github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-23 22:21:55 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #13572] Ollama embedding model bge-m3 produces Nan output in some seemingly unrelated cases #34697

New Issue
Closed
opened 2026-04-22 18:27:41 -05:00 by GiteaMirror · 17 comments
GiteaMirror commented 2026-04-22 18:27:41 -05:00
Owner
Copy Link

Originally created by @JuergenMS on GitHub (Dec 26, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/13572

What is the issue?

The following command produces NaN:
curl http://localhost:11434/api/embed -d '{"model":"bge-m3","input":"Titel: Geschichte der Quantentheorie -- Thema: Wissenschaft- und Technikgeschichte - Physik -- Anmerkungen: Darstellung der Strukturen der Quantentheorie von Plank bis zu den Anfängen der Quantenfeldtheorie"}'
{"error":"failed to encode response: json: unsupported value: NaN"}

Removing the last word "Quantentheorie" or replacing it with "A" returns an embedding.

This behavior occurs with 76 texts out of 1217.
I cannot detect any pattern, I checked for special characters and I also reinstalled the model with no effect. The 76 cases are either English or German and have different lengths.
I attached the commands and the debug logs for the success and the NaN case.
Ollama version is 0.13.5, bge-me is latest.

Ollama-OK.txt
Ollama-Nan.txt

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @JuergenMS on GitHub (Dec 26, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/13572 ### What is the issue? The following command produces NaN: curl http://localhost:11434/api/embed -d '{"model":"bge-m3","input":"Titel: Geschichte der Quantentheorie -- Thema: Wissenschaft- und Technikgeschichte - Physik -- Anmerkungen: Darstellung der Strukturen der Quantentheorie von Plank bis zu den Anfängen der Quantenfeldtheorie"}' {"error":"failed to encode response: json: unsupported value: NaN"} Removing the last word "Quantentheorie" or replacing it with "A" returns an embedding. This behavior occurs with 76 texts out of 1217. I cannot detect any pattern, I checked for special characters and I also reinstalled the model with no effect. The 76 cases are either English or German and have different lengths. I attached the commands and the debug logs for the success and the NaN case. Ollama version is 0.13.5, bge-me is latest. [Ollama-OK.txt](https://github.com/user-attachments/files/24350523/Ollama-OK.txt) [Ollama-Nan.txt](https://github.com/user-attachments/files/24350522/Ollama-Nan.txt) ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-22 18:27:41 -05:00
GiteaMirror commented 2026-04-22 18:27:43 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Dec 26, 2025):

btw: other models like nomic-embed-text:latest work but produce lower quality results since they do not support German very well.

<!-- gh-comment-id:3693320724 --> @JuergenMS commented on GitHub (Dec 26, 2025): btw: other models like nomic-embed-text:latest work but produce lower quality results since they do not support German very well.
GiteaMirror commented 2026-04-22 18:27:44 -05:00
Author
Owner
Copy Link

@anumukul commented on GitHub (Dec 26, 2025):

Hi, I’d like to take this issue and can deliver a fix within 24 hours.
I’ve worked on similar projects before and have relevant experience, so I should be able to handle this efficiently.

<!-- gh-comment-id:3693470724 --> @anumukul commented on GitHub (Dec 26, 2025): Hi, I’d like to take this issue and can deliver a fix within 24 hours. I’ve worked on similar projects before and have relevant experience, so I should be able to handle this efficiently.
GiteaMirror commented 2026-04-22 18:27:45 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Dec 27, 2025):

Thank you very much for your help!!

<!-- gh-comment-id:3694274379 --> @JuergenMS commented on GitHub (Dec 27, 2025): Thank you very much for your help!!
GiteaMirror commented 2026-04-22 18:27:45 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Jan 18, 2026):

I tested this bug with the data and config used above using ollama 014.3.-rc0. I still get the same NaN results.

<!-- gh-comment-id:3765402972 --> @JuergenMS commented on GitHub (Jan 18, 2026): I tested this bug with the data and config used above using ollama 014.3.-rc0. I still get the same NaN results.
GiteaMirror commented 2026-04-22 18:27:46 -05:00
Author
Owner
Copy Link

@krusmir commented on GitHub (Jan 20, 2026):

Root Cause Analysis

I traced this bug to the F32→F16 cast in llama.cpp/src/llama-graph.cpp (lines 1431-1437).

When flash attention is enabled for embedding models (which don't use KV cache), K and V tensors are F32 and get cast to F16 before ggml_flash_attn_ext:

// this can happen when KV cache is not used (e.g. an embedding model with non-causal attn)
if (k->type == GGML_TYPE_F32) {
    k = ggml_cast(ctx0, k, GGML_TYPE_F16);
}
if (v->type == GGML_TYPE_F32) {
    v = ggml_cast(ctx0, v, GGML_TYPE_F16);
}

F16 max value is ~65504. When intermediate values exceed this during cast, they overflow to Inf, which propagates through softmax and becomes NaN.

Proposed Fix

Skip flash attention when K is F32 (embedding models without KV cache):

const bool use_flash_attn = cparams.flash_attn && kq_b == nullptr && k->type != GGML_TYPE_F32;

This makes embedding models use the standard attention path which keeps F32 precision throughout.

Workaround

Setting OLLAMA_FLASH_ATTENTION=false resolves the issue by disabling flash attention entirely.

<!-- gh-comment-id:3774192946 --> @krusmir commented on GitHub (Jan 20, 2026): ## Root Cause Analysis I traced this bug to the F32→F16 cast in `llama.cpp/src/llama-graph.cpp` (lines 1431-1437). When flash attention is enabled for embedding models (which don't use KV cache), K and V tensors are F32 and get cast to F16 before `ggml_flash_attn_ext`: ```cpp // this can happen when KV cache is not used (e.g. an embedding model with non-causal attn) if (k->type == GGML_TYPE_F32) { k = ggml_cast(ctx0, k, GGML_TYPE_F16); } if (v->type == GGML_TYPE_F32) { v = ggml_cast(ctx0, v, GGML_TYPE_F16); } ``` F16 max value is ~65504. When intermediate values exceed this during cast, they overflow to `Inf`, which propagates through softmax and becomes `NaN`. ## Proposed Fix Skip flash attention when K is F32 (embedding models without KV cache): ```cpp const bool use_flash_attn = cparams.flash_attn && kq_b == nullptr && k->type != GGML_TYPE_F32; ``` This makes embedding models use the standard attention path which keeps F32 precision throughout. ## Workaround Setting `OLLAMA_FLASH_ATTENTION=false` resolves the issue by disabling flash attention entirely.
GiteaMirror commented 2026-04-22 18:27:47 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Jan 20, 2026):

I run ollama as systemd and set "Environment=OLLAMA_FLASH_ATTENTION=0" in the override config. I still get the same error.

<!-- gh-comment-id:3774957403 --> @JuergenMS commented on GitHub (Jan 20, 2026): I run ollama as systemd and set "Environment=OLLAMA_FLASH_ATTENTION=0" in the override config. I still get the same error.
GiteaMirror commented 2026-04-22 18:27:47 -05:00
Author
Owner
Copy Link

@krusmir commented on GitHub (Jan 20, 2026):

Confirmed: Bug introduced in v0.13.5

Decided to test a few ollama versions (v.0.13.3, v0.13.4, v0.13.5 up to v0.14.0) and the bug doesn't exist in versions < v0.13.5. The bug was introduced in v0.13.5 when "bert" was added to the architectures that auto-enable flash attention.

Version comparison:

Version bert in FlashAttention() Bug Present
v0.13.3 No No
v0.13.4 No No
v0.13.5 Yes Yes
v0.14.0 Yes Yes

I tested both texts (yours and one I additionally found) on v0.13.5 and confirmed the NaN error:

# English text
curl -X POST "http://localhost:11434/api/embed" \
  -H "Content-Type: application/json" \
  -d '{"model":"bge-m3","input":"To train models for this task, there are currently two large datasets available to the community,"}'
# Returns: {"error":"failed to encode response: json: unsupported value: NaN"}

# German text (your example)
curl -X POST "http://localhost:11434/api/embed" \
  -H "Content-Type: application/json" \
  -d '{"model":"bge-m3","input":"Titel: Geschichte der Quantentheorie -- Thema: Wissenschaft- und Technikgeschichte - Physik -- Anmerkungen: Darstellung der Strukturen der Quantentheorie von Plank bis zu den Anfängen der Quantenfeldtheorie"}'
# Returns: {"error":"failed to encode response: json: unsupported value: NaN"}

For your issue with OLLAMA_FLASH_ATTENTION=0 not working:

  1. Verify the env var is actually loaded:

    systemctl show ollama --property=Environment

    or

    cat /proc/$(pgrep ollama)/environ | tr '\0' '\n' | grep FLASH

  2. Try these alternatives:

    • Use false instead of 0: OLLAMA_FLASH_ATTENTION=false
    • Make sure you ran systemctl daemon-reload after editing the override

  3. What GPU are you using? The bug may have GPU-specific behavior.

  4. Alternative workaround: Downgrade to v0.13.4 where bert wasn't added to the flash attention list yet.

Root Cause (for reference)

In v0.13.5, "bert" was added to FlashAttention() in fs/ggml/ggml.go:

func (f GGML) FlashAttention() bool {
    return slices.Contains([]string{
        "bert",  // <-- Added in v0.13.5, causes the bug for bge-m3
        "gemma3",
        ...
    }, f.KV().String("general.architecture"))
}

This auto-enables flash attention for BERT-based models like bge-m3, triggering the F32→F16 overflow issue I described in my previous comment.

<!-- gh-comment-id:3775230501 --> @krusmir commented on GitHub (Jan 20, 2026): ## Confirmed: Bug introduced in v0.13.5 Decided to test a few ollama versions (`v.0.13.3`, `v0.13.4`, `v0.13.5` up to `v0.14.0`) and the bug doesn't exist in versions < **v0.13.5**. The bug was introduced in **v0.13.5** when `"bert"` was added to the architectures that auto-enable flash attention. **Version comparison:** | Version | `bert` in FlashAttention() | Bug Present | |---------|---------------------------|-------------| | v0.13.3 | No | No | | v0.13.4 | No | No | | **v0.13.5** | **Yes** | **Yes** | | v0.14.0 | Yes | Yes | I tested both texts (yours and one I additionally found) on v0.13.5 and confirmed the NaN error: ```bash # English text curl -X POST "http://localhost:11434/api/embed" \ -H "Content-Type: application/json" \ -d '{"model":"bge-m3","input":"To train models for this task, there are currently two large datasets available to the community,"}' # Returns: {"error":"failed to encode response: json: unsupported value: NaN"} # German text (your example) curl -X POST "http://localhost:11434/api/embed" \ -H "Content-Type: application/json" \ -d '{"model":"bge-m3","input":"Titel: Geschichte der Quantentheorie -- Thema: Wissenschaft- und Technikgeschichte - Physik -- Anmerkungen: Darstellung der Strukturen der Quantentheorie von Plank bis zu den Anfängen der Quantenfeldtheorie"}' # Returns: {"error":"failed to encode response: json: unsupported value: NaN"} ``` --- ### For your issue with `OLLAMA_FLASH_ATTENTION=0` not working: 1. **Verify the env var is actually loaded:** ```bash systemctl show ollama --property=Environment ``` or ```bash cat /proc/$(pgrep ollama)/environ | tr '\0' '\n' | grep FLASH ``` 2. **Try these alternatives:** - Use `false` instead of `0`: `OLLAMA_FLASH_ATTENTION=false` - Make sure you ran `systemctl daemon-reload` after editing the override 3. **What GPU are you using?** The bug may have GPU-specific behavior. 4. **Alternative workaround:** Downgrade to **v0.13.4** where `bert` wasn't added to the flash attention list yet. --- ### Root Cause (for reference) In v0.13.5, `"bert"` was added to `FlashAttention()` in `fs/ggml/ggml.go`: ```go func (f GGML) FlashAttention() bool { return slices.Contains([]string{ "bert", // <-- Added in v0.13.5, causes the bug for bge-m3 "gemma3", ... }, f.KV().String("general.architecture")) } ``` This auto-enables flash attention for BERT-based models like bge-m3, triggering the F32→F16 overflow issue I described in my previous comment.
GiteaMirror commented 2026-04-22 18:27:48 -05:00
Author
Owner
Copy Link

@sleepyddl commented on GitHub (Jan 21, 2026):

OLLAMA_FLASH_ATTENTION=false ， it works in my computer. My version is 14.2

<!-- gh-comment-id:3776017599 --> @sleepyddl commented on GitHub (Jan 21, 2026): OLLAMA_FLASH_ATTENTION=false ， it works in my computer. My version is 14.2
GiteaMirror commented 2026-04-22 18:27:48 -05:00
Author
Owner
Copy Link

@morioka commented on GitHub (Jan 21, 2026):

OLLAMA_FLASH_ATTENTION=false ，jeffh/intfloat-multilingual-e5-large:f16 works well on 0.14.3. Thanks you.

<!-- gh-comment-id:3776348111 --> @morioka commented on GitHub (Jan 21, 2026): OLLAMA_FLASH_ATTENTION=false ，jeffh/intfloat-multilingual-e5-large:f16 works well on 0.14.3. Thanks you.
GiteaMirror commented 2026-04-22 18:27:49 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Jan 21, 2026):

It is also working on my machine. Thank you!

<!-- gh-comment-id:3777963157 --> @JuergenMS commented on GitHub (Jan 21, 2026): It is also working on my machine. Thank you!
GiteaMirror commented 2026-04-22 18:27:50 -05:00
Author
Owner
Copy Link

@danielguerra69 commented on GitHub (Jan 27, 2026):

When I use , OLLAMA_FLASH_ATTENTION=false , ollama uses more memory for non embedding models

<!-- gh-comment-id:3802769630 --> @danielguerra69 commented on GitHub (Jan 27, 2026): When I use , OLLAMA_FLASH_ATTENTION=false , ollama uses more memory for non embedding models
GiteaMirror commented 2026-04-22 18:27:50 -05:00
Author
Owner
Copy Link

@danielguerra69 commented on GitHub (Jan 27, 2026):

Ok, for me it is a memory issue. I was running qwen3-embedding:8b-fp16 on 2X16GGPU which uses 17G, should be no problem on a 32G system. Then i went back to a smaller model embeddinggemma:300m, and no problem. Now I use qwen3-embedding:4b-fp16 that uses 9.6 GB, so it fits on 1 card. That solved the problem for me, without the FLASH_ATTENTION environment (like i said that setting increases about 30% of the memory on most regular models)

<!-- gh-comment-id:3806916925 --> @danielguerra69 commented on GitHub (Jan 27, 2026): Ok, for me it is a memory issue. I was running qwen3-embedding:8b-fp16 on 2X16GGPU which uses 17G, should be no problem on a 32G system. Then i went back to a smaller model embeddinggemma:300m, and no problem. Now I use qwen3-embedding:4b-fp16 that uses 9.6 GB, so it fits on 1 card. That solved the problem for me, without the FLASH_ATTENTION environment (like i said that setting increases about 30% of the memory on most regular models)
GiteaMirror commented 2026-04-22 18:27:51 -05:00
Author
Owner
Copy Link

@shihkauskas commented on GitHub (Mar 3, 2026):

OLLAMA_FLASH_ATTENTION=false

work with ollama version is 0.17.0

<!-- gh-comment-id:3990216438 --> @shihkauskas commented on GitHub (Mar 3, 2026): OLLAMA_FLASH_ATTENTION=false work with ollama version is 0.17.0
GiteaMirror commented 2026-04-22 18:27:52 -05:00
Author
Owner
Copy Link

@telunyang commented on GitHub (Mar 9, 2026):

In my case, after setting OLLAMA_FLASH_ATTENTION=false, the PROCESSOR using gpt-oss:20b runs at 100% CPU. Has anyone encountered this issue?

<!-- gh-comment-id:4023558575 --> @telunyang commented on GitHub (Mar 9, 2026): In my case, after setting `OLLAMA_FLASH_ATTENTION=false`, the PROCESSOR using `gpt-oss:20b` runs at `100% CPU`. Has anyone encountered this issue?
GiteaMirror commented 2026-04-22 18:27:53 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Mar 15, 2026):

After I upgraded to 0.18.0 I get the error again (even with FLASH_ATTENTION=false). Since I could not downgrade to 0.14.3-rc1 again I tried 0.14.3 and even 0.13.4 (where bert had not been included)
I am kind of lost and consider to move away from ollama altogether. Does any body know what else to do else?

<!-- gh-comment-id:4063062084 --> @JuergenMS commented on GitHub (Mar 15, 2026): After I upgraded to 0.18.0 I get the error again (even with FLASH_ATTENTION=false). Since I could not downgrade to 0.14.3-rc1 again I tried 0.14.3 and even 0.13.4 (where bert had not been included) I am kind of lost and consider to move away from ollama altogether. Does any body know what else to do else?
GiteaMirror commented 2026-04-22 18:27:53 -05:00
Author
Owner
Copy Link

@lukestanley commented on GitHub (Mar 18, 2026):

@JuergenMS I had the same problem with an older version. After upgrading to 0.18.1, the problem was solved for me for all tests I did. Without turning off Flash Attention.

<!-- gh-comment-id:4085142903 --> @lukestanley commented on GitHub (Mar 18, 2026): @JuergenMS I had the same problem with an older version. After upgrading to 0.18.1, the problem was solved for me for all tests I did. Without turning off Flash Attention.
GiteaMirror commented 2026-04-22 18:27:54 -05:00
Author
Owner
Copy Link

@JuergenMS commented on GitHub (Mar 18, 2026):

@lukestanley Thank you for your comment.
0.18.1. is working now with flash attention=false., however, it still produces the error if true.

<!-- gh-comment-id:4086208348 --> @JuergenMS commented on GitHub (Mar 18, 2026): @lukestanley Thank you for your comment. 0.18.1. is working now with flash attention=false., however, it still produces the error if true.
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#34697
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?