What causes "Unable to process key {key}" in hiyouga/LlamaFactory?

Switch model_name_or_path to the llava-hf equivalent, e.g. llava-hf/llava-1.5-7b-hf

hiyouga/LlamaFactory · error · KeyError

Unable to process key {key}

Error message

Unable to process key {key}

What it means

config patching rejects checkpoints whose architectures contain LlavaLlamaForCausalLM (patcher.py:410). That architecture name is the original liuhaotian/llava repository's custom class, which LlamaFactory cannot load; it requires the llava-hf conversions that use transformers' native LlavaForConditionalGeneration.

Source

Thrown at scripts/convert_ckpt/llamafy_qwen.py:83

            elif "attn.c_proj" in key:
                llama_state_dict[key.replace("attn.c_proj", "self_attn.o_proj")] = value
                llama_state_dict[key.replace("attn.c_proj.weight", "self_attn.o_proj.bias")] = torch.zeros_like(
                    value[:, 0]
                ).squeeze()
            elif "ln_1" in key:
                llama_state_dict[key.replace("ln_1", "input_layernorm")] = value
            elif "ln_2" in key:
                llama_state_dict[key.replace("ln_2", "post_attention_layernorm")] = value
            elif "mlp.w1" in key:
                llama_state_dict[key.replace("mlp.w1", "mlp.up_proj")] = value
            elif "mlp.w2" in key:
                llama_state_dict[key.replace("mlp.w2", "mlp.gate_proj")] = value
            elif "mlp.c_proj" in key:
                llama_state_dict[key.replace("mlp.c_proj", "mlp.down_proj")] = value
            elif "lm_head" in key:
                llama_state_dict[key] = value
            else:
                raise KeyError(f"Unable to process key {key}")

    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
    state_dict_split = split_torch_state_dict_into_shards(
        llama_state_dict, filename_pattern=filename_pattern, max_shard_size=shard_size
    )
    for shard_file, tensors in tqdm(state_dict_split.filename_to_tensors.items(), desc="Save weights"):
        shard = {tensor: llama_state_dict[tensor].contiguous() for tensor in tensors}
        if save_safetensors:
            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
        else:
            torch.save(shard, os.path.join(output_dir, shard_file))

    if not state_dict_split.is_sharded:
        print(f"Model weights saved in {os.path.join(output_dir, weights_name)}.")
    else:
        index = {
            "metadata": state_dict_split.metadata,

View on GitHub (pinned to f28afaf635)

Solutions

Switch model_name_or_path to the llava-hf equivalent, e.g. llava-hf/llava-1.5-7b-hf
Re-publish/convert legacy checkpoints with the llava-hf conversion scripts if the weights are custom

Example fix

# before
model_name_or_path: liuhaotian/llava-v1.5-7b

# after
model_name_or_path: llava-hf/llava-1.5-7b-hf

Defensive patterns

Strategy: validation

Validate before calling

from transformers import AutoConfig
arch = getattr(AutoConfig.from_pretrained(model_path), 'architectures', []) or []
assert 'LlavaLlamaForCausalLM' not in arch, 'Use llava-hf/* checkpoints (e.g. llava-hf/llava-1.5-7b-hf)'

Type guard

def is_hf_llava(model_name_or_path: str) -> bool:
    cfg = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=False)
    archs = getattr(cfg, 'architectures', None) or []
    return 'LlavaLlamaForCausalLM' not in archs and any(a.startswith('Llava') for a in archs)

Try / catch

try:
    run_sft(train_args)
except ValueError as e:
    if 'llava' in str(e):
        raise SystemExit('Switch to llava-hf/... model id') from e
    raise

Prevention

Always source LLaVA model ids from the llava-hf organization
Check architectures before adding models to training configs

When it happens

Trigger: model_name_or_path points at liuhaotian/llava-v1.5-7b (or a finetune derived from it); AutoConfig reports LlavaLlamaForCausalLM in architectures and the loader raises ValueError immediately.

Common situations: Copying a model id from the original LLaVA paper/repo instead of the llava-hf org; using community merges or LoRA bases published in the legacy format.

Related errors

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/98a021f8c61efabe. Report an issue: GitHub.