support microsoft/bitnet-b1.58-2B-4T-bf16

[csabakecskemeti]

I found the followings:

• MS named it's butnet architecture slightly different, so we need to update the converter script to support their naming:
  @Model.register("BitnetForCausalLM", "BitNetForCausalLM")

• They also use slightly different layer names, I've manually edited the gguf-py/gguf/tensor_mapping.py file:

        #MODEL_TENSOR.ATTN_SUB_NORM: (
        #    "model.layers.{bid}.self_attn.inner_attn_ln",  # bitnet
        #),
        
        MODEL_TENSOR.ATTN_SUB_NORM: (
            "model.layers.{bid}.self_attn.attn_sub_norm",  # bitnet
        ),
        
        #MODEL_TENSOR.FFN_SUB_NORM: ( 
        #    "model.layers.{bid}.mlp.ffn_layernorm",  # bitnet
        #), 
            
        MODEL_TENSOR.FFN_SUB_NORM: (      
            "model.layers.{bid}.mlp.ffn_sub_norm",  # bitnet
        ),  

Now I've reached the state where the converter looks for the tokenizer model
raise FileNotFoundError(f"File not found: {tokenizer_path}")

(I've get to the same point with their own BitNet.cpp utils/convert-hf-to-gguf-bitnet.py )

Any hit how to move forward?

Related change
e112b61