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

[GH-ISSUE #10128] Incorrect VRAM estimation #53159

New Issue
Open
opened 2026-04-29 02:09:15 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-29 02:09:15 -05:00
Owner
Copy Link

Originally created by @Khawn2u on GitHub (Apr 4, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10128

What is the issue?

I run have a batch file to run ollama for windows that looks like the following:

set OLLAMA_FLASH_ATTENTION=1
set OLLAMA_KV_CACHE_TYPE=q8_0
set CUDA_VISIBLE_DEVICES=1,2
ollama serve

But when I load a model "Qwen2.5 7b 1M" with 262144 context window. Ollama only uses 10GB (across my 2 24GB Nvidia P40s, so 5GB on both) out of my 48GB total (across both cards)

Here is a more visual explanation:

1 Tesla P40:
6097MiB / 24576MiB
2 Tesla P40:
5419MiB / 24576MiB

But ollama reports:
hf.co/bartowski/Qwen2.5-7B-Instruct-1M-GGUF:Q4_K_M 311926e8160a 52 GB 4%/96% CPU/GPU 4 minutes from now

This is really bad, as the model could fit on even 1 of my GPUs. I assume this is because the VRAM estimation does not account for quantized KV cache. This should be an easy fix in that case.

Relevant log output

hf.co/bartowski/Qwen2.5-7B-Instruct-1M-GGUF:Q4_K_M    311926e8160a    52 GB    4%/96% CPU/GPU    4 minutes from now

OS

Windows

GPU

Nvidia

CPU

AMD

Ollama version

0.6.3

Originally created by @Khawn2u on GitHub (Apr 4, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10128 ### What is the issue? I run have a batch file to run ollama for windows that looks like the following: set OLLAMA_FLASH_ATTENTION=1 set OLLAMA_KV_CACHE_TYPE=q8_0 set CUDA_VISIBLE_DEVICES=1,2 ollama serve But when I load a model "Qwen2.5 7b 1M" with 262144 context window. Ollama only uses 10GB (across my 2 24GB Nvidia P40s, so 5GB on both) out of my 48GB total (across both cards) Here is a more visual explanation: 1 Tesla P40: 6097MiB / 24576MiB 2 Tesla P40: 5419MiB / 24576MiB But ollama reports: hf.co/bartowski/Qwen2.5-7B-Instruct-1M-GGUF:Q4_K_M 311926e8160a 52 GB 4%/96% CPU/GPU 4 minutes from now This is really bad, as the model could fit on even 1 of my GPUs. I assume this is because the VRAM estimation does not account for quantized KV cache. This should be an easy fix in that case. ### Relevant log output ```shell hf.co/bartowski/Qwen2.5-7B-Instruct-1M-GGUF:Q4_K_M 311926e8160a 52 GB 4%/96% CPU/GPU 4 minutes from now ``` ### OS Windows ### GPU Nvidia ### CPU AMD ### Ollama version 0.6.3
GiteaMirror added the bug label 2026-04-29 02:09:15 -05:00
GiteaMirror commented 2026-04-29 02:09:19 -05:00
Author
Owner
Copy Link

@chigkim commented on GitHub (Apr 5, 2025):

if you search, you'll find a lot of issues regarding this unfortunately.

<!-- gh-comment-id:2781115950 --> @chigkim commented on GitHub (Apr 5, 2025): if you search, you'll find a lot of issues regarding this unfortunately.
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#53159
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?