Skip to content

Instantly share code, notes, and snippets.

@mtisz
Created May 9, 2024 14:57
Show Gist options
  • Save mtisz/5cd0e72844e552fd06e77535c81bbfae to your computer and use it in GitHub Desktop.
Save mtisz/5cd0e72844e552fd06e77535c81bbfae to your computer and use it in GitHub Desktop.
Axolotl Config for Mixtral-8x22B
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>"
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment