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 #10141] [MERGED] models: llama4 multimodal #75449

New Issue
Closed
opened 2026-05-05 07:52:59 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-05-05 07:52:59 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/10141
Author: @mxyng
Created: 4/5/2025
Status: ✅ Merged
Merged: 4/25/2025
Merged by: @mxyng

Base: main ← Head: mxyng/llama4

📝 Commits (6)

📊 Changes

17 files changed (+1513 additions, -19 deletions)

View changed files

📝 convert/convert.go (+2 -0)
📝 convert/convert_llama.go (+10 -3)
➕ convert/convert_llama4.go (+169 -0)
📝 convert/reader.go (+9 -7)
📝 convert/reader_safetensors.go (+15 -0)
📝 convert/reader_torch.go (+11 -0)
📝 fs/ggml/ggml.go (+6 -2)
📝 kvcache/causal.go (+13 -0)
📝 kvcache/causal_test.go (+68 -2)
📝 ml/backend.go (+4 -0)
📝 ml/backend/ggml/ggml.go (+34 -5)
➕ model/models/llama4/model.go (+189 -0)
➕ model/models/llama4/model_text.go (+259 -0)
➕ model/models/llama4/model_vision.go (+256 -0)
➕ model/models/llama4/process_image.go (+167 -0)
➕ model/models/llama4/process_image_test.go (+300 -0)
📝 model/models/models.go (+1 -0)

📄 Description

This change adds Meta's Llama 4 model to Ollama. This has mainly been tested with the 16E Scout model.

Key model features:

  • Mixture of expert
  • Multimodal (vision, text)

Of note, there are a couple things of note:

  • The vision model is trained on tiled images which requires special handling for both the pixel values and rotary embeddings. Both require 5D tensors which is not supported by ggml so the operation has been split into multiple operations
  • The text model uses chunked attention mask for select layers and causal attention for other layers. This requires a minor change to kvcache.Causal similar to sliding window attention

🔄 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/10141 **Author:** [@mxyng](https://github.com/mxyng) **Created:** 4/5/2025 **Status:** ✅ Merged **Merged:** 4/25/2025 **Merged by:** [@mxyng](https://github.com/mxyng) **Base:** `main` ← **Head:** `mxyng/llama4` --- ### 📝 Commits (6) - [`729ab3b`](https://github.com/ollama/ollama/commit/729ab3b699fa85d5320790cc4003bfae54976195) llama4 - [`4de5ec2`](https://github.com/ollama/ollama/commit/4de5ec242634e447736005841f97cccccac07c40) image processing - [`34f93a3`](https://github.com/ollama/ollama/commit/34f93a3e7f6b144c86dcb6d074f65ef5e26f10b7) connect vision to text - [`30af287`](https://github.com/ollama/ollama/commit/30af287fc28227d28771026e28462102599065db) chunked attention - [`8b97f8a`](https://github.com/ollama/ollama/commit/8b97f8adb60e60859d256671ba6b8fb01f2008a6) fixes for maverick - [`e0c0311`](https://github.com/ollama/ollama/commit/e0c031124916bb48b24a4c9fc3738eb56b28b761) memory ### 📊 Changes **17 files changed** (+1513 additions, -19 deletions) <details> <summary>View changed files</summary> 📝 `convert/convert.go` (+2 -0) 📝 `convert/convert_llama.go` (+10 -3) ➕ `convert/convert_llama4.go` (+169 -0) 📝 `convert/reader.go` (+9 -7) 📝 `convert/reader_safetensors.go` (+15 -0) 📝 `convert/reader_torch.go` (+11 -0) 📝 `fs/ggml/ggml.go` (+6 -2) 📝 `kvcache/causal.go` (+13 -0) 📝 `kvcache/causal_test.go` (+68 -2) 📝 `ml/backend.go` (+4 -0) 📝 `ml/backend/ggml/ggml.go` (+34 -5) ➕ `model/models/llama4/model.go` (+189 -0) ➕ `model/models/llama4/model_text.go` (+259 -0) ➕ `model/models/llama4/model_vision.go` (+256 -0) ➕ `model/models/llama4/process_image.go` (+167 -0) ➕ `model/models/llama4/process_image_test.go` (+300 -0) 📝 `model/models/models.go` (+1 -0) </details> ### 📄 Description This change adds Meta's [Llama 4](https://huggingface.co/collections/meta-llama/llama-4-67f0c30d9fe03840bc9d0164) model to Ollama. This has mainly been tested with the 16E Scout model. Key model features: * Mixture of expert * Multimodal (vision, text) Of note, there are a couple things of note: * The vision model is trained on tiled images which requires special handling for both the pixel values and rotary embeddings. Both require 5D tensors which is not supported by ggml so the operation has been split into multiple operations * The text model uses chunked attention mask for select layers and causal attention for other layers. This requires a minor change to `kvcache.Causal` similar to sliding window attention --- <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 07:52:59 -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#75449
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?