fix (#107)

Author: tastelikefeet

cookbook/client/twinkle/self_congnition.py

@@ -21,8 +21,7 @@
 
 logger = get_logger()
 
-# Whether to use Megatron for training
-use_megatron = True
+
 # Step 2: Initialize the Twinkle client to communicate with the remote server.
 # - base_url: the address of the running Twinkle server
 # - api_key: authentication token (loaded from environment variable)
@@ -88,8 +87,7 @@ def train():
     model.set_optimizer('Adam', lr=1e-4)
 
     # Use a linear learning rate scheduler (Do not support LR scheduler if server use megatron)
-    if not use_megatron:
-        model.set_lr_scheduler('LinearLR')
+    model.set_lr_scheduler('LinearLR')
 
     # Step 6: Optionally resume from a previous checkpoint
     if resume_path:

docs/source_en/Usage Guide/Introduction-with-Qwen3.5.md

@@ -1,6 +1,6 @@
-# Qwen3.5-4B Training Best Practices
+# Qwen3.5 Training Best Practices
 
-Using Qwen3.5-4B as an example, this guide demonstrates the core capability of the Twinkle framework: **one codebase, from single GPU training to Client-Server mode**.
+Using Qwen3.5-4B as an example, this guide demonstrates the core capability of the Twinkle framework: **one component-based code, used from single GPU training to Client-Server mode**.
 
 ---
 
@@ -364,7 +364,6 @@ Twinkle Client provides an API nearly identical to local training, ideal for sce
 import dotenv
 dotenv.load_dotenv('.env')
 
-import os
 from peft import LoraConfig
 
 from twinkle import get_logger
@@ -376,9 +375,6 @@ from twinkle_client.model import MultiLoraTransformersModel
 
 logger = get_logger()
 
-# Whether to use the Megatron backend
-use_megatron = True
-
 # Initialize the Twinkle client
 client = init_twinkle_client(base_url='http://127.0.0.1:8000', api_key='EMPTY_TOKEN')
 
@@ -410,11 +406,8 @@ def train():
     model.set_template('Template')
     model.set_processor('InputProcessor', padding_side='right')
     model.set_loss('CrossEntropyLoss')
-    model.set_optimizer('Adam', lr=1e-4)
-
-    # LR scheduler not supported with Megatron backend
-    if not use_megatron:
-        model.set_lr_scheduler('LinearLR')
+    model.set_optimizer('AdamW', lr=1e-4)
+    model.set_lr_scheduler('LinearLR')
 
     # Resume from checkpoint if available
     if resume_path:
@@ -473,15 +466,15 @@ init_tinker_client()
 from tinker import ServiceClient
 
 # Base model
-base_model = 'Qwen/Qwen3-30B-A3B-Instruct-2507'
+base_model = 'Qwen/Qwen3.5-4B'
 base_url = 'http://www.modelscope.cn/twinkle'
 
 
 def train():
     # Prepare dataset
     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 model', 'twinkle team'), load_from_cache_file=False)
+    dataset.map(SelfCognitionProcessor('Twinkle Model', 'ModelScope Team'), load_from_cache_file=False)
     dataset.encode(batched=True, load_from_cache_file=False)
     dataloader = DataLoader(dataset=dataset, batch_size=8)
 

docs/source_zh/使用指引/Qwen3.5最佳实践.md

@@ -1,6 +1,6 @@
-# Qwen3.5-4B 训练最佳实践
+# Qwen3.5 训练最佳实践
 
-本文以 Qwen3.5-4B 为例，演示 Twinkle 框架的核心能力：**一套代码，从单卡训练到Client-Server环境**。
+本文以 Qwen3.5-4B 为例，演示 Twinkle 框架的核心能力：**一套组件化代码，从单卡训练到Client-Server环境**。
 
 ---
 
@@ -364,7 +364,6 @@ Twinkle Client 提供与本地训练几乎完全一致的 API，适合需要精
 import dotenv
 dotenv.load_dotenv('.env')
 
-import os
 from peft import LoraConfig
 
 from twinkle import get_logger
@@ -376,9 +375,6 @@ from twinkle_client.model import MultiLoraTransformersModel
 
 logger = get_logger()
 
-# 是否使用 Megatron 后端
-use_megatron = True
-
 # 初始化 Twinkle 客户端
 client = init_twinkle_client(base_url='http://127.0.0.1:8000', api_key='EMPTY_TOKEN')
 
@@ -410,11 +406,8 @@ def train():
     model.set_template('Template')
     model.set_processor('InputProcessor', padding_side='right')
     model.set_loss('CrossEntropyLoss')
-    model.set_optimizer('Adam', lr=1e-4)
-
-    # Megatron 后端暂不支持 LR 调度器
-    if not use_megatron:
-        model.set_lr_scheduler('LinearLR')
+    model.set_optimizer('AdamW', lr=1e-4)
+    model.set_lr_scheduler('LinearLR')
 
     # 恢复训练（如有检查点）
     if resume_path:
@@ -473,15 +466,15 @@ init_tinker_client()
 from tinker import ServiceClient
 
 # 基座模型
-base_model = 'Qwen/Qwen3-30B-A3B-Instruct-2507'
+base_model = 'Qwen/Qwen3.5-4B'
 base_url = 'http://www.modelscope.cn/twinkle'
 
 
 def train():
     # 准备数据集
     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模型', 'ModelScope团队'), load_from_cache_file=False)
     dataset.encode(batched=True, load_from_cache_file=False)
     dataloader = DataLoader(dataset=dataset, batch_size=8)