Axolotl Config for Mixtral-8x22B

mixtral-8x22B.yaml

base_model: mistral-community/Mixtral-8x22B-v0.1
model_type: MixtralForCausalLM
tokenizer_type: AutoTokenizer
is_mistral_derived_model: false
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

model_config:
  output_router_logits: true

datasets:
  - path: /home/migel/ai_datasets/tess-v1.5b.jsonl
    type:
      system_prompt: "SYSTEM: {system}"
      field_system: system
      format: "USER: {instruction}\nASSISTANT: "
      no_input_format: "USER: {instruction}\nASSISTANT: "
      field_output: response               

dataset_prepared_path:
val_set_size: 0
output_dir: /home/migel/Tess-2-Mixtral-22B

adapter:  qlora
lora_model_dir:

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project: axolotl-tess
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 8
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00001

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

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

warmup_steps: 10
evals_per_epoch:
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.0
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"