github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[PR #13042] [MERGED] Prefer dedicated GPUs over iGPUs when offloading #39918

New Issue
Closed
opened 2026-04-23 00:56:14 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-23 00:56:14 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/13042
Author: @jessegross
Created: 11/11/2025
Status: ✅ Merged
Merged: 11/11/2025
Merged by: @jessegross

Base: main ← Head: jessegross/igpu

📝 Commits (4)

  • c0b1f79 llamarunner: Respect device ordering for offloaded layers
  • a51cff6 llm: Use Ollama engine memory layouts for both old and new engines
  • 84ff4dc llm: Separate llamaServer and ollamaServer code paths
  • fe5220b llm: Prefer dedicated GPUs over iGPUs when allocating memory

📊 Changes

9 files changed (+437 additions, -1007 deletions)

View changed files

📝 fs/ggml/ggml.go (+0 -67)
📝 llama/llama.go (+29 -5)
➖ llm/memory.go (+0 -516)
➖ llm/memory_test.go (+0 -130)
📝 llm/server.go (+269 -254)
📝 llm/server_test.go (+60 -29)
📝 ml/device.go (+50 -0)
📝 runner/llamarunner/runner.go (+12 -6)
📝 server/sched.go (+17 -0)

📄 Description

Currently, when we split models over multiple GPUs it is proportional to the amount of free VRAM. This works well if the GPUs have roughly equal performance. However, it does not work well with a mix of dedicated GPUs and iGPUs because iGPUs often have large amounts of memory (system RAM) but much slower performance. Instead, we should prefer the dedicated GPUs and only use iGPUs for layers that would otherwise go onto the CPU.

In support of this, the PR also updates the llama engine to use the same layout code as the Ollama engine. It is still subject to inaccurate inputs as it continues to use estimates rather than allocation data. However, the actual layout is more accurate, especially with multi-GPU setups. Sharing the same layout code also means that it gets the iGPU logic automatically.

🔄 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/13042 **Author:** [@jessegross](https://github.com/jessegross) **Created:** 11/11/2025 **Status:** ✅ Merged **Merged:** 11/11/2025 **Merged by:** [@jessegross](https://github.com/jessegross) **Base:** `main` ← **Head:** `jessegross/igpu` --- ### 📝 Commits (4) - [`c0b1f79`](https://github.com/ollama/ollama/commit/c0b1f798908c9ad245b406e2ed008ec70f058ece) llamarunner: Respect device ordering for offloaded layers - [`a51cff6`](https://github.com/ollama/ollama/commit/a51cff6272cb9f365c5e44e4fab85abcdc14ea83) llm: Use Ollama engine memory layouts for both old and new engines - [`84ff4dc`](https://github.com/ollama/ollama/commit/84ff4dc946ef11cf38fa1e75419a3fb74662c222) llm: Separate llamaServer and ollamaServer code paths - [`fe5220b`](https://github.com/ollama/ollama/commit/fe5220ba6b189b192df5b0e315092a0a5f6f5881) llm: Prefer dedicated GPUs over iGPUs when allocating memory ### 📊 Changes **9 files changed** (+437 additions, -1007 deletions) <details> <summary>View changed files</summary> 📝 `fs/ggml/ggml.go` (+0 -67) 📝 `llama/llama.go` (+29 -5) ➖ `llm/memory.go` (+0 -516) ➖ `llm/memory_test.go` (+0 -130) 📝 `llm/server.go` (+269 -254) 📝 `llm/server_test.go` (+60 -29) 📝 `ml/device.go` (+50 -0) 📝 `runner/llamarunner/runner.go` (+12 -6) 📝 `server/sched.go` (+17 -0) </details> ### 📄 Description Currently, when we split models over multiple GPUs it is proportional to the amount of free VRAM. This works well if the GPUs have roughly equal performance. However, it does not work well with a mix of dedicated GPUs and iGPUs because iGPUs often have large amounts of memory (system RAM) but much slower performance. Instead, we should prefer the dedicated GPUs and only use iGPUs for layers that would otherwise go onto the CPU. In support of this, the PR also updates the llama engine to use the same layout code as the Ollama engine. It is still subject to inaccurate inputs as it continues to use estimates rather than allocation data. However, the actual layout is more accurate, especially with multi-GPU setups. Sharing the same layout code also means that it gets the iGPU logic automatically. --- <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 2026-04-23 00:56:14 -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#39918
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?