MoonBit 是一个用于云和边缘计算的 WebAssembly 端到端编程语言工具链。MoonBit Pilot:为 MoonBit 项目提供的人工智能开发助手。这次测试几个大模型在moon pilot里的表现。

先上结论:暂时都没法用!

AIStudio星河社区支持部署的大模型

AIStudio星河社区支持部署大模型,网址:飞桨AI Studio星河社区-人工智能学习与实训社区 分别部署了这三个模型进行测试:

unsloth/GLM-4-32B-0414-GGUF

 GLM是国内非常优秀的大模型,它是混合结构,开创性的将GPT和Bird架构进行了结合,在文字处理尤其是中文处理方面一直领军。

PaddlePaddle/ERNIE-4.5-21B-A3B-Paddle

 百度的PaddlePaddle团队开发的ERNIE(Enhanced Representation through Knowledge Integration)系列模型,属于“ERNIE 4.5”版本,参数达到21B(210亿),是大型预训练语言模型,非常强大。

deepseek-ai/DeepSeek-R1-Distill-Llama-70B

deepseek,不用多说了,大家都懂。参数达70B(700亿),属于超大模型。

部署unsloth/GLM-4-32B-0414-GGUF

这是GLM4模型,理论上这个模型应该非常优秀才对。

但是部署完成后,问什么都是只输出GGGG

data: {"id":"chatcmpl-155","object":"chat.completion.chunk","created":1756943731,"model":"unsloth/GLM-4-32B-0414-GGUF","system_fingerprint":"fp_ollama","choices":[{"index":0,"delta":{"role":"assistant","content":"G"},"finish_reason":null}]}

data: {"id":"chatcmpl-155","object":"chat.completion.chunk","created":1756943731,"model":"unsloth/GLM-4-32B-0414-GGUF","system_fingerprint":"fp_ollama","choices":[{"index":0,"delta":{"role":"assistant","content":"G"},"finish_reason":null}]}

data: {"id":"chatcmpl-155","object":"chat.completion.chunk","created":1756943731,"model":"unsloth/GLM-4-32B-0414-GGUF","system_fingerprint":"fp_ollama","choices":[{"index":0,"delta":{"role":"assistant","content":"G"},"finish_reason":null}]}

第二个问题是因为模型较大,所以这个部署比较慢,有好几次都没等到它部署成功,唯一一次部署成功,用了11分钟

2025-09-03T23:30:07.265986896Z 
2025-09-03T23:30:07.266166290Z ==========
2025-09-03T23:30:07.266189643Z == CUDA ==
2025-09-03T23:30:07.266192558Z ==========
2025-09-03T23:30:07.275857280Z 
2025-09-03T23:30:07.275877659Z CUDA Version 11.8.0
2025-09-03T23:30:07.278614230Z 
2025-09-03T23:30:07.278638541Z Container image Copyright (c) 2016-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2025-09-03T23:30:07.280487675Z 
2025-09-03T23:30:07.280500753Z This container image and its contents are governed by the NVIDIA Deep Learning Container License.
2025-09-03T23:30:07.280503542Z By pulling and using the container, you accept the terms and conditions of this license:
2025-09-03T23:30:07.280506007Z https://developer.nvidia.com/ngc/nvidia-deep-learning-container-license
2025-09-03T23:30:07.280508100Z 
2025-09-03T23:30:07.280510466Z A copy of this license is made available in this container at /NGC-DL-CONTAINER-LICENSE for your convenience.
2025-09-03T23:30:07.348971052Z 
2025-09-03T23:30:10.969388961Z 开始下载模型
2025-09-03T23:31:02.496047700Z pulling ca2a22975dfe:  10% ▕█                 ▏ 1.9 GB/ 19 GB   38 MB/s   7m41s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:31:52.296341305Z pulling ca2a22975dfe:  20% ▕███               ▏ 3.8 GB/ 19 GB   40 MB/s   6m27s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:32:40.696221794Z pulling ca2a22975dfe:  30% ▕█████             ▏ 5.8 GB/ 19 GB   40 MB/s   5m40s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:33:29.996418787Z pulling ca2a22975dfe:  40% ▕███████           ▏ 7.8 GB/ 19 GB   41 MB/s   4m48s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:34:18.496956064Z pulling ca2a22975dfe:  50% ▕████████          ▏ 9.7 GB/ 19 GB   41 MB/s   3m58s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:35:08.396748289Z pulling ca2a22975dfe:  60% ▕██████████        ▏  11 GB/ 19 GB   40 MB/s   3m14s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:35:59.096924535Z pulling ca2a22975dfe:  70% ▕████████████      ▏  13 GB/ 19 GB   39 MB/s   2m32s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:36:49.496996263Z pulling ca2a22975dfe:  80% ▕██████████████    ▏  15 GB/ 19 GB   40 MB/s   1m40s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:37:40.796713526Z pulling ca2a22975dfe:  90% ▕████████████████  ▏  17 GB/ 19 GB   39 MB/s     52s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:38:54.396324277Z pulling ca2a22975dfe: 100% ▕█████████████████ ▏  19 GB/ 19 GB   16 MB/s      5s[K[?25h[?2026l[?2026h[?25l[A[1Gpulling manifest [K
2025-09-03T23:41:43.252256882Z 模型下载成功
2025-09-03T23:41:43.458208773Z [?2026h[?25l[1Ggathering model components [K
2025-09-03T23:41:43.458249031Z using existing layer sha256:ca2a22975dfe2a4e3e19986bac1436527176b886bb613339f1c76965ddfcd764 [K
2025-09-03T23:41:43.458252906Z using existing layer sha256:aaa93f59d003a2a3202b09f283857659f5ee5ca789ea364a85b211f1f0bc38f1 [K
2025-09-03T23:41:43.458256051Z creating new layer sha256:7d2bdeed3b43f190eac48e16e0dc7a30aba406d974f9a3d6b3995e6e2c54662d [K
2025-09-03T23:41:43.510516337Z writing manifest ⠋ [K[?25h[?2026l[?2026h[?25l[A[A[A[A[1Ggathering model components [K
2025-09-03T23:41:43.510531414Z using existing layer sha256:ca2a22975dfe2a4e3e19986bac1436527176b886bb613339f1c76965ddfcd764 [K
2025-09-03T23:41:43.510535294Z using existing layer sha256:aaa93f59d003a2a3202b09f283857659f5ee5ca789ea364a85b211f1f0bc38f1 [K
2025-09-03T23:41:43.510540381Z creating new layer sha256:7d2bdeed3b43f190eac48e16e0dc7a30aba406d974f9a3d6b3995e6e2c54662d [K
2025-09-03T23:41:43.510556177Z writing manifest [K
2025-09-03T23:41:43.510560612Z success [K[?25h[?2026l
2025-09-03T23:41:51.861506381Z 模型加载完毕

测试失败。

PaddlePaddle/ERNIE-4.5-21B-A3B-Paddle

这是飞桨特有的文心开源大模型,也是非常不错的模型,部署起来速度快。

但是问题是它不兼容openai api的调用,或者说兼容,但是需要调整一下输入....使用moon pilot调用的报错信息:

│ Error: [LLM] request failed with status 400: {"object":"error","message":"can only concatenate str (not \"list\") │
│  to str","code":400}

要解决这个错误 "can only concatenate str (not 'list') to str",需要确保在字符串拼接(+ 操作)时,所有操作数都是字符串类型,而不是列表(list)。

算了,不去解决它了。也就是它也不能在moon pilot里使用。

deepseek-ai/DeepSeek-R1-Distill-Llama-70B

DeepSeek-R1 70B这个模型不支持tools

│ ❌ Error                                                                                                         │
├───────────────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ [LLM] request failed with status 400: {"error":{"message":"registry.ollama.ai/library/deepseek-r1:70b does not    │
│ support tools","type":"api_error","param":null,"code":null}}

这样它也不能在moon pilot里使用。

总结

本来想薅羊毛,使用星河社区的模型,在moon pilot里使用,目前看来不改代码,或者不加中继,是无法使用的。

朋友们如果有谁测试成功了,告知一下哦!

Logo

聚焦RDK与具身智能,汇集机器人开发教程、开源案例和技术交流

更多推荐