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 #14957] Format output for qwen 3.5 35b model does not count thinking tokens as eval #9617

New Issue
Open
opened 2026-04-12 22:31:12 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-12 22:31:12 -05:00
Owner
Copy Link

Originally created by @BigArty on GitHub (Mar 19, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14957

What is the issue?

Format output for qwen 3.5 35b model does not count thinking tokens as eval, but adds them as prompt_eval.

This is the request params:

response = client_llm.chat(model="qwen3.5:35b",
messages=messages,
context_length= 40000,
top_p = 0.95,
top_k = 20,
temperature = 1,
repeat_penalty=1.5,
max_tokens=6000,
stream=True,
format=ProcessedTranscription.model_json_schema(),
)

The problem can be identified by changing prompt_eval for the same request running several times.

Relevant log output

OS

Linux

GPU

Nvidia

CPU

Intel

Ollama version

1.18.1

Originally created by @BigArty on GitHub (Mar 19, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14957 ### What is the issue? Format output for qwen 3.5 35b model does not count thinking tokens as eval, but adds them as prompt_eval. This is the request params: response = client_llm.chat(model="qwen3.5:35b", messages=messages, context_length= 40000, top_p = 0.95, top_k = 20, temperature = 1, repeat_penalty=1.5, max_tokens=6000, stream=True, format=ProcessedTranscription.model_json_schema(), ) The problem can be identified by changing prompt_eval for the same request running several times. ### Relevant log output ```shell ``` ### OS Linux ### GPU Nvidia ### CPU Intel ### Ollama version 1.18.1
GiteaMirror added the bug label 2026-04-12 22:31:12 -05:00
GiteaMirror commented 2026-04-12 22:31:13 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 19, 2026):

Inference done with thinking models and structured output is done in two passes. The first accumulates thinking tokens without applying the structured output restriction, the second takes the output of the first and runs inference while applying the structured output restriction. The code should save the prompt_eval_count/eval_count fields after the first pass, and then add the eval_count of the second pass to the stored value. The stored values are then used to update the usage statistics of the generation request.

<!-- gh-comment-id:4092611145 --> @rick-github commented on GitHub (Mar 19, 2026): Inference done with thinking models and structured output is done in two passes. The first accumulates thinking tokens without applying the structured output restriction, the second takes the output of the first and runs inference while applying the structured output restriction. The [code](https://github.com/ollama/ollama/blob/126d8db7f3ad151ead9fb588f3a56cd5f5c9c13b/server/routes.go#L2374) should save the prompt_eval_count/eval_count fields after the first pass, and then add the eval_count of the second pass to the stored value. The stored values are then used to update the usage statistics of the generation request.
GiteaMirror commented 2026-04-12 22:31:13 -05:00
Author
Owner
Copy Link

@BigArty commented on GitHub (Mar 20, 2026):

The main problem for me is that I can't estimate the reasonable max_tokens limit for the generation. Based on the run with 12000 I saw the maximum of 40 requests to be at 3k. So I lowered the max tokens to 6k and the next run I get more then 50% empty replies because the model have not finished thinking.

<!-- gh-comment-id:4096528284 --> @BigArty commented on GitHub (Mar 20, 2026): The main problem for me is that I can't estimate the reasonable max_tokens limit for the generation. Based on the run with 12000 I saw the maximum of 40 requests to be at 3k. So I lowered the max tokens to 6k and the next run I get more then 50% empty replies because the model have not finished thinking.
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#9617
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?