File size: 12,051 Bytes
4e046d9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 | # Copyright 2025 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
from dataclasses import asdict, dataclass
from typing import TYPE_CHECKING, Any, Optional
from ...extras import logging
from ...extras.constants import IGNORE_INDEX
from .processor_utils import DatasetProcessor, greedy_knapsack, infer_seqlen
if TYPE_CHECKING:
from ..mm_plugin import AudioInput, ImageInput, VideoInput
logger = logging.get_logger(__name__)
MAX_SU_SEQ_IDX = 2**32 # maximum sub-sequence index
@dataclass
class PackingParams:
r"""Metadata for a packed sequence: sub-sequence boundaries and multimodal data indices.
- sequence_boundaries: cumulative token positions, e.g. [0, 100, 250, 512] means 3 sub-seqs
with token ranges [0,100), [100,250), [250,512). Length = num_sub_seqs + 1.
- image_subseq_ids / video_subseq_ids / audio_subseq_ids: for each mm item, the 0-based
sub-sequence index it belongs to. Length = total number of that mm type in the packed sample.
"""
sequence_boundaries: list[int]
image_subseq_ids: list[int]
video_subseq_ids: list[int]
audio_subseq_ids: list[int]
right_padding_length: int
@dataclass
class SupervisedDatasetProcessor(DatasetProcessor):
def _encode_data_example(
self,
prompt: list[dict[str, str]],
response: list[dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: list["ImageInput"],
videos: list["VideoInput"],
audios: list["AudioInput"],
) -> tuple[list[int], list[int]]:
messages = self.template.mm_plugin.process_messages(prompt + response, images, videos, audios, self.processor)
input_ids, labels = self.template.mm_plugin.process_token_ids(
[], [], images, videos, audios, self.tokenizer, self.processor
)
discarding_history_cot = self.data_args.mask_history and not self.template.preserve_thinking
encoded_pairs = self.template.encode_multiturn(self.tokenizer, messages, system, tools, discarding_history_cot)
total_length = len(input_ids) + (1 if self.template.efficient_eos else 0)
if self.data_args.mask_history:
encoded_pairs = encoded_pairs[::-1] # high priority for last turns
for turn_idx, (source_ids, target_ids) in enumerate(encoded_pairs):
if total_length >= self.data_args.cutoff_len:
break
source_len, target_len = infer_seqlen(
len(source_ids), len(target_ids), self.data_args.cutoff_len - total_length
)
source_ids = source_ids[:source_len]
target_ids = target_ids[:target_len]
total_length += source_len + target_len
if self.data_args.train_on_prompt:
source_label = source_ids
elif self.template.efficient_eos and turn_idx != 0:
source_label = [self.tokenizer.eos_token_id] + [IGNORE_INDEX] * (source_len - 1)
else:
source_label = [IGNORE_INDEX] * source_len
if self.data_args.mask_history and turn_idx != 0: # train on the last turn only
target_label = [IGNORE_INDEX] * target_len
else:
target_label = target_ids
if self.data_args.mask_history: # reversed sequences
input_ids = source_ids + target_ids + input_ids
labels = source_label + target_label + labels
else:
input_ids += source_ids + target_ids
labels += source_label + target_label
if self.template.efficient_eos:
input_ids += [self.tokenizer.eos_token_id]
labels += [self.tokenizer.eos_token_id]
return input_ids, labels
def preprocess_dataset(self, examples: dict[str, list[Any]]) -> dict[str, list[Any]]:
# build inputs with format `<bos> X Y <eos>` and labels with format `<ignore> ... <ignore> Y <eos>`
# for multiturn examples, we only mask the prompt part in each prompt-response pair.
model_inputs = defaultdict(list)
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) != 1:
logger.warning_rank0(
"Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
)
continue
input_ids, labels = self._encode_data_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
audios=examples["_audios"][i] or [],
)
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
model_inputs["images"].append(examples["_images"][i])
model_inputs["videos"].append(examples["_videos"][i])
model_inputs["audios"].append(examples["_audios"][i])
return model_inputs
def print_data_example(self, example: dict[str, list[int]]) -> None:
valid_labels = list(filter(lambda x: x != IGNORE_INDEX, example["labels"]))
print("input_ids:\n{}".format(example["input_ids"]))
print("inputs:\n{}".format(self.tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
print("label_ids:\n{}".format(example["labels"]))
print(f"labels:\n{self.tokenizer.decode(valid_labels, skip_special_tokens=False)}")
@dataclass
class PackedSupervisedDatasetProcessor(SupervisedDatasetProcessor):
def preprocess_dataset(self, examples: dict[str, list[Any]]) -> dict[str, list[Any]]:
# TODO: use `position_ids` to achieve packing
# build inputs with format `<bos> X1 Y1 <eos> <bos> X2 Y2 <eos>`
# and labels with format `<ignore> ... <ignore> Y1 <eos> <ignore> ... <ignore> Y2 <eos>`
valid_num = 0
batch_input_ids, batch_labels, batch_images, batch_videos, batch_audios = [], [], [], [], []
lengths = []
length2indexes = defaultdict(list)
for i in range(len(examples["_prompt"])):
if len(examples["_prompt"][i]) % 2 != 1 or len(examples["_response"][i]) != 1:
logger.warning_rank0(
"Dropped invalid example: {}".format(examples["_prompt"][i] + examples["_response"][i])
)
continue
input_ids, labels = self._encode_data_example(
prompt=examples["_prompt"][i],
response=examples["_response"][i],
system=examples["_system"][i],
tools=examples["_tools"][i],
images=examples["_images"][i] or [],
videos=examples["_videos"][i] or [],
audios=examples["_audios"][i] or [],
)
length = len(input_ids)
if length > self.data_args.cutoff_len:
logger.warning_rank0(f"Dropped lengthy example with length {length} > {self.data_args.cutoff_len}.")
else:
lengths.append(length)
length2indexes[length].append(valid_num)
batch_input_ids.append(input_ids)
batch_labels.append(labels)
batch_images.append(examples["_images"][i] or [])
batch_videos.append(examples["_videos"][i] or [])
batch_audios.append(examples["_audios"][i] or [])
valid_num += 1
model_inputs = defaultdict(list)
requires_packing_params = self.data_args.neat_packing
knapsacks = greedy_knapsack(lengths, self.data_args.cutoff_len)
for knapsack in knapsacks:
packed_input_ids, packed_attention_masks, packed_position_ids, packed_labels = [], [], [], []
packed_images, packed_videos, packed_audios = [], [], []
if requires_packing_params:
sequence_boundaries = [0]
image_subseq_ids: list[int] = []
video_subseq_ids: list[int] = []
audio_subseq_ids: list[int] = []
for i, length in enumerate(knapsack):
index = length2indexes[length].pop()
packed_input_ids += batch_input_ids[index]
packed_position_ids += list(range(len(batch_input_ids[index]))) # NOTE: pad_to_multiple_of ignore this
packed_labels += batch_labels[index]
packed_images += batch_images[index]
packed_videos += batch_videos[index]
packed_audios += batch_audios[index]
if requires_packing_params:
n_img = len(batch_images[index])
n_vid = len(batch_videos[index])
n_aud = len(batch_audios[index])
sequence_boundaries.append(sequence_boundaries[-1] + len(batch_input_ids[index]))
image_subseq_ids.extend([i] * n_img)
video_subseq_ids.extend([i] * n_vid)
audio_subseq_ids.extend([i] * n_aud)
if self.data_args.neat_packing:
packed_attention_masks += [i + 1] * len(batch_input_ids[index]) # start from 1
else:
packed_attention_masks += [1] * len(batch_input_ids[index])
if len(packed_input_ids) < self.data_args.cutoff_len + 1: # avoid flash_attn drops attn mask
pad_length = self.data_args.cutoff_len - len(packed_input_ids) + 1
packed_input_ids += [self.tokenizer.pad_token_id] * pad_length
packed_position_ids += [0] * pad_length
packed_labels += [IGNORE_INDEX] * pad_length
if self.data_args.neat_packing:
packed_attention_masks += [0] * pad_length
else:
packed_attention_masks += [1] * pad_length # more efficient flash_attn
if requires_packing_params:
sequence_boundaries.append(sequence_boundaries[-1] + pad_length)
if len(packed_input_ids) != self.data_args.cutoff_len + 1:
raise ValueError("The length of packed example should be identical to the cutoff length.")
model_inputs["input_ids"].append(packed_input_ids)
if requires_packing_params:
packing_params = PackingParams(
sequence_boundaries=sequence_boundaries,
image_subseq_ids=image_subseq_ids or [MAX_SU_SEQ_IDX], # avoid dataset concat error
video_subseq_ids=video_subseq_ids or [MAX_SU_SEQ_IDX],
audio_subseq_ids=audio_subseq_ids or [MAX_SU_SEQ_IDX],
right_padding_length=pad_length,
)
model_inputs["packing_params"].append(asdict(packing_params))
model_inputs["attention_mask"].append(packed_attention_masks)
model_inputs["position_ids"].append(packed_position_ids)
model_inputs["labels"].append(packed_labels)
model_inputs["images"].append(packed_images or None)
model_inputs["videos"].append(packed_videos or None)
model_inputs["audios"].append(packed_audios or None)
return model_inputs
|