github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #13840] Generation stops after tool call with Ollama (GLM-4.7-Flash) #9061

New Issue
Closed
opened 2026-04-12 21:53:35 -05:00 by GiteaMirror · 13 comments
GiteaMirror commented 2026-04-12 21:53:35 -05:00
Owner
Copy Link

Originally created by @HuysArthur on GitHub (Jan 22, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/13840

Originally assigned to: @jmorganca on GitHub.

What is the issue?

When running glm-4.7-flash:bf16 locally via Ollama (with context_length: 80000) and connecting it to OpenCode, the model halts generation immediately after outputting any tool call. I have to manually type "continue" every single time to resume.
The same model hosted through vLLM (instead of Ollama) works flawlessly in OpenCode — tool calls are handled normally without interruption.
This seems specific to Ollama's handling of this GLM-4.7 Flash variant during tool use. Has anyone else run into this?

Relevant log output

OS

linux arm

GPU

NVIDIA GH200

CPU

No response

Ollama version

0.14.3

Originally created by @HuysArthur on GitHub (Jan 22, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/13840 Originally assigned to: @jmorganca on GitHub. ### What is the issue? When running glm-4.7-flash:bf16 locally via Ollama (with context_length: 80000) and connecting it to OpenCode, the model halts generation immediately after outputting any tool call. I have to manually type "continue" every single time to resume. The same model hosted through vLLM (instead of Ollama) works flawlessly in OpenCode — tool calls are handled normally without interruption. This seems specific to Ollama's handling of this GLM-4.7 Flash variant during tool use. Has anyone else run into this? ### Relevant log output ```shell ``` ### OS linux arm ### GPU NVIDIA GH200 ### CPU _No response_ ### Ollama version `0.14.3`
GiteaMirror added the bug label 2026-04-12 21:53:35 -05:00
GiteaMirror commented 2026-04-12 21:53:38 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 22, 2026):

Sounds like OpenCode is not returning the tool result to ollama. Running the agent loop shows that glm-4.7-flash:bf16 will process tools results.

$ ollama run glm-4.7-flash:bf16 --experimental --experimental-yolo

This experimental version of Ollama has the bash tool enabled.
Models can read files on your computer, or run commands (after you allow them).

warning: yolo mode - all tool approvals will be skipped
>>> what is the time?
Thinking...
The user is asking "what is the time?" - a simple question about the current time. To answer this, I need to use a tool to get the current time from the system. I have the bash tool available which can run 
system commands. I can use the `date` command to get the current time.
...done thinking.


running: Bash: date
  Thu Jan 22 10:50:28 AM CET 2026
  

Thinking...
The `date` command returned the current time as "Thu Jan 22 10:50:28 AM CET 2026". This tells me:

- It's Thursday, January 22nd, 2026
- The time is 10:50:28 AM
- It's in Central European Time (CET) timezone

I should provide this information in a clear, readable format for the user.
...done thinking.

The current time is **10:50:28 AM** on Thursday, January 22nd, 2026 (CET).

Does OpenCode have logs?

Server log from ollama with OLLAMA_DEBUG=2 may help in debugging. Note that this will include the prompt so be aware of PII.

<!-- gh-comment-id:3783509718 --> @rick-github commented on GitHub (Jan 22, 2026): Sounds like OpenCode is not returning the tool result to ollama. Running the agent loop shows that glm-4.7-flash:bf16 will process tools results. ```console $ ollama run glm-4.7-flash:bf16 --experimental --experimental-yolo This experimental version of Ollama has the bash tool enabled. Models can read files on your computer, or run commands (after you allow them). warning: yolo mode - all tool approvals will be skipped >>> what is the time? Thinking... The user is asking "what is the time?" - a simple question about the current time. To answer this, I need to use a tool to get the current time from the system. I have the bash tool available which can run system commands. I can use the `date` command to get the current time. ...done thinking. running: Bash: date Thu Jan 22 10:50:28 AM CET 2026 Thinking... The `date` command returned the current time as "Thu Jan 22 10:50:28 AM CET 2026". This tells me: - It's Thursday, January 22nd, 2026 - The time is 10:50:28 AM - It's in Central European Time (CET) timezone I should provide this information in a clear, readable format for the user. ...done thinking. The current time is **10:50:28 AM** on Thursday, January 22nd, 2026 (CET). ``` Does OpenCode have logs? [Server log](https://docs.ollama.com/troubleshooting) from ollama with `OLLAMA_DEBUG=2` may help in debugging. Note that this will include the prompt so be aware of PII.
GiteaMirror commented 2026-04-12 21:53:39 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 22, 2026):

Just installed OpenCode and gave it a try, tool calls seem to work fine:

$ opencode run "Create the file 'README' with the contents 'hello world'" --model ollama/glm-4.7-flash-c198k
|  Write    README

File ' README' created successfully.

$ ls -al
total 12
drwxrwxr-x 2 rick rick 4096 Jan 22 18:42 .
drwxrwxr-x 3 rick rick 4096 Jan 22 18:42 ..
-rw-rw-r-- 1 rick rick   11 Jan 22 18:42 README
$ cat README 
hello world
<!-- gh-comment-id:3786083907 --> @rick-github commented on GitHub (Jan 22, 2026): Just installed OpenCode and gave it a try, tool calls seem to work fine: ```console $ opencode run "Create the file 'README' with the contents 'hello world'" --model ollama/glm-4.7-flash-c198k | Write README File ' README' created successfully. $ ls -al total 12 drwxrwxr-x 2 rick rick 4096 Jan 22 18:42 . drwxrwxr-x 3 rick rick 4096 Jan 22 18:42 .. -rw-rw-r-- 1 rick rick 11 Jan 22 18:42 README $ cat README hello world ```
GiteaMirror commented 2026-04-12 21:53:40 -05:00
Author
Owner
Copy Link

@sohelzerdoumi commented on GitHub (Jan 22, 2026):

I’m seeing the same issue with Ollama when using tools with glm-4.7-flash:latest.

"glm-4.6 tool call parsing failed" error="failed to parse XML: XML syntax error on line 1: element <tool_call> closed by </arg_key>"

The error only appears once the context grows beyond ~5k tokens and multiple tool calls are chained. With shorter contexts, everything works fine. When it fails, the model outputs malformed XML (e.g. mismatched <tool_call> / <arg_key> tags), which breaks parsing.

My suspicion is that this is related to quantization.

<!-- gh-comment-id:3786658905 --> @sohelzerdoumi commented on GitHub (Jan 22, 2026): I’m seeing the same issue with Ollama when using tools with glm-4.7-flash:latest. ``` "glm-4.6 tool call parsing failed" error="failed to parse XML: XML syntax error on line 1: element <tool_call> closed by </arg_key>" ``` The error only appears once the context grows beyond ~5k tokens and multiple tool calls are chained. With shorter contexts, everything works fine. When it fails, the model outputs malformed XML (e.g. mismatched <tool_call> / <arg_key> tags), which breaks parsing. My suspicion is that this is related to quantization.
GiteaMirror commented 2026-04-12 21:53:41 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 22, 2026):

What context size have you set for the model?

<!-- gh-comment-id:3786683399 --> @rick-github commented on GitHub (Jan 22, 2026): What context size have you set for the model?
GiteaMirror commented 2026-04-12 21:53:42 -05:00
Author
Owner
Copy Link

@sohelzerdoumi commented on GitHub (Jan 22, 2026):

I have 30 000 tokens context length and 40G Vram. The is enough to fit.
Environment="OLLAMA_CONTEXT_LENGTH=30000"

<!-- gh-comment-id:3786692120 --> @sohelzerdoumi commented on GitHub (Jan 22, 2026): I have 30 000 tokens context length and 40G Vram. The is enough to fit. Environment="OLLAMA_CONTEXT_LENGTH=30000"
GiteaMirror commented 2026-04-12 21:53:43 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 22, 2026):

Server log from ollama with OLLAMA_DEBUG=2 may help in debugging. Note that this will include the prompt so be aware of PII.

<!-- gh-comment-id:3786697721 --> @rick-github commented on GitHub (Jan 22, 2026): [Server log](https://docs.ollama.com/troubleshooting) from ollama with `OLLAMA_DEBUG=2` may help in debugging. Note that this will include the prompt so be aware of PII.
GiteaMirror commented 2026-04-12 21:53:43 -05:00
Author
Owner
Copy Link

@sohelzerdoumi commented on GitHub (Jan 22, 2026):

I increased max_tokens and context to 40,000 tokens each, and it worked. My bad.

Here is my trace from when it failed.

janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=bytepairencoding.go:270 msg=decoded string=</tool_call> from=[154844]
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:657 msg="computeBatch: outputs are ready" batchID=6298
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:652 msg="computeBatch: inputs are ready" batchID=6299
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=routes.go:2257 msg="builtin parser input" parser=glm-4.7 content=</tool_call>
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=glm46.go:118 msg="glm-4.6 events parsed" events="[{raw:write_file</arg_key><arg_value>[REDACTED]</arg_value><arg_key>content</arg_key><arg_value>[REDACTED]</arg_value>}]" state=4 buffer=""
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=WARN source=glm46.go:89 msg="glm-4.6 tool call parsing failed" error="failed to parse XML: XML syntax error on line 1: element <tool_call> closed by </arg_key>"
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:725 msg="computeBatch: signaling computeStartedCh" batchID=6299
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:476 msg="forwardBatch compute started, setting up next batch" pendingBatch.id=6299 id=6300
janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:598 msg="forwardBatch iBatch" batchID=6300 seqIdx=0 seq.iBatch=0 i+1=1 len(seq.inputs)=1
<!-- gh-comment-id:3786768534 --> @sohelzerdoumi commented on GitHub (Jan 22, 2026): I increased max_tokens and context to 40,000 tokens each, and it worked. My bad. Here is my trace from when it failed. ``` janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=bytepairencoding.go:270 msg=decoded string=</tool_call> from=[154844] janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:657 msg="computeBatch: outputs are ready" batchID=6298 janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:652 msg="computeBatch: inputs are ready" batchID=6299 janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=routes.go:2257 msg="builtin parser input" parser=glm-4.7 content=</tool_call> janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=glm46.go:118 msg="glm-4.6 events parsed" events="[{raw:write_file</arg_key><arg_value>[REDACTED]</arg_value><arg_key>content</arg_key><arg_value>[REDACTED]</arg_value>}]" state=4 buffer="" janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=WARN source=glm46.go:89 msg="glm-4.6 tool call parsing failed" error="failed to parse XML: XML syntax error on line 1: element <tool_call> closed by </arg_key>" janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:725 msg="computeBatch: signaling computeStartedCh" batchID=6299 janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:476 msg="forwardBatch compute started, setting up next batch" pendingBatch.id=6299 id=6300 janv. 22 22:13:56 gpu-server ollama[3374697]: time=2026-01-22T22:13:56.576+01:00 level=TRACE source=runner.go:598 msg="forwardBatch iBatch" batchID=6300 seqIdx=0 seq.iBatch=0 i+1=1 len(seq.inputs)=1 ```
GiteaMirror commented 2026-04-12 21:53:44 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 22, 2026):

OpenCode requires a lot of context, just the base level tools and instructions needs ~11k tokens. I asked it to create a golang program to print "hello world" and add tests, and the context grew to 25k tokens.

<!-- gh-comment-id:3786973148 --> @rick-github commented on GitHub (Jan 22, 2026): OpenCode requires a lot of context, just the base level tools and instructions needs ~11k tokens. I asked it to create a golang program to print "hello world" and add tests, and the context grew to 25k tokens.
GiteaMirror commented 2026-04-12 21:53:45 -05:00
Author
Owner
Copy Link

@HuysArthur commented on GitHub (Jan 23, 2026):

Turns out the issue isn't with Ollama itself.
I was routing OpenCode → Open WebUI API → Ollama.
GLM-4.7-Flash works perfectly when served directly via vLLM, and other models (e.g. gpt-oss) work fine through the same WebUI → Ollama path.
So the problem seems specific to how GLM-4.7-Flash behaves when going through Ollama + WebUI as a middleman.

<!-- gh-comment-id:3789066808 --> @HuysArthur commented on GitHub (Jan 23, 2026): Turns out the issue isn't with Ollama itself. I was routing OpenCode → Open WebUI API → Ollama. GLM-4.7-Flash works perfectly when served directly via vLLM, and other models (e.g. gpt-oss) work fine through the same WebUI → Ollama path. So the problem seems specific to how GLM-4.7-Flash behaves when going through Ollama + WebUI as a middleman.
GiteaMirror commented 2026-04-12 21:53:46 -05:00
Author
Owner
Copy Link

@HuysArthur commented on GitHub (Jan 23, 2026):

Here is a clear difference in behavior:

user@host dir % opencode run "Create the file 'README' with the contents 'hello world', let me know if it succeeded" --model webui/glm-4.7-flash:bf16-80k
|  Write    Users/user/dir/README
user@host dir % rm README                                                                                                                               
user@host dir % opencode run "Create the file 'README' with the contents 'hello world', let me know if it succeeded" --model ollama/glm-4.7-flash:bf16-80k     

I'll create the README file with the specified content.

|  Write    Users/user/dir/README

README created successfully with content "hello world".

I didn't change any of the model configuration in the webUI model config:

glm-4.7-flash_bf16-80k-1769159953751.json

<!-- gh-comment-id:3789249677 --> @HuysArthur commented on GitHub (Jan 23, 2026): Here is a clear difference in behavior: ``` user@host dir % opencode run "Create the file 'README' with the contents 'hello world', let me know if it succeeded" --model webui/glm-4.7-flash:bf16-80k | Write Users/user/dir/README user@host dir % rm README user@host dir % opencode run "Create the file 'README' with the contents 'hello world', let me know if it succeeded" --model ollama/glm-4.7-flash:bf16-80k I'll create the README file with the specified content. | Write Users/user/dir/README README created successfully with content "hello world". ``` I didn't change any of the model configuration in the webUI model config: [glm-4.7-flash_bf16-80k-1769159953751.json](https://github.com/user-attachments/files/24817410/glm-4.7-flash_bf16-80k-1769159953751.json)
GiteaMirror commented 2026-04-12 21:53:49 -05:00
Author
Owner
Copy Link

@bitflower commented on GitHub (Jan 25, 2026):

I run this very model with claude code using ollama launch claude. When I prompt, it seems to lose all context with the next prompt. Version 0.15.0.

<!-- gh-comment-id:3797220485 --> @bitflower commented on GitHub (Jan 25, 2026): I run this very model with claude code using `ollama launch claude`. When I prompt, it seems to lose all context with the next prompt. Version 0.15.0.
GiteaMirror commented 2026-04-12 21:53:51 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 25, 2026):

Increase the size of the context buffer in the server. The documentation recommends 64k.

<!-- gh-comment-id:3797236849 --> @rick-github commented on GitHub (Jan 25, 2026): Increase the size of the [context buffer](https://github.com/ollama/ollama/blob/main/docs/faq.mdx#how-can-i-specify-the-context-window-size) in the server. The documentation [recommends](https://docs.ollama.com/integrations/claude-code#manual-setup) 64k.
GiteaMirror commented 2026-04-12 21:53:52 -05:00
Author
Owner
Copy Link

@HuysArthur commented on GitHub (Jan 27, 2026):

Found out this is a OpenWebUI issue, so no longer relevant here.

<!-- gh-comment-id:3804292587 --> @HuysArthur commented on GitHub (Jan 27, 2026): Found out this is a OpenWebUI issue, so no longer relevant here.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#9061
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?