ChatGLM3, an open-source bilingual conversational language model developed by Zhipu AI and Tsinghua University's KEG Lab, has seen no new pull requests or commits in the last 30 days, indicating a potential pause in active development.
The absence of recent pull requests and commits suggests a lull in development activities. This stagnation could be due to various factors such as resource allocation, strategic planning, or awaiting further community feedback. The last significant activities included addressing performance issues and enhancing documentation, but no new features or bug fixes have been introduced recently.
| Timespan | Opened | Closed | Comments | Labeled | Milestones |
|---|---|---|---|---|---|
| 7 Days | 0 | 0 | 0 | 0 | 0 |
| 30 Days | 6 | 7 | 6 | 6 | 1 |
| 90 Days | 29 | 23 | 37 | 29 | 1 |
| All Time | 774 | 755 | - | - | - |
Like all software activity quantification, these numbers are imperfect but sometimes useful. Comments, Labels, and Milestones refer to those issues opened in the timespan in question.
The GitHub repository for ChatGLM3 has recently seen a total of 19 open issues, with several users reporting various technical challenges related to model training, inference, and API usage. Notably, there are recurring themes around model performance, memory management during training, and issues with specific functionalities like tool invocation and multi-GPU support.
Several issues indicate a lack of clarity in documentation or implementation details, particularly regarding the expected behavior of the models when handling different input formats or configurations. For example, users have reported problems with the model's ability to handle long prompts and maintain coherent responses without falling into repetitive loops.
Issue #1314: openai_api_request.py请求超时
Issue #1312: 使用ptuning_v2微调过程中出现报错ValueError: Hypothesis is empty.
Issue #1311: tools_using_demo中的openai_api_demo.py，stream等于True的时候 finish_reason 不出现"function_call"
Issue #1310: The figure in execute function will be valued as None
Issue #1308: 在ChatGLM3-6B的微调过程中，遇到如下报错（ImportError: cannot import name 'log' from 'torch.distributed.elastic.agent.server.api'）
The issue regarding the ValueError: Hypothesis is empty (#1312) suggests potential problems with data formatting or preprocessing steps during fine-tuning, which could lead to incomplete or improperly structured training data.
The timeout issue in openai_api_request.py (#1314) indicates that users may be facing network-related problems or resource limitations when attempting to invoke the model's API.
The repeated mention of errors related to memory management and CUDA device allocation across multiple issues highlights a common pain point for users working with large models like ChatGLM3.
There are multiple reports of users experiencing difficulties with multi-GPU setups, particularly regarding configuration settings that do not seem to yield the expected performance improvements.
Users are also struggling with understanding how to properly format their input data for both training and inference, leading to confusion and errors during execution.
The community appears active in discussing solutions and workarounds for these issues, but there remains a need for clearer documentation and examples to guide users through common pitfalls.
Overall, the recent activity on the ChatGLM3 repository reflects a vibrant user community engaged in troubleshooting and optimizing their experiences with the model, while also indicating areas where additional support and resources could enhance usability.
The repository THUDM/ChatGLM3 has a total of 168 closed pull requests (PRs), with no open PRs at the moment. The majority of these PRs have been merged, indicating active development and maintenance of the project.
PR #1287: discord
Closed 66 days ago. This PR involved minor updates to the README files, with a net reduction of lines, indicating a cleanup or restructuring of documentation.
PR #1283: [fix#1272] Fix performance of resume from checkpoint
Closed 72 days ago. This PR addressed performance issues related to resuming from checkpoints, which is critical for model training efficiency.
PR #1274: update
Closed 86 days ago. Similar to PR #1287, this involved updates to README files, suggesting ongoing improvements in documentation.
PR #1269: fix #1268
Closed 90 days ago. This PR fixed a specific issue (#1268) in the codebase, showing responsiveness to reported bugs.
PR #1260: GLM-4 update
Closed 97 days ago. This PR included significant updates to the README and related documentation, likely reflecting new features or enhancements in the GLM-4 model.
PR #1251: Update ChatGLM3.py fix index error
Closed 99 days ago. This PR addressed an index error in the main script, indicating ongoing debugging efforts.
PR #1245: fix finetune req
Closed 107 days ago. Specific details are not provided, but it suggests improvements related to fine-tuning requirements.
PR #1235: 更新协议
Closed 112 days ago. The title suggests updates to protocols, likely related to API or model usage guidelines.
PR #1229: 多个文档和demo调整
Closed 116 days ago. This PR involved multiple adjustments to documentation and demos, emphasizing community engagement and usability.
PR #1227: Feat/modelscope 310
Closed 115 days ago but not merged. Indicates potential new features that were not accepted into the main branch.
PR #1226: Feat/main 310
Similar to PR #1227, closed without merging, suggesting either lack of consensus or further work needed on proposed features.
PR #1223: Add RAGFlow in README
Closed 117 days ago. This PR added information about RAGFlow support in the README, enhancing user awareness of available functionalities.
PR #1212: agent-chat-openai 更新
Closed 123 days ago with unspecified changes; indicates ongoing enhancements to the agent-chat functionality.
PR #1211: agent-chat openai能力完全恢复，完全参数对齐。兼容两种模式。
Not merged; suggests significant changes that may not have met project standards or requirements.
PR #1210: agent-chat openai能力完全恢复，添加说明，完全参数对齐。兼容两种模式。
Not merged; similar issues as PR #1211 regarding compatibility and parameter alignment.
The analysis of the pull requests reveals several key themes and patterns that highlight both the strengths and potential weaknesses within the development process of ChatGLM3.
The high number of closed pull requests (168) indicates a robust development cycle where issues are regularly addressed and features are actively integrated into the main branch. The fact that most PRs are merged suggests a well-functioning review process led by Yuxuan.Zhang (zRzRzRzRzRzRzR), who appears to be a primary maintainer or contributor in this repository.
A notable trend is the frequent updates to README files and other documentation (e.g., PRs #1287, #1274, and #1260). This emphasis on improving documentation is critical for user engagement and helps ensure that developers can effectively utilize the model's capabilities without confusion or ambiguity.
Several pull requests focus specifically on bug fixes (e.g., PRs #1283 and #1251) and performance enhancements (e.g., PR #1283 addressing resume from checkpoint). This reflects a proactive approach to maintaining software quality and performance, which is essential for user satisfaction and reliability in production environments.
There are instances where feature proposals were submitted but not merged (e.g., PRs #1227 and #1226). This could indicate either that these features did not align with project goals or that they required further refinement before acceptance. It raises questions about communication within the team regarding feature expectations and acceptance criteria.
The inclusion of community contributors (e.g., various authors for different PRs) highlights an effort to engage with external developers and foster collaboration within the open-source community. However, there are also indications that some contributions may have been rejected or left unmerged due to various reasons, which could affect morale or participation levels among contributors if not managed effectively.
While there is a healthy volume of activity in merging pull requests, the absence of open pull requests may suggest a bottleneck where contributors are waiting for feedback or approval on their submissions before proceeding with new work. This could lead to stagnation if contributors feel discouraged by delayed responses or unmerged work.
In conclusion, while THUDM/ChatGLM3 exhibits strong development practices characterized by active maintenance and community involvement, attention should be given to managing feature proposals effectively and ensuring timely feedback on contributions to maintain momentum within the project.
Yuxuan Zhang:
Pandong:
Jin Hai:
Lilong BaiYang:
Ethan Yang:
Kero Yang:
Longman:
The development team is currently active with a clear focus on resolving bugs, enhancing existing functionalities, and improving documentation. The collaborative nature of the commits indicates a well-coordinated effort towards maintaining and advancing the ChatGLM3 project.
Summary
Quantified Reports
Detailed Reports