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 #5476] Scheduler attempts to load model split over cuda + rocm GPUs #3423

New Issue
Closed
opened 2026-04-12 14:04:37 -05:00 by GiteaMirror · 5 comments
GiteaMirror commented 2026-04-12 14:04:37 -05:00
Owner
Copy Link

Originally created by @sksonic on GitHub (Jul 4, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/5476

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

I have two GPUs mixed nvidia (P40) + amd (RTX7900XTX).

I am able to load smaller models - these go to the P40 first. When loading a larger model than can fit on the P40, it seems the malloc operation is trying to allocate the full model size on the first GPU despite the "offload to cuda" logs pointing to a spread across the two GPUs.

Error is: allocating 39979.48 MiB on device 0: cudaMalloc failed: out of memory

ollama[8945]: time=2024-07-03T23:34:38.947+04:00 level=INFO source=memory.go:309 msg="offload to cuda" layers.requested=-1 layers.model=81 layers.offload=81 layers.split=41,40 memory.available="[23.7 GiB 23.5 GiB]" memory.required.full="44.4 GiB" memory.required.partial="44.4 GiB" memory.required.kv="1.2 GiB" memory.required.allocations="[22.6 GiB 21.8 GiB]" memory.weights.total="39.5 GiB" memory.weights.repeating="38.7 GiB" memory.weights.nonrepeating="822.0 MiB" memory.graph.full="1.1 GiB" memory.graph.partial="1.1 GiB"
ollama[8945]: time=2024-07-03T23:34:38.947+04:00 level=INFO source=server.go:368 msg="starting llama server" cmd="/tmp/ollama3425135909/runners/cuda_v11/ollama_llama_server --model /usr/share/ollama/.ollama/models/blobs/sha256-6baa2a027ec7595d421d151fec74dd338a15acebb83e52510a67e08fa4dd7b71 --ctx-size 4000 --batch-size 512 --embedding --log-disable --n-gpu-layers 81 --parallel 1 --tensor-split 41,40 --tensor-split 41,40 --port 43421
...
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 39979.48 MiB on device 0: cudaMalloc failed: out of memory```

### OS

Linux

### GPU

Nvidia, AMD

### CPU

AMD

### Ollama version

0.1.48 (also had issue on previous version)
Originally created by @sksonic on GitHub (Jul 4, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/5476 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? I have two GPUs mixed nvidia (P40) + amd (RTX7900XTX). I am able to load smaller models - these go to the P40 first. When loading a larger model than can fit on the P40, it seems the malloc operation is trying to allocate the full model size on the first GPU despite the "offload to cuda" logs pointing to a spread across the two GPUs. Error is: allocating 39979.48 MiB on device 0: cudaMalloc failed: out of memory ``` ollama[8945]: time=2024-07-03T23:34:38.947+04:00 level=INFO source=memory.go:309 msg="offload to cuda" layers.requested=-1 layers.model=81 layers.offload=81 layers.split=41,40 memory.available="[23.7 GiB 23.5 GiB]" memory.required.full="44.4 GiB" memory.required.partial="44.4 GiB" memory.required.kv="1.2 GiB" memory.required.allocations="[22.6 GiB 21.8 GiB]" memory.weights.total="39.5 GiB" memory.weights.repeating="38.7 GiB" memory.weights.nonrepeating="822.0 MiB" memory.graph.full="1.1 GiB" memory.graph.partial="1.1 GiB" ollama[8945]: time=2024-07-03T23:34:38.947+04:00 level=INFO source=server.go:368 msg="starting llama server" cmd="/tmp/ollama3425135909/runners/cuda_v11/ollama_llama_server --model /usr/share/ollama/.ollama/models/blobs/sha256-6baa2a027ec7595d421d151fec74dd338a15acebb83e52510a67e08fa4dd7b71 --ctx-size 4000 --batch-size 512 --embedding --log-disable --n-gpu-layers 81 --parallel 1 --tensor-split 41,40 --tensor-split 41,40 --port 43421 ... ggml_backend_cuda_buffer_type_alloc_buffer: allocating 39979.48 MiB on device 0: cudaMalloc failed: out of memory``` ### OS Linux ### GPU Nvidia, AMD ### CPU AMD ### Ollama version 0.1.48 (also had issue on previous version)
GiteaMirror added the gpunvidiaamdbug labels 2026-04-12 14:04:37 -05:00
GiteaMirror commented 2026-04-12 14:04:39 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Jul 4, 2024):

Sorry about this - will work on fixing it

<!-- gh-comment-id:2208109469 --> @jmorganca commented on GitHub (Jul 4, 2024): Sorry about this - will work on fixing it
GiteaMirror commented 2026-04-12 14:04:41 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Jul 5, 2024):

Can you share a server log with OLLAMA_DEBUG=1 set when it tries to load across both cards? (It's not supposed to, but it sounds like there's a logic error someplace leading it to make a mistake during scheduling)

<!-- gh-comment-id:2211363178 --> @dhiltgen commented on GitHub (Jul 5, 2024): Can you share a server log with OLLAMA_DEBUG=1 set when it tries to load across both cards? (It's not supposed to, but it sounds like there's a logic error someplace leading it to make a mistake during scheduling)
GiteaMirror commented 2026-04-12 14:04:41 -05:00
Author
Owner
Copy Link

@sksonic commented on GitHub (Jul 6, 2024):

server.log
Attached the server log.

<!-- gh-comment-id:2211692839 --> @sksonic commented on GitHub (Jul 6, 2024): [server.log](https://github.com/user-attachments/files/16115182/server.log) Attached the server log.
GiteaMirror commented 2026-04-12 14:04:42 -05:00
Author
Owner
Copy Link

@sksonic commented on GitHub (Jul 10, 2024):

This is still happening on the latest release.

<!-- gh-comment-id:2219606937 --> @sksonic commented on GitHub (Jul 10, 2024): This is still happening on the latest release.
GiteaMirror commented 2026-04-12 14:04:43 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Jul 22, 2024):

There's a logic flaw somewhere in the scheduler where we're accidentally trying to split a model over mixed GPUs types. We can only spread on the same brand of GPUs. This setup with mixed brands can be used for different models, but not with a single model.

<!-- gh-comment-id:2243542079 --> @dhiltgen commented on GitHub (Jul 22, 2024): There's a logic flaw somewhere in the scheduler where we're accidentally trying to split a model over mixed GPUs types. We can only spread on the same brand of GPUs. This setup with mixed brands can be used for different models, but not with a single model.
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 amd bug gpu nvidia
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#3423
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?