fix sharegpt tokenization, refactor tokenization debugging
Browse files- scripts/finetune.py +4 -31
- src/axolotl/prompters.py +16 -5
- src/axolotl/utils/models.py +5 -5
- src/axolotl/utils/tokenization.py +33 -0
- src/axolotl/utils/trainer.py +5 -0
scripts/finetune.py
CHANGED
|
@@ -11,6 +11,8 @@ import yaml
|
|
| 11 |
from attrdict import AttrDefault
|
| 12 |
|
| 13 |
# add src to the pythonpath so we don't need to pip install this
|
|
|
|
|
|
|
| 14 |
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
|
| 15 |
src_dir = os.path.join(project_root, "src")
|
| 16 |
sys.path.insert(0, src_dir)
|
|
@@ -42,36 +44,6 @@ def choose_device(cfg):
|
|
| 42 |
cfg.device_map = {"": cfg.device}
|
| 43 |
|
| 44 |
|
| 45 |
-
def check_dataset_labels(dataset, tokenizer):
|
| 46 |
-
from termcolor import colored
|
| 47 |
-
|
| 48 |
-
# the dataset is already shuffled, so let's just check the first 5 elements
|
| 49 |
-
for idx in range(5):
|
| 50 |
-
# Get the input_ids, labels, and attention_mask from the dataset
|
| 51 |
-
input_ids = dataset[idx]["input_ids"]
|
| 52 |
-
labels = dataset[idx]["labels"]
|
| 53 |
-
attention_mask = dataset[idx]["attention_mask"]
|
| 54 |
-
|
| 55 |
-
# You can compare the input_ids and labels element-wise
|
| 56 |
-
# Remember to ignore positions with IGNORE_TOKEN_ID (if you use it) or attention_mask equal to 0
|
| 57 |
-
colored_tokens = []
|
| 58 |
-
for i, (input_id, label_id, mask) in enumerate(
|
| 59 |
-
zip(input_ids, labels, attention_mask)
|
| 60 |
-
):
|
| 61 |
-
decoded_input_token = tokenizer.decode(input_id)
|
| 62 |
-
# Choose the color based on whether the label has the ignore value or not
|
| 63 |
-
color = (
|
| 64 |
-
"red" if label_id == -100 else ("yellow" if label_id == 0 else "green")
|
| 65 |
-
)
|
| 66 |
-
colored_token = colored(decoded_input_token, color) + colored(
|
| 67 |
-
f"({label_id}, {mask})", "white"
|
| 68 |
-
)
|
| 69 |
-
colored_tokens.append(colored_token)
|
| 70 |
-
|
| 71 |
-
logging.info(" ".join(colored_tokens))
|
| 72 |
-
logging.info("\n\n\n")
|
| 73 |
-
|
| 74 |
-
|
| 75 |
def do_inference(cfg, model, tokenizer):
|
| 76 |
tokenizer.add_special_tokens({"unk_token": "<unk>"})
|
| 77 |
tokenizer.add_special_tokens({"bos_token": "<s>"})
|
|
@@ -199,8 +171,9 @@ def train(
|
|
| 199 |
return
|
| 200 |
|
| 201 |
if cfg.debug:
|
|
|
|
| 202 |
check_dataset_labels(
|
| 203 |
-
train_dataset.select([random.randrange(0, len(train_dataset) - 1)]),
|
| 204 |
tokenizer,
|
| 205 |
)
|
| 206 |
|
|
|
|
| 11 |
from attrdict import AttrDefault
|
| 12 |
|
| 13 |
# add src to the pythonpath so we don't need to pip install this
|
| 14 |
+
from axolotl.utils.tokenization import check_dataset_labels
|
| 15 |
+
|
| 16 |
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
|
| 17 |
src_dir = os.path.join(project_root, "src")
|
| 18 |
sys.path.insert(0, src_dir)
|
|
|
|
| 44 |
cfg.device_map = {"": cfg.device}
|
| 45 |
|
| 46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
def do_inference(cfg, model, tokenizer):
|
| 48 |
tokenizer.add_special_tokens({"unk_token": "<unk>"})
|
| 49 |
tokenizer.add_special_tokens({"bos_token": "<s>"})
|
|
|
|
| 171 |
return
|
| 172 |
|
| 173 |
if cfg.debug:
|
| 174 |
+
logging.info("check_dataset_labels...")
|
| 175 |
check_dataset_labels(
|
| 176 |
+
train_dataset.select([random.randrange(0, len(train_dataset) - 1) for i in range(5)]),
|
| 177 |
tokenizer,
|
| 178 |
)
|
| 179 |
|
src/axolotl/prompters.py
CHANGED
|
@@ -127,7 +127,7 @@ conv_vicuna_v1_1 = Conversation(
|
|
| 127 |
|
| 128 |
|
| 129 |
class ShareGPTPrompter:
|
| 130 |
-
def build_prompt(self, source, tokenizer):
|
| 131 |
# ignore the system prompt if provided
|
| 132 |
if source[0]["from"] == "system":
|
| 133 |
source.pop(0)
|
|
@@ -157,13 +157,14 @@ class ShareGPTPrompter:
|
|
| 157 |
role = roles[sentence["from"]]
|
| 158 |
assert role == conv.roles[j % 2]
|
| 159 |
conv.append_message(role, sentence["value"])
|
|
|
|
| 160 |
conversation = conv.get_prompt()
|
| 161 |
|
| 162 |
# Tokenize conversations
|
| 163 |
tokenized_result = tokenizer(
|
| 164 |
conversation,
|
| 165 |
truncation=True,
|
| 166 |
-
max_length=
|
| 167 |
padding=False,
|
| 168 |
return_tensors=None,
|
| 169 |
)
|
|
@@ -173,7 +174,9 @@ class ShareGPTPrompter:
|
|
| 173 |
sep = conv.sep + conv.roles[1] + ": "
|
| 174 |
|
| 175 |
rounds = conversation.split(conv.sep2)
|
|
|
|
| 176 |
cur_len = 1
|
|
|
|
| 177 |
for i, rou in enumerate(rounds):
|
| 178 |
if rou == "":
|
| 179 |
break
|
|
@@ -182,19 +185,27 @@ class ShareGPTPrompter:
|
|
| 182 |
if len(parts) != 2:
|
| 183 |
break
|
| 184 |
parts[0] += sep
|
| 185 |
-
round_len = len(tokenizer(rou)["input_ids"])
|
| 186 |
-
|
|
|
|
| 187 |
target[cur_len : cur_len + instruction_len] = [
|
| 188 |
IGNORE_TOKEN_ID
|
| 189 |
] * instruction_len
|
| 190 |
|
| 191 |
cur_len += round_len
|
| 192 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
attention_mask = [
|
| 194 |
1 if x != tokenizer.pad_token_id else 0
|
| 195 |
for x in tokenized_result["input_ids"]
|
| 196 |
]
|
| 197 |
|
|
|
|
| 198 |
return dict(
|
| 199 |
input_ids=tokenized_result["input_ids"],
|
| 200 |
labels=target,
|
|
|
|
| 127 |
|
| 128 |
|
| 129 |
class ShareGPTPrompter:
|
| 130 |
+
def build_prompt(self, source, tokenizer, sequence_len=2048):
|
| 131 |
# ignore the system prompt if provided
|
| 132 |
if source[0]["from"] == "system":
|
| 133 |
source.pop(0)
|
|
|
|
| 157 |
role = roles[sentence["from"]]
|
| 158 |
assert role == conv.roles[j % 2]
|
| 159 |
conv.append_message(role, sentence["value"])
|
| 160 |
+
# TODO, this concatenates everything, but doesn't seem to properly add the eos_token_id, as the eos_token gets split up
|
| 161 |
conversation = conv.get_prompt()
|
| 162 |
|
| 163 |
# Tokenize conversations
|
| 164 |
tokenized_result = tokenizer(
|
| 165 |
conversation,
|
| 166 |
truncation=True,
|
| 167 |
+
max_length=sequence_len, # FIXME
|
| 168 |
padding=False,
|
| 169 |
return_tensors=None,
|
| 170 |
)
|
|
|
|
| 174 |
sep = conv.sep + conv.roles[1] + ": "
|
| 175 |
|
| 176 |
rounds = conversation.split(conv.sep2)
|
| 177 |
+
rounds = [r + conv.sep2 for r in rounds]
|
| 178 |
cur_len = 1
|
| 179 |
+
target[0] = IGNORE_TOKEN_ID # mask out the bos
|
| 180 |
for i, rou in enumerate(rounds):
|
| 181 |
if rou == "":
|
| 182 |
break
|
|
|
|
| 185 |
if len(parts) != 2:
|
| 186 |
break
|
| 187 |
parts[0] += sep
|
| 188 |
+
round_len = len(tokenizer(rou)["input_ids"]) - 1 # -1 ignores the bos_token generated for this
|
| 189 |
+
# we have to strip the initial part, any dangling whitespace creates an additional ghost token
|
| 190 |
+
instruction_len = len(tokenizer(parts[0].strip())["input_ids"]) - 1 # -1 ignores the bos_token generated for this
|
| 191 |
target[cur_len : cur_len + instruction_len] = [
|
| 192 |
IGNORE_TOKEN_ID
|
| 193 |
] * instruction_len
|
| 194 |
|
| 195 |
cur_len += round_len
|
| 196 |
+
if cur_len >= sequence_len:
|
| 197 |
+
break
|
| 198 |
+
|
| 199 |
+
# Fix: Truncate the target to have the same length as input_ids
|
| 200 |
+
target = target[:len(tokenized_result["input_ids"])]
|
| 201 |
+
# target[cur_len:] = [IGNORE_TOKEN_ID] * (len(target) - cur_len)
|
| 202 |
+
|
| 203 |
attention_mask = [
|
| 204 |
1 if x != tokenizer.pad_token_id else 0
|
| 205 |
for x in tokenized_result["input_ids"]
|
| 206 |
]
|
| 207 |
|
| 208 |
+
# TODO truncate len to sequence_len
|
| 209 |
return dict(
|
| 210 |
input_ids=tokenized_result["input_ids"],
|
| 211 |
labels=target,
|
src/axolotl/utils/models.py
CHANGED
|
@@ -53,7 +53,7 @@ def load_model(
|
|
| 53 |
logging.info("patching with xformers attention")
|
| 54 |
hijack_llama_attention()
|
| 55 |
|
| 56 |
-
torch_dtype =
|
| 57 |
try:
|
| 58 |
if cfg.load_4bit:
|
| 59 |
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
|
|
@@ -161,11 +161,11 @@ def load_model(
|
|
| 161 |
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
|
| 162 |
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
| 163 |
|
| 164 |
-
if cfg.
|
| 165 |
-
for k, v in cfg.
|
| 166 |
-
|
| 167 |
|
| 168 |
-
if load_in_8bit and
|
| 169 |
logging.info("converting model w/ prepare_model_for_int8_training")
|
| 170 |
model = prepare_model_for_int8_training(model)
|
| 171 |
|
|
|
|
| 53 |
logging.info("patching with xformers attention")
|
| 54 |
hijack_llama_attention()
|
| 55 |
|
| 56 |
+
torch_dtype = torch.float16 if cfg.load_in_8bit or cfg.fp16 or cfg.bf16 else torch.float32
|
| 57 |
try:
|
| 58 |
if cfg.load_4bit:
|
| 59 |
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
|
|
|
|
| 161 |
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
|
| 162 |
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
| 163 |
|
| 164 |
+
if cfg.tokens:
|
| 165 |
+
for k, v in cfg.tokens.items():
|
| 166 |
+
tokenizer.add_special_tokens({k: v})
|
| 167 |
|
| 168 |
+
if load_in_8bit and cfg.load_4bit:
|
| 169 |
logging.info("converting model w/ prepare_model_for_int8_training")
|
| 170 |
model = prepare_model_for_int8_training(model)
|
| 171 |
|
src/axolotl/utils/tokenization.py
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from termcolor import colored
|
| 2 |
+
import logging
|
| 3 |
+
|
| 4 |
+
def check_dataset_labels(dataset, tokenizer):
|
| 5 |
+
# the dataset is already shuffled, so let's just check the first 5 elements
|
| 6 |
+
for idx in range(5):
|
| 7 |
+
check_example_labels(dataset[idx], tokenizer)
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
def check_example_labels(example, tokenizer):
|
| 11 |
+
# Get the input_ids, labels, and attention_mask from the dataset
|
| 12 |
+
input_ids = example["input_ids"]
|
| 13 |
+
labels = example["labels"]
|
| 14 |
+
attention_mask =example["attention_mask"]
|
| 15 |
+
|
| 16 |
+
# You can compare the input_ids and labels element-wise
|
| 17 |
+
# Remember to ignore positions with IGNORE_TOKEN_ID (if you use it) or attention_mask equal to 0
|
| 18 |
+
colored_tokens = []
|
| 19 |
+
for i, (input_id, label_id, mask) in enumerate(
|
| 20 |
+
zip(input_ids, labels, attention_mask)
|
| 21 |
+
):
|
| 22 |
+
decoded_input_token = tokenizer.decode(input_id)
|
| 23 |
+
# Choose the color based on whether the label has the ignore value or not
|
| 24 |
+
color = (
|
| 25 |
+
"red" if label_id == -100 else ("yellow" if label_id == 0 else "green")
|
| 26 |
+
)
|
| 27 |
+
colored_token = colored(decoded_input_token, color) + colored(
|
| 28 |
+
f"({label_id}, {mask}, {input_id})", "white"
|
| 29 |
+
)
|
| 30 |
+
colored_tokens.append(colored_token)
|
| 31 |
+
|
| 32 |
+
logging.info(" ".join(colored_tokens))
|
| 33 |
+
logging.info("\n\n\n")
|
src/axolotl/utils/trainer.py
CHANGED
|
@@ -61,6 +61,11 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
|
|
| 61 |
group_by_length=cfg.group_by_length,
|
| 62 |
report_to="wandb" if cfg.use_wandb else None,
|
| 63 |
run_name=cfg.wandb_run_id if cfg.use_wandb else None,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
**training_arguments_kwargs,
|
| 65 |
)
|
| 66 |
|
|
|
|
| 61 |
group_by_length=cfg.group_by_length,
|
| 62 |
report_to="wandb" if cfg.use_wandb else None,
|
| 63 |
run_name=cfg.wandb_run_id if cfg.use_wandb else None,
|
| 64 |
+
optim=cfg.optimizer if cfg.optimizer != "adam8bit" else cfg.optimizer,
|
| 65 |
+
lr_scheduler_type=cfg.lr_scheduler if cfg.lr_scheduler else None,
|
| 66 |
+
weight_decay=cfg.weight_decay if cfg.weight_decay else 0.0,
|
| 67 |
+
fsdp=cfg.fsdp.split(" ") if cfg.fsdp else None,
|
| 68 |
+
fsdp_transformer_layer_cls_to_wrap=cfg.fsdp_transformer_layer_cls_to_wrap if cfg.fsdp_transformer_layer_cls_to_wrap else None,
|
| 69 |
**training_arguments_kwargs,
|
| 70 |
)
|
| 71 |
|