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 #3418] [MERGED] Request and model concurrency #21688

New Issue
Closed
opened 2026-04-19 15:47:14 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-19 15:47:14 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/3418
Author: @dhiltgen
Created: 3/30/2024
Status: ✅ Merged
Merged: 4/23/2024
Merged by: @dhiltgen

Base: main ← Head: concurrency

📝 Commits (2)

  • 34b9db5 Request and model concurrency
  • f2ea847 Local unicode test case

📊 Changes

30 files changed (+2591 additions, -1363 deletions)

View changed files

📝 api/client.go (+7 -0)
📝 format/bytes.go (+1 -0)
📝 gpu/amd_common.go (+56 -14)
📝 gpu/amd_hip_windows.go (+2 -2)
📝 gpu/amd_linux.go (+174 -285)
📝 gpu/amd_windows.go (+80 -74)
📝 gpu/assets.go (+2 -2)
➕ gpu/cuda_common.go (+22 -0)
📝 gpu/gpu.go (+82 -145)
📝 gpu/gpu_darwin.go (+23 -34)
📝 gpu/gpu_info.h (+8 -4)
📝 gpu/gpu_info_cpu.c (+6 -2)
📝 gpu/gpu_info_cudart.c (+63 -82)
📝 gpu/gpu_info_cudart.h (+96 -10)
➖ gpu/gpu_info_nvml.c (+0 -221)
➖ gpu/gpu_info_nvml.h (+0 -57)
📝 gpu/gpu_test.go (+6 -13)
📝 gpu/types.go (+43 -6)
📝 integration/basic_test.go (+44 -2)
➕ integration/concurrency_test.go (+225 -0)

...and 10 more files

📄 Description

This change adds support for multiple concurrent requests, as well as loading multiple models by spawning multiple runners. This change is designed to be "opt in" initially, so the default behavior mimics the current sequential implementation (1 request at a time, and only a single model), but can be changed by setting environment variables for the server. In the future we will adjust the default settings to enable concurrency by default.

By default, this change supports 1 concurrent requests to a loaded model, which can be adjusted via OLLAMA_NUM_PARALLEL.

By default, this change supports 1 loaded model at a time, which can be adjusted via OLLAMA_MAX_LOADED_MODELS . Set to zero for fully dynamic based on VRAM capacity, or a fixed number greater than 1 to limit the total number of loaded models regardless of VRAM capacity. In the >1 scenario, we'll still perform VRAM prediction calculations to see if the new model will completely fit into available VRAM, but will prevent loading more than the specified number of runners even if they would have fit.

This change also adjusts our GPU selection algorithm in multi-GPU scenarios. If we can fit the model into a single GPU, we'll favor that over spreading it to multiple GPUs.

Note: system memory capacity is not taken into consideration in this change, so for CPU mode, setting max runners to zero could result in memory exhaustion and paging/swapping.

Fixes #2109
Fixes #1656
Fixes #1514
Fixes #3304
Fixes #961
Fixes #3507

🔄 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/3418 **Author:** [@dhiltgen](https://github.com/dhiltgen) **Created:** 3/30/2024 **Status:** ✅ Merged **Merged:** 4/23/2024 **Merged by:** [@dhiltgen](https://github.com/dhiltgen) **Base:** `main` ← **Head:** `concurrency` --- ### 📝 Commits (2) - [`34b9db5`](https://github.com/ollama/ollama/commit/34b9db5afc43b352c5ef04fe6ef52684bfdd57b5) Request and model concurrency - [`f2ea847`](https://github.com/ollama/ollama/commit/f2ea8470e5e11f4f251ad0a59bd218a3bc972c24) Local unicode test case ### 📊 Changes **30 files changed** (+2591 additions, -1363 deletions) <details> <summary>View changed files</summary> 📝 `api/client.go` (+7 -0) 📝 `format/bytes.go` (+1 -0) 📝 `gpu/amd_common.go` (+56 -14) 📝 `gpu/amd_hip_windows.go` (+2 -2) 📝 `gpu/amd_linux.go` (+174 -285) 📝 `gpu/amd_windows.go` (+80 -74) 📝 `gpu/assets.go` (+2 -2) ➕ `gpu/cuda_common.go` (+22 -0) 📝 `gpu/gpu.go` (+82 -145) 📝 `gpu/gpu_darwin.go` (+23 -34) 📝 `gpu/gpu_info.h` (+8 -4) 📝 `gpu/gpu_info_cpu.c` (+6 -2) 📝 `gpu/gpu_info_cudart.c` (+63 -82) 📝 `gpu/gpu_info_cudart.h` (+96 -10) ➖ `gpu/gpu_info_nvml.c` (+0 -221) ➖ `gpu/gpu_info_nvml.h` (+0 -57) 📝 `gpu/gpu_test.go` (+6 -13) 📝 `gpu/types.go` (+43 -6) 📝 `integration/basic_test.go` (+44 -2) ➕ `integration/concurrency_test.go` (+225 -0) _...and 10 more files_ </details> ### 📄 Description This change adds support for multiple concurrent requests, as well as loading multiple models by spawning multiple runners. This change is designed to be "opt in" initially, so the default behavior mimics the current sequential implementation (1 request at a time, and only a single model), but can be changed by setting environment variables for the server. In the future we will adjust the default settings to enable concurrency by default. By default, this change supports 1 concurrent requests to a loaded model, which can be adjusted via `OLLAMA_NUM_PARALLEL`. By default, this change supports 1 loaded model at a time, which can be adjusted via `OLLAMA_MAX_LOADED_MODELS `. Set to zero for fully dynamic based on VRAM capacity, or a fixed number greater than 1 to limit the total number of loaded models regardless of VRAM capacity. In the >1 scenario, we'll still perform VRAM prediction calculations to see if the new model will completely fit into available VRAM, but will prevent loading more than the specified number of runners even if they would have fit. This change also adjusts our GPU selection algorithm in multi-GPU scenarios. If we can fit the model into a single GPU, we'll favor that over spreading it to multiple GPUs. Note: system memory capacity is not taken into consideration in this change, so for CPU mode, setting max runners to zero could result in memory exhaustion and paging/swapping. Fixes #2109 Fixes #1656 Fixes #1514 Fixes #3304 Fixes #961 Fixes #3507 --- <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 2026-04-19 15:47:14 -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#21688
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?