github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 17:49:24 -05:00
Code Issues 14.6k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #6558] Multiple GPU´s Nvidia 56GB VRAM gemma2:27b #4127

New Issue
Closed
opened 2026-04-12 15:01:28 -05:00 by GiteaMirror · 15 comments
GiteaMirror commented 2026-04-12 15:01:28 -05:00
Owner
Copy Link

Originally created by @paulopais on GitHub (Aug 29, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/6558

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

Hi,
Error: cudaMalloc failed: out of memory

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.3.8

Originally created by @paulopais on GitHub (Aug 29, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/6558 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? Hi, Error: cudaMalloc failed: out of memory ### OS Windows ### GPU Nvidia ### CPU Intel ### Ollama version 0.3.8
GiteaMirror added the bug label 2026-04-12 15:01:28 -05:00
GiteaMirror commented 2026-04-12 15:01:29 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

Server logs will aid in debugging.

<!-- gh-comment-id:2318258735 --> @rick-github commented on GitHub (Aug 29, 2024): [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) will aid in debugging.
GiteaMirror commented 2026-04-12 15:01:29 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

there: last nvidia drivers
server.log

<!-- gh-comment-id:2318269676 --> @paulopais commented on GitHub (Aug 29, 2024): there: last nvidia drivers [server.log](https://github.com/user-attachments/files/16802113/server.log)
GiteaMirror commented 2026-04-12 15:01:30 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

gpu.go didn't report one of the GPUs but it showed up later, which seems odd. Please add OLLAMA_DEBUG=1 to the server environment, try reloading and then post the log.

<!-- gh-comment-id:2318327930 --> @rick-github commented on GitHub (Aug 29, 2024): gpu.go didn't report one of the GPUs but it showed up later, which seems odd. Please [add](https://github.com/ollama/ollama/blob/main/docs/faq.md#setting-environment-variables-on-windows) `OLLAMA_DEBUG=1` to the server environment, try reloading and then post the log.
GiteaMirror commented 2026-04-12 15:01:30 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

Debug Enable - server.log
Some GPUS disable - server_GPUdisable.log

<!-- gh-comment-id:2318423520 --> @paulopais commented on GitHub (Aug 29, 2024): Debug Enable - [server.log](https://github.com/user-attachments/files/16802963/server.log) Some GPUS disable - [server_GPUdisable.log](https://github.com/user-attachments/files/16802964/server_GPUdisable.log)
GiteaMirror commented 2026-04-12 15:01:30 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

time=2024-08-29T18:17:32.866+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-4b2ad278-bccb-6196-270c-82a4fdd95065 library=cuda compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" overhead="867.3 MiB"
time=2024-08-29T18:17:33.036+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-ab3f1087-a352-d62c-5af9-f6037c3b3117 library=cuda compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" overhead="811.9 MiB"
time=2024-08-29T18:17:33.178+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-68b31f94-a055-1bd7-76ba-ef3538fec9c3 library=cuda compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" overhead="602.9 MiB"
time=2024-08-29T18:17:33.572+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-0eb3a589-a697-56c2-6d04-0e35193a44ff library=cuda compute=7.5 driver=12.6 name="NVIDIA GeForce RTX 2070 SUPER" overhead="4.9 GiB"
time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-4b2ad278-bccb-6196-270c-82a4fdd95065 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB"
time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-ab3f1087-a352-d62c-5af9-f6037c3b3117 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB"
time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-68b31f94-a055-1bd7-76ba-ef3538fec9c3 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB"
time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-90919ffc-17c6-1bb8-338c-20346ee76233 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB"
time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-0eb3a589-a697-56c2-6d04-0e35193a44ff library=cuda variant=v12 compute=7.5 driver=12.6 name="NVIDIA GeForce RTX 2070 SUPER" total="8.0 GiB" available="7.0 GiB"

time=2024-08-29T18:17:34.479+01:00 level=DEBUG source=server.go:101 msg="system memory" total="3.9 GiB" free="2.0 GiB" free_swap="8.0 GiB"
time=2024-08-29T18:17:34.479+01:00 level=DEBUG source=memory.go:101 msg=evaluating library=cuda gpu_count=5 available="[11.0 GiB 11.0 GiB 11.0 GiB 7.8 GiB 7.0 GiB]"
time=2024-08-29T18:17:34.480+01:00 level=INFO source=memory.go:309 msg="offload to cuda" layers.requested=-1 layers.model=47 layers.offload=47 layers.split=10,10,9,9,9 memory.available="[11.0 GiB 11.0 GiB 11.0 GiB 7.8 GiB 7.0 GiB]" memory.required.full="28.4 GiB" memory.required.partial="28.4 GiB" memory.required.kv="2.9 GiB" memory.required.allocations="[5.8 GiB 6.3 GiB 5.4 GiB 5.4 GiB 5.4 GiB]" memory.weights.total="16.5 GiB" memory.weights.repeating="15.6 GiB" memory.weights.nonrepeating="922.9 MiB" memory.graph.full="1.4 GiB" memory.graph.partial="1.4 GiB"

time=2024-08-29T18:17:34.490+01:00 level=INFO source=server.go:391 msg="starting llama server" cmd="C:\\Users\\PP\\AppData\\Local\\Programs\\Ollama\\lib\\ollama\\runners\\cuda_v12\\ollama_llama_server.exe --model C:\\Users\\PP\\.ollama\\models\\blobs\\sha256-d7e4b00a7d7a8d03d4eed9b0f3f61a427e9f0fc5dea6aeb414e41dee23dc8ecc --ctx-size 8192 --batch-size 512 --embedding --log-disable --n-gpu-layers 47 --verbose --no-mmap --parallel 4 --tensor-split 10,10,9,9,9 --port 50268"

ggml_cuda_init: found 5 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
  Device 2: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
  Device 3: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
  Device 4: NVIDIA GeForce RTX 2070 SUPER, compute capability 7.5, VMM: yes
llm_load_tensors: ggml ctx size =    1.36 MiB
ggml_cuda_host_malloc: failed to allocate 922.85 MiB of pinned memory: out of memory
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 2734.38 MiB on device 3: cudaMalloc failed: out of memory
llama_model_load: error loading model: unable to allocate backend buffer

Somebody with more familiarity with Windows will have better insight. The ggml_cuda_host_malloc error would seem to indicate that your system RAM is tight, but the logs show 2G free. The runner is using --no-mmap which may be a factor, what happens if you run ollama run gemma2:27b and then when you get a >>> prompt, enter /set parameter use_mmap true and try asking a question.

The ggml_backend_cuda_buffer_type_alloc_buffer error indicates OOM on the GPU, which again from the logs looks like it should have plenty, It may be a cascading failure from the earlier ggml_cuda_host_malloc error though.

If it's memory pressure, you can try loading less layers on the GPU (set num_gpu in an API call or through /set parameter) or reducing the KV cache size by setting OLLAMA_NUM_PARALLEL=1 in the server environment.

<!-- gh-comment-id:2318531358 --> @rick-github commented on GitHub (Aug 29, 2024): ``` time=2024-08-29T18:17:32.866+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-4b2ad278-bccb-6196-270c-82a4fdd95065 library=cuda compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" overhead="867.3 MiB" time=2024-08-29T18:17:33.036+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-ab3f1087-a352-d62c-5af9-f6037c3b3117 library=cuda compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" overhead="811.9 MiB" time=2024-08-29T18:17:33.178+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-68b31f94-a055-1bd7-76ba-ef3538fec9c3 library=cuda compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" overhead="602.9 MiB" time=2024-08-29T18:17:33.572+01:00 level=INFO source=gpu.go:292 msg="detected OS VRAM overhead" id=GPU-0eb3a589-a697-56c2-6d04-0e35193a44ff library=cuda compute=7.5 driver=12.6 name="NVIDIA GeForce RTX 2070 SUPER" overhead="4.9 GiB" time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-4b2ad278-bccb-6196-270c-82a4fdd95065 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB" time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-ab3f1087-a352-d62c-5af9-f6037c3b3117 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB" time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-68b31f94-a055-1bd7-76ba-ef3538fec9c3 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB" time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-90919ffc-17c6-1bb8-338c-20346ee76233 library=cuda variant=v12 compute=8.6 driver=12.6 name="NVIDIA GeForce RTX 3060" total="12.0 GiB" available="11.0 GiB" time=2024-08-29T18:17:33.574+01:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-0eb3a589-a697-56c2-6d04-0e35193a44ff library=cuda variant=v12 compute=7.5 driver=12.6 name="NVIDIA GeForce RTX 2070 SUPER" total="8.0 GiB" available="7.0 GiB" time=2024-08-29T18:17:34.479+01:00 level=DEBUG source=server.go:101 msg="system memory" total="3.9 GiB" free="2.0 GiB" free_swap="8.0 GiB" time=2024-08-29T18:17:34.479+01:00 level=DEBUG source=memory.go:101 msg=evaluating library=cuda gpu_count=5 available="[11.0 GiB 11.0 GiB 11.0 GiB 7.8 GiB 7.0 GiB]" time=2024-08-29T18:17:34.480+01:00 level=INFO source=memory.go:309 msg="offload to cuda" layers.requested=-1 layers.model=47 layers.offload=47 layers.split=10,10,9,9,9 memory.available="[11.0 GiB 11.0 GiB 11.0 GiB 7.8 GiB 7.0 GiB]" memory.required.full="28.4 GiB" memory.required.partial="28.4 GiB" memory.required.kv="2.9 GiB" memory.required.allocations="[5.8 GiB 6.3 GiB 5.4 GiB 5.4 GiB 5.4 GiB]" memory.weights.total="16.5 GiB" memory.weights.repeating="15.6 GiB" memory.weights.nonrepeating="922.9 MiB" memory.graph.full="1.4 GiB" memory.graph.partial="1.4 GiB" time=2024-08-29T18:17:34.490+01:00 level=INFO source=server.go:391 msg="starting llama server" cmd="C:\\Users\\PP\\AppData\\Local\\Programs\\Ollama\\lib\\ollama\\runners\\cuda_v12\\ollama_llama_server.exe --model C:\\Users\\PP\\.ollama\\models\\blobs\\sha256-d7e4b00a7d7a8d03d4eed9b0f3f61a427e9f0fc5dea6aeb414e41dee23dc8ecc --ctx-size 8192 --batch-size 512 --embedding --log-disable --n-gpu-layers 47 --verbose --no-mmap --parallel 4 --tensor-split 10,10,9,9,9 --port 50268" ggml_cuda_init: found 5 CUDA devices: Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes Device 1: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes Device 2: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes Device 3: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes Device 4: NVIDIA GeForce RTX 2070 SUPER, compute capability 7.5, VMM: yes llm_load_tensors: ggml ctx size = 1.36 MiB ggml_cuda_host_malloc: failed to allocate 922.85 MiB of pinned memory: out of memory ggml_backend_cuda_buffer_type_alloc_buffer: allocating 2734.38 MiB on device 3: cudaMalloc failed: out of memory llama_model_load: error loading model: unable to allocate backend buffer ``` Somebody with more familiarity with Windows will have better insight. The `ggml_cuda_host_malloc` error would seem to indicate that your system RAM is tight, but the logs show 2G free. The runner is using `--no-mmap` which may be a factor, what happens if you run `ollama run gemma2:27b` and then when you get a `>>>` prompt, enter `/set parameter use_mmap true` and try asking a question. The `ggml_backend_cuda_buffer_type_alloc_buffer` error indicates OOM on the GPU, which again from the logs looks like it should have plenty, It may be a cascading failure from the earlier `ggml_cuda_host_malloc` error though. If it's memory pressure, you can try loading less layers on the GPU (set `num_gpu` in an API call or through `/set parameter`) or reducing the KV cache size by [setting](https://github.com/ollama/ollama/blob/main/docs/faq.md#setting-environment-variables-on-windows) `OLLAMA_NUM_PARALLEL=1` in the server environment.
GiteaMirror commented 2026-04-12 15:01:31 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

set parameter use_mmap true -> error is on start never reach the prompt

cascading failure -> for me is this the problem

OLLAMA_NUM_PARALLEL=1 -> same error

<!-- gh-comment-id:2318568023 --> @paulopais commented on GitHub (Aug 29, 2024): set parameter use_mmap true -> error is on start never reach the prompt cascading failure -> for me is this the problem OLLAMA_NUM_PARALLEL=1 -> same error
GiteaMirror commented 2026-04-12 15:01:32 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

set parameter use_mmap true -> error is on start never reach the prompt

OK, I thought you would be able to skip over the initial failure and then set the parameter, but that doesn't work then try calling the API directly:

curl http://localhost:11434/api/generate -d "{\"model\":\"gemma2:27b\",\"options\":{\"use_mmap\":true},\"prompt\":\"hello\",\"stream\":false}"
<!-- gh-comment-id:2318582484 --> @rick-github commented on GitHub (Aug 29, 2024): > set parameter use_mmap true -> error is on start never reach the prompt OK, I thought you would be able to skip over the initial failure and then set the parameter, but that doesn't work then try calling the API directly: ``` curl http://localhost:11434/api/generate -d "{\"model\":\"gemma2:27b\",\"options\":{\"use_mmap\":true},\"prompt\":\"hello\",\"stream\":false}" ```
GiteaMirror commented 2026-04-12 15:01:32 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

same error ..

<!-- gh-comment-id:2318587275 --> @paulopais commented on GitHub (Aug 29, 2024): same error ..
GiteaMirror commented 2026-04-12 15:01:32 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

What happens if you reduce the number of layers offloaded to the GPU:

curl http://localhost:11434/api/generate -d "{\"model\":\"gemma2:27b\",\"options\":{\"num_gpu\":40},\"prompt\":\"hello\",\"stream\":false}"
<!-- gh-comment-id:2318600967 --> @rick-github commented on GitHub (Aug 29, 2024): What happens if you reduce the number of layers offloaded to the GPU: ``` curl http://localhost:11434/api/generate -d "{\"model\":\"gemma2:27b\",\"options\":{\"num_gpu\":40},\"prompt\":\"hello\",\"stream\":false}" ```
GiteaMirror commented 2026-04-12 15:01:33 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

{"error":"llama runner process has terminated: CUDA error"}

<!-- gh-comment-id:2318611993 --> @paulopais commented on GitHub (Aug 29, 2024): {"error":"llama runner process has terminated: CUDA error"}
GiteaMirror commented 2026-04-12 15:01:33 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

Can you post the logs from the num_gpu failure?

<!-- gh-comment-id:2318616040 --> @rick-github commented on GitHub (Aug 29, 2024): Can you post the logs from the `num_gpu` failure?
GiteaMirror commented 2026-04-12 15:01:34 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

server.log
this error only occurs if i call over the network and not on the localhost:
curl http://192.168.5.12:11434/api/generate -d "{"model":"gemma2:27b","options":{"use_mmap":true},"prompt":"hello","stream":false}"

<!-- gh-comment-id:2318656907 --> @paulopais commented on GitHub (Aug 29, 2024): [server.log](https://github.com/user-attachments/files/16804401/server.log) this error only occurs if i call over the network and not on the localhost: curl http://192.168.5.12:11434/api/generate -d "{\"model\":\"gemma2:27b\",\"options\":{\"use_mmap\":true},\"prompt\":\"hello\",\"stream\":false}"
GiteaMirror commented 2026-04-12 15:01:34 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 29, 2024):

The log is from the use_mmap attempt, do you have the logs from the num_gpu attempt?

<!-- gh-comment-id:2318888640 --> @rick-github commented on GitHub (Aug 29, 2024): The log is from the `use_mmap` attempt, do you have the logs from the `num_gpu` attempt?
GiteaMirror commented 2026-04-12 15:01:35 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

server.log

<!-- gh-comment-id:2318896085 --> @paulopais commented on GitHub (Aug 29, 2024): [server.log](https://github.com/user-attachments/files/16805577/server.log)
GiteaMirror commented 2026-04-12 15:01:35 -05:00
Author
Owner
Copy Link

@paulopais commented on GitHub (Aug 29, 2024):

Problem solved. You need to have virtual memory > than the model.

<!-- gh-comment-id:2318983458 --> @paulopais commented on GitHub (Aug 29, 2024): Problem solved. You need to have virtual memory > than the 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 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#4127
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?