github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 17:49:24 -05:00
Code Issues 14.5k Packages Projects Releases 122 Wiki Activity

[PR #9378] [MERGED] ml: split model across backends #18205

New Issue
Closed
opened 2026-04-16 06:28:15 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-16 06:28:15 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/9378
Author: @mxyng
Created: 2/26/2025
Status: ✅ Merged
Merged: 3/7/2025
Merged by: @mxyng

Base: main ← Head: mxyng/gpu-split

📝 Commits (10+)

  • 69e5454 ml/backend/ggml: update model loading for hybrid/multi backends
  • 5094161 model: load non-repeated tensors into multiple backends
  • f1ee773 kvcache: create cache ctx per layer
  • 90fe069 ml/backend/ggml: create tensor on specific backend
  • 83ea2f0 kvcache: update tests
  • e3593c9 ml/backend/ggml: set cpu n_threads
  • de03ba8 ml/backend/ggml: handle user specified cpu offloading
  • 948443b ml/backend/ggml: handle tensor split
  • ef791bf ml/backend/ggml: offload vision to cpu
  • 77add8d ml/backend/ggml: clean up

📊 Changes

8 files changed (+513 additions, -229 deletions)

View changed files

📝 kvcache/causal.go (+34 -22)
📝 kvcache/causal_test.go (+9 -1)
📝 kvcache/encoder.go (+22 -14)
📝 ml/backend.go (+11 -1)
📝 ml/backend/ggml/ggml.go (+408 -164)
📝 model/models/llama/model.go (+11 -11)
📝 model/models/mllama/model.go (+6 -6)
📝 model/models/mllama/model_text.go (+12 -10)

📄 Description

Assign model weights to backends using a strategy similar to llama.cpp as well as assigning inputs, outputs, and cache tensors to their respective backends.

In order to accurately assign tensors, two additional changes have been made to cache and models:

  1. Cache now create separate contexts for each layer. This allows the key and value tensors to be created next to their
  2. Move as a model-level tensor to a layer-level tensor. This ensures it is using the correct backend for the layer

TODO:

  • Accept ml.BackendParams to configure GPU split
  • Check buffer type supports tensor type before assigning
  • Fix multimodal models

🔄 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/9378 **Author:** [@mxyng](https://github.com/mxyng) **Created:** 2/26/2025 **Status:** ✅ Merged **Merged:** 3/7/2025 **Merged by:** [@mxyng](https://github.com/mxyng) **Base:** `main` ← **Head:** `mxyng/gpu-split` --- ### 📝 Commits (10+) - [`69e5454`](https://github.com/ollama/ollama/commit/69e5454c9bb0a7cacbcbf54e042398487002fa39) ml/backend/ggml: update model loading for hybrid/multi backends - [`5094161`](https://github.com/ollama/ollama/commit/509416139d4c3cbde5d2763733a1efd77b8a5d0d) model: load non-repeated tensors into multiple backends - [`f1ee773`](https://github.com/ollama/ollama/commit/f1ee773fdacbff3c8cf0162dbd7f4efbef96ee42) kvcache: create cache ctx per layer - [`90fe069`](https://github.com/ollama/ollama/commit/90fe0696b69657ec671791ed64e5e22a35e39794) ml/backend/ggml: create tensor on specific backend - [`83ea2f0`](https://github.com/ollama/ollama/commit/83ea2f0fc29d049a70d34237d3d72c378cb248af) kvcache: update tests - [`e3593c9`](https://github.com/ollama/ollama/commit/e3593c94daa73466fa05b23861a1390d2ea7b00c) ml/backend/ggml: set cpu n_threads - [`de03ba8`](https://github.com/ollama/ollama/commit/de03ba81bb32d2d88594c44fe8d7d65967e86769) ml/backend/ggml: handle user specified cpu offloading - [`948443b`](https://github.com/ollama/ollama/commit/948443b4ef52bd0e6f3dda910c23525e66998827) ml/backend/ggml: handle tensor split - [`ef791bf`](https://github.com/ollama/ollama/commit/ef791bfa458ddac2d0f1a7cf3036db034a85d33d) ml/backend/ggml: offload vision to cpu - [`77add8d`](https://github.com/ollama/ollama/commit/77add8d604a32d58b27144a556078d805d8dbc8d) ml/backend/ggml: clean up ### 📊 Changes **8 files changed** (+513 additions, -229 deletions) <details> <summary>View changed files</summary> 📝 `kvcache/causal.go` (+34 -22) 📝 `kvcache/causal_test.go` (+9 -1) 📝 `kvcache/encoder.go` (+22 -14) 📝 `ml/backend.go` (+11 -1) 📝 `ml/backend/ggml/ggml.go` (+408 -164) 📝 `model/models/llama/model.go` (+11 -11) 📝 `model/models/mllama/model.go` (+6 -6) 📝 `model/models/mllama/model_text.go` (+12 -10) </details> ### 📄 Description Assign model weights to backends using a strategy similar to `llama.cpp` as well as assigning inputs, outputs, and cache tensors to their respective backends. In order to accurately assign tensors, two additional changes have been made to cache and models: 1. Cache now create separate contexts for each layer. This allows the key and value tensors to be created next to their 2. Move as a model-level tensor to a layer-level tensor. This ensures it is using the correct backend for the layer TODO: - [x] Accept `ml.BackendParams` to configure GPU split - [x] Check buffer type supports tensor type before assigning - [x] Fix multimodal models --- <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 2026-04-16 06:28:15 -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#18205
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?