Vram fix attempt (#1164) [skip ci]
Browse files* revert order of filter/drop_long step and handle calc for max_input_len only during preprocessing
* revert some changes to preparing for packing to allow more flexibility
* prepare dataset for packing during pre-processing step
* prepare dataset hash based on sample packing too
* enclose none check
* just cast straight to string for ds hash
- src/axolotl/utils/data.py +7 -1
- src/axolotl/utils/samplers/utils.py +6 -6
- src/axolotl/utils/trainer.py +27 -27
src/axolotl/utils/data.py
CHANGED
@@ -116,6 +116,12 @@ def load_tokenized_prepared_datasets(
|
|
116 |
(
|
117 |
str(cfg.sequence_len)
|
118 |
+ "@"
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
+ "|".join(
|
120 |
sorted(
|
121 |
[
|
@@ -162,7 +168,7 @@ def load_tokenized_prepared_datasets(
|
|
162 |
LOG.info("Loading raw datasets...")
|
163 |
if not cfg.is_preprocess:
|
164 |
LOG.warning(
|
165 |
-
"Processing datasets during training can lead to VRAM instability. Please pre-process your dataset"
|
166 |
)
|
167 |
|
168 |
if cfg.seed:
|
|
|
116 |
(
|
117 |
str(cfg.sequence_len)
|
118 |
+ "@"
|
119 |
+
+ str(cfg.sample_packing)
|
120 |
+
+ "@"
|
121 |
+
+ str(cfg.eval_sample_packing)
|
122 |
+
+ "@"
|
123 |
+
+ str(cfg.group_by_length)
|
124 |
+
+ "@"
|
125 |
+ "|".join(
|
126 |
sorted(
|
127 |
[
|
|
|
168 |
LOG.info("Loading raw datasets...")
|
169 |
if not cfg.is_preprocess:
|
170 |
LOG.warning(
|
171 |
+
"Processing datasets during training can lead to VRAM instability. Please pre-process your dataset."
|
172 |
)
|
173 |
|
174 |
if cfg.seed:
|
src/axolotl/utils/samplers/utils.py
CHANGED
@@ -7,11 +7,11 @@ import numpy as np
|
|
7 |
def get_dataset_lengths(dataset):
|
8 |
if "length" in dataset.data.column_names:
|
9 |
lengths = np.array(dataset.data.column("length"))
|
|
|
|
|
|
|
10 |
else:
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
.apply(lambda x: x[-1] + 1)
|
15 |
-
.values
|
16 |
-
)
|
17 |
return lengths
|
|
|
7 |
def get_dataset_lengths(dataset):
|
8 |
if "length" in dataset.data.column_names:
|
9 |
lengths = np.array(dataset.data.column("length"))
|
10 |
+
elif "position_ids" in dataset.data.column_names:
|
11 |
+
position_ids = dataset.data.column("position_ids")
|
12 |
+
lengths = np.array([x[-1] + 1 for x in position_ids])
|
13 |
else:
|
14 |
+
input_ids = dataset.data.column("input_ids")
|
15 |
+
lengths = np.vectorize(len)(np.array(input_ids, dtype=object))
|
16 |
+
return lengths
|
|
|
|
|
|
|
17 |
return lengths
|
src/axolotl/utils/trainer.py
CHANGED
@@ -109,6 +109,33 @@ def drop_long_seq(sample, sequence_len=2048):
|
|
109 |
def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
|
110 |
drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
|
111 |
with zero_first(is_main_process()):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
if cfg.group_by_length:
|
113 |
train_dataset = train_dataset.map(
|
114 |
add_length,
|
@@ -130,33 +157,6 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
|
|
130 |
load_from_cache_file=not cfg.is_preprocess,
|
131 |
)
|
132 |
|
133 |
-
if cfg.group_by_length or cfg.sample_packing:
|
134 |
-
max_input_len = np.max(get_dataset_lengths(train_dataset))
|
135 |
-
LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
|
136 |
-
|
137 |
-
train_dataset = train_dataset.filter(
|
138 |
-
drop_long,
|
139 |
-
num_proc=cfg.dataset_processes,
|
140 |
-
load_from_cache_file=not cfg.is_preprocess,
|
141 |
-
)
|
142 |
-
if eval_dataset:
|
143 |
-
eval_dataset = eval_dataset.filter(
|
144 |
-
drop_long,
|
145 |
-
num_proc=cfg.dataset_processes,
|
146 |
-
load_from_cache_file=not cfg.is_preprocess,
|
147 |
-
)
|
148 |
-
|
149 |
-
# Phi doesn't want the attention_mask feature when training
|
150 |
-
if (
|
151 |
-
"CodeGenTokenizer" in tokenizer.__class__.__name__
|
152 |
-
or (cfg.is_mistral_derived_model and cfg.flash_attention)
|
153 |
-
or cfg.model_config_type == "mamba"
|
154 |
-
):
|
155 |
-
LOG.info("dropping attention_mask column")
|
156 |
-
train_dataset = train_dataset.remove_columns("attention_mask")
|
157 |
-
if eval_dataset:
|
158 |
-
eval_dataset = eval_dataset.remove_columns("attention_mask")
|
159 |
-
|
160 |
return train_dataset, eval_dataset
|
161 |
|
162 |
|
|
|
109 |
def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
|
110 |
drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
|
111 |
with zero_first(is_main_process()):
|
112 |
+
if cfg.is_preprocess:
|
113 |
+
max_input_len = np.max(get_dataset_lengths(train_dataset))
|
114 |
+
LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
|
115 |
+
|
116 |
+
# Phi doesn't want the attention_mask feature when training
|
117 |
+
if (
|
118 |
+
"CodeGenTokenizer" in tokenizer.__class__.__name__
|
119 |
+
or (cfg.is_mistral_derived_model and cfg.flash_attention)
|
120 |
+
or cfg.model_config_type == "mamba"
|
121 |
+
):
|
122 |
+
LOG.info("dropping attention_mask column")
|
123 |
+
train_dataset = train_dataset.remove_columns("attention_mask")
|
124 |
+
if eval_dataset:
|
125 |
+
eval_dataset = eval_dataset.remove_columns("attention_mask")
|
126 |
+
|
127 |
+
train_dataset = train_dataset.filter(
|
128 |
+
drop_long,
|
129 |
+
num_proc=cfg.dataset_processes,
|
130 |
+
load_from_cache_file=not cfg.is_preprocess,
|
131 |
+
)
|
132 |
+
if eval_dataset:
|
133 |
+
eval_dataset = eval_dataset.filter(
|
134 |
+
drop_long,
|
135 |
+
num_proc=cfg.dataset_processes,
|
136 |
+
load_from_cache_file=not cfg.is_preprocess,
|
137 |
+
)
|
138 |
+
|
139 |
if cfg.group_by_length:
|
140 |
train_dataset = train_dataset.map(
|
141 |
add_length,
|
|
|
157 |
load_from_cache_file=not cfg.is_preprocess,
|
158 |
)
|
159 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
return train_dataset, eval_dataset
|
161 |
|
162 |
|