782b6a4216
* set fp16 to false if bf16, update bf16: auto in example YAMLs * unset fp16 so that it fallsback properly if bf16 isn't available * Update README.md [skip-ci] Co-authored-by: NanoCode012 <kevinvong@rocketmail.com> * test that bf16 disables fp16 --------- Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>
77 lines
1.2 KiB
YAML
77 lines
1.2 KiB
YAML
base_model: 01-ai/Yi-34B-Chat
|
|
model_type: LlamaForCausalLM
|
|
tokenizer_type: LlamaTokenizer
|
|
is_mistral_derived_model: false
|
|
is_llama_derived_model: true
|
|
load_in_8bit: false
|
|
load_in_4bit: true
|
|
strict: false
|
|
sequence_len: 1024
|
|
bf16: auto
|
|
fp16:
|
|
tf32: false
|
|
flash_attention: true
|
|
special_tokens:
|
|
bos_token: "<|startoftext|>"
|
|
eos_token: "<|endoftext|>"
|
|
unk_token: "<unk>"
|
|
|
|
# Data
|
|
datasets:
|
|
- path: mhenrichsen/alpaca_2k_test
|
|
type: alpaca
|
|
warmup_steps: 10
|
|
|
|
# Iterations
|
|
num_epochs: 1
|
|
|
|
# Evaluation
|
|
val_set_size: 0.1
|
|
evals_per_epoch: 5
|
|
eval_table_size:
|
|
eval_table_max_new_tokens: 128
|
|
eval_sample_packing: false
|
|
eval_batch_size: 1
|
|
|
|
# LoRA
|
|
output_dir: ./qlora-out
|
|
adapter: qlora
|
|
lora_model_dir:
|
|
lora_r: 32
|
|
lora_alpha: 16
|
|
lora_dropout: 0.05
|
|
lora_target_linear: true
|
|
lora_fan_in_fan_out:
|
|
lora_target_modules:
|
|
|
|
# Sampling
|
|
sample_packing: false
|
|
pad_to_sequence_len: false
|
|
|
|
# Batching
|
|
gradient_accumulation_steps: 4
|
|
micro_batch_size: 1
|
|
gradient_checkpointing: true
|
|
|
|
# wandb
|
|
wandb_project:
|
|
|
|
# Optimizer
|
|
optimizer: paged_adamw_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 0.0002
|
|
|
|
# Misc
|
|
train_on_inputs: false
|
|
group_by_length: false
|
|
early_stopping_patience:
|
|
resume_from_checkpoint:
|
|
local_rank:
|
|
logging_steps: 1
|
|
xformers_attention:
|
|
debug:
|
|
deepspeed:
|
|
weight_decay: 0
|
|
fsdp:
|
|
fsdp_config:
|