github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #13687] qwen2.5vl:3b no longer runs on 8GB GPUs since Ollama 0.13.4 due to compute graph memory estimation #8986

New Issue
Closed
opened 2026-04-12 21:49:02 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-12 21:49:02 -05:00
Owner
Copy Link

Originally created by @dav-ctrl on GitHub (Jan 12, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/13687

What is the issue?

<html>

qwen2.5vl:3b no longer runs on 8GB GPUs since Ollama 0.13.4 due to compute graph memory estimation

Summary

Starting with Ollama 0.13.4, the model qwen2.5vl:3b no longer runs on GPU on an 8GB VRAM GPU, despite working correctly in Ollama 0.13.3 with identical configuration.

The issue appears to be caused by a significant change in compute graph construction and memory estimation, which increases the required memory from ~1.8 GiB to ~6.7 GiB before layer offloading, causing Ollama to fall back entirely to CPU.

This occurs even when Flash Attention is explicitly disabled.

Environment

  • OS: Windows 11

  • GPU: NVIDIA RTX 4060 Laptop (8 GB VRAM)

  • CUDA compute capability: 8.9

  • Model: qwen2.5vl:3b

  • Quantization: Q4_K_M

  • Context length: 4096

  • Flash Attention: Disabled (OLLAMA_FLASH_ATTENTION=false)

  • Batch size: 512

  • KV cache: default

Ollama Versions Tested

Version | Result -- | -- 0.13.3 | ✅ Fully runs on GPU 0.13.4+ (0.13.5 tested) | ❌ Falls back to CPU

Observed Behavior (Ollama 0.13.5)

When loading the model, Ollama estimates a large compute graph size and decides not to offload any layers to the GPU:

compute graph device=CPU size="6.7 GiB"
model weights device=CPU size="3.2 GiB"
total memory size="10.1 GiB"
offloaded 0/37 layers to GPU

As a result, the model runs entirely on CPU.

Expected Behavior (Ollama 0.13.3)

With the same model and configuration, Ollama successfully offloads all layers to the GPU:

offloaded 37/37 layers to GPU
model weights device=CUDA0 size="3.0 GiB"
kv cache device=CUDA0 size="144.0 MiB"
compute graph device=CUDA0 size="1.8 GiB"
total memory size="5.2 GiB"

The model runs normally on GPU.

Notes

  • The issue reproduces with Flash Attention disabled

  • Reducing context length does not change the behavior

  • Reducing batch size does not change the behavior

  • Limiting GPU layers does not change the behavior

  • The decision to avoid GPU offloading appears to be made before any layers are loaded

  • The regression is reproducible by comparing Ollama 0.13.3 vs 0.13.4+ on the same hardware

</html>

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @dav-ctrl on GitHub (Jan 12, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/13687 ### What is the issue? <html> <body> <!--StartFragment--><h2 data-start="212" data-end="330"><code data-start="230" data-end="244">qwen2.5vl:3b</code> no longer runs on 8GB GPUs since Ollama 0.13.4 due to compute graph memory estimation</h2> <h3 data-start="332" data-end="343">Summary</h3> <p data-start="345" data-end="529">Starting with <strong data-start="359" data-end="376">Ollama 0.13.4</strong>, the model <strong data-start="388" data-end="406"><code data-start="390" data-end="404">qwen2.5vl:3b</code></strong> no longer runs on GPU on an <strong data-start="435" data-end="451">8GB VRAM GPU</strong>, despite working correctly in <strong data-start="482" data-end="499">Ollama 0.13.3</strong> with identical configuration.</p> <p data-start="531" data-end="778">The issue appears to be caused by a <strong data-start="567" data-end="641">significant change in compute graph construction and memory estimation</strong>, which increases the required memory from ~1.8 GiB to ~6.7 GiB <strong data-start="705" data-end="732">before layer offloading</strong>, causing Ollama to fall back entirely to CPU.</p> <p data-start="780" data-end="845">This occurs <strong data-start="792" data-end="844">even when Flash Attention is explicitly disabled</strong>.</p> <hr data-start="847" data-end="850"> <h3 data-start="852" data-end="867">Environment</h3> <ul data-start="869" data-end="1181"> <li data-start="869" data-end="891"> <p data-start="871" data-end="891"><strong data-start="871" data-end="877">OS</strong>: Windows 11</p> </li> <li data-start="892" data-end="939"> <p data-start="894" data-end="939"><strong data-start="894" data-end="901">GPU</strong>: NVIDIA RTX 4060 Laptop (8 GB VRAM)</p> </li> <li data-start="940" data-end="976"> <p data-start="942" data-end="976"><strong data-start="942" data-end="969">CUDA compute capability</strong>: 8.9</p> </li> <li data-start="977" data-end="1006"> <p data-start="979" data-end="1006"><strong data-start="979" data-end="988">Model</strong>: <code data-start="990" data-end="1004">qwen2.5vl:3b</code></p> </li> <li data-start="1007" data-end="1035"> <p data-start="1009" data-end="1035"><strong data-start="1009" data-end="1025">Quantization</strong>: Q4_K_M</p> </li> <li data-start="1036" data-end="1064"> <p data-start="1038" data-end="1064"><strong data-start="1038" data-end="1056">Context length</strong>: 4096</p> </li> <li data-start="1065" data-end="1131"> <p data-start="1067" data-end="1131"><strong data-start="1067" data-end="1086">Flash Attention</strong>: Disabled (<code data-start="1098" data-end="1128">OLLAMA_FLASH_ATTENTION=false</code>)</p> </li> <li data-start="1132" data-end="1155"> <p data-start="1134" data-end="1155"><strong data-start="1134" data-end="1148">Batch size</strong>: 512</p> </li> <li data-start="1156" data-end="1181"> <p data-start="1158" data-end="1181"><strong data-start="1158" data-end="1170">KV cache</strong>: default</p> </li> </ul> <hr data-start="1183" data-end="1186"> <h3 data-start="1188" data-end="1214">Ollama Versions Tested</h3> <div class="TyagGW_tableContainer"><div tabindex="-1" class="group TyagGW_tableWrapper flex flex-col-reverse w-fit"> Version | Result -- | -- 0.13.3 | ✅ Fully runs on GPU 0.13.4+ (0.13.5 tested) | ❌ Falls back to CPU </div></div> <hr data-start="1345" data-end="1348"> <h3 data-start="1350" data-end="1387">Observed Behavior (Ollama 0.13.5)</h3> <p data-start="1389" data-end="1506">When loading the model, Ollama estimates a large compute graph size and decides not to offload any layers to the GPU:</p> <pre class="overflow-visible! px-0!" data-start="1508" data-end="1657"><div class="contain-inline-size rounded-2xl corner-superellipse/1.1 relative bg-token-sidebar-surface-primary"><div class="sticky top-[calc(--spacing(9)+var(--header-height))] @w-xl/main:top-9"><div class="absolute end-0 bottom-0 flex h-9 items-center pe-2"><div class="bg-token-bg-elevated-secondary text-token-text-secondary flex items-center gap-4 rounded-sm px-2 font-sans text-xs"></div></div></div><div class="overflow-y-auto p-4" dir="ltr"><code class="whitespace-pre! language-text"><span><span>compute graph device=CPU size="6.7 GiB" model weights device=CPU size="3.2 GiB" total memory size="10.1 GiB" offloaded 0/37 layers to GPU </span></span></code></div></div></pre> <p data-start="1659" data-end="1703">As a result, the model runs entirely on CPU.</p> <hr data-start="1705" data-end="1708"> <h3 data-start="1710" data-end="1747">Expected Behavior (Ollama 0.13.3)</h3> <p data-start="1749" data-end="1839">With the same model and configuration, Ollama successfully offloads all layers to the GPU:</p> <pre class="overflow-visible! px-0!" data-start="1841" data-end="2033"><div class="contain-inline-size rounded-2xl corner-superellipse/1.1 relative bg-token-sidebar-surface-primary"><div class="sticky top-[calc(--spacing(9)+var(--header-height))] @w-xl/main:top-9"><div class="absolute end-0 bottom-0 flex h-9 items-center pe-2"><div class="bg-token-bg-elevated-secondary text-token-text-secondary flex items-center gap-4 rounded-sm px-2 font-sans text-xs"></div></div></div><div class="overflow-y-auto p-4" dir="ltr"><code class="whitespace-pre! language-text"><span><span>offloaded 37/37 layers to GPU model weights device=CUDA0 size="3.0 GiB" kv cache device=CUDA0 size="144.0 MiB" compute graph device=CUDA0 size="1.8 GiB" total memory size="5.2 GiB" </span></span></code></div></div></pre> <p data-start="2035" data-end="2066">The model runs normally on GPU.</p> <hr data-start="2068" data-end="2071"> <h3 data-start="2073" data-end="2082">Notes</h3> <ul data-start="2084" data-end="2484"> <li data-start="2084" data-end="2140"> <p data-start="2086" data-end="2140">The issue reproduces with <strong data-start="2112" data-end="2140">Flash Attention disabled</strong></p> </li> <li data-start="2141" data-end="2195"> <p data-start="2143" data-end="2195">Reducing context length does not change the behavior</p> </li> <li data-start="2196" data-end="2246"> <p data-start="2198" data-end="2246">Reducing batch size does not change the behavior</p> </li> <li data-start="2247" data-end="2297"> <p data-start="2249" data-end="2297">Limiting GPU layers does not change the behavior</p> </li> <li data-start="2298" data-end="2388"> <p data-start="2300" data-end="2388">The decision to avoid GPU offloading appears to be made <strong data-start="2356" data-end="2388">before any layers are loaded</strong></p> </li> <li data-start="2389" data-end="2484"> <p data-start="2391" data-end="2484">The regression is reproducible by comparing Ollama <strong data-start="2442" data-end="2463">0.13.3 vs 0.13.4+</strong> on the same hardware</p></li></ul><!--EndFragment--> </body> </html> ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-12 21:49:02 -05:00
GiteaMirror commented 2026-04-12 21:49:03 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 12, 2026):

https://github.com/ollama/ollama/pull/13486

<!-- gh-comment-id:3739604231 --> @rick-github commented on GitHub (Jan 12, 2026): https://github.com/ollama/ollama/pull/13486
GiteaMirror commented 2026-04-12 21:49:04 -05:00
Author
Owner
Copy Link

@popkc3 commented on GitHub (Feb 18, 2026):

It’s still not working and the GPU isn’t being used (qwen2.5-vl:7b). Why was this issue closed?

<!-- gh-comment-id:3919461274 --> @popkc3 commented on GitHub (Feb 18, 2026): It’s still not working and the GPU isn’t being used (qwen2.5-vl:7b). Why was this issue closed?
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#8986
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?