From 5aa6751e52b5c2e06727c50e60218226b146b7bf Mon Sep 17 00:00:00 2001
From: "yuze.zyz" <yuze.zyz@alibaba-inc.com>
Date: Fri, 1 Dec 2023 16:11:30 +0800
Subject: [PATCH] add readme

---
 README.md                        | 29 +++++++++++++++++++++++++++++
 README_zh.md                     | 29 +++++++++++++++++++++++++++++
 src/llmtuner/extras/constants.py |  2 +-
 src/llmtuner/model/loader.py     |  4 ++--
 4 files changed, 61 insertions(+), 3 deletions(-)

diff --git a/README.md b/README.md
index 7ef3b6f1..0a03d61c 100644
--- a/README.md
+++ b/README.md
@@ -51,6 +51,8 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Changelog
 
+[23/12/01] We supported **[ModelScope Hub](https://www.modelscope.cn/models)** to accelerate model downloading. Add environment variable `USE_MODELSCOPE_HUB=1` to your command line, then you can use the model-id of ModelScope Hub.
+
 [23/10/21] We supported **[NEFTune](https://arxiv.org/abs/2310.05914)** trick for fine-tuning. Try `--neft_alpha` argument to activate NEFTune, e.g., `--neft_alpha 5`.
 
 [23/09/27] We supported **$S^2$-Attn** proposed by [LongLoRA](https://github.com/dvlab-research/LongLoRA) for the LLaMA models. Try `--shift_attn` argument to enable shift short attention.
@@ -227,6 +229,33 @@ If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you wi
 pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
 ```
 
+### Use ModelScope Models
+
+If you have trouble with downloading models from HuggingFace, we have supported ModelScope Hub. To use LLaMA-Factory together with ModelScope, please add a environment variable:
+
+```shell
+export USE_MODELSCOPE_HUB=1
+```
+
+> [!NOTE]
+>
+> Please use integers only. 0 or not set for using HuggingFace hub. Other values will be treated as use ModelScope hub.
+
+Then you can use LLaMA-Factory with ModelScope model-ids:
+
+```shell
+python src/train_bash.py \
+    --model_name_or_path ZhipuAI/chatglm3-6b \
+    ... other arguments
+# You can find all model ids in this link: https://www.modelscope.cn/models
+```
+
+Web demo also supports ModelScope, after setting the environment variable please run with this command:
+
+```shell
+CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+```
+
 ### Train on a single GPU
 
 > [!IMPORTANT]
diff --git a/README_zh.md b/README_zh.md
index d6f01d2f..9c92c261 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -51,6 +51,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 ## 更新日志
 
+[23/12/01] 我们支持了 **[魔搭ModelHub](https://www.modelscope.cn/models)** 进行模型下载加速。在启动命令前环境变量中增加 `USE_MODELSCOPE_HUB=1` 即可开启。
+
 [23/10/21] 我们支持了 **[NEFTune](https://arxiv.org/abs/2310.05914)** 训练技巧。请使用 `--neft_alpha` 参数启用 NEFTune，例如 `--neft_alpha 5`。
 
 [23/09/27] 我们针对 LLaMA 模型支持了 [LongLoRA](https://github.com/dvlab-research/LongLoRA) 提出的 **$S^2$-Attn**。请使用 `--shift_attn` 参数以启用该功能。
@@ -227,6 +229,33 @@ pip install -r requirements.txt
 pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
 ```
 
+### 使用魔搭的模型
+
+如果下载HuggingFace模型存在问题，我们已经支持了魔搭的ModelHub，只需要添加一个环境变量：
+
+```shell
+export USE_MODELSCOPE_HUB=1
+```
+
+> [!NOTE]
+>
+> 该环境变量仅支持整数，0或者不设置代表使用HuggingFace，其他值代表使用ModelScope
+
+之后就可以在命令行中指定魔搭的模型id：
+
+```shell
+python src/train_bash.py \
+    --model_name_or_path ZhipuAI/chatglm3-6b \
+    ... other arguments
+# 在这个链接中可以看到所有可用模型: https://www.modelscope.cn/models
+```
+
+Web demo目前也支持了魔搭, 在设置环境变量后即可使用：
+
+```shell
+CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+```
+
 ### 单 GPU 训练
 
 > [!IMPORTANT]
diff --git a/src/llmtuner/extras/constants.py b/src/llmtuner/extras/constants.py
index 3930a2e3..c865b4f8 100644
--- a/src/llmtuner/extras/constants.py
+++ b/src/llmtuner/extras/constants.py
@@ -43,7 +43,7 @@ def register_model_group(
         else:
             assert prefix == name.split("-")[0], "prefix should be identical."
 
-        if not os.environ.get('USE_MODELSCOPE_HUB', False):
+        if not int(os.environ.get('USE_MODELSCOPE_HUB', '0')):
             # If path is a string, we treat it as a huggingface model-id by default.
             SUPPORTED_MODELS[name] = path["hf"] if isinstance(path, dict) else path
         elif isinstance(path, dict) and "ms" in path:
diff --git a/src/llmtuner/model/loader.py b/src/llmtuner/model/loader.py
index 85e7ed31..122cd7f2 100644
--- a/src/llmtuner/model/loader.py
+++ b/src/llmtuner/model/loader.py
@@ -235,7 +235,7 @@ def load_model_and_tokenizer(
 
 
 def try_download_model_from_ms(model_args):
-    if os.environ.get('USE_MODELSCOPE_HUB', False) and not os.path.exists(model_args.model_name_or_path):
+    if int(os.environ.get('USE_MODELSCOPE_HUB', '0')) and not os.path.exists(model_args.model_name_or_path):
         try:
             from modelscope import snapshot_download
             revision = model_args.model_revision
@@ -243,5 +243,5 @@ def try_download_model_from_ms(model_args):
                 revision = 'master'
             model_args.model_name_or_path = snapshot_download(model_args.model_name_or_path, revision)
         except ImportError as e:
-            raise ImportError(f'You are using `USE_MODELSCOPE_HUB=True` but you have no modelscope sdk installed. '
+            raise ImportError(f'You are using `USE_MODELSCOPE_HUB=1` but you have no modelscope sdk installed. '
                               f'Please install it by `pip install modelscope -U`') from e