[model] support glm-4.5 agent #5305

laocheyujie · 2025-08-08T05:40:37Z

PR type

Bug Fix
New Feature
Document Updates
More Models or Datasets Support

PR information

This PR adds support for GLM-4.5 agent (function call) training in ms-swift.

Key updates:

Added GLM4_5AgentTemplate in plugin/agent_template/glm4.py to support training GLM-4.5 function call agents.
Enhanced suffix appending logic: in the final generation round, check if the output already ends with a stop word, and skip appending the suffix if so to avoid redundancy.

Experiment results

Training successfully runs on 8×A100 (80GB) with GLM-4.5 agent configuration.

gemini-code-assist

Summary of Changes

Hello @laocheyujie, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

I've implemented support for GLM-4.5 agent training, specifically enabling function call capabilities within the ms-swift project. This involved integrating a new agent template and refining the text generation process to ensure more accurate and efficient output by preventing redundant suffixes. The changes aim to expand the framework's compatibility with advanced large language models and their interactive features.

Highlights

GLM-4.5 Agent Support: I've added comprehensive support for GLM-4.5 agent (function call) training within the ms-swift framework. This includes defining a new template and agent-specific logic for GLM-4.5 agents.
Suffix Appending Logic Enhancement: I've refined the suffix appending logic during the final generation round. The system now checks if the output already ends with a stop word, preventing redundant suffix additions and improving output quality.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request adds support for the GLM-4.5 agent, including a new agent template and updates to the suffix appending logic. The changes are well-structured and address the goal of the PR. I've provided a few suggestions to improve code clarity, style, and performance.

swift/llm/template/base.py

swift/plugin/agent_template/glm4.py

Jintao-Huang · 2025-08-13T08:12:11Z

swift/llm/template/base.py

@@ -1135,7 +1135,8 @@ def _swift_encode(self, inputs: StdTemplateInputs):
                context_list.append('{{RESPONSE}}')
                # self.is_training needed because we may want to continue generation from
                # the current response
-                if self.is_training and not sep_token or self.task_type == 'embedding':
+                string_stop_words = tuple(s for s in template_meta.stop_words if isinstance(s, str))


What is the reason for this modification here?

Originally, I made that change to prevent models like GLM-4.5 from producing <|observation|><|user|> when the final round was a tool_call. But I just realized this can be handled more cleanly by overriding _swift_encode in the GLM4_5Template class, so I removed that part of the modification in my latest commit.

Hello!
Here, <|user|> (suffix) is only added during training. Could you please clarify: in what training scenarios would the final round include <|observation|> but not tool_response?

Jintao-Huang · 2025-08-24T12:50:39Z

Hello, please run the following code to pass the lint test.

pip install pre-commit
pre-commit run --all-files

Jintao-Huang · 2025-08-25T06:10:07Z

/gemini review

gemini-code-assist

Code Review

This pull request introduces support for GLM-4.5 agent function call training. The changes include a new GLM4_5AgentTemplate for handling the specific tool call format of GLM-4.5, along with corresponding template metadata. The logic for suffix appending is also enhanced to prevent redundant tokens. The implementation is well-structured and follows existing patterns in the codebase. I've identified a few areas for improvement regarding code robustness and readability in the new agent template implementation.

gemini-code-assist · 2025-08-25T06:11:42Z

swift/plugin/agent_template/glm4.py

+            for arg_key, arg_value in tool_call['arguments'].items():
+                tool_calls.append(f'<arg_key>{arg_key}</arg_key>\n<arg_value>{arg_value}</arg_value>')


The code assumes that tool_call['arguments'] is always a dictionary, which may not be guaranteed by _parse_tool_call. If arguments is not a dictionary, tool_call['arguments'].items() will raise a runtime error. It's safer to check if arguments is a dictionary before iterating over it to make the function more robust against malformed inputs.

Suggested change

for arg_key, arg_value in tool_call['arguments'].items():

tool_calls.append(f'<arg_key>{arg_key}</arg_key>\n<arg_value>{arg_value}</arg_value>')

arguments = tool_call.get('arguments')

if isinstance(arguments, dict):

for arg_key, arg_value in arguments.items():

tool_calls.append(f'<arg_key>{arg_key}</arg_key>\n<arg_value>{arg_value}</arg_value>')

gemini-code-assist · 2025-08-25T06:11:42Z

swift/plugin/agent_template/glm4.py

+        functions = []
+        for toolcall in toolcall_list:
+            function = self._find_function_call(toolcall)
+            if function:
+                functions.append(function)


The loop for collecting functions can be made more concise and Pythonic by using a list comprehension with a walrus operator (PEP 572). This can improve readability and reduce boilerplate code.

Suggested change

functions = []

for toolcall in toolcall_list:

function = self._find_function_call(toolcall)

if function:

functions.append(function)

functions = [func for toolcall in toolcall_list if (func := self._find_function_call(toolcall))]

gemini-code-assist · 2025-08-25T06:11:43Z

swift/plugin/agent_template/glm4.py

+        if with_action:
+            return super()._format_tool_responses(assistant_content, tool_messages)
+        res = []
+        for _, tool_message in enumerate(tool_messages):


The use of enumerate here is unnecessary as the index _ is not used. You can directly iterate over tool_messages for cleaner and more idiomatic code.

Suggested change

for _, tool_message in enumerate(tool_messages):

for tool_message in tool_messages:

gemini-code-assist bot reviewed Aug 8, 2025

View reviewed changes

Jintao-Huang reviewed Aug 13, 2025

View reviewed changes

[template] add GLM4_5AgentTemplate & update GLM4_5Template

36c0c8a

laocheyujie force-pushed the glm_4_5 branch from 8d8a462 to 36c0c8a Compare August 22, 2025 02:34

laocheyujie requested a review from Jintao-Huang August 22, 2025 08:55

Jintao-Huang approved these changes Aug 22, 2025

View reviewed changes

[template] add GLM4_5AgentTemplate & update GLM4_5Template

28936c8

Jintao-Huang merged commit f9a925a into modelscope:main Aug 25, 2025
1 of 2 checks passed

gemini-code-assist bot reviewed Aug 25, 2025

View reviewed changes

Jintao-Huang mentioned this pull request Aug 25, 2025

[template] update glm4.5 agent template #5518

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[model] support glm-4.5 agent #5305

[model] support glm-4.5 agent #5305

Uh oh!

laocheyujie commented Aug 8, 2025

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Jintao-Huang Aug 13, 2025

Uh oh!

laocheyujie Aug 21, 2025 •

edited

Loading

Uh oh!

Jintao-Huang Aug 25, 2025

Uh oh!

Jintao-Huang commented Aug 24, 2025 •

edited

Loading

Uh oh!

Uh oh!

Jintao-Huang commented Aug 25, 2025

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Aug 25, 2025

Uh oh!

gemini-code-assist bot Aug 25, 2025

Uh oh!

gemini-code-assist bot Aug 25, 2025

Uh oh!

Uh oh!

		for arg_key, arg_value in tool_call['arguments'].items():
		tool_calls.append(f'<arg_key>{arg_key}</arg_key>\n<arg_value>{arg_value}</arg_value>')

	for _, tool_message in enumerate(tool_messages):
	for tool_message in tool_messages:

[model] support glm-4.5 agent #5305

[model] support glm-4.5 agent #5305

Uh oh!

Conversation

laocheyujie commented Aug 8, 2025

PR type

PR information

Experiment results

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Jintao-Huang Aug 13, 2025

Choose a reason for hiding this comment

Uh oh!

laocheyujie Aug 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Jintao-Huang Aug 25, 2025

Choose a reason for hiding this comment

Uh oh!

Jintao-Huang commented Aug 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Jintao-Huang commented Aug 25, 2025

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Aug 25, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Aug 25, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Aug 25, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

laocheyujie Aug 21, 2025 •

edited

Loading

Jintao-Huang commented Aug 24, 2025 •

edited

Loading