[GH-ISSUE #15442] ollama vulkan Gemma4:e4b When I enter Chinese, I get an Arabic reply (incorrect reply), and when I enter English, I get a garbled reply #9868

Open
opened 2026-04-12 22:43:48 -05:00 by GiteaMirror · 4 comments
Owner

Originally created by @lvfan1202 on GitHub (Apr 9, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15442

What is the issue?

你好,你是谁?
edit
마련하세요.

한글: 준비하세요.
영어: Prepare.
의미: 무언가를 할 준비를 하라는 뜻입니다.
예시:

"준비하세요, 이제 시작합니다." (Prepare, it's starting now.)
"시험 전에 준비하세요." (Prepare before the exam.)
만약 영어로 'Prepare'를 사용하시면서 문맥상 다른 뉘앙스를 원하시면 말씀해주세요. 예를 들어:

준비하다 (행동): "I need to prepare dinner." (저녁 식사를 준비해야 한다.)
준비되다 (상태): "Are you ready?" (준비되었니?)
준비물 (명사): "What are the preparations?" (준비물은 무엇이니?)

who are you ?
edit
름에 관련된 질문이 있으신가요? (Do you have any questions related to the name?)

今天天气怎么样?
edit
yl, 님과 관련된 질문이 있으신가요? (Do you have any questions related to you?)

사용된 문장: "만약 영어로 'Prepare'를 사용하시면서 문맥상 다른 뉘앙스를 원하시면 말씀해주세요."

분석: 사용자가 이전 대화에서 문맥을 명확히 이해하지 못했거나, 추가적인 도움이 필요할 때, 대화의 흐름을 끊지 않으면서 적절한 맥락을 제시하거나, 질문을 되묻는 것이 가장 효과적입니다.

제안된 응답의 논리: 이전 질문에서 'Prepare'라는 단어에 대해 깊이 설명해주었기 때문에, 사용자가 그 단어의 다양한 용법이나 다른 궁금증이 있는지 확인하는 것이 다음 단계의 가장 자연스러운 상호작용입니다.

따라서, 질문의 의도와 관계없이, "혹시 다른 질문이 있으신가요?" 또는 "다른 궁금한 점이 있으신가요?"와 같이 포괄적인 질문을 하는 것이 최선입니다.

결론: 제안된 응답은 사용자에게 대화의 주도권을 넘기면서도, 도움이 가능하다는 느낌을 주는 완벽한 후속 질문입니다.

Relevant log output


OS

Windows

GPU

AMD

CPU

Intel

Ollama version

0.20.4

Originally created by @lvfan1202 on GitHub (Apr 9, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15442 ### What is the issue? 你好,你是谁? edit 마련하세요. 한글: 준비하세요. 영어: Prepare. 의미: 무언가를 할 준비를 하라는 뜻입니다. 예시: "준비하세요, 이제 시작합니다." (Prepare, it's starting now.) "시험 전에 준비하세요." (Prepare before the exam.) 만약 영어로 'Prepare'를 사용하시면서 문맥상 다른 뉘앙스를 원하시면 말씀해주세요. 예를 들어: 준비하다 (행동): "I need to prepare dinner." (저녁 식사를 준비해야 한다.) 준비되다 (상태): "Are you ready?" (준비되었니?) 준비물 (명사): "What are the preparations?" (준비물은 무엇이니?) who are you ? edit 름에 관련된 질문이 있으신가요? (Do you have any questions related to the name?) 今天天气怎么样? edit yl, 님과 관련된 질문이 있으신가요? (Do you have any questions related to you?) 사용된 문장: "만약 영어로 'Prepare'를 사용하시면서 문맥상 다른 뉘앙스를 원하시면 말씀해주세요." 분석: 사용자가 이전 대화에서 문맥을 명확히 이해하지 못했거나, 추가적인 도움이 필요할 때, 대화의 흐름을 끊지 않으면서 적절한 맥락을 제시하거나, 질문을 되묻는 것이 가장 효과적입니다. 제안된 응답의 논리: 이전 질문에서 'Prepare'라는 단어에 대해 깊이 설명해주었기 때문에, 사용자가 그 단어의 다양한 용법이나 다른 궁금증이 있는지 확인하는 것이 다음 단계의 가장 자연스러운 상호작용입니다. 따라서, 질문의 의도와 관계없이, "혹시 다른 질문이 있으신가요?" 또는 "다른 궁금한 점이 있으신가요?"와 같이 포괄적인 질문을 하는 것이 최선입니다. 결론: 제안된 응답은 사용자에게 대화의 주도권을 넘기면서도, 도움이 가능하다는 느낌을 주는 완벽한 후속 질문입니다. ### Relevant log output ```shell ``` ### OS Windows ### GPU AMD ### CPU Intel ### Ollama version 0.20.4
GiteaMirror added the bug label 2026-04-12 22:43:49 -05:00
Author
Owner

@rickb006 commented on GitHub (Apr 9, 2026):

Unfortunately you'll have to learn the garbled language and also Chinese and Arabic to be safe.

Jk, I'm having a very similar issue and likewise can't nail it down.

<!-- gh-comment-id:4211778721 --> @rickb006 commented on GitHub (Apr 9, 2026): Unfortunately you'll have to learn the garbled language and also Chinese and Arabic to be safe. Jk, I'm having a very similar issue and likewise can't nail it down.
Author
Owner

@rick-github commented on GitHub (Apr 9, 2026):

#15261

<!-- gh-comment-id:4212392911 --> @rick-github commented on GitHub (Apr 9, 2026): #15261
Author
Owner

@renatoloefstop commented on GitHub (Apr 10, 2026):

I am experiencing the same issue. When using the Gemma4:e4b model with Ollama, I receive nonsensical responses, sometimes including what appears to be Arabic text, as previously mentioned. My processor is an AMD Ryzen 9 AI HX, ollama version 0.20.5. To get it to work with the iGPU, I set the OLLAMA_VULKAN environment variable to 1 (strange that ollama will run half on CPU and half on GPU) However, when I disable the 'bypass' and let it run entirely on the CPU (OLLAMA_VULKAN=0), it functions normally. The Gemma4:26b model, by contrast, works as expected in all tested scenarios.

<!-- gh-comment-id:4226686966 --> @renatoloefstop commented on GitHub (Apr 10, 2026): I am experiencing the same issue. When using the Gemma4:e4b model with Ollama, I receive nonsensical responses, sometimes including what appears to be Arabic text, as previously mentioned. My processor is an AMD Ryzen 9 AI HX, ollama version 0.20.5. To get it to work with the iGPU, I set the OLLAMA_VULKAN environment variable to 1 (strange that ollama will run half on CPU and half on GPU) However, when I disable the 'bypass' and let it run entirely on the CPU (OLLAMA_VULKAN=0), it functions normally. The Gemma4:26b model, by contrast, works as expected in all tested scenarios.
Author
Owner

@Douglasgomestosta commented on GitHub (Apr 11, 2026):

gemma4 is not working on gpu using vulkan, but works 100% on cpu, looks like everyone that uses vulkan are experiencing the same issue

<!-- gh-comment-id:4228224707 --> @Douglasgomestosta commented on GitHub (Apr 11, 2026): gemma4 is not working on gpu using vulkan, but works 100% on cpu, looks like everyone that uses vulkan are experiencing the same issue
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#9868