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 #12400] [MERGED] Preallocate CUDA pool memory #76105

New Issue
Closed
opened 2026-05-05 08:33:48 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-05-05 08:33:48 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/12400
Author: @jessegross
Created: 9/24/2025
Status: ✅ Merged
Merged: 9/30/2025
Merged by: @jessegross

Base: main ← Head: jessegross/memory

📝 Commits (3)

  • f4aba08 ggml: Remove allocation status reporting
  • cf954fe ggml: Backport scale kernel fixes
  • 0b076e8 ggml: Preallocate CUDA pool memory

📊 Changes

15 files changed (+1077 additions, -333 deletions)

View changed files

📝 llama/patches/0014-graph-memory-reporting-on-failure.patch (+22 -37)
📝 llama/patches/0022-ggml-No-alloc-mode.patch (+586 -33)
➕ llama/patches/0026-ggml-Backport-scale-kernel-fixes.patch (+57 -0)
📝 llm/server.go (+19 -19)
📝 llm/server_test.go (+5 -5)
📝 ml/backend.go (+22 -73)
📝 ml/backend/ggml/ggml.go (+56 -88)
📝 ml/backend/ggml/ggml/include/ggml-alloc.h (+1 -6)
📝 ml/backend/ggml/ggml/include/ggml-backend.h (+2 -7)
📝 ml/backend/ggml/ggml/src/ggml-alloc.c (+3 -5)
📝 ml/backend/ggml/ggml/src/ggml-backend-impl.h (+14 -0)
📝 ml/backend/ggml/ggml/src/ggml-backend.cpp (+57 -9)
📝 ml/backend/ggml/ggml/src/ggml-cuda/common.cuh (+46 -2)
📝 ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu (+177 -40)
📝 ml/backend/ggml/ggml/src/ggml-cuda/scale.cu (+10 -9)

📄 Description

The GGML CUDA backend allocates additional memory for intermediate results during calculation. This memory isn't currently allocated during worst case graph reservation and therefore not included in scheduling. This means that as these buffers potentially grow with context length, we could crash.

This extends the memory allocation system down layer from the GGML graph to the CUDA layer, preallocating the worst case memory there as well.

🔄 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/12400 **Author:** [@jessegross](https://github.com/jessegross) **Created:** 9/24/2025 **Status:** ✅ Merged **Merged:** 9/30/2025 **Merged by:** [@jessegross](https://github.com/jessegross) **Base:** `main` ← **Head:** `jessegross/memory` --- ### 📝 Commits (3) - [`f4aba08`](https://github.com/ollama/ollama/commit/f4aba085a19f342c67164d9beb43d4b9079ea45d) ggml: Remove allocation status reporting - [`cf954fe`](https://github.com/ollama/ollama/commit/cf954fec8b761ba4b3110c3d35bb217a1fa0a5b6) ggml: Backport scale kernel fixes - [`0b076e8`](https://github.com/ollama/ollama/commit/0b076e86d6f429675c0be05ea953a5cececc5662) ggml: Preallocate CUDA pool memory ### 📊 Changes **15 files changed** (+1077 additions, -333 deletions) <details> <summary>View changed files</summary> 📝 `llama/patches/0014-graph-memory-reporting-on-failure.patch` (+22 -37) 📝 `llama/patches/0022-ggml-No-alloc-mode.patch` (+586 -33) ➕ `llama/patches/0026-ggml-Backport-scale-kernel-fixes.patch` (+57 -0) 📝 `llm/server.go` (+19 -19) 📝 `llm/server_test.go` (+5 -5) 📝 `ml/backend.go` (+22 -73) 📝 `ml/backend/ggml/ggml.go` (+56 -88) 📝 `ml/backend/ggml/ggml/include/ggml-alloc.h` (+1 -6) 📝 `ml/backend/ggml/ggml/include/ggml-backend.h` (+2 -7) 📝 `ml/backend/ggml/ggml/src/ggml-alloc.c` (+3 -5) 📝 `ml/backend/ggml/ggml/src/ggml-backend-impl.h` (+14 -0) 📝 `ml/backend/ggml/ggml/src/ggml-backend.cpp` (+57 -9) 📝 `ml/backend/ggml/ggml/src/ggml-cuda/common.cuh` (+46 -2) 📝 `ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu` (+177 -40) 📝 `ml/backend/ggml/ggml/src/ggml-cuda/scale.cu` (+10 -9) </details> ### 📄 Description The GGML CUDA backend allocates additional memory for intermediate results during calculation. This memory isn't currently allocated during worst case graph reservation and therefore not included in scheduling. This means that as these buffers potentially grow with context length, we could crash. This extends the memory allocation system down layer from the GGML graph to the CUDA layer, preallocating the worst case memory there as well. --- <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 08:33:48 -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#76105
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?