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 #8776] Over VRAM allocation - Error in DeekSeek R1:671B model distribution on MI210 #67754

New Issue
Open
opened 2026-05-04 11:34:58 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-05-04 11:34:58 -05:00
Owner
Copy Link

Originally created by @itej89 on GitHub (Feb 2, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/8776

What is the issue?

Hi, I would like to report an issue while running DeepSeek R1 on 3X MI210 GPU configuration.

When I try to run DeekSeek R1:671B on 3X MI210 GPUs (each with 64GB), ollama is trying to allocation 70GB on GPU0 and failing. Please find the error logs below

--------------------------------------------------------------------------------------------------------------------------------------------
_load_backend: loaded ROCm backend from /home/master/vpolamre/ollama/ollama/build/lib/ollama/libggml-hip.so
llama_load_model_from_file: using device ROCm1 (AMD Instinct MI210) - 65430 MiB free
llama_load_model_from_file: using device ROCm2 (AMD Instinct MI210) - 65430 MiB free
llama_load_model_from_file: using device ROCm0 (AMD Instinct MI210) - 64468 MiB free
---
---
llm_load_print_meta: expert_weights_norm  = 1
llm_load_print_meta: expert_gating_func   = sigmoid
llm_load_print_meta: rope_yarn_log_mul    = 0.1000
llm_load_tensors: tensor 'token_embd.weight' (q4_K) (and 531 others) cannot be used with preferred buffer type ROCm_Host, using CPU instead
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 70139.08 MiB on device 0: cudaMalloc failed: out of memory
llama_model_load: error loading model: unable to allocate ROCm0 buffer
llama_load_model_from_file: failed to load model
panic: unable to load model: /root/.ollama/models/blobs/sha256-9801e7fce27dbf3d0bfb468b7b21f1d132131a546dfc43e50518631b8b1800a9_
------------------------------------------------------------------------------------------------------------------------------------------------

I tried debugging issue by printing the GPUs memory computed during layercount calculation and the output is as follows

time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=1 "LAYER SIZE"=530100224 "GPU ID"=0 "REQUIRED MEMORY in GB"=2 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=2 "LAYER SIZE"=530100224 "GPU ID"=1 "REQUIRED MEMORY in GB"=2 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=3 "LAYER SIZE"=530100224 "GPU ID"=2 "REQUIRED MEMORY in GB"=2 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=4 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=9 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=5 "LAYER SIZE"=7619654656 "GPU ID"=1 "REQUIRED MEMORY in GB"=9 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=6 "LAYER SIZE"=7619654656 "GPU ID"=2 "REQUIRED MEMORY in GB"=9 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=7 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=16 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=8 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=15 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=9 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=15 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=10 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=23 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=11 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=21 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=12 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=21 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=13 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=30 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=14 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=28 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=15 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=28 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=16 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=37 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=17 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=34 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=18 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=34 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=19 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=45 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=20 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=40 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=21 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=40 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=22 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=52 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=23 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=46 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=24 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=46 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=25 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=59 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=26 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=52 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=27 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=52 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=28 "LAYER SIZE"=7619654656 "GPU ID"=1 "REQUIRED MEMORY in GB"=59 "GPU FREE MEMORY in GB"=63
time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=29 "LAYER SIZE"=6646985728 "GPU ID"=0 "REQUIRED MEMORY in GB"=59 "GPU FREE MEMORY in GB"=63

As per my experiment I noticed only a total of 24 layers can be successfully offloaded to on to these GPUs through Ollama and llama.cpp.

Ollama MI210 3X GPU Memory allocation fail.txt

MI210 3xGPU - Force 24 layercount.txt

Please let me know if I can share any more information that can help in resolving the issue. Appreciate your support!

OS

Linux

GPU

AMD

CPU

AMD

Ollama version

0.5.7

Originally created by @itej89 on GitHub (Feb 2, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/8776 ### What is the issue? Hi, I would like to report an issue while running DeepSeek R1 on 3X MI210 GPU configuration. When I try to run DeekSeek R1:671B on 3X MI210 GPUs (each with 64GB), ollama is trying to allocation 70GB on GPU0 and failing. Please find the error logs below ``` -------------------------------------------------------------------------------------------------------------------------------------------- _load_backend: loaded ROCm backend from /home/master/vpolamre/ollama/ollama/build/lib/ollama/libggml-hip.so llama_load_model_from_file: using device ROCm1 (AMD Instinct MI210) - 65430 MiB free llama_load_model_from_file: using device ROCm2 (AMD Instinct MI210) - 65430 MiB free llama_load_model_from_file: using device ROCm0 (AMD Instinct MI210) - 64468 MiB free --- --- llm_load_print_meta: expert_weights_norm = 1 llm_load_print_meta: expert_gating_func = sigmoid llm_load_print_meta: rope_yarn_log_mul = 0.1000 llm_load_tensors: tensor 'token_embd.weight' (q4_K) (and 531 others) cannot be used with preferred buffer type ROCm_Host, using CPU instead ggml_backend_cuda_buffer_type_alloc_buffer: allocating 70139.08 MiB on device 0: cudaMalloc failed: out of memory llama_model_load: error loading model: unable to allocate ROCm0 buffer llama_load_model_from_file: failed to load model panic: unable to load model: /root/.ollama/models/blobs/sha256-9801e7fce27dbf3d0bfb468b7b21f1d132131a546dfc43e50518631b8b1800a9_ ------------------------------------------------------------------------------------------------------------------------------------------------ ``` I tried debugging issue by printing the GPUs memory computed during layercount calculation and the output is as follows ``` time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=1 "LAYER SIZE"=530100224 "GPU ID"=0 "REQUIRED MEMORY in GB"=2 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=2 "LAYER SIZE"=530100224 "GPU ID"=1 "REQUIRED MEMORY in GB"=2 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=3 "LAYER SIZE"=530100224 "GPU ID"=2 "REQUIRED MEMORY in GB"=2 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=4 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=9 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=5 "LAYER SIZE"=7619654656 "GPU ID"=1 "REQUIRED MEMORY in GB"=9 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=6 "LAYER SIZE"=7619654656 "GPU ID"=2 "REQUIRED MEMORY in GB"=9 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=7 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=16 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=8 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=15 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=9 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=15 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=10 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=23 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=11 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=21 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=12 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=21 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=13 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=30 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=14 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=28 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=15 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=28 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=16 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=37 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=17 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=34 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=18 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=34 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=19 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=45 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=20 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=40 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=21 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=40 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=22 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=52 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=23 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=46 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=24 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=46 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=25 "LAYER SIZE"=7619654656 "GPU ID"=0 "REQUIRED MEMORY in GB"=59 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=26 "LAYER SIZE"=6646985728 "GPU ID"=1 "REQUIRED MEMORY in GB"=52 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=27 "LAYER SIZE"=6646985728 "GPU ID"=2 "REQUIRED MEMORY in GB"=52 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=28 "LAYER SIZE"=7619654656 "GPU ID"=1 "REQUIRED MEMORY in GB"=59 "GPU FREE MEMORY in GB"=63 time=2025-02-02T19:15:01.440Z level=DEBUG source=memory.go:235 msg="CUSTOM DEBUG PRINT:" "LAYER COUNT"=29 "LAYER SIZE"=6646985728 "GPU ID"=0 "REQUIRED MEMORY in GB"=59 "GPU FREE MEMORY in GB"=63 ``` As per my experiment I noticed only a total of 24 layers can be successfully offloaded to on to these GPUs through Ollama and llama.cpp. [Ollama MI210 3X GPU Memory allocation fail.txt](https://github.com/user-attachments/files/18633774/Ollama.MI210.3X.GPU.Memory.allocation.fail.txt) [MI210 3xGPU - Force 24 layercount.txt](https://github.com/user-attachments/files/18633779/MI210.3xGPU.-.Force.24.layercount.txt) Please let me know if I can share any more information that can help in resolving the issue. Appreciate your support! ### OS Linux ### GPU AMD ### CPU AMD ### Ollama version 0.5.7
GiteaMirror added the bug label 2026-05-04 11:34:58 -05:00
GiteaMirror commented 2026-05-04 11:34:59 -05:00
Author
Owner
Copy Link

@mikessb commented on GitHub (Feb 17, 2025):

Have you resolved this issue? I encountered the same problem on the ollama version 0.5.9！

<!-- gh-comment-id:2661732653 --> @mikessb commented on GitHub (Feb 17, 2025): Have you resolved this issue? I encountered the same problem on the ollama version 0.5.9！
GiteaMirror commented 2026-05-04 11:35:00 -05:00
Author
Owner
Copy Link

@itej89 commented on GitHub (Feb 17, 2025):

Have you resolved this issue? I encountered the same problem on the ollama version 0.5.9！

I am still debugging this one. My initial observations are Ollama is following different schemes for calculating the tensor split sizes and actual loading the tensors. Anyone please correct me if I am understanding it the right way.

Scheme 1: Tensor split calcualtion
https://github.com/ollama/ollama/blob/main/llm/memory.go#L213

Here the code is calculating the GPU offloadable layercount by allocating alternative layers to each gpu till all gpu's are exhausted of space. It starts with Block one.

Scheme2: Loading of tensors
https://github.com/ollama/ollama/blob/main/llama/llama.cpp/src/llama.cpp#L330

Here the code is assigning final layers into the GPU, instead of initial layers used above for the layer count calculation .

As you can see from my first comment, the first 3 layers are of 530100224 bytes but all last layers are either 6646985728 or 7619654656. Since the initial calculation is using the first 3 layers it is estimating more number of layers can be fit.

When I change Scheme1 to use last layers for computation the model does load with 26 layers. But I am facing a different ROCm error. during inference which I am debugging.

<!-- gh-comment-id:2661855768 --> @itej89 commented on GitHub (Feb 17, 2025): > Have you resolved this issue? I encountered the same problem on the ollama version 0.5.9！ I am still debugging this one. My initial observations are Ollama is following different schemes for calculating the tensor split sizes and actual loading the tensors. Anyone please correct me if I am understanding it the right way. Scheme 1: Tensor split calcualtion https://github.com/ollama/ollama/blob/main/llm/memory.go#L213 Here the code is calculating the GPU offloadable layercount by allocating alternative layers to each gpu till all gpu's are exhausted of space. **It starts with Block one**. Scheme2: Loading of tensors https://github.com/ollama/ollama/blob/main/llama/llama.cpp/src/llama.cpp#L330 Here the **code is assigning final layers** into the GPU, instead of initial layers used above for the layer count calculation . As you can see from my first comment, the first 3 layers are of 530100224 bytes but all last layers are either 6646985728 or 7619654656. Since the initial calculation is using the first 3 layers it is estimating more number of layers can be fit. When I change Scheme1 to use last layers for computation the model does load with 26 layers. But I am facing a different ROCm error. during inference which I am debugging.
GiteaMirror commented 2026-05-04 11:35:00 -05:00
Author
Owner
Copy Link

@itej89 commented on GitHub (Feb 20, 2025):

Have you resolved this issue? I encountered the same problem on the ollama version 0.5.9！

I have created this PR to resolve the issue.
https://github.com/ollama/ollama/pull/9243

<!-- gh-comment-id:2671853280 --> @itej89 commented on GitHub (Feb 20, 2025): > Have you resolved this issue? I encountered the same problem on the ollama version 0.5.9！ I have created this PR to resolve the issue. https://github.com/ollama/ollama/pull/9243
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#67754
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?