keep_alive: -1 tells Ollama to keep a model loaded, but the available evidence does not show that this setting by itself deadlocks model switching on a 6 GB GPU. Persistent models can contribute to VRAM pressure and eviction; a separate open Ollama report describes a scheduler hang during a concurrent eviction race, but on a 32 GB RTX 5090. The exact 6 GB scenario and its cause remain unconfirmed.
What `keep_alive: -1` does—and does not mean
Ollama’s FAQ says the default idle residency is five minutes. A negative `keep_alive` value, including `-1`, keeps the model in memory; `0` unloads it after the response. The API parameter overrides the server-wide `OLLAMA_KEEP_ALIVE` setting. You can also unload a model with ollama stop <model>.
As an Amazon Associate I earn from qualifying purchases.
Keeping a model resident uses memory that may otherwise be available for another model or workload. That can lead to ordinary memory pressure or model eviction when Ollama needs room. It is not, by itself, evidence of a deadlock.
Why a 6 GB GPU can run into model-loading limits
Ollama documents that multiple models may be loaded concurrently when memory allows, and that concurrent GPU model loads require the models to fit entirely in VRAM. Nominal GPU capacity alone does not establish whether a workload fits: other GPU users, model requirements, context length, and parallel requests matter too.
#1 Best Overall
- Axial-tech fans now feature a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- 2.5-slot design allows for greater build compatibility while maintaining cooling performance
- 0dB technology lets you enjoy light gaming in relative silence
- Dual BIOS switch lets you toggle between Quiet and Performance BIOS profiles
- Dual ball fan bearings last up to twice as long as sleeve bearing designs
The FAQ documents OLLAMA_NUM_PARALLEL as the maximum number of parallel requests per model (default 1), and says required RAM scales with the number of parallel requests multiplied by context length. It also describes OLLAMA_MAX_LOADED_MODELS as a cap conditional on available memory and OLLAMA_MAX_QUEUE as a queue limit (default 512). These settings affect loading and queuing behavior; raising them is not an established fix for a hang.
What the open scheduler-hang report says
In issue #17408, filed July 26, 2026, a user reports silent hangs when a new model load takes an eviction path and the selected eviction target receives a concurrent request. The report says later cold /api/generate loads hung without logs, while requests to already-loaded models, /api/ps, /api/tags, and /api/embed continued working; restarting the server restored service.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5070 Ti
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
The reporter’s environment was Ollama 0.31.1 on Ubuntu 24.04.4, kernel 6.8.0-107-generic, with an NVIDIA RTX 5090 32 GB. The setup included OLLAMA_KEEP_ALIVE=-1, OLLAMA_NUM_PARALLEL=2, context length 32768, Flash Attention, and an f16 KV cache. The completion model was gemma4:26b Q4_K_M; an embedding runner was CPU-only and pinned with keep_alive: -1.
Free tools Windows power users keep installed
One-click scans. No signup required.
The issue author proposes that a concurrent request can overwrite an eviction mark, leaving a scheduler operation waiting indefinitely. That is the reporter’s code analysis, not an upstream-confirmed root cause. The report is a user account on different hardware, not verification of the 6 GB symptom, and the available evidence does not establish a fix for that exact scenario.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5060
- Integrated with 8GB GDDR7 128bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
How to narrow down your own hang
First distinguish memory pressure or slow loading from a scheduler stall. Ollama’s FAQ recommends ollama ps to see which models are resident and whether they are using CPU or GPU. Record the Ollama version, GPU and backend, available VRAM, model names and sizes, context settings, parallelism, other GPU workloads, and the request sequence immediately before the stall.
- Check residency: run
ollama psand note whether the model that appears stuck is loaded on GPU, CPU, or both. - Check the request pattern: determine whether only cold requests that need a new model to load hang, or whether requests to already-loaded models also stop responding. Note whether requests overlap and whether a model is selected for eviction.
- Compare memory conditions: record free VRAM and other GPU users, along with model, context length, and parallel request settings. Do not treat the GPU’s advertised capacity as a complete fit calculation.
- Capture diagnostics: follow Ollama’s troubleshooting guidance for GPU discovery and initialization. It recommends debug and system diagnostics when those are suspect; for AMD it specifically names
OLLAMA_DEBUG=1and checking system logs for driver errors. Its GPU troubleshooting also covers NVIDIA discovery and container access.
GPU discovery diagnostics can help separate an initialization problem from a scheduler issue, but they do not independently identify the cause. If the process is stuck, preserve server logs and a minimal sequence that reproduces the behavior before restarting, if practical; restarting recovered service in the issue reporter’s account, but does not diagnose the underlying cause.
Rank #4
- Powered by Radeon RX 9070 XT
- WINDFORCE Cooling System
- Hawk Fan
- Server-grade Thermal Conductive Gel
- RGB Lighting
How to interpret the symptom
- Models fail to fit or are repeatedly unloaded: investigate memory demand, context, parallelism, and other GPU workloads. This is consistent with documented memory constraints, not proof of a deadlock.
- Cold model loads stop while already-loaded requests and inspection endpoints still work: that resembles the behavior in issue #17408, but resemblance does not prove the same scheduler race is present.
- GPU discovery or initialization fails: use the relevant Ollama troubleshooting steps and system logs to investigate the GPU/backend path before attributing the failure to model eviction.
“Every other model” and “silent deadlock on a 6 GB GPU” should therefore be treated as reported symptoms, not established behavior across Ollama versions or GPUs. No prevalence figure or universal 6 GB failure threshold is established by the available documentation and report.
Quick Recap
Best Value
- Axial-tech fans now feature a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- Phase-change GPU thermal pad helps ensure optimal heat transfer, lowering GPU temperatures for enhanced performance and reliability
- 2.5-slot design allows for greater build compatibility while maintaining cooling performance
- Dual-ball fan bearings last up to twice as long as standard conventional sleeve bearings designs
- 0dB technology lets you enjoy light gaming in relative silence
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
