[Fix] Fix MMLU evaluation (#208)

* add default_map_fn

* fix bug

* add mmlu cfg

* update

* remove prompt_template

* update hook

* fix test bug

* Update runtime.txt

requirements/runtime.txt

@@ -1,6 +1,7 @@
 bitsandbytes>=0.40.0
 datasets
 einops
+fsspec<=2023.6.0
 lagent>=0.1.2
 mmengine>=0.9.0
 modelscope

xtuner/configs/internlm/internlm_7b/internlm_7b_qlora_oasst1_mmlu_e3.py

@@ -0,0 +1,237 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from bitsandbytes.optim import PagedAdamW32bit
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR
+from peft import LoraConfig
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn, mmlu_collate_fn
+from xtuner.dataset.map_fns import (default_map_fn, oasst1_map_fn,
+                                    template_map_fn_factory)
+from xtuner.engine import DatasetInfoHook, EvaluateChatHook
+from xtuner.evaluation import MMLUMetric
+from xtuner.model import SupervisedFinetune
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm-7b'
+
+# Data
+data_path = 'timdettmers/openassistant-guanaco'
+prompt_template = PROMPT_TEMPLATE.internlm_chat
+max_length = 2048
+pack_to_max_length = True
+
+# Val/Test data
+# Download from https://github.com/artidoro/qlora/tree/main/data/mmlu
+mmlu_data_root = './data/mmlu/'
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = PagedAdamW32bit
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=data_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=oasst1_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn))
+
+mmlu_fs_dataset = dict(
+    type=load_dataset,
+    path='json',
+    data_files=dict(
+        val=mmlu_data_root + 'five_shot_mmlu_val.json',
+        test=mmlu_data_root + 'five_shot_mmlu_test.json'))
+
+val_mmlu_fs = dict(
+    type=process_hf_dataset,
+    dataset=mmlu_fs_dataset,
+    tokenizer=tokenizer,
+    dataset_map_fn=default_map_fn,
+    max_length=max_length,
+    input_ids_with_output=False,
+    pack_to_max_length=False,
+    split='val')
+
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=0,
+    dataset=val_mmlu_fs,
+    sampler=dict(type=DefaultSampler, shuffle=False),
+    collate_fn=dict(type=mmlu_collate_fn))
+
+val_evaluator = dict(
+    type=MMLUMetric, tokenizer=tokenizer, prefix='mmlu_fs_val')
+
+test_mmlu_fs = dict(
+    type=process_hf_dataset,
+    dataset=mmlu_fs_dataset,
+    tokenizer=tokenizer,
+    dataset_map_fn=default_map_fn,
+    max_length=max_length,
+    input_ids_with_output=False,
+    pack_to_max_length=False,
+    split='test')
+
+test_dataloader = dict(
+    batch_size=1,
+    num_workers=0,
+    dataset=test_mmlu_fs,
+    sampler=dict(type=DefaultSampler, shuffle=False),
+    collate_fn=dict(type=mmlu_collate_fn))
+
+test_evaluator = dict(
+    type=MMLUMetric, tokenizer=tokenizer, prefix='mmlu_fs_test')
+
+#######################################################################
+#                    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=CosineAnnealingLR,
+    eta_min=lr * 0.1,
+    by_epoch=True,
+    T_max=max_epochs,
+    convert_to_iter_based=True)
+
+# train, val, test setting
+train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)
+val_cfg = dict(type='ValLoop')
+test_cfg = dict(type='TestLoop')
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 100 iterations.
+    logger=dict(type=LoggerHook, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per epoch.
+    checkpoint=dict(type=CheckpointHook, interval=1),
+    # 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)

xtuner/dataset/collate_fns/mmlu_collate_fn.py

@@ -36,4 +36,4 @@ def mmlu_collate_fn(instances: Sequence[Dict],
     if return_hf_format:
         return data_dict
     else:
-        return {'data': data_dict, 'data_samples': None}
+        return {'data': data_dict, 'data_samples': data_samples}

xtuner/dataset/map_fns/dataset_map_fns/__init__.py

@@ -5,6 +5,7 @@
 from .code_alpaca_map_fn import code_alpaca_map_fn
 from .colors_map_fn import colors_map_fn
 from .crime_kg_assitant_map_fn import crime_kg_assitant_map_fn
+from .default_map_fn import default_map_fn
 from .law_reference_map_fn import law_reference_map_fn
 from .medical_map_fn import medical_map_fn
 from .msagent_map_fn import msagent_react_map_fn
@@ -23,5 +24,5 @@
     'tiny_codes_map_fn', 'colors_map_fn', 'law_reference_map_fn',
     'crime_kg_assitant_map_fn', 'sql_map_fn', 'openai_map_fn',
     'wizardlm_map_fn', 'stack_exchange_map_fn', 'msagent_react_map_fn',
-    'pretrain_map_fn'
+    'default_map_fn', 'pretrain_map_fn'
 ]

xtuner/dataset/map_fns/dataset_map_fns/default_map_fn.py

@@ -0,0 +1,8 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+def default_map_fn(example):
+    return {
+        'conversation': [{
+            'input': example['input'],
+            'output': example['output']
+        }]
+    }

xtuner/engine/hooks/dataset_info_hook.py

@@ -14,7 +14,7 @@ def log(self, runner, dataset, mode='train'):
         runner.logger.info(f'{mode} example:')
         runner.logger.info(self.tokenizer.decode(dataset[0]['input_ids']))
 
-    def before_run(self, runner) -> None:
+    def before_train(self, runner) -> None:
         do_train = runner.train_loop is not None
         do_eval = runner.val_loop is not None
         do_test = runner.test_loop is not None
@@ -27,3 +27,11 @@ def before_run(self, runner) -> None:
         if do_test:
             test_dataset = runner.test_dataloader.dataset
             self.log(runner, test_dataset, mode='test')
+
+    def before_val(self, runner) -> None:
+        eval_dataset = runner.val_dataloader.dataset
+        self.log(runner, eval_dataset, mode='eval')
+
+    def before_test(self, runner) -> None:
+        test_dataset = runner.test_dataloader.dataset
+        self.log(runner, test_dataset, mode='test')

xtuner/tools/test.py

@@ -4,6 +4,7 @@
 import os.path as osp
 from types import FunctionType
 
+import torch
 from mmengine.config import Config, DictAction
 from mmengine.registry import RUNNERS
 from mmengine.runner import Runner
@@ -12,6 +13,29 @@
 from xtuner.registry import MAP_FUNC
 
 
+def guess_load_checkpoint(pth_model):
+    if os.path.isfile(pth_model):
+        state_dict = torch.load(pth_model, map_location='cpu')
+        if 'state_dict' in state_dict:
+            state_dict = state_dict['state_dict']
+    elif os.path.isdir(pth_model):
+        try:
+            from deepspeed.utils.zero_to_fp32 import \
+                get_fp32_state_dict_from_zero_checkpoint
+        except ImportError:
+            raise ImportError(
+                'The provided PTH model appears to be a DeepSpeed checkpoint. '
+                'However, DeepSpeed library is not detected in current '
+                'environment. This suggests that DeepSpeed may not be '
+                'installed or is incorrectly configured. Please verify your '
+                'setup.')
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(
+            os.path.dirname(pth_model), os.path.basename(pth_model))
+    else:
+        raise FileNotFoundError(f'Cannot find {pth_model}')
+    return state_dict
+
+
 def parse_args():
     parser = argparse.ArgumentParser(description='Test model')
     parser.add_argument('config', help='config file name or path.')
@@ -85,9 +109,6 @@ def main():
         cfg.work_dir = osp.join('./work_dirs',
                                 osp.splitext(osp.basename(args.config))[0])
 
-    if args.checkpoint is not None:
-        cfg.load_from = args.checkpoint
-
     # build the runner from config
     if 'runner_type' not in cfg:
         # build the default runner
@@ -97,6 +118,10 @@ def main():
         # if 'runner_type' is set in the cfg
         runner = RUNNERS.build(cfg)
 
+    state_dict = guess_load_checkpoint(args.checkpoint)
+    runner.model.load_state_dict(state_dict, strict=False)
+    runner.logger.info(f'Load checkpoint from {args.checkpoint}')
+
     # start testing
     runner.test()