fixes w/ example for super basic lora starter
Browse files- examples/lora-alpaca-7b/config.yml +67 -0
- src/axolotl/prompters.py +1 -1
- src/axolotl/utils/data.py +6 -4
examples/lora-alpaca-7b/config.yml
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
base_model: huggyllama/llama-7b
|
| 2 |
+
base_model_config: huggyllama/llama-7b
|
| 3 |
+
model_type: LlamaForCausalLM
|
| 4 |
+
tokenizer_type: LlamaTokenizer
|
| 5 |
+
load_in_8bit: true
|
| 6 |
+
load_in_4bit: false
|
| 7 |
+
strict: false
|
| 8 |
+
push_dataset_to_hub:
|
| 9 |
+
datasets:
|
| 10 |
+
- path: teknium/GPT4-LLM-Cleaned
|
| 11 |
+
type: alpaca
|
| 12 |
+
dataset_prepared_path: last_run_prepared
|
| 13 |
+
val_set_size: 0.02
|
| 14 |
+
adapter: lora
|
| 15 |
+
lora_model_dir:
|
| 16 |
+
sequence_len: 512
|
| 17 |
+
max_packed_sequence_len:
|
| 18 |
+
lora_r: 8
|
| 19 |
+
lora_alpha: 16
|
| 20 |
+
lora_dropout: 0.0
|
| 21 |
+
lora_target_modules:
|
| 22 |
+
- gate_proj
|
| 23 |
+
- down_proj
|
| 24 |
+
- up_proj
|
| 25 |
+
- q_proj
|
| 26 |
+
- v_proj
|
| 27 |
+
- k_proj
|
| 28 |
+
- o_proj
|
| 29 |
+
lora_fan_in_fan_out:
|
| 30 |
+
wandb_project:
|
| 31 |
+
wandb_watch:
|
| 32 |
+
wandb_run_id:
|
| 33 |
+
wandb_log_model:
|
| 34 |
+
output_dir: ./lora-out
|
| 35 |
+
batch_size: 4
|
| 36 |
+
micro_batch_size: 1
|
| 37 |
+
num_epochs: 4
|
| 38 |
+
optimizer: adamw_bnb_8bit
|
| 39 |
+
torchdistx_path:
|
| 40 |
+
lr_scheduler: cosine
|
| 41 |
+
learning_rate: 0.0002
|
| 42 |
+
train_on_inputs: false
|
| 43 |
+
group_by_length: false
|
| 44 |
+
bf16: false
|
| 45 |
+
fp16: true
|
| 46 |
+
tf32: true
|
| 47 |
+
gradient_checkpointing: true
|
| 48 |
+
early_stopping_patience:
|
| 49 |
+
resume_from_checkpoint:
|
| 50 |
+
local_rank:
|
| 51 |
+
logging_steps: 1
|
| 52 |
+
xformers_attention: true
|
| 53 |
+
flash_attention:
|
| 54 |
+
gptq_groupsize:
|
| 55 |
+
gptq_model_v1:
|
| 56 |
+
warmup_steps: 10
|
| 57 |
+
eval_steps: 50
|
| 58 |
+
save_steps:
|
| 59 |
+
debug:
|
| 60 |
+
deepspeed:
|
| 61 |
+
weight_decay: 0.0
|
| 62 |
+
fsdp:
|
| 63 |
+
fsdp_config:
|
| 64 |
+
special_tokens:
|
| 65 |
+
bos_token: "<s>"
|
| 66 |
+
eos_token: "</s>"
|
| 67 |
+
unk_token: "<unk>"
|
src/axolotl/prompters.py
CHANGED
|
@@ -18,7 +18,7 @@ class AlpacaPrompter:
|
|
| 18 |
prompt_style = None
|
| 19 |
|
| 20 |
def __init__(self, prompt_style="instruct"):
|
| 21 |
-
self.prompt_style = prompt_style
|
| 22 |
self.match_prompt_style()
|
| 23 |
|
| 24 |
def match_prompt_style(self):
|
|
|
|
| 18 |
prompt_style = None
|
| 19 |
|
| 20 |
def __init__(self, prompt_style="instruct"):
|
| 21 |
+
self.prompt_style = prompt_style if prompt_style else PromptStyle.instruct.value
|
| 22 |
self.match_prompt_style()
|
| 23 |
|
| 24 |
def match_prompt_style(self):
|
src/axolotl/utils/data.py
CHANGED
|
@@ -60,10 +60,12 @@ def load_tokenized_prepared_datasets(
|
|
| 60 |
else Path(default_dataset_prepared_path) / ds_hash
|
| 61 |
)
|
| 62 |
dataset = None
|
|
|
|
| 63 |
try:
|
| 64 |
if cfg.push_dataset_to_hub:
|
|
|
|
| 65 |
dataset = load_dataset(
|
| 66 |
-
f"{cfg.push_dataset_to_hub}/{ds_hash}", use_auth_token=
|
| 67 |
)
|
| 68 |
dataset = dataset["train"]
|
| 69 |
except:
|
|
@@ -83,7 +85,7 @@ def load_tokenized_prepared_datasets(
|
|
| 83 |
ds = None
|
| 84 |
ds_from_hub = False
|
| 85 |
try:
|
| 86 |
-
load_dataset(d.path, streaming=True, use_auth_token=
|
| 87 |
ds_from_hub = True
|
| 88 |
except FileNotFoundError:
|
| 89 |
pass
|
|
@@ -99,10 +101,10 @@ def load_tokenized_prepared_datasets(
|
|
| 99 |
d.path,
|
| 100 |
streaming=False,
|
| 101 |
data_files=d.data_files,
|
| 102 |
-
use_auth_token=
|
| 103 |
)
|
| 104 |
else:
|
| 105 |
-
ds = load_dataset(d.path, streaming=False, use_auth_token=
|
| 106 |
else:
|
| 107 |
fp = hf_hub_download(
|
| 108 |
repo_id=d.path, repo_type="dataset", filename=d.data_files
|
|
|
|
| 60 |
else Path(default_dataset_prepared_path) / ds_hash
|
| 61 |
)
|
| 62 |
dataset = None
|
| 63 |
+
use_auth_token = False
|
| 64 |
try:
|
| 65 |
if cfg.push_dataset_to_hub:
|
| 66 |
+
use_auth_token = True
|
| 67 |
dataset = load_dataset(
|
| 68 |
+
f"{cfg.push_dataset_to_hub}/{ds_hash}", use_auth_token=use_auth_token
|
| 69 |
)
|
| 70 |
dataset = dataset["train"]
|
| 71 |
except:
|
|
|
|
| 85 |
ds = None
|
| 86 |
ds_from_hub = False
|
| 87 |
try:
|
| 88 |
+
load_dataset(d.path, streaming=True, use_auth_token=use_auth_token)
|
| 89 |
ds_from_hub = True
|
| 90 |
except FileNotFoundError:
|
| 91 |
pass
|
|
|
|
| 101 |
d.path,
|
| 102 |
streaming=False,
|
| 103 |
data_files=d.data_files,
|
| 104 |
+
use_auth_token=use_auth_token,
|
| 105 |
)
|
| 106 |
else:
|
| 107 |
+
ds = load_dataset(d.path, streaming=False, use_auth_token=use_auth_token)
|
| 108 |
else:
|
| 109 |
fp = hf_hub_download(
|
| 110 |
repo_id=d.path, repo_type="dataset", filename=d.data_files
|