Merge branch 'dev' into fix_moe

Author: Yunnglin

.github/copilot-instructions.md

@@ -60,4 +60,4 @@ These instructions help AI agents work productively in this repo. Focus on concr
 - **Sample with vLLM:** Configure `vLLMSampler`, set `Template`/`Processor`, then `sample()` on `Trajectory` list; see [src/twinkle/sampler/vllm_sampler.py](src/twinkle/sampler/vllm_sampler.py).
 
 ---
-Questions or gaps? Tell us where guidance is unclear (e.g., missing run scripts, Ray cluster setup), and we’ll refine this document.
+Questions or gaps? Tell us where guidance is unclear (e.g., missing run scripts, Ray cluster setup), and we’ll refine this document.

.gitignore

@@ -152,4 +152,4 @@ megatron_output/
 ast_index_file.py
 test_cookbook/
 /test*.py
-swanlog/
+swanlog/

.pre-commit-config.yaml

@@ -1,38 +1,44 @@
 repos:
-  - repo: https://github.com/pycqa/flake8.git
+  - repo: https://github.com/pycqa/flake8
     rev: 7.3.0
     hooks:
       - id: flake8
-        exclude: |
-            (?x)^(
-                examples/|
-                cookbook/|
-                client_tools/|
-                src/twinkle_client|
-                
-            )$
-  - repo: https://github.com/PyCQA/isort.git
+        exclude: ^(examples/|cookbook/|client_tools/|src/twinkle_client/)
+
+  - repo: https://github.com/PyCQA/isort
     rev: 7.0.0
     hooks:
       - id: isort
-        exclude: |
-            (?x)^(
-                examples/|
-                cookbook/|
-                client_tools/|
-                src/twinkle_client|
-            )$
+        exclude: ^(examples/|cookbook/|client_tools/|src/twinkle_client/)
+
   - repo: https://github.com/google/yapf
     rev: v0.43.0
     hooks:
       - id: yapf
-        exclude: |
-            (?x)^(
-                examples/|
-                cookbook/|
-                client_tools/|
-                src/twinkle_client|
-            )$
-  - repo: https://github.com/pre-commit/pre-commit-hooks.git
+        exclude: ^(examples/|cookbook/|client_tools/|src/twinkle_client/)
+
+  - repo: https://github.com/asottile/pyupgrade
+    rev: v3.19.1
+    hooks:
+      - id: pyupgrade
+        args: [--py38-plus]
+        exclude: ^client_tools/
+
+  - repo: https://github.com/pre-commit/pre-commit-hooks
     rev: v6.0.0
     hooks:
+      - id: trailing-whitespace
+        exclude: ^client_tools/
+      - id: check-yaml
+        exclude: ^client_tools/
+      - id: end-of-file-fixer
+        exclude: ^client_tools/
+      - id: requirements-txt-fixer
+        exclude: ^client_tools/
+      - id: double-quote-string-fixer
+        exclude: ^client_tools/
+      - id: check-merge-conflict
+        exclude: ^client_tools/
+      - id: mixed-line-ending
+        args: ["--fix=lf"]
+        exclude: ^client_tools/

.pre-commit-config_local.yaml

@@ -64,4 +64,4 @@ twinkle runs two types of tests after a PR is submitted:
 
 Before submitting a PR, please ensure your development code is protected by test cases. For example, smoke tests for new features, or unit tests for various edge cases. Reviewers will also pay attention to this during code review. Additionally, a dedicated service will run CI Tests, executing all test cases. Code can only be merged after all test cases pass.
 
-Please ensure these tests pass successfully.
+Please ensure these tests pass successfully.

CONTRIBUTING.md

@@ -77,6 +77,12 @@ pip install -e .
 
 - 🎉2026-02-10 Initial version of Twinkle✨ released, including SFT/PT/RL for text models and serverless training capabilities on [ModelScope](https://modelscope.cn).
 
+# ModelScope Community
+
+## ModelScope Official Environment
+
+The ModelScope community provides an official environment for running Twinkle. The API endpoint is: [base_url](https://www.modelscope.cn/twinkle). Developers can refer to our [documentation](docs/source_en/Usage%20Guide/ModelScope-Official-Resources.md) for usage instructions.
+
 ## Supported Hardware
 
 | Hardware Environment | Notes                                                            |
@@ -181,6 +187,7 @@ if __name__ == '__main__':
 ### Tinker-Like Remote API
 
 ```python
+import os
 from tqdm import tqdm
 from tinker import types
 from twinkle_client import init_tinker_compat_client
@@ -191,21 +198,20 @@ from twinkle.server.tinker.common import input_feature_to_datum
 
 base_model = "Qwen/Qwen2.5-0.5B-Instruct"
 
-# 使用 Twinkle 的 Dataset 组件加载和预处理数据
+# Use twinkle dataset to load the data
 dataset = Dataset(dataset_meta=DatasetMeta('ms://swift/self-cognition', data_slice=range(500)))
 dataset.set_template('Template', model_id=f'ms://{base_model}', max_length=256)
-dataset.map(SelfCognitionProcessor('twinkle模型', 'twinkle团队'), load_from_cache_file=False)
+dataset.map(SelfCognitionProcessor('twinkle Model', 'twinkle Team'), load_from_cache_file=False)
 dataset.encode(batched=True, load_from_cache_file=False)
 dataloader = DataLoader(dataset=dataset, batch_size=8)
 
-# 初始化 Tinker 兼容客户端
-service_client = init_tinker_compat_client(base_url='http://localhost:8000')
+# Initialize tinker client
+service_client = init_tinker_compat_client(base_url='http://www.modelscope.cn/twinkle', api_key=os.environ.get('MODELSCOPE_SDK_TOKEN'))
 training_client = service_client.create_lora_training_client(base_model=base_model, rank=16)
 
-# 训练循环：使用 input_feature_to_datum 转换数据格式
+# Training loop: use input_feature_to_datum to transfer the input format
 for epoch in range(3):
     for step, batch in tqdm(enumerate(dataloader)):
-        # 将 Twinkle 的 InputFeature 转换为 Tinker 的 Datum
         input_datum = [input_feature_to_datum(input_feature) for input_feature in batch]
 
         fwdbwd_future = training_client.forward_backward(input_datum, "cross_entropy")

README.md

@@ -67,6 +67,10 @@ pip install -e .
 
 - 🎉2026-02-10 Twinkle✨ 初始版本发布，包含文本模型的 SFT/PT/RL 以及在 [ModelScope](https://modelscope.cn) 上的无服务器训练能力。
 
+## 魔搭社区官方环境
+
+魔搭社区提供了Twinkle运行的官方环境，调用端点为:[base_url](https://www.modelscope.cn/twinkle)，开发者可以参考我们的[文档](docs/source_zh/使用指引/魔搭官方环境.md)来进行使用。
+
 ## 支持的硬件
 
 | 硬件环境 | 备注                                                            |
@@ -169,6 +173,7 @@ if __name__ == '__main__':
 ### Tinker兼容的远程训练
 
 ```python
+import os
 from tqdm import tqdm
 from tinker import types
 from twinkle_client import init_tinker_compat_client
@@ -187,7 +192,7 @@ dataset.encode(batched=True, load_from_cache_file=False)
 dataloader = DataLoader(dataset=dataset, batch_size=8)
 
 # 初始化 Tinker 兼容客户端
-service_client = init_tinker_compat_client(base_url='http://localhost:8000')
+service_client = init_tinker_compat_client(base_url='http://www.modelscope.cn/twinkle', api_key=os.environ.get('MODELSCOPE_SDK_TOKEN'))
 training_client = service_client.create_lora_training_client(base_model=base_model, rank=16)
 
 # 训练循环：使用 input_feature_to_datum 转换数据格式
@@ -320,4 +325,4 @@ python3 train.py
 
 | 组件类型 | 组件链接                                                                                           | 组件功能                                                                      | 作者              |
 | -------------- | -------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------- | ------------------- |
-| Patch          | [qwen3_moe_transformers4_patch](https://www.modelscope.cn/models/twinkle-kit/qwen3_moe_transformers4_patch) | 修复 Qwen3 MoE 模型在 FSDP2 训练期间的挂起问题，对 transformers==4.x 有效 | ModelScope 官方 |
+| Patch          | [qwen3_moe_transformers4_patch](https://www.modelscope.cn/models/twinkle-kit/qwen3_moe_transformers4_patch) | 修复 Qwen3 MoE 模型在 FSDP2 训练期间的挂起问题，对 transformers==4.x 有效 | ModelScope 官方 |

README_ZH.md

@@ -86,4 +86,3 @@
 - [ ] Support for batched preprocessing
 
 ### Networking Capabilities
-