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

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

Author: compilade

convert-hf-to-gguf.py

@@ -218,7 +218,7 @@ def from_model_architecture(model_architecture):
             return BertModel
         if model_architecture == "NomicBertModel":
             return NomicBertModel
-        if model_architecture == "MambaForCausalLM":
+        if model_architecture in ("MambaForCausalLM", "MambaLMHeadModel"):
             return MambaModel
         return Model
 
@@ -279,7 +279,7 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH:
             return gguf.MODEL_ARCH.BERT
         if arch == "NomicBertModel":
             return gguf.MODEL_ARCH.NOMIC_BERT
-        if arch == "MambaForCausalLM":
+        if arch in ("MambaForCausalLM", "MambaLMHeadModel"):
             return gguf.MODEL_ARCH.MAMBA
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')

gguf-py/gguf/constants.py

@@ -205,7 +205,6 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down.{xid}",
     MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up.{xid}",
     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_CONV1D:      "blk.{bid}.ssm_conv1d",
     MODEL_TENSOR.SSM_X:           "blk.{bid}.ssm_x",

llama.cpp

@@ -393,8 +393,6 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_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_CONV1D,
     LLM_TENSOR_SSM_X,