ms-swift 4.5.0.dev0__tar.gz → 4.5.2__tar.gz

{ms_swift-4.5.0.dev0/ms_swift.egg-info → ms_swift-4.5.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ms_swift
-Version: 4.5.0.dev0
+Version: 4.5.2
 Summary: Swift: Scalable lightWeight Infrastructure for Fine-Tuning
 Home-page: https://github.com/modelscope/ms-swift
 Author: DAMO ModelScope teams
@@ -53,13 +53,13 @@ Requires-Dist: sortedcontainers>=1.5.9
 Requires-Dist: tensorboard
 Requires-Dist: tiktoken
 Requires-Dist: tqdm
-Requires-Dist: transformers<5.13.0,>=4.33
+Requires-Dist: transformers<5.17.0,>=4.33
 Requires-Dist: transformers_stream_generator
 Requires-Dist: trl<1.0,>=0.15
 Requires-Dist: uvicorn
 Requires-Dist: zstandard
 Provides-Extra: megatron
-Requires-Dist: mcore-bridge>=1.5.0; extra == "megatron"
+Requires-Dist: mcore-bridge>=1.6.2; extra == "megatron"
 Requires-Dist: megatron-core>=0.16; extra == "megatron"
 Requires-Dist: peft>=0.15; extra == "megatron"
 Provides-Extra: eval
@@ -105,7 +105,7 @@ Requires-Dist: sortedcontainers>=1.5.9; extra == "all"
 Requires-Dist: tensorboard; extra == "all"
 Requires-Dist: tiktoken; extra == "all"
 Requires-Dist: tqdm; extra == "all"
-Requires-Dist: transformers<5.13.0,>=4.33; extra == "all"
+Requires-Dist: transformers<5.17.0,>=4.33; extra == "all"
 Requires-Dist: transformers_stream_generator; extra == "all"
 Requires-Dist: trl<1.0,>=0.15; extra == "all"
 Requires-Dist: uvicorn; extra == "all"
@@ -195,6 +195,7 @@ You can contact us and communicate with us by adding our group:
 
 
 ## 🎉 News
+- 🎁 2026.08.11: Support inference and training for the multimodal model [Muse-Glimmer-30B](https://modelscope.cn/models/meta-models/Muse-Glimmer-30B), with Megatron-SWIFT support as well.
 - 🎁 2026.07.22: Support inference for the moonshotai multimodal model [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3), including the XTML chat template, thinking channel (`reasoning_effort`) and tool calling (`--agent_template kimi_k3`).
 - 🎁 2026.06.10: Megatron-Ray now supports GRPO and GKD training. See [docs](./docs/source_en/Instruction/Ray.md) and [examples](examples/ray).
 - 🎁 2026.03.03: **ms-swift v4.0** major version is officially released. For release notes, please refer to [here](https://github.com/modelscope/ms-swift/releases/tag/v4.0.0). You can provide your suggestions to us in [this issue](https://github.com/modelscope/ms-swift/issues/7250). Thank you for your support.
@@ -222,7 +223,7 @@ You can contact us and communicate with us by adding our group:
 - 🎁 2025.02.21: The GRPO algorithm now supports LMDeploy, with the training script available [here](examples/train/grpo/internal/full_lmdeploy.sh). Additionally, the performance of the GRPO algorithm has been tested, achieving a training speed increase of up to 300% using various tricks. Please check the WanDB table [here](https://wandb.ai/tastelikefeet/grpo_perf_test?nw=nwuseryuzezyz).
 - 🎁 2025.02.21: The `swift sample` command is now supported. The reinforcement fine-tuning script can be found [here](docs/source_en/Instruction/Reinforced-Fine-tuning.md), and the large model API distillation sampling script is available [here](examples/sampler/distill/distill.sh).
 - 🔥 2025.02.12: Support for the GRPO (Group Relative Policy Optimization) training algorithm has been added. Documentation is available [here](docs/source_en/Instruction/GRPO/GetStarted/GRPO.md).
-- 🎁 2024.12.04: Major update to **ms-swift 3.0**. Please refer to the [release notes and changes](docs/source_en/Instruction/ReleaseNote3.0.md).
+- 🎁 2024.12.04: Major update to **ms-swift 3.0**. Please refer to the [release notes and changes](https://github.com/modelscope/ms-swift/releases/tag/v3.0.0).
 - 🎉 2024.08.12: The ms-swift paper has been published on arXiv and can be read [here](https://arxiv.org/abs/2408.05517).
 - 🔥 2024.08.05: Support for using [evalscope](https://github.com/modelscope/evalscope/) as a backend for evaluating large models and multimodal models.
 - 🔥 2024.07.29: Support for using [vllm](https://github.com/vllm-project/vllm) and [lmdeploy](https://github.com/InternLM/lmdeploy) to accelerate inference for large models and multimodal models. When performing infer/deploy/eval, you can specify `--infer_backend vllm/lmdeploy`.
@@ -264,7 +265,7 @@ Running Environment:
 | transformers | >=4.33       | 4.57.6/5.12.1              |                          |
 | modelscope   | >=1.23       |                     |                                           |
 | datasets     | >=3.0,<4.8.5 | 3.6.0/4.8.4         |                    |
-| peft         | >=0.11,<0.20 |                     |                                           |
+| peft         | >=0.11,<0.21 |                     |                                           |
 | flash_attn   |              | 2.8.3/4.0.0b15 |                                           |
 | trl          | >=0.15,<1.0 | 0.29.1              | RLHF                                      |
 | deepspeed    | >=0.14       | 0.18.9              | Training                                  |
@@ -624,4 +625,4 @@ This framework is licensed under the [Apache License (Version 2.0)](https://gith
 
 ## Star History
 
-[![Star History Chart](https://api.star-history.com/svg?repos=modelscope/ms-swift&type=Date)](https://star-history.com/#modelscope/ms-swift&Date)
+[![Star History Chart](https://star-history.dera.page/svg?repos=modelscope/ms-swift&type=Date)](https://star-history.dera.page/#modelscope/ms-swift&Date)

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/README.md

@@ -77,6 +77,7 @@ You can contact us and communicate with us by adding our group:
 
 
 ## 🎉 News
+- 🎁 2026.08.11: Support inference and training for the multimodal model [Muse-Glimmer-30B](https://modelscope.cn/models/meta-models/Muse-Glimmer-30B), with Megatron-SWIFT support as well.
 - 🎁 2026.07.22: Support inference for the moonshotai multimodal model [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3), including the XTML chat template, thinking channel (`reasoning_effort`) and tool calling (`--agent_template kimi_k3`).
 - 🎁 2026.06.10: Megatron-Ray now supports GRPO and GKD training. See [docs](./docs/source_en/Instruction/Ray.md) and [examples](examples/ray).
 - 🎁 2026.03.03: **ms-swift v4.0** major version is officially released. For release notes, please refer to [here](https://github.com/modelscope/ms-swift/releases/tag/v4.0.0). You can provide your suggestions to us in [this issue](https://github.com/modelscope/ms-swift/issues/7250). Thank you for your support.
@@ -104,7 +105,7 @@ You can contact us and communicate with us by adding our group:
 - 🎁 2025.02.21: The GRPO algorithm now supports LMDeploy, with the training script available [here](examples/train/grpo/internal/full_lmdeploy.sh). Additionally, the performance of the GRPO algorithm has been tested, achieving a training speed increase of up to 300% using various tricks. Please check the WanDB table [here](https://wandb.ai/tastelikefeet/grpo_perf_test?nw=nwuseryuzezyz).
 - 🎁 2025.02.21: The `swift sample` command is now supported. The reinforcement fine-tuning script can be found [here](docs/source_en/Instruction/Reinforced-Fine-tuning.md), and the large model API distillation sampling script is available [here](examples/sampler/distill/distill.sh).
 - 🔥 2025.02.12: Support for the GRPO (Group Relative Policy Optimization) training algorithm has been added. Documentation is available [here](docs/source_en/Instruction/GRPO/GetStarted/GRPO.md).
-- 🎁 2024.12.04: Major update to **ms-swift 3.0**. Please refer to the [release notes and changes](docs/source_en/Instruction/ReleaseNote3.0.md).
+- 🎁 2024.12.04: Major update to **ms-swift 3.0**. Please refer to the [release notes and changes](https://github.com/modelscope/ms-swift/releases/tag/v3.0.0).
 - 🎉 2024.08.12: The ms-swift paper has been published on arXiv and can be read [here](https://arxiv.org/abs/2408.05517).
 - 🔥 2024.08.05: Support for using [evalscope](https://github.com/modelscope/evalscope/) as a backend for evaluating large models and multimodal models.
 - 🔥 2024.07.29: Support for using [vllm](https://github.com/vllm-project/vllm) and [lmdeploy](https://github.com/InternLM/lmdeploy) to accelerate inference for large models and multimodal models. When performing infer/deploy/eval, you can specify `--infer_backend vllm/lmdeploy`.
@@ -146,7 +147,7 @@ Running Environment:
 | transformers | >=4.33       | 4.57.6/5.12.1              |                          |
 | modelscope   | >=1.23       |                     |                                           |
 | datasets     | >=3.0,<4.8.5 | 3.6.0/4.8.4         |                    |
-| peft         | >=0.11,<0.20 |                     |                                           |
+| peft         | >=0.11,<0.21 |                     |                                           |
 | flash_attn   |              | 2.8.3/4.0.0b15 |                                           |
 | trl          | >=0.15,<1.0 | 0.29.1              | RLHF                                      |
 | deepspeed    | >=0.14       | 0.18.9              | Training                                  |
@@ -506,4 +507,4 @@ This framework is licensed under the [Apache License (Version 2.0)](https://gith
 
 ## Star History
 
-[![Star History Chart](https://api.star-history.com/svg?repos=modelscope/ms-swift&type=Date)](https://star-history.com/#modelscope/ms-swift&Date)
+[![Star History Chart](https://star-history.dera.page/svg?repos=modelscope/ms-swift&type=Date)](https://star-history.dera.page/#modelscope/ms-swift&Date)

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2/ms_swift.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ms_swift
-Version: 4.5.0.dev0
+Version: 4.5.2
 Summary: Swift: Scalable lightWeight Infrastructure for Fine-Tuning
 Home-page: https://github.com/modelscope/ms-swift
 Author: DAMO ModelScope teams
@@ -53,13 +53,13 @@ Requires-Dist: sortedcontainers>=1.5.9
 Requires-Dist: tensorboard
 Requires-Dist: tiktoken
 Requires-Dist: tqdm
-Requires-Dist: transformers<5.13.0,>=4.33
+Requires-Dist: transformers<5.17.0,>=4.33
 Requires-Dist: transformers_stream_generator
 Requires-Dist: trl<1.0,>=0.15
 Requires-Dist: uvicorn
 Requires-Dist: zstandard
 Provides-Extra: megatron
-Requires-Dist: mcore-bridge>=1.5.0; extra == "megatron"
+Requires-Dist: mcore-bridge>=1.6.2; extra == "megatron"
 Requires-Dist: megatron-core>=0.16; extra == "megatron"
 Requires-Dist: peft>=0.15; extra == "megatron"
 Provides-Extra: eval
@@ -105,7 +105,7 @@ Requires-Dist: sortedcontainers>=1.5.9; extra == "all"
 Requires-Dist: tensorboard; extra == "all"
 Requires-Dist: tiktoken; extra == "all"
 Requires-Dist: tqdm; extra == "all"
-Requires-Dist: transformers<5.13.0,>=4.33; extra == "all"
+Requires-Dist: transformers<5.17.0,>=4.33; extra == "all"
 Requires-Dist: transformers_stream_generator; extra == "all"
 Requires-Dist: trl<1.0,>=0.15; extra == "all"
 Requires-Dist: uvicorn; extra == "all"
@@ -195,6 +195,7 @@ You can contact us and communicate with us by adding our group:
 
 
 ## 🎉 News
+- 🎁 2026.08.11: Support inference and training for the multimodal model [Muse-Glimmer-30B](https://modelscope.cn/models/meta-models/Muse-Glimmer-30B), with Megatron-SWIFT support as well.
 - 🎁 2026.07.22: Support inference for the moonshotai multimodal model [Kimi-K3](https://modelscope.cn/models/moonshotai/Kimi-K3), including the XTML chat template, thinking channel (`reasoning_effort`) and tool calling (`--agent_template kimi_k3`).
 - 🎁 2026.06.10: Megatron-Ray now supports GRPO and GKD training. See [docs](./docs/source_en/Instruction/Ray.md) and [examples](examples/ray).
 - 🎁 2026.03.03: **ms-swift v4.0** major version is officially released. For release notes, please refer to [here](https://github.com/modelscope/ms-swift/releases/tag/v4.0.0). You can provide your suggestions to us in [this issue](https://github.com/modelscope/ms-swift/issues/7250). Thank you for your support.
@@ -222,7 +223,7 @@ You can contact us and communicate with us by adding our group:
 - 🎁 2025.02.21: The GRPO algorithm now supports LMDeploy, with the training script available [here](examples/train/grpo/internal/full_lmdeploy.sh). Additionally, the performance of the GRPO algorithm has been tested, achieving a training speed increase of up to 300% using various tricks. Please check the WanDB table [here](https://wandb.ai/tastelikefeet/grpo_perf_test?nw=nwuseryuzezyz).
 - 🎁 2025.02.21: The `swift sample` command is now supported. The reinforcement fine-tuning script can be found [here](docs/source_en/Instruction/Reinforced-Fine-tuning.md), and the large model API distillation sampling script is available [here](examples/sampler/distill/distill.sh).
 - 🔥 2025.02.12: Support for the GRPO (Group Relative Policy Optimization) training algorithm has been added. Documentation is available [here](docs/source_en/Instruction/GRPO/GetStarted/GRPO.md).
-- 🎁 2024.12.04: Major update to **ms-swift 3.0**. Please refer to the [release notes and changes](docs/source_en/Instruction/ReleaseNote3.0.md).
+- 🎁 2024.12.04: Major update to **ms-swift 3.0**. Please refer to the [release notes and changes](https://github.com/modelscope/ms-swift/releases/tag/v3.0.0).
 - 🎉 2024.08.12: The ms-swift paper has been published on arXiv and can be read [here](https://arxiv.org/abs/2408.05517).
 - 🔥 2024.08.05: Support for using [evalscope](https://github.com/modelscope/evalscope/) as a backend for evaluating large models and multimodal models.
 - 🔥 2024.07.29: Support for using [vllm](https://github.com/vllm-project/vllm) and [lmdeploy](https://github.com/InternLM/lmdeploy) to accelerate inference for large models and multimodal models. When performing infer/deploy/eval, you can specify `--infer_backend vllm/lmdeploy`.
@@ -264,7 +265,7 @@ Running Environment:
 | transformers | >=4.33       | 4.57.6/5.12.1              |                          |
 | modelscope   | >=1.23       |                     |                                           |
 | datasets     | >=3.0,<4.8.5 | 3.6.0/4.8.4         |                    |
-| peft         | >=0.11,<0.20 |                     |                                           |
+| peft         | >=0.11,<0.21 |                     |                                           |
 | flash_attn   |              | 2.8.3/4.0.0b15 |                                           |
 | trl          | >=0.15,<1.0 | 0.29.1              | RLHF                                      |
 | deepspeed    | >=0.14       | 0.18.9              | Training                                  |
@@ -624,4 +625,4 @@ This framework is licensed under the [Apache License (Version 2.0)](https://gith
 
 ## Star History
 
-[![Star History Chart](https://api.star-history.com/svg?repos=modelscope/ms-swift&type=Date)](https://star-history.com/#modelscope/ms-swift&Date)
+[![Star History Chart](https://star-history.dera.page/svg?repos=modelscope/ms-swift&type=Date)](https://star-history.dera.page/#modelscope/ms-swift&Date)

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/ms_swift.egg-info/SOURCES.txt

@@ -150,6 +150,7 @@ swift/loss_scale/config/agentflan.json
 swift/loss_scale/config/alpha_umi.json
 swift/loss_scale/config/hermes.json
 swift/loss_scale/config/ignore_empty_think.json
+swift/loss_scale/config/ignore_think_prefix.json
 swift/loss_scale/config/qwen.json
 swift/loss_scale/config/react.json
 swift/megatron/__init__.py
@@ -240,6 +241,7 @@ swift/model/models/mistral.py
 swift/model/models/mllm.py
 swift/model/models/moonshot.py
 swift/model/models/mplug.py
+swift/model/models/nvidia.py
 swift/model/models/openbuddy.py
 swift/model/models/qwen.py
 swift/model/models/seed.py
@@ -403,6 +405,7 @@ swift/template/templates/llm.py
 swift/template/templates/megrez.py
 swift/template/templates/microsoft.py
 swift/template/templates/midashenglm.py
+swift/template/templates/mimo.py
 swift/template/templates/minicpm.py
 swift/template/templates/minimax.py
 swift/template/templates/minimind.py
@@ -410,6 +413,7 @@ swift/template/templates/mistral.py
 swift/template/templates/molmo.py
 swift/template/templates/moonshot.py
 swift/template/templates/mplug.py
+swift/template/templates/museglimmer.py
 swift/template/templates/openbuddy.py
 swift/template/templates/pixtral.py
 swift/template/templates/qwen.py

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/ms_swift.egg-info/requires.txt

@@ -32,7 +32,7 @@ sortedcontainers>=1.5.9
 tensorboard
 tiktoken
 tqdm
-transformers<5.13.0,>=4.33
+transformers<5.17.0,>=4.33
 transformers_stream_generator
 trl<1.0,>=0.15
 uvicorn
@@ -73,7 +73,7 @@ sortedcontainers>=1.5.9
 tensorboard
 tiktoken
 tqdm
-transformers<5.13.0,>=4.33
+transformers<5.17.0,>=4.33
 transformers_stream_generator
 trl<1.0,>=0.15
 uvicorn
@@ -90,7 +90,7 @@ evalscope[opencompass]
 evalscope[vlmeval]
 
 [megatron]
-mcore-bridge>=1.5.0
+mcore-bridge>=1.6.2
 megatron-core>=0.16
 peft>=0.15
 

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/requirements/framework.txt

@@ -32,7 +32,7 @@ sortedcontainers>=1.5.9
 tensorboard
 tiktoken
 tqdm
-transformers>=4.33,<5.13.0
+transformers>=4.33,<5.17.0
 transformers_stream_generator
 trl>=0.15,<1.0
 uvicorn

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/requirements/megatron.txt

@@ -1,3 +1,3 @@
-mcore-bridge>=1.5.0
+mcore-bridge>=1.6.2
 megatron-core>=0.16
 peft>=0.15

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/agent_template/base.py

@@ -109,6 +109,19 @@ class ReactCompatMixin:
                 res.append(tool_message['content'])
         return assistant_content, res
 
+    def _format_react_standalone_tool_responses(self, tool_messages) -> 'Prompt':
+        assert len(tool_messages) > 0
+        res = []
+        for tool_message in tool_messages:
+            tool_content = tool_message['content']
+            res.extend(['\n', self.keyword.observation, '\n', tool_content])
+        return res
+
+    def _format_standalone_tool_responses(self, tool_messages) -> 'Prompt':
+        """Format observations that do not follow an assistant tool call."""
+        raise NotImplementedError(
+            f'Agent template {self.__class__.__name__} does not support standalone tool responses.')
+
     @staticmethod
     def _parse_tool_call(content) -> Dict[str, Any]:
         obj = BaseAgentTemplate._parse_json(content)

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/agent_template/deepseek_v4.py

@@ -92,7 +92,14 @@ class DeepSeekV4AgentTemplate(BaseAgentTemplate):
         return functions
 
     def _get_tool_responses(self, tool_messages):
-        return ''.join(f'<tool_result>{tool_message["content"]}</tool_result>' for tool_message in tool_messages)
+        # The official encoding merges tool results into one user turn, joining the
+        # `<tool_result>` blocks with a blank line.
+        return '\n\n'.join(f'<tool_result>{tool_message["content"]}</tool_result>' for tool_message in tool_messages)
+
+    def _add_tool_call_prefix(self, tool_content: str, pre_message=None) -> str:
+        # The official encoding always renders the tool_calls block as `\n\n` + block,
+        # right after the assistant's (possibly empty) textual content.
+        return '\n\n' + tool_content
 
     def _format_tool_responses(
         self,

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/agent_template/glm4.py

@@ -62,6 +62,12 @@ class ChatGLM4AgentTemplate(BaseAgentTemplate):
         res.append('<|assistant|>\n')
         return assistant_content, res
 
+    def _format_standalone_tool_responses(self, tool_messages) -> 'Prompt':
+        res = []
+        for tool_message in tool_messages:
+            res.extend(['<|observation|>\n', tool_message['content']])
+        return res
+
     def _format_tool_calls(self, tool_call_messages) -> str:
         tool_calls = []
         for message in tool_call_messages:
@@ -154,6 +160,16 @@ class GLM4_5AgentTemplate(BaseAgentTemplate):
             res.append('<|assistant|>')
         return assistant_content, res
 
+    def _format_standalone_tool_responses(self, tool_messages) -> 'Prompt':
+        res = ['<|observation|>']
+        for tool_message in tool_messages:
+            tool_content = tool_message['content']
+            if self.model_type == 'glm4_5':
+                res.append(f'\n<tool_response>\n{tool_content}\n</tool_response>')
+            elif self.model_type in {'glm4_7', 'glm5_1'}:
+                res.append(f'<tool_response>{tool_content}</tool_response>')
+        return res
+
     def _format_tool_calls(self, tool_call_messages) -> str:
         tool_calls = []
         for message in tool_call_messages:

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/agent_template/hermes.py

@@ -54,6 +54,23 @@ class HermesAgentTemplate(BaseAgentTemplate):
             res.append(context)
         return assistant_content, res
 
+    def _format_standalone_tool_responses(self, tool_messages) -> 'Prompt':
+        """Render standalone results as a native tool-response user turn."""
+        if not hasattr(self, 'template_meta'):
+            return super()._format_standalone_tool_responses(tool_messages)
+
+        res = (self.template_meta.chat_sep or []).copy()
+        total_tool = self._get_tool_responses(tool_messages)
+        for context in self.template_meta.prompt:
+            if isinstance(context, str) and '{{QUERY}}' in context:
+                query_prefix = context.split('{{QUERY}}', maxsplit=1)[0]
+                if query_prefix:
+                    res.append(query_prefix)
+                res.append(total_tool)
+                return res
+            res.append(context)
+        raise ValueError(f'Template prompt does not contain {{{{QUERY}}}}: {self.template_meta.prompt}')
+
     def _format_tools(self, tools: List[Union[str, dict]], system: Optional[str] = None, user_message=None) -> str:
         tool_descs = [json.dumps(self.wrap_tool(tool), ensure_ascii=False) for tool in tools]
         system = system or ''

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/agent_template/react.py

@@ -6,6 +6,9 @@ from .base import BaseAgentTemplate
 
 class ReactEnAgentTemplate(BaseAgentTemplate):
 
+    def _format_standalone_tool_responses(self, tool_messages):
+        return self._format_react_standalone_tool_responses(tool_messages)
+
     def _format_tools(self, tools: List[Union[str, dict]], system: Optional[str] = None, user_message=None) -> str:
         tool_names = []
         tool_descs = []
@@ -38,6 +41,9 @@ Begin!
 
 class ReactZnAgentTemplate(BaseAgentTemplate):
 
+    def _format_standalone_tool_responses(self, tool_messages):
+        return self._format_react_standalone_tool_responses(tool_messages)
+
     def _format_tools(self, tools: List[Union[str, dict]], system: Optional[str] = None, user_message=None) -> str:
         tool_names = []
         tool_descs = []

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/arguments/base_args/template_args.py

@@ -64,7 +64,8 @@ class TemplateArguments:
             has no preprocessing overhead, but packing offers faster training speeds and more stable memory usage.
         loss_scale (str): Loss weight configuration for training tokens. Default is `'default'`.
             loss_scale includes 3 basic strategies: 'default', 'last_round', 'all', and other strategies:
-            'ignore_empty_think' and agent-specific ones: 'react', 'hermes', 'qwen', 'agentflan', 'alpha_umi', etc.
+            'ignore_empty_think', 'ignore_think_prefix', and agent-specific ones: 'react', 'hermes', 'qwen',
+            'agentflan', 'alpha_umi', etc.
             For available options, refer to
             [loss_scale module](https://github.com/modelscope/ms-swift/blob/main/swift/loss_scale/mapping.py).
             ms-swift supports mixing basic strategies with other strategies,
@@ -83,7 +84,9 @@ class TemplateArguments:
             means all content after the last "user". (**Default value for RLHF**)
             - 'all': Calculate loss for all tokens. (**Default value for `swift pt`**)
             - 'ignore_empty_think': Ignore loss computation for empty `'<think>\n\n</think>\n\n'`
-            (as long as it matches the regex `'<think>\\s*</think>\\s*'`).
+              (as long as it matches the regex `'<think>\\s*</think>\\s*'`).
+            - 'ignore_think_prefix': Ignore loss computation for `'<think>\n'` at the start of a response
+              (or `'<think>'` when no newline follows). The reasoning content and `'</think>'` remain trainable.
             - 'react', 'hermes', 'qwen': Adjust the loss weight of the `tool_call` part to 2.
         sequence_parallel_size (int): The size of sequence parallelism. Defaults to 1. Currently supported for CPT,
             SFT, DPO, and GRPO.

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/arguments/export_args.py

@@ -37,6 +37,9 @@ class ExportArguments(MergeArguments, BaseArguments):
         test_convert_precision (bool): Whether to test the precision error of weight conversion between Hugging Face
             and Megatron-Core formats. Defaults to False.
         test_convert_dtype (str): The dtype to use for the conversion precision test. Defaults to 'float32'.
+        save_missing_weights (bool): Whether to copy weights that exist in the source checkpoint but are not
+            supported by Megatron-Core (e.g. the `mtp.*` DSpark weights of DeepSeek-V4-Flash) into the exported
+            model, so that they are not lost. Defaults to False.
         push_to_hub (bool): Whether to push the output to the Model Hub. Defaults to False. See examples for more
             details.
         hub_model_id (Optional[str]): The model ID for pushing to the Hub (e.g., 'user_name/repo_name' or 'repo_name').
@@ -71,6 +74,7 @@ class ExportArguments(MergeArguments, BaseArguments):
     thread_count: Optional[int] = None
     test_convert_precision: bool = False
     test_convert_dtype: str = 'float32'
+    save_missing_weights: bool = False
 
     # push to ms hub
     push_to_hub: bool = False

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/arguments/sft_args.py

@@ -380,6 +380,9 @@ class SftArguments(SwanlabArguments, TunerArguments, BaseArguments, Seq2SeqTrain
             self.logging_dir = f'{self.output_dir}/runs'
 
         self.logging_dir = to_abspath(self.logging_dir)
+        # transformers>=5.15 dropped `TrainingArguments.logging_dir`; its TensorBoardCallback reads the log
+        # dir from this environment variable instead. Harmless on older versions, which ignore it.
+        os.environ.setdefault('TENSORBOARD_LOGGING_DIR', self.logging_dir)
         os.makedirs(self.output_dir, exist_ok=True)
 
         if self.run_name is None:

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/callbacks/activation_cpu_offload.py

@@ -510,10 +510,13 @@ class ActivationHandler:
 
         @functools.wraps(orig_method)
         def wrapped_method(model_self, *args, **kwargs):
+            if not model_self.training or not torch.is_grad_enabled():
+                return orig_method(*args, **kwargs)
             handler.pre_forward(model_self)
-            out = handler.forward(model_self, orig_method, *args, **kwargs)
-            handler.post_forward(model_self)
-            return out
+            try:
+                return handler.forward(model_self, orig_method, *args, **kwargs)
+            finally:
+                handler.post_forward(model_self)
 
         module.forward = wrapped_method.__get__(module, type(module))
 

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/dataloader/dispatcher.py

@@ -37,6 +37,11 @@ class DataLoaderDispatcher:
             for _ in tqdm(range(self.skip_batches), dynamic_ncols=True, desc='Skip Batches: '):
                 [next(base_iter) for _ in range(self.world_size)]
 
+    def set_epoch(self, epoch: int):
+        dataset = self.base_dataloader.dataset
+        if hasattr(dataset, 'set_epoch'):
+            dataset.set_epoch(epoch)
+
     def __iter__(self):
         base_iter = iter(self.base_dataloader)
         self._skip_batches(base_iter)

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/dataset/__init__.py

@@ -6,8 +6,8 @@ from . import dataset
 from .loader import DATASET_TYPE, DatasetLoader, DatasetSyntax, load_dataset
 from .media import MediaResource
 from .packing import IterablePackingDataset, PackingDataset
-from .preprocessor import (AlpacaPreprocessor, AutoPreprocessor, MessagesPreprocessor, ResponsePreprocessor,
-                           RowPreprocessor)
+from .preprocessor import (AlpacaPreprocessor, AnthropicMessagesPreprocessor, AutoPreprocessor, MessagesPreprocessor,
+                           OpenAIMessagesPreprocessor, ResponsePreprocessor, RowPreprocessor)
 from .register import (DATASET_MAPPING, DatasetMeta, SubsetDataset, get_dataset_list, register_dataset,
                        register_dataset_info)
 from .utils import (AddLengthPreprocessor, EncodePreprocessor, LazyLLMDataset, get_temporary_cache_files_directory,

{ms_swift-4.5.0.dev0 → ms_swift-4.5.2}/swift/dataset/packing.py

@@ -205,6 +205,10 @@ class IterablePackingDataset(IterableDataset):
             for x in iterable:
                 yield x
 
+    def set_epoch(self, epoch: int):
+        if hasattr(self.dataset, 'set_epoch'):
+            self.dataset.set_epoch(epoch)
+
     def __iter__(self):
         try:
             next(iter(self.dataset))

ms_swift-4.5.2/swift/dataset/preprocessor/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (c) ModelScope Contributors. All rights reserved.
+from .core import (DATASET_TYPE, AlpacaPreprocessor, AnthropicMessagesPreprocessor, AutoPreprocessor, ClsPreprocessor,
+                   MessagesPreprocessor, OpenAIMessagesPreprocessor, ResponsePreprocessor, RowPreprocessor)
+from .extra import ClsGenerationPreprocessor, GroundingMixin, TextGenerationPreprocessor