update MindSpore Transformers LLM Course

02.MindSpore_Transformer_LLM_Course/Chapter03/3-MindSpore Transformers预训练与微调实践.md

@@ -0,0 +1,349 @@
+# MindSpore Transformers预训练与微调实践
+
+## 环境与目录
+
+* 建议目录：`/home/mindspore/work/demo`
+* 建议 Python ≥ 3.9，已安装 MindSpore + MindFormers
+* 可选：配置国内镜像
+
+```bash
+mkdir -p /home/mindspore/work/demo && cd /home/mindspore/work/demo
+pip install -U huggingface_hub
+export HF_ENDPOINT=https://hf-mirror.com   # 可按需要注释/取消
+```
+
+## 准备权重、数据集与MindSpore Transformers
+
+Qwen3-0.6B: [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B)
+Qwen3-1.7B: [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B)
+
+预训练数据集：[wikitext-2-v1](https://modelscope-open.oss-cn-hangzhou.aliyuncs.com/wikitext/wikitext-2-v1.zip)
+
+> 预训练数据集参考[社区issue](https://gitee.com/mindspore/mindformers/issues/IBV35D)获取
+
+微调数据集: [llm-wizard/alpaca-gpt4-data-zh](https://huggingface.co/datasets/llm-wizard/alpaca-gpt4-data-zh)
+
+```bash
+# 下载权重
+mkdir -p /home/mindspore/work/demo
+cd /home/mindspore/work/demo
+pip install -U huggingface_hub
+export HF_ENDPOINT=https://hf-mirror.com
+huggingface-cli download --resume-download Qwen/Qwen3-0.6B --local-dir Qwen3-0.6B
+huggingface-cli download --resume-download Qwen/Qwen3-1.7B --local-dir Qwen3-1.7B
+
+# 下载预训练数据集
+wget https://modelscope-open.oss-cn-hangzhou.aliyuncs.com/wikitext/wikitext-2-v1.zip
+unzip wikitext-2-v1.zip
+
+# 下载微调数据集
+huggingface-cli download --repo-type dataset --resume-download llm-wizard/alpaca-gpt4-data-zh --local-dir alpaca-gpt4-data-zh
+
+# 下载MindSpore Transformers
+git clone https://gitee.com/mindspore/mindformers.git
+cd mindformers
+git checkout 5a12973fb38bfd5b504240334492f4fb7ff7f7a6
+pip install -r requirements.txt
+
+# 升级MindSpore
+pip install https://repo.mindspore.cn/mindspore/mindspore/version/202509/20250917/master_20250917220006_52c46b3bfd9e9d50b2334d764afc80a6d7b56e90_newest/unified/aarch64/mindspore-2.7.1-cp39-cp39-linux_aarch64.whl
+
+```
+
+> 所有需要的权重和数据集都应当挂载到/home/mindspore/work/demo
+
+最后的文件结构应为：
+
+```plaintext
+/home/mindspore/work/demo
+├── Qwen3-0.6B/
+├── Qwen3-1.7B/
+├── wikitext-2-v1/
+├── alpaca-gpt4-data-zh/
+└── mindformers/
+```
+
+## 数据预处理
+
+### wiki => json
+
+脚本来源：[社区issue #ICOKGY](https://gitee.com/mindspore/mindformers/issues/ICOKGY)
+
+将脚本复制到`/home/mindspore/work/demo/mindformers/gen_wiki_json.py`
+
+```bash
+cd /home/mindspore/work/demo/mindformers
+python gen_wiki_json.py --input /home/mindspore/work/demo/wikitext-2/wiki.train.tokens  --output /home/mindspore/work/demo/wikitext-2/wiki.jsonl
+```
+
+### json => megatron
+
+先取前1000条数据
+
+```bash
+head -n 1000 /home/mindspore/work/demo/wikitext-2/wiki.jsonl > /home/mindspore/work/demo/wikitext-2/wiki.less.jsonl
+```
+
+```shell
+# cd /home/mindspore/work/demo/mindformers
+mkdir -p /home/mindspore/work/demo/megatron_data
+python toolkit/data_preprocess/megatron/preprocess_indexed_dataset.py \
+  --input /home/mindspore/work/demo/wikitext-2/wiki.less.jsonl \
+  --output-prefix /home/mindspore/work/demo/megatron_data/wikitext-2-v1-qwen3_text_document \
+  --tokenizer-type HuggingFaceTokenizer \
+  --tokenizer-dir /home/mindspore/work/demo/Qwen3-0.6B/ \
+  --workers 64
+```
+
+## 准备Qwen3-0.6B模型的yaml
+
+MindSpore Transformers源码仓库中提供了Qwen3-32B的预训练模型[yaml配置文件](https://gitee.com/mindspore/mindformers/blob/master/configs/qwen3/pretrain_qwen3_32b_4k.yaml)，我们可以参考该文件来创建Qwen3-0.6B的yaml文件。
+
+首先复制32B配置并重命名为0.6b的配置
+
+```bash
+cp configs/qwen3/pretrain_qwen3_32b_4k.yaml configs/qwen3/pretrain_qwen3_0_6b_4k.yaml
+```
+
+接下来需要对`configs/qwen3/pretrain_qwen3_0_6b_4k.yaml`做出以下修改：
+
+```yaml
+# Model configuration
+model:
+  model_config:
+    # Configurations from Hugging Face
+    ...
+    hidden_size: 1024
+    intermediate_size: 3072
+    num_hidden_layers: 28
+    num_attention_heads: 16
+    num_key_value_heads: 8
+    head_dim: 128
+    ...
+    # Configurations from MindFormers
+    offset: 0
+```
+
+## 预训练
+
+### 修改配置以启动预训练任务
+
+修改训练epoch数
+
+```yaml
+# runner config
+runner_config:
+  epochs: 1
+```
+
+修改数据集配置
+
+```yaml
+# dataset
+train_dataset: &train_dataset
+  data_loader:
+    ...
+    sizes:
+      - 400   # 训练集数据样本数
+      - 0    # 测试集数据样本数，当前不支持配置
+      - 0    # 评测集数据样本数，当前不支持配置
+    config:  # GPTDataset配置项
+      ...
+      data_path:  # Megatron数据集采样比例以及路径
+        - '1'
+        - "/home/mindspore/work/demo/megatron_data/wikitext-2-v1-qwen3_text_document"
+```
+
+添加以下参数以开启TensorBoard监控
+
+```yaml
+monitor_config:
+    monitor_on: True
+    dump_path: './dump'
+    target: ['layers.0', 'layers.1'] # 只监控第一、二层的参数
+    invert: False
+    step_interval: 1
+    local_loss_format: ['tensorboard']
+    device_local_loss_format: ['tensorboard']
+    local_norm_format: ['tensorboard']
+    device_local_norm_format: ['tensorboard']
+    optimizer_state_format: null
+    weight_state_format: null
+    throughput_baseline: null
+    print_struct: False
+    check_for_global_norm: False
+    global_norm_spike_threshold: 1.0
+    global_norm_spike_count_threshold: 10
+
+tensorboard:
+    tensorboard_dir: 'worker/tensorboard'
+    tensorboard_queue_size: 10
+    log_loss_scale_to_tensorboard: True
+    log_timers_to_tensorboard: True
+```
+
+
+
+添加以下配置以开启权重去冗余
+
+```yaml
+callbacks:
+  - type: MFLossMonitor  # Prints training progress information
+  - type: CheckpointMonitor  # Saves model weights during training
+    ...
+    save_checkpoint_steps: 50  # Interval steps for saving model weights
+    keep_checkpoint_max: 2  # Maximum number of saved model weight files
+    ...
+    remove_redundancy: True
+```
+
+修改以下配置以切换并行配置
+
+```yaml
+parallel_config:
+  data_parallel: &dp 2
+  model_parallel: 2
+  pipeline_stage: 2
+  micro_batch_num: &micro_batch_num 2
+```
+
+### 启动预训练任务
+
+```bash
+rm -rf ouput/checkpoint
+bash scripts/msrun_launcher.sh "run_mindformer.py \
+--config configs/qwen3/pretrain_qwen3_0_6b_4k.yaml \
+--auto_trans_ckpt False \
+--use_parallel True \
+--run_mode train \
+--recompute_config.recompute False"
+```
+
+### 启动断点续训的预训练任务
+
+修改以下配置以开启断点续训功能
+
+```yaml
+resume_training: True
+```
+
+备份上一次任务日志，用于查看续训是否成功
+
+```bash
+mv ouput/msrun_log output/msrun_log_bak
+```
+
+```bash
+bash scripts/msrun_launcher.sh "run_mindformer.py \
+--config configs/qwen3/pretrain_qwen3_0_6b_4k.yaml \
+--auto_trans_ckpt False \
+--use_parallel True \
+--run_mode train \
+--load_checkpoint "output/checkpoint_50_step" \
+--recompute_config.recompute False"
+```
+
+## 全参微调
+
+### 修改配置以启动全参微调任务
+
+MindSpore Transformers源码仓库中提供了Qwen3类不同规模的微调模型[yaml配置文件](https://gitee.com/mindspore/mindformers/blob/master/configs/qwen3/finetune_qwen3.yaml)，不同规模的模型可以通过`--pretrained_model_dir /path/to/Qwen3-0.6B/`来指定。我们可以参考该文件来创建Qwen3-0.6B的全参微调的yaml文件。
+
+
+需要对`configs/qwen3/finetune_qwen3.yaml`做出以下修改：
+
+修改以下配置项以设置数据集大小
+
+```yaml
+# Dataset configuration
+train_dataset: &train_dataset
+  ...
+  data_loader:
+  type: HFDataLoader
+    path: "llm-wizard/alpaca-gpt4-data-zh"
+    ...
+    # dataset process arguments
+    handler:
+      - type: take
+        n: 1000
+```
+
+修改以下配置以切换并行配置
+
+```yaml
+parallel_config:
+  data_parallel: &dp 8
+  model_parallel: 1
+  pipeline_stage: 1
+  micro_batch_num: &micro_batch_num 1
+  ...
+  use_seq_parallel: False
+```
+
+### 启动全参微调任务
+
+启动Qwen3-0.6B的全参微调任务
+
+```bash
+bash scripts/msrun_launcher.sh "run_mindformer.py \
+--config configs/qwen3/finetune_qwen3.yaml \
+--auto_trans_ckpt True \
+--use_parallel True \
+--run_mode train \
+--pretrained_model_dir /home/mindspore/work/demo/Qwen3-0.6B/ \
+--recompute_config.recompute False"
+```
+
+## LoRA微调
+
+MindSpore Transformers源码仓库中提供了Qwen3类不同规模的微调模型[yaml配置文件](https://gitee.com/mindspore/mindformers/blob/master/configs/qwen3/finetune_qwen3.yaml)，不同规模的模型可以通过`--pretrained_model_dir /path/to/Qwen3-1.7B/`来指定。我们可以参考该文件来创建Qwen3-1.7B的LoRA微调的yaml文件。
+
+首先复制微调配置并重命名为lora的配置
+
+```bash
+cp configs/qwen3/finetune_qwen3.yaml configs/qwen3/finetune_qwen3_lora.yaml
+```
+
+### 修改配置以启动LoRA微调任务
+
+添加以下参数以启动LoRA微调
+
+```yaml
+# Model configuration
+model:
+  model_config:
+    # Configurations from Hugging Face
+    ...
+    pet_config:
+      pet_type: lora
+      lora_rank: 8
+      lora_alpha: 16
+      lora_dropout: 0.1
+      lora_a_init: 'normal'
+      lora_b_init: 'zeros'
+      target_modules: '.*word_embeddings|.*linear_qkv|.*linear_proj|.*linear_fc1|.*linear_fc2'
+      freeze_include: ['*']
+      freeze_exclude: ['*lora*']
+```
+
+修改以下配置以切换并行配置
+
+```yaml
+parallel_config:
+  data_parallel: &dp 8
+  model_parallel: 1
+  pipeline_stage: 1
+  micro_batch_num: &micro_batch_num 1
+```
+
+### 启动LoRA微调任务
+
+```bash
+bash scripts/msrun_launcher.sh "run_mindformer.py \
+--config configs/qwen3/finetune_qwen3_lora.yaml \
+--auto_trans_ckpt True \
+--use_parallel True \
+--run_mode train \
+--pretrained_model_dir /home/mindspore/work/demo/Qwen3-1.7B/ \
+--recompute_config.recompute False"
+```

02.MindSpore_Transformer_LLM_Course/Chapter05/5-instructions.md

@@ -0,0 +1,45 @@
+# 使用指令汇总
+
+## 获取 Hugging Face 权重
+```bash
+git clone https://hf-mirror.com/Qwen/Qwen3-0.6B
+```
+
+## 获取数据集
+```bash
+mkdir qwen-datasets
+cd qwen-datasets
+
+wget https://atp-modelzoo-wlcb-pai.oss-cn-wulanchabu.aliyuncs.com/release/models/pai-megatron-patch/qwen-datasets/mmap_qwen3_datasets_text_document.bin
+wget https://atp-modelzoo-wlcb-pai.oss-cn-wulanchabu.aliyuncs.com/release/models/pai-megatron-patch/qwen-datasets/mmap_qwen3_datasets_text_document.idx
+```
+
+## 运行任务
+```bash
+bash scripts/msrun_launcher.sh "run_mindformer.py --config xx.yaml"
+```
+
+## 合并权重
+```bash
+python toolkit/safetensors/unified_safetensors.py \
+  --src_strategy_dirs ./output/strategy \
+  --mindspore_ckpt_dir ./output/checkpoint \
+  --output_dir ./unified_trained_qwen3_60_sf \
+  --file_suffix "60_1" \
+  --has_redundancy True \
+  --filter_out_param_prefix "adam_" \
+  --max_process_num 16
+```
+
+## qwen3-0.6B 反转权重
+```bash
+python toolkit/weight_convert/qwen3/reverse_mcore_qwen3_weight_to_hf.py \
+  --mindspore_ckpt_path ./unified_trained_qwen3_60_sf/60_1_ckpt_convert/unified_safe \
+  --huggingface_ckpt_path ./hf_sf \
+  --num_layers 28 \
+  --num_attention_heads 16 \
+  --num_query_groups 8 \
+  --kv_channels 128 \
+  --ffn_hidden_size 3072 \
+  --dtype 'bf16'
+```