failed to run on colab: ModulesToSaveWrapper has no attribute embed_tokens

[Vostredamus]

I am trying to train a lora in colab based on the example provided in the repository. My current yml is as follows:

base_model: /content/drive/MyDrive/models/ehartforddolphin-2.2.1-mistral-7b
base_model_config: /content/drive/MyDrive/models/ehartforddolphin-2.2.1-mistral-7b
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:

• path: /content/drive/MyDrive/samantha.json
  type: sharegpt
  conversation: chatml

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: /content/drive/MyDrive/samantha

adapter: lora
lora_model_dir:

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

lora_r: 512
lora_alpha: 512
lora_dropout: 0.05
lora_target_modules: q_proj, k_proj, v_proj, o_proj, gate_proj
lora_target_linear: true
lora_modules_to_save: embed_tokens, lm_head

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

mlflow_experiment_name: test-test

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 4
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: false
fp16: true
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false

warmup_steps: 50
eval_steps: 0.05
eval_table_size:
eval_table_max_new_tokens:
saves_per_epoch: 10
save_steps:
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
bos_token: ""
eos_token: "<|im_end|>"
unk_token: ""
tokens:

• "<|im_start|>"
• "<|im_end|>"

Many thanks in advance!