[Feature] Support balanced dataset to speed-up VL training

xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_finetune_balance.py

@@ -0,0 +1,179 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset, BalancedDataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook, VarlenAttnArgsToMessageHubHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = '/model/internvl'
+
+# Data
+data_path='pack_internvl_sft_1.2M.json'
+'''about the data pack_internvl_sft_1.2M.json
+use the scripts data_preprocess_stastics.sh in the xtuner/tools
+to generate pack_internvl_sft_1.2M.json
+'''
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 4096
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 8
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=False,  # or False
+    use_varlen_attn=True
+)
+
+evaluation_freq = 0
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=BalancedDataset,
+    model_path=path,
+    data_path=data_path,
+    vit_packed_length=9,  # The value for vit packed length
+    llm_packed_length=4096,  # The value for llm packed length
+    llm_thresh=4068,  # The value for llm thresh
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=True, balance_data=True))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]  # vallen_attention 依赖的 Hook
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=1),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
+

xtuner/dataset/__init__.py

@@ -14,6 +14,7 @@
 from .refcoco_json import (InvRefCOCOJsonDataset, RefCOCOJsonDataset,
                            RefCOCOJsonEvalDataset)
 from .utils import decode_base64_to_image, expand2square, load_image
+from .fast_dataset import BalancedDataset
 
 # ignore FutureWarning in hf datasets
 warnings.simplefilter(action='ignore', category=FutureWarning)
@@ -25,5 +26,5 @@
     'load_intern_repo_tokenized_dataset',
     'load_intern_repo_untokenized_dataset', 'build_packed_dataset',
     'RefCOCOJsonDataset', 'RefCOCOJsonEvalDataset', 'InvRefCOCOJsonDataset',
-    'load_json_file', 'InternVL_V1_5_Dataset'
+    'load_json_file', 'InternVL_V1_5_Dataset', 'BalancedDataset'
 ]

xtuner/dataset/collate_fns/default_collate_fn.py

@@ -12,7 +12,8 @@
 def default_collate_fn(instances: Sequence[Dict],
                        pad_index: int = DEFAULT_PAD_TOKEN_INDEX,
                        return_hf_format: bool = False,
-                       use_varlen_attn: bool = False):
+                       use_varlen_attn: bool = False,
+                       balance_data: bool = False):
     seq_parallel_world_size = get_sequence_parallel_world_size()
 
     input_ids, labels = [], []
@@ -22,7 +23,7 @@ def default_collate_fn(instances: Sequence[Dict],
         assert len(instances) == 1, (
             f'If utilizing varlen attention, the batch size should be'
             f' set to 1, but got {len(instances)}')
-        assert not has_image, 'Currently, it is not configured to '
+        assert not has_image or balance_data, 'Currently, it is not configured to '
         'accommodate the use of varlen Attention in multimodal training'
 
     if has_image:
@@ -39,6 +40,7 @@ def default_collate_fn(instances: Sequence[Dict],
             pixel_values.append(example['pixel_values'])
 
     ori_length = [len(ids) for ids in input_ids]
+
     if len(instances) > 1:
         input_ids = pad_sequence(
             input_ids, batch_first=True, padding_value=pad_index)