Skip to content

[Lora] Cannot convert falcon lora to GGUF. #3319

Closed
@Ph0rk0z

Description

@Ph0rk0z

Expected Behavior

convert-lora-to-gguf.py converts a HF lora adapter bin to GGUF format.

Current Behavior

Spits out the error Error: unrecognized tensor

Environment and Context

I think at least some stuff has to be added to:

HF_SUBLAYER_TO_GGML = {
    "self_attn.q_proj": "attn_q",
    "self_attn.k_proj": "attn_k",
    "self_attn.v_proj": "attn_v",
    "self_attn.o_proj": "attn_output",
    "mlp.gate_proj": "ffn_gate",
    "mlp.down_proj": "ffn_down",
    "mlp.up_proj": "ffn_up",
    "input_layernorm": "attn_norm",
    "post_attention_layernorm": "ffn_norm",
}

And maybe some other llama specific things changed down the line but I have no idea of what the names are supposed to be or where to find them.

Now that merging adapters is possible, this would let all of the falcon 180b tunes work without downloading 100GB+ each.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions