch = gguf.MODEL_ARCH.MAMBA model_arch = gguf.MODEL_ARCH.MAMBA

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

@ -211,7 +211,6 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down.{xid}", MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down.{xid}",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up.{xid}", MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up.{xid}",
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm", MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
# FIXME: NAMES FOR MAMBA ARE NOT FINAL
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in", MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d", MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x", MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x",

2
llama.cpp
View file

@ -401,8 +401,6 @@ enum llm_tensor {
LLM_TENSOR_ATTN_Q_NORM, LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM, LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM, LLM_TENSOR_LAYER_OUT_NORM,
// TODO: maybe use longer names?
// TODO: can the in_proj and/or the out_proj instead re-use some of the above types?
LLM_TENSOR_SSM_IN, LLM_TENSOR_SSM_IN,
LLM_TENSOR_SSM_CONV1D, LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_X, LLM_TENSOR_SSM_X,