Update internlm2 math (#349)

Author: Jintao-Huang

README.md

@@ -234,9 +234,10 @@ app_ui_main(infer_args)
     - SQL: text2sql-en, 🔥sql-create-context-en.
     - Text Generation: 🔥advertise-gen-zh, 🔥dureader-robust-zh.
     - Classification: cmnli-zh, 🔥cmnli-mini-zh, 🔥jd-sentiment-zh, 🔥hc3-zh, 🔥hc3-en.
+    - RLHF: 🔥hh-rlhf, stack-exchange-paired.
     - Other: finance-en, poetry-zh, webnovel-zh, generated-chat-zh, cls-fudan-news-zh, ner-jave-zh.
   - Multi-Modal:
-    - Vision: coco-en, 🔥coco-mini-en.
+    - Vision: coco-en, 🔥coco-mini-en, coco-mini-en-2, capcha-images.
     - Audio: aishell1-zh, 🔥aishell1-mini-zh.
   - Custom Dataset
 - Supported Templates:

README_CN.md

@@ -234,9 +234,10 @@ app_ui_main(infer_args)
     - SQL: text2sql-en, 🔥sql-create-context-en.
     - 文本生成: 🔥advertise-gen-zh, 🔥dureader-robust-zh.
     - 分类: cmnli-zh, 🔥cmnli-mini-zh, 🔥jd-sentiment-zh, 🔥hc3-zh, 🔥hc3-en.
+    - RLHF: 🔥hh-rlhf, stack-exchange-paired.
     - 其他: finance-en, poetry-zh, webnovel-zh, generated-chat-zh, cls-fudan-news-zh, ner-jave-zh.
   - 多模态:
-    - 视觉: coco-en, 🔥coco-mini-en.
+    - 视觉: coco-en, 🔥coco-mini-en, coco-mini-en-2, capcha-images.
     - 音频: aishell1-zh, 🔥aishell1-mini-zh.
   - 自定义数据集
 - 支持的对话模板:

docs/source/LLM/LLM微调文档.md

@@ -146,7 +146,7 @@ cd examples/pytorch/llm
 - 如果你使用的是**V100**等较老的GPU, 你需要设置`--dtype AUTO`或者`--dtype fp16`, 因为其不支持bf16.
 - 如果你的机器是A100等高性能显卡, 且使用的是qwen系列模型, 推荐你安装[**flash-attn**](https://github.com/Dao-AILab/flash-attention), 这将会加快训练和推理的速度以及显存占用(A10, 3090, V100等显卡不支持flash-attn进行训练). 支持flash-attn的模型可以查看[LLM支持的模型](./支持的模型和数据集.md#模型)
 - 如果你要进行**二次预训练**, **多轮对话**, 你可以参考[自定义与拓展](./自定义与拓展.md#注册数据集的方式)
-- 如果你需要断网进行训练, 请使用`--model_cache_dir`和设置`--check_model_is_latest false`. 具体参数含义请查看[命令行参数](./命令行参数.md).
+- 如果你需要**断网**进行训练, 请使用`--model_cache_dir`和设置`--check_model_is_latest false`. 具体参数含义请查看[命令行参数](./命令行参数.md).
 - 如果你想在训练时, 将权重push到ModelScope Hub中, 你需要设置`--push_to_hub true`.
 - 如何你想要在推理时, 合并LoRA权重并保存，你需要设置`--merge_lora_and_save true`. **不推荐对qlora训练的模型进行merge**, 这会存在精度损失.
 - 以下提供了可以直接运行的`qwen_7b_chat`的sh脚本(你只需要在推理时指定`--ckpt_dir`即可顺利执行). 更多模型的scripts脚本, 可以查看[scripts文件夹](https://github.com/modelscope/swift/tree/main/examples/pytorch/llm/scripts). 如果你想要**自定义sh脚本**, 推荐你参考`scripts/qwen_7b_chat`中的脚本进行书写.

swift/llm/utils/model.py

@@ -88,6 +88,7 @@ class ModelType:
     internlm_7b_chat_8k = 'internlm-7b-chat-8k'
     internlm_20b = 'internlm-20b'
     internlm_20b_chat = 'internlm-20b-chat'
+    # internlm2
     internlm2_7b_base = 'internlm2-7b-base'
     internlm2_7b = 'internlm2-7b'
     internlm2_7b_sft_chat = 'internlm2-7b-sft-chat'
@@ -96,6 +97,11 @@ class ModelType:
     internlm2_20b = 'internlm2-20b'
     internlm2_20b_sft_chat = 'internlm2-20b-sft-chat'
     internlm2_20b_chat = 'internlm2-20b-chat'
+    # internlm2-math
+    internlm2_math_7b_chat = 'internlm2-math-7b-chat'
+    internlm2_math_7b = 'internlm2-math-7b'
+    internlm2_math_20b_chat = 'internlm2-math-20b-chat'
+    internlm2_math_20b = 'internlm2-math-20b'
     # deepseek
     deepseek_7b = 'deepseek-7b'
     deepseek_7b_chat = 'deepseek-7b-chat'
@@ -120,6 +126,7 @@ class ModelType:
     baichuan_7b = 'baichuan-7b'
     baichuan_13b = 'baichuan-13b'
     baichuan_13b_chat = 'baichuan-13b-chat'
+    # baichuan2
     baichuan2_7b = 'baichuan2-7b'
     baichuan2_7b_chat = 'baichuan2-7b-chat'
     baichuan2_7b_chat_int4 = 'baichuan2-7b-chat-int4'
@@ -911,6 +918,32 @@ def get_model_tokenizer_with_flash_attn(model_dir: str,
                                          load_model, model_config, **kwargs)
 
 
+@register_model(
+    ModelType.internlm2_math_7b,
+    'Shanghai_AI_Laboratory/internlm2-math-base-7b',
+    LoRATM.internlm2,
+    TemplateType.default_generation_bos,
+    support_flash_attn=True)
+@register_model(
+    ModelType.internlm2_math_20b,
+    'Shanghai_AI_Laboratory/internlm2-math-base-20b',
+    LoRATM.internlm2,
+    TemplateType.default_generation_bos,
+    support_flash_attn=True)
+@register_model(
+    ModelType.internlm2_math_7b_chat,
+    'Shanghai_AI_Laboratory/internlm2-math-7b',
+    LoRATM.internlm2,
+    TemplateType.internlm2,
+    eos_token='<|im_end|>',
+    support_flash_attn=True)
+@register_model(
+    ModelType.internlm2_math_20b_chat,
+    'Shanghai_AI_Laboratory/internlm2-math-20b',
+    LoRATM.internlm2,
+    TemplateType.internlm2,
+    eos_token='<|im_end|>',
+    support_flash_attn=True)
 @register_model(
     ModelType.internlm2_7b_sft_chat,
     'Shanghai_AI_Laboratory/internlm2-chat-7b-sft',
@@ -986,9 +1019,7 @@ def get_model_tokenizer_internlm2(model_dir: str,
         if getattr(tokenizer.__class__.eos_token_id, 'fset', None) is None:
             del tokenizer.__class__.eos_token_id
         tokenizer.eos_token = eos_token
-    if model is not None and use_flash_attn:
-        # fix AttributeError: no attribute 'attention_dropout'
-        model.model.layers[0].attention.__class__.attention_dropout = 0.
+
     return model, tokenizer
 
 

swift/llm/utils/template.py

@@ -318,7 +318,7 @@ def _encode(
         if self.max_length is not None:
             if truncation_strategy == 'delete' and len(
                     input_ids) > self.max_length:
-                return None
+                return {}, {}
             input_ids = input_ids[-self.max_length:]
             if labels is not None:
                 labels = labels[-self.max_length:]

swift/llm/utils/utils.py

@@ -218,7 +218,7 @@ def __len__(self) -> int:
 def _single_map(d: Dict[str, Any],
                 map_func: MapFunc) -> Optional[Dict[str, Any]]:
     d = map_func(d)
-    if d is None:
+    if len(d[0]) == 0:
         return None
     return d