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 #14130] v0.15.5: OLLAMA_NUM_CTX ignored, causes OOM on systems with ≥48GB VRAM #55730

New Issue
Closed
opened 2026-04-29 09:39:32 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-29 09:39:32 -05:00
Owner
Copy Link

Originally created by @Bunkmil on GitHub (Feb 7, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14130

What is the issue?

Bug Report: v0.15.5 Ignores OLLAMA_NUM_CTX and Forces Excessive Context Size

Summary

Ollama v0.15.5 ignores the OLLAMA_NUM_CTX and OLLAMA_MAX_CONTEXT environment variables and automatically sets extremely large context sizes based on detected VRAM, causing OOM kills on systems with limited RAM.

Environment

  • Ollama Version: 0.15.5
  • OS: CachyOS (Arch Linux)
  • GPU: AMD Radeon 890M (64GB shared VRAM)
  • System RAM: 56GB available
  • Backend: Vulkan / ROCm

Expected Behavior

When OLLAMA_NUM_CTX=6144 is explicitly set in the systemd service configuration, Ollama should respect this value and allocate memory accordingly.

Actual Behavior

Ollama v0.15.5 detects 64GB VRAM and automatically sets default_num_ctx=262144 (262K tokens), completely ignoring the configured OLLAMA_NUM_CTX value. This results in:

  • KV cache allocation of ~96GB (48GB GPU + 48GB CPU)
  • Total memory requirement of 118GB for an 18GB model
  • Immediate OOM kills when attempting to load models
  • Service crashes with oom-kill status

Steps to Reproduce

  1. Configure Ollama service with explicit context limit:
# /etc/systemd/system/ollama.service.d/override.conf
[Service]
Environment="OLLAMA_NUM_CTX=6144"
Environment="OLLAMA_MAX_CONTEXT=6144"
  1. Reload and restart service:
sudo systemctl daemon-reload
sudo systemctl restart ollama
  1. Attempt to load a 30B Q4 model:
ollama run hf.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF:Q4_K_M

Logs

time=2026-02-06T17:33:34.142-05:00 level=INFO source=routes.go:1739 msg="vram-based default context" total_vram="64.0 GiB" default_num_ctx=262144

llama_context: n_ctx = 811008
llama_context: n_ctx_seq = 202752
llama_kv_cache: CPU KV buffer size = 30294.00 MiB
llama_kv_cache: Vulkan0 KV buffer size = 48807.00 MiB
total memory size="118.3 GiB"

radv/amdgpu: Failed to allocate a buffer:
systemd[1]: ollama.service: The kernel OOM killer killed some processes in this unit.
ollama.service: Failed with result 'oom-kill'.

Root Cause Analysis

The changelog for v0.15.5 states:

Ollama will now default to the following context lengths based on VRAM:

  • < 24 GiB VRAM: 4,096 context
  • 24-48 GiB VRAM: 32,768 context

  • = 48 GiB VRAM: 262,144 context

This auto-detection appears to override user-configured OLLAMA_NUM_CTX values instead of being used as a fallback when the variable is not set.

Workaround

Downgrade to v0.15.4:

sudo pacman -U https://archive.archlinux.org/packages/o/ollama/ollama-0.15.4-2-x86_64.pkg.tar.zst

v0.15.4 correctly respects OLLAMA_NUM_CTX and models load successfully.

Impact

This bug makes Ollama v0.15.5 completely unusable on systems with:

  • Shared VRAM/RAM architectures (iGPUs)
  • Systems with ≥48GB VRAM but limited total system memory
  • Any configuration requiring explicit context size control

Proposed Fix

The VRAM-based auto-detection should only apply when OLLAMA_NUM_CTX is not explicitly set. User-configured values should always take precedence over auto-detection.

Suggested logic:

if OLLAMA_NUM_CTX is set:
    use OLLAMA_NUM_CTX
else:
    use VRAM-based auto-detection

Additional Context

  • Models tested: Qwen3-30B-A3B, GLM-4.7-Flash-REAP-23B-A3B, Ministral-3-14B
  • All models work perfectly on v0.15.4 with OLLAMA_NUM_CTX=6144
  • All models fail with OOM on v0.15.5 despite identical configuration
  • The issue affects both Vulkan and ROCm backends

System Configuration Files

Systemd override used (works on v0.15.4, ignored on v0.15.5):

[Service]
Environment="OLLAMA_VULKAN=1"
Environment="VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.x86_64.json"
Environment="OLLAMA_HOST=0.0.0.0:11434"
Environment="OLLAMA_ORIGINS=*"
Environment="OLLAMA_KEEP_ALIVE=24h"
Environment="OLLAMA_MAX_LOADED_MODELS=1"
Environment="OLLAMA_NUM_GPU=1"
Environment="OLLAMA_NUM_PARALLEL=4"
Environment="OLLAMA_NUM_CTX=6144"
Environment="OLLAMA_MAX_CONTEXT=6144"
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="OLLAMA_NUM_THREAD=12"
Environment="OLLAMA_MAX_VRAM=24576"
Environment="OLLAMA_NUM_BATCH=512"

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @Bunkmil on GitHub (Feb 7, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14130 ### What is the issue? # Bug Report: v0.15.5 Ignores OLLAMA_NUM_CTX and Forces Excessive Context Size ## Summary Ollama v0.15.5 ignores the `OLLAMA_NUM_CTX` and `OLLAMA_MAX_CONTEXT` environment variables and automatically sets extremely large context sizes based on detected VRAM, causing OOM kills on systems with limited RAM. ## Environment - **Ollama Version**: 0.15.5 - **OS**: CachyOS (Arch Linux) - **GPU**: AMD Radeon 890M (64GB shared VRAM) - **System RAM**: 56GB available - **Backend**: Vulkan / ROCm ## Expected Behavior When `OLLAMA_NUM_CTX=6144` is explicitly set in the systemd service configuration, Ollama should respect this value and allocate memory accordingly. ## Actual Behavior Ollama v0.15.5 detects 64GB VRAM and automatically sets `default_num_ctx=262144` (262K tokens), completely ignoring the configured `OLLAMA_NUM_CTX` value. This results in: - KV cache allocation of ~96GB (48GB GPU + 48GB CPU) - Total memory requirement of 118GB for an 18GB model - Immediate OOM kills when attempting to load models - Service crashes with `oom-kill` status ## Steps to Reproduce 1. Configure Ollama service with explicit context limit: ```bash # /etc/systemd/system/ollama.service.d/override.conf [Service] Environment="OLLAMA_NUM_CTX=6144" Environment="OLLAMA_MAX_CONTEXT=6144" ``` 2. Reload and restart service: ```bash sudo systemctl daemon-reload sudo systemctl restart ollama ``` 3. Attempt to load a 30B Q4 model: ```bash ollama run hf.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF:Q4_K_M ``` ## Logs ``` time=2026-02-06T17:33:34.142-05:00 level=INFO source=routes.go:1739 msg="vram-based default context" total_vram="64.0 GiB" default_num_ctx=262144 llama_context: n_ctx = 811008 llama_context: n_ctx_seq = 202752 llama_kv_cache: CPU KV buffer size = 30294.00 MiB llama_kv_cache: Vulkan0 KV buffer size = 48807.00 MiB total memory size="118.3 GiB" radv/amdgpu: Failed to allocate a buffer: systemd[1]: ollama.service: The kernel OOM killer killed some processes in this unit. ollama.service: Failed with result 'oom-kill'. ``` ## Root Cause Analysis The changelog for v0.15.5 states: > Ollama will now default to the following context lengths based on VRAM: > - < 24 GiB VRAM: 4,096 context > - 24-48 GiB VRAM: 32,768 context > - >= 48 GiB VRAM: 262,144 context This auto-detection appears to **override** user-configured `OLLAMA_NUM_CTX` values instead of being used as a fallback when the variable is not set. ## Workaround Downgrade to v0.15.4: ```bash sudo pacman -U https://archive.archlinux.org/packages/o/ollama/ollama-0.15.4-2-x86_64.pkg.tar.zst ``` v0.15.4 correctly respects `OLLAMA_NUM_CTX` and models load successfully. ## Impact This bug makes Ollama v0.15.5 **completely unusable** on systems with: - Shared VRAM/RAM architectures (iGPUs) - Systems with ≥48GB VRAM but limited total system memory - Any configuration requiring explicit context size control ## Proposed Fix The VRAM-based auto-detection should only apply when `OLLAMA_NUM_CTX` is **not explicitly set**. User-configured values should always take precedence over auto-detection. Suggested logic: ``` if OLLAMA_NUM_CTX is set: use OLLAMA_NUM_CTX else: use VRAM-based auto-detection ``` ## Additional Context - Models tested: Qwen3-30B-A3B, GLM-4.7-Flash-REAP-23B-A3B, Ministral-3-14B - All models work perfectly on v0.15.4 with `OLLAMA_NUM_CTX=6144` - All models fail with OOM on v0.15.5 despite identical configuration - The issue affects both Vulkan and ROCm backends ## System Configuration Files Systemd override used (works on v0.15.4, ignored on v0.15.5): ```ini [Service] Environment="OLLAMA_VULKAN=1" Environment="VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.x86_64.json" Environment="OLLAMA_HOST=0.0.0.0:11434" Environment="OLLAMA_ORIGINS=*" Environment="OLLAMA_KEEP_ALIVE=24h" Environment="OLLAMA_MAX_LOADED_MODELS=1" Environment="OLLAMA_NUM_GPU=1" Environment="OLLAMA_NUM_PARALLEL=4" Environment="OLLAMA_NUM_CTX=6144" Environment="OLLAMA_MAX_CONTEXT=6144" Environment="OLLAMA_FLASH_ATTENTION=1" Environment="OLLAMA_NUM_THREAD=12" Environment="OLLAMA_MAX_VRAM=24576" Environment="OLLAMA_NUM_BATCH=512" ``` ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-29 09:39:32 -05:00
GiteaMirror commented 2026-04-29 09:39:34 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Feb 7, 2026):

Set OLLAMA_CONTEXT_LENGTH.

The following are not Ollama configuration variables:

OLLAMA_NUM_GPU
OLLAMA_NUM_CTX
OLLAMA_MAX_CONTEXT
OLLAMA_NUM_THREAD
OLLAMA_MAX_VRAM
OLLAMA_NUM_BATCH

<!-- gh-comment-id:3863233020 --> @rick-github commented on GitHub (Feb 7, 2026): Set [`OLLAMA_CONTEXT_LENGTH`](https://github.com/ollama/ollama/blob/main/docs/faq.mdx#how-can-i-specify-the-context-window-size). The following are not Ollama configuration variables: OLLAMA_NUM_GPU OLLAMA_NUM_CTX OLLAMA_MAX_CONTEXT OLLAMA_NUM_THREAD OLLAMA_MAX_VRAM OLLAMA_NUM_BATCH
GiteaMirror commented 2026-04-29 09:39:34 -05:00
Author
Owner
Copy Link

@Bunkmil commented on GitHub (Feb 7, 2026):

Set OLLAMA_CONTEXT_LENGTH.

The following are not Ollama configuration variables:

OLLAMA_NUM_GPU OLLAMA_NUM_CTX OLLAMA_MAX_CONTEXT OLLAMA_NUM_THREAD OLLAMA_MAX_VRAM OLLAMA_NUM_BATCH

You're right. Those must be artifacts of me messing around. It didn't bother until now...
Thanks!

<!-- gh-comment-id:3863341682 --> @Bunkmil commented on GitHub (Feb 7, 2026): > Set [`OLLAMA_CONTEXT_LENGTH`](https://github.com/ollama/ollama/blob/main/docs/faq.mdx#how-can-i-specify-the-context-window-size). > > The following are not Ollama configuration variables: > > OLLAMA_NUM_GPU OLLAMA_NUM_CTX OLLAMA_MAX_CONTEXT OLLAMA_NUM_THREAD OLLAMA_MAX_VRAM OLLAMA_NUM_BATCH You're right. Those must be artifacts of me messing around. It didn't bother until now... Thanks!
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 bug
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#55730
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?