[Feature] Support finetune Deepseek v2 (InternLM#663)

* support deepseek v2 * fix dispatch * refactor deepseek v2 * fix lint * fix bugs * fix bugs * delete useless codes * refactor deepseek config * rewrite DeepseekV2PreTrainedModel.from_pretrained * revert sft.py to main * delete useless codes * add deepseek v2 config * add deepseek readme * add HFCheckpointHook * optimize mixtral moe * fix bugs * delete useless codes * delete evalchathook * fix bugs * fix bugs * add moe SUPPORT_MODELS and fix HFCheckpointHook * add moe SUPPORT_MODELS and fix HFCheckpointHook * fix bugs * refactor modeling_deepseek * update deepseek readme * support deepseek v2 lite * fix bugs
llkn-2 · Jun 13, 2024 · 35fdf40 · 35fdf40
1 parent 83de829
commit 35fdf40
Show file tree

Hide file tree

Showing 21 changed files with 5,518 additions and 2 deletions.
diff --git a/.pre-commit-config-zh-cn.yaml b/.pre-commit-config-zh-cn.yaml
@@ -4,6 +4,7 @@ repos:
     rev: 5.0.4
     hooks:
       - id: flake8
+        args: ["--exclude=xtuner/model/transformers_models/*"]
   - repo: https://gitee.com/openmmlab/mirrors-isort
     rev: 5.11.5
     hooks:

diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
@@ -4,6 +4,7 @@ repos:
     rev: 5.0.4
     hooks:
       - id: flake8
+        args: ["--exclude=xtuner/model/transformers_models/*"]
   - repo: https://github.com/PyCQA/isort
     rev: 5.11.5
     hooks:

diff --git a/xtuner/configs/deepseek/README.md b/xtuner/configs/deepseek/README.md
@@ -0,0 +1,59 @@
+# DeepSeek V2
+
+## Install
+
+```bash
+# Git clone the latest xtuner
+git clone https://github.com/InternLM/xtuner.git
+
+# Install the latest xtuner
+cd xtuner
+pip install -e '.[all]'
+
+# Mixtral requires flash-attn
+pip install flash-attn
+
+# install the latest transformers
+pip install -U transformers
+```
+
+## Full Parameter Fine-tune
+
+Full parameter fine-tune DeepSeek V2 236B needs at least 64 A100-80G. The full-tuned model will be saved to `${WORK_DIRS}/hf_model` by `HFCheckpointHook`.
+
+### slurm
+
+Note: `$PARTITION` means the virtual partition of slurm.
+
+```bash
+srun -p $PARTITION --job-name=mixtral --nodes=8 --gres=gpu:8 --ntasks-per-node=8 xtuner train deepseek_v2_chat_full_alpaca_e3 --deepspeed deepspeed_zero3 --launcher slurm
+```
+
+### torchrun
+
+Note: `$NODE_0_ADDR` means the ip address of the node_0 machine.
+
+```bash
+# excuete on node 0
+NPROC_PER_NODE=8 NNODES=8 PORT=29600 ADDR=$NODE_0_ADDR NODE_RANK=0 xtuner train deepseek_v2_chat_full_alpaca_e3 --deepspeed deepspeed_zero3 --launcher pytorch
+
+# excuete on node 1
+NPROC_PER_NODE=8 NNODES=8 PORT=29600 ADDR=$NODE_0_ADDR NODE_RANK=1 xtuner train deepseek_v2_chat_full_alpaca_e3 --deepspeed deepspeed_zero3 --launcher pytorch
+
+# excuete on node 2, 3, ..., 7
+```
+
+### Speed
+
+128 * A100 80G:
+
+|         Model          | Sequence Length | Use Varlen Attn | Sequence Parallel World Size | Tokens per Second |
+| :--------------------: | :-------------: | :-------------: | :--------------------------: | :---------------: |
+|     deepseek v2 hf     |       8k        |      False      |              1               |        60         |
+| **deepseek v2 XTuner** |     **8k**      |    **False**    |            **1**             |   **120 (2x)**    |
+|     deepseek v2 hf     |       8k        |      True       |              1               |        60         |
+| **deepseek v2 XTuner** |     **8k**      |    **True**     |            **1**             |  **130 (2.2x)**   |
+|     deepseek v2 hf     |       16k       |      False      |              1               |        OOM        |
+| **deepseek v2 XTuner** |     **16k**     |    **False**    |            **1**             |      **148**      |
+|     deepseek v2 hf     |       16k       |      True       |              1               |        95         |
+| **deepseek v2 XTuner** |     **16k**     |    **True**     |            **1**             |  **180 (1.9x)**   |
diff --git a/xtuner/configs/deepseek/deepseek_v2_chat/deepseek_v2_chat_full_alpaca_e3.py b/xtuner/configs/deepseek/deepseek_v2_chat/deepseek_v2_chat_full_alpaca_e3.py
@@ -0,0 +1,198 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+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 torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, HFCheckpointHook,
+                                 ThroughputHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.model.transformers_models.deepseek_v2 import DeepseekV2ForCausalLM
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'deepseek-ai/DeepSeek-V2-Chat'
+use_varlen_attn = False
+
+# Data
+data_path = 'tatsu-lab/alpaca'
+prompt_template = PROMPT_TEMPLATE.deepseek_v2
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1  # bs per device 1 * acc 1 * 128 gpus = 128 total bs
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 4
+max_epochs = 3
+optim_type = AdamW
+lr = 1e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 50
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+# Save the optimizer states of deepseek v2 236B will require a lot of
+# storage space. It is recommended to set `save_optimizer` to False
+# (The training phase can not be resumed.)
+save_optimizer = True
+
+# Evaluate the generation performance during the training
+evaluation_freq = 25
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+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,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        # Only full-finetune is supported in `DeepseekV2ForCausalLM``, XTuner.
+        # Please use `AutoModelForCausalLM` for lora or qlora finetune.
+        type=DeepseekV2ForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        moe_implementation='shard',
+        expert_in_one_shard=10,
+        trust_remote_code=True))
+
+#######################################################################
+#                      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=alpaca_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,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    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=0.0,
+        by_epoch=True,
+        begin=0,
+        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
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(type=ThroughputHook),
+    dict(type=HFCheckpointHook)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# 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,
+        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, window_size=1)