github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-21 13:11:39 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[PR #15935] Fix intermittent gemma4 crash from split image group (#15929) #77659

New Issue
Open
opened 2026-05-05 10:20:09 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-05-05 10:20:09 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/15935
Author: @ssam18
Created: 5/2/2026
Status: 🔄 Open

Base: main ← Head: fix-gemma-image-batch-split

📝 Commits (1)

  • df2bc6e Fix intermittent gemma4 crash from split image group (#15929)

📊 Changes

3 files changed (+193 additions, -5 deletions)

View changed files

📝 model/models/gemma3/model.go (+7 -3)
📝 model/models/gemma4/model.go (+11 -2)
➕ model/models/gemma4/model_posttokenize_test.go (+175 -0)

📄 Description

The cache prefix matcher in runner/ollamarunner/cache.go compares only Token and MultimodalHash, so it can match through the image begin token (which carries no hash) and leave the placeholder as the first uncached input. With nothing forcing the image group to stay together from the placeholder onward, the runner can pack another sequence's tokens before it under parallel load, leave the placeholder at a row position where 256 vision tokens no longer fit, and then crash inside hiddenState.View. The fix puts SameBatch on the placeholder itself in both gemma4 and gemma3 so the group stays in one batch regardless of where the cache prefix ends. Added a regression test in model/models/gemma4/model_posttokenize_test.go covering the 256 token case from the crash report and three other PostTokenize paths. Closes #15929

🔄 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/15935 **Author:** [@ssam18](https://github.com/ssam18) **Created:** 5/2/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `fix-gemma-image-batch-split` --- ### 📝 Commits (1) - [`df2bc6e`](https://github.com/ollama/ollama/commit/df2bc6e9acdfdf7fc99d0e3f6b7e5b65af682221) Fix intermittent gemma4 crash from split image group (#15929) ### 📊 Changes **3 files changed** (+193 additions, -5 deletions) <details> <summary>View changed files</summary> 📝 `model/models/gemma3/model.go` (+7 -3) 📝 `model/models/gemma4/model.go` (+11 -2) ➕ `model/models/gemma4/model_posttokenize_test.go` (+175 -0) </details> ### 📄 Description The cache prefix matcher in runner/ollamarunner/cache.go compares only Token and MultimodalHash, so it can match through the image begin token (which carries no hash) and leave the placeholder as the first uncached input. With nothing forcing the image group to stay together from the placeholder onward, the runner can pack another sequence's tokens before it under parallel load, leave the placeholder at a row position where 256 vision tokens no longer fit, and then crash inside hiddenState.View. The fix puts SameBatch on the placeholder itself in both gemma4 and gemma3 so the group stays in one batch regardless of where the cache prefix ends. Added a regression test in model/models/gemma4/model_posttokenize_test.go covering the 256 token case from the crash report and three other PostTokenize paths. Closes #15929 --- <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 10:20:09 -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#77659
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?