github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[PR #14739] server: handle NaN values in embedding responses #77108

New Issue
Open
opened 2026-05-05 09:48:25 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-05-05 09:48:25 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/14739
Author: @mvanhorn
Created: 3/9/2026
Status: 🔄 Open

Base: main ← Head: osc/14657-fix-embedding-nan

📝 Commits (2)

  • 1f90db4 server: handle NaN values in embedding responses
  • 6aa8c5a server: consolidate embedding validation to single ValidateEmbedding call

📊 Changes

5 files changed (+107 additions, -1 deletions)

View changed files

➕ llm/embedding_test.go (+79 -0)
📝 llm/server.go (+12 -0)
📝 runner/llamarunner/runner.go (+4 -0)
📝 runner/ollamarunner/runner.go (+7 -1)
📝 server/routes.go (+5 -0)

📄 Description

Fixes #14657

Summary

  • Added ValidateEmbedding function in the llm package to detect NaN/Inf values before JSON serialization
  • Applied validation in both runner-level embedding handlers (ollamarunner and llamarunner) where the crash originates
  • Also added NaN/Inf check in the deprecated EmbeddingsHandler endpoint which was missing the validation that EmbedHandler already had via normalize()
  • Returns a clear error message ("model produced invalid embedding values (NaN or Inf)") instead of crashing with json: unsupported value: NaN

Context

Go's encoding/json does not support NaN or Inf float values. When a model (e.g., bge-m3 with certain inputs) produces NaN values in its embeddings, the JSON encoder crashes with an unhelpful 500 error. The EmbedHandler path already catches this via the normalize() function, but the runner-level handlers and the deprecated EmbeddingsHandler did not have this protection.

The workaround OLLAMA_FLASH_ATTENTION=false mentioned in the issue suggests the root cause may be in flash attention computation, which could be investigated separately as a deeper fix.

Testing

  • Added TestValidateEmbedding with coverage for valid embeddings, NaN, positive/negative Inf, empty/nil slices, and edge cases
  • Existing TestNormalize continues to pass

This contribution was developed with AI assistance (Claude Code).

🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/ollama/ollama/pull/14739 **Author:** [@mvanhorn](https://github.com/mvanhorn) **Created:** 3/9/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `osc/14657-fix-embedding-nan` --- ### 📝 Commits (2) - [`1f90db4`](https://github.com/ollama/ollama/commit/1f90db4b93ae1c412ef4e2eceb085ee39e5d8ae8) server: handle NaN values in embedding responses - [`6aa8c5a`](https://github.com/ollama/ollama/commit/6aa8c5aeb6986bc51b556456b1e148e9bb2d6fc3) server: consolidate embedding validation to single ValidateEmbedding call ### 📊 Changes **5 files changed** (+107 additions, -1 deletions) <details> <summary>View changed files</summary> ➕ `llm/embedding_test.go` (+79 -0) 📝 `llm/server.go` (+12 -0) 📝 `runner/llamarunner/runner.go` (+4 -0) 📝 `runner/ollamarunner/runner.go` (+7 -1) 📝 `server/routes.go` (+5 -0) </details> ### 📄 Description Fixes #14657 ## Summary - Added `ValidateEmbedding` function in the `llm` package to detect NaN/Inf values before JSON serialization - Applied validation in both runner-level embedding handlers (`ollamarunner` and `llamarunner`) where the crash originates - Also added NaN/Inf check in the deprecated `EmbeddingsHandler` endpoint which was missing the validation that `EmbedHandler` already had via `normalize()` - Returns a clear error message (`"model produced invalid embedding values (NaN or Inf)"`) instead of crashing with `json: unsupported value: NaN` ## Context Go's `encoding/json` does not support NaN or Inf float values. When a model (e.g., bge-m3 with certain inputs) produces NaN values in its embeddings, the JSON encoder crashes with an unhelpful 500 error. The `EmbedHandler` path already catches this via the `normalize()` function, but the runner-level handlers and the deprecated `EmbeddingsHandler` did not have this protection. The workaround `OLLAMA_FLASH_ATTENTION=false` mentioned in the issue suggests the root cause may be in flash attention computation, which could be investigated separately as a deeper fix. ## Testing - Added `TestValidateEmbedding` with coverage for valid embeddings, NaN, positive/negative Inf, empty/nil slices, and edge cases - Existing `TestNormalize` continues to pass This contribution was developed with AI assistance (Claude Code). --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
GiteaMirror added the pull-request label 2026-05-05 09:48:25 -05:00
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 pull-request
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#77108
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?