github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 08:02:14 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15822] MLX runner failed with qwen3.6:35b-a3b-coding-bf16 format=json #56595

New Issue
Open
opened 2026-04-29 11:04:27 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-29 11:04:27 -05:00
Owner
Copy Link

Originally created by @dujeonglee on GitHub (Apr 26, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15822

What is the issue?

curl http://localhost:11434/api/chat -d '{
"model": "qwen3.6:35b-a3b-coding-bf16",
"messages": [{"role":"user","content":"hi"}],
"format": "json",
"stream": false
}'

Relevant log output

{"error":"mlx runner failed: time=2026-04-26T21:12:59.383+09:00 level=INFO source=pipeline.go:129 msg=\"Prompt processing progress\" processed=10 total=11"}

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

ollama version is 0.21.2

Originally created by @dujeonglee on GitHub (Apr 26, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15822 ### What is the issue? curl http://localhost:11434/api/chat -d '{ "model": "qwen3.6:35b-a3b-coding-bf16", "messages": [{"role":"user","content":"hi"}], "format": "json", "stream": false }' ### Relevant log output ```shell {"error":"mlx runner failed: time=2026-04-26T21:12:59.383+09:00 level=INFO source=pipeline.go:129 msg=\"Prompt processing progress\" processed=10 total=11"} ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version ollama version is 0.21.2
GiteaMirror added the bug label 2026-04-29 11:04:27 -05:00
GiteaMirror commented 2026-04-29 11:04:30 -05:00
Author
Owner
Copy Link

@dujeonglee commented on GitHub (Apr 26, 2026):

it is 100% reproduced with the above query with qwen3.6:35b-a3b-coding-bf16 model only.

<!-- gh-comment-id:4322031308 --> @dujeonglee commented on GitHub (Apr 26, 2026): it is 100% reproduced with the above query with qwen3.6:35b-a3b-coding-bf16 model only.
GiteaMirror commented 2026-04-29 11:04:32 -05:00
Author
Owner
Copy Link

@andreinknv commented on GitHub (Apr 26, 2026):

https://github.com/ollama/ollama/pull/15793 check if this fix helps you, or it is different bug

<!-- gh-comment-id:4322309072 --> @andreinknv commented on GitHub (Apr 26, 2026): https://github.com/ollama/ollama/pull/15793 check if this fix helps you, or it is different bug
GiteaMirror commented 2026-04-29 11:04:33 -05:00
Author
Owner
Copy Link

@andreinknv commented on GitHub (Apr 28, 2026):

Following up — I dug into this further and isolated a separate root cause that affects all qwen3.6:35b-a3b-* (and qwen3.5:35b-a3b-*) variants on the MLX runner: the gated_delta_step Metal kernel writes the recurrent state back as InT (the input dtype, bf16 here) instead of StT (fp32). The reference mlx_lm.models.gated_delta uses a separate StT template arg and keeps state in fp32 between recurrent steps; without that, every decode step round-trips state through bf16 and the linear-attention layers' state degrades quickly.

Filed as a focused report with patch + before/after evidence here: #15865.

For the format=json case in this thread specifically: I expect the json mode triggers a different sampling/grammar path and may be a separate bug on top, but worth retesting once the kernel patch lands.

<!-- gh-comment-id:4339464922 --> @andreinknv commented on GitHub (Apr 28, 2026): Following up — I dug into this further and isolated a separate root cause that affects all `qwen3.6:35b-a3b-*` (and `qwen3.5:35b-a3b-*`) variants on the MLX runner: the `gated_delta_step` Metal kernel writes the recurrent state back as **`InT`** (the input dtype, bf16 here) instead of **`StT`** (fp32). The reference `mlx_lm.models.gated_delta` uses a separate `StT` template arg and keeps state in fp32 between recurrent steps; without that, every decode step round-trips state through bf16 and the linear-attention layers' state degrades quickly. Filed as a focused report with patch + before/after evidence here: **#15865**. For the `format=json` case in this thread specifically: I expect the json mode triggers a different sampling/grammar path and may be a separate bug on top, but worth retesting once the kernel patch lands.
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#56595
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?