github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 08:30:05 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #14872] 🚀 Architecture Proposal: Swarm Memory (Zero-Copy KV-Cache Sharing Across Heterogeneous Models) #56102

New Issue
Closed
opened 2026-04-29 10:16:17 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-29 10:16:17 -05:00
Owner
Copy Link

Originally created by @allornothingai on GitHub (Mar 16, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14872

🚀 Architecture Proposal: Swarm Memory (Zero-Copy KV-Cache Sharing Across Heterogeneous Models)

Hi @jmorganca and the Ollama team. Ollama has won the local inference war, but as we move from single-chat interfaces to autonomous multi-agent swarms (running Qwen for code, LLaVA for vision, and Mistral for routing concurrently), we are hitting a hard VRAM wall.

The core bottleneck isn't the model weights; it's the redundant KV-Cache.

I propose a fundamental architectural shift for the Ollama engine: Swarm Memory.

The Problem: Redundant Context Processing

Right now, if I have three agents analyzing the same 20,000-token codebase, I have to pass that 20k token prompt to three separate models. Ollama computes the KV-cache three separate times, occupying 3x the VRAM and burning 3x the compute, even though the prefix context is mathematically identical.

The Vision: A Unified, Zero-Copy Memory Bus

Ollama should not just be a model runner; it should act as a central Memory Server.

When a user defines a system prompt or a large context window, Ollama computes the KV-cache once and allocates it to a shared memory region in Metal/CUDA. When subsequent models are called with the same prefix context, Ollama detects the exact hash match of the tokens.

Instead of recomputing the cache, the inference engine simply passes a zero-copy pointer to the shared memory region.

// Proposed Architecture Flow:
// 1. User loads large codebase into context.
// 2. Ollama routes to SwarmMemoryManager.
// 3. Cache generated -> Shared VRAM allocation (Address: 0xFA88...)
// 4. Agent A (Qwen) starts generation -> Reads prefix from 0xFA88...
// 5. Agent B (Llama3) starts generation -> Reads prefix from 0xFA88...

Why this changes everything:

If implemented, a 64GB Mac Studio could run 10-15 autonomous agents concurrently operating on the exact same massive context window without ever crashing, because the context is only stored in VRAM once.

It fundamentally shifts Ollama from a "local ChatGPT" to an "Enterprise Swarm OS."

Next Steps

If the core team is interested in tackling this memory optimization, my engineering team (allornothingai) has extensive experience in C++/Metal shared memory architectures. We are prepared to write the memory routing layer and submit a PR.

Let's build the ultimate swarm architecture.

Originally created by @allornothingai on GitHub (Mar 16, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14872 # 🚀 Architecture Proposal: Swarm Memory (Zero-Copy KV-Cache Sharing Across Heterogeneous Models) Hi @jmorganca and the Ollama team. Ollama has won the local inference war, but as we move from single-chat interfaces to autonomous multi-agent swarms (running Qwen for code, LLaVA for vision, and Mistral for routing concurrently), we are hitting a hard VRAM wall. The core bottleneck isn't the model weights; it's the **redundant KV-Cache**. I propose a fundamental architectural shift for the Ollama engine: **Swarm Memory**. ### The Problem: Redundant Context Processing Right now, if I have three agents analyzing the same 20,000-token codebase, I have to pass that 20k token prompt to three separate models. Ollama computes the KV-cache three separate times, occupying 3x the VRAM and burning 3x the compute, even though the prefix context is mathematically identical. ### The Vision: A Unified, Zero-Copy Memory Bus Ollama should not just be a model runner; it should act as a central **Memory Server**. When a user defines a system prompt or a large context window, Ollama computes the KV-cache once and allocates it to a shared memory region in Metal/CUDA. When subsequent models are called with the same prefix context, Ollama detects the exact hash match of the tokens. Instead of recomputing the cache, the inference engine simply passes a **zero-copy pointer** to the shared memory region. ```go // Proposed Architecture Flow: // 1. User loads large codebase into context. // 2. Ollama routes to SwarmMemoryManager. // 3. Cache generated -> Shared VRAM allocation (Address: 0xFA88...) // 4. Agent A (Qwen) starts generation -> Reads prefix from 0xFA88... // 5. Agent B (Llama3) starts generation -> Reads prefix from 0xFA88... ``` ### Why this changes everything: If implemented, a 64GB Mac Studio could run 10-15 autonomous agents concurrently operating on the *exact same* massive context window without ever crashing, because the context is only stored in VRAM once. It fundamentally shifts Ollama from a "local ChatGPT" to an "Enterprise Swarm OS." ### Next Steps If the core team is interested in tackling this memory optimization, my engineering team (`allornothingai`) has extensive experience in C++/Metal shared memory architectures. We are prepared to write the memory routing layer and submit a PR. Let's build the ultimate swarm architecture.
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
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#56102
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?