github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-16 13:19:40 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #15153] GPU is there but fails to be registered #87390

New Issue
Closed
opened 2026-05-10 05:28:01 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-05-10 05:28:01 -05:00
Owner
Copy Link

Originally created by @nexxyrus on GitHub (Mar 30, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15153

What is the issue?

Installed rocm and librocdxg via the instructions but it won't work.

Relevant log output

time=2026-03-31T00:00:26.751+02:00 level=DEBUG source=ggml.go:94 msg="ggml backend load all from path" path=/usr/lib/ollama/rocm
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
ggml_cuda_init: initializing rocBLAS on device 0
pid:588 tid:0x7f3c0edfb6c0 [hsakmt_hsa_signal_store_screlease] hsa_signal_store_screlease not found - /opt/rocm-7.2.1/lib/librocdxg.so: undefined symbol: hsa_signal_store_screlease
time=2026-03-31T00:00:56.727+02:00 level=INFO source=runner.go:464 msg="failure during GPU discovery" OLLAMA_LIBRARY_PATH="[/usr/lib/ollama /usr/lib/ollama/rocm]" extra_envs="map[GGML_CUDA_INIT:1 ROCR_VISIBLE_DEVICES:GPU-3ae513de0b7ca695]" error="failed to finish discovery before timeout"
time=2026-03-31T00:00:56.727+02:00 level=TRACE source=runner.go:467 msg="runner enumerated devices" OLLAMA_LIBRARY_PATH="[/usr/lib/ollama /usr/lib/ollama/rocm]" devices=[]
time=2026-03-31T00:00:56.727+02:00 level=DEBUG source=runner.go:437 msg="bootstrap discovery took" duration=30.001268652s OLLAMA_LIBRARY_PATH="[/usr/lib/ollama /usr/lib/ollama/rocm]" extra_envs="map[GGML_CUDA_INIT:1 ROCR_VISIBLE_DEVICES:GPU-3ae513de0b7ca695]"
time=2026-03-31T00:00:56.727+02:00 level=DEBUG source=runner.go:153 msg="filtering device which didn't fully initialize" id=GPU-3ae513de0b7ca695 libdir=/usr/lib/ollama/rocm pci_id=0000:03:00.0 library=ROCm
time=2026-03-31T00:00:56.727+02:00 level=TRACE source=runner.go:174 msg="supported GPU library combinations before filtering" supported=map[]
time=2026-03-31T00:00:56.727+02:00 level=TRACE source=runner.go:183 msg="removing unsupported or overlapping GPU combination" libDir=/usr/lib/ollama/rocm description="AMD Radeon RX 9060 XT" compute=gfx1200 pci_id=0000:03:00.0
time=2026-03-31T00:00:56.727+02:00 level=DEBUG source=runner.go:40 msg="GPU bootstrap discovery took" duration=30.150257536s
time=2026-03-31T00:00:56.727+02:00 level=INFO source=types.go:60 msg="inference compute" id=cpu library=cpu compute="" name=cpu description=cpu libdirs=ollama driver="" pci_id="" type="" total="15.6 GiB" available="15.0 GiB"
time=2026-03-31T00:00:56.727+02:00 level=INFO source=routes.go:1850 msg="vram-based default context" total_vram="0 B" default_num_ctx=4096

OS

WSL2

GPU

AMD

CPU

AMD

Ollama version

0.19.0

Originally created by @nexxyrus on GitHub (Mar 30, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15153 ### What is the issue? Installed rocm and librocdxg via the instructions but it won't work. ### Relevant log output ```shell time=2026-03-31T00:00:26.751+02:00 level=DEBUG source=ggml.go:94 msg="ggml backend load all from path" path=/usr/lib/ollama/rocm ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ggml_cuda_init: found 1 ROCm devices: ggml_cuda_init: initializing rocBLAS on device 0 pid:588 tid:0x7f3c0edfb6c0 [hsakmt_hsa_signal_store_screlease] hsa_signal_store_screlease not found - /opt/rocm-7.2.1/lib/librocdxg.so: undefined symbol: hsa_signal_store_screlease time=2026-03-31T00:00:56.727+02:00 level=INFO source=runner.go:464 msg="failure during GPU discovery" OLLAMA_LIBRARY_PATH="[/usr/lib/ollama /usr/lib/ollama/rocm]" extra_envs="map[GGML_CUDA_INIT:1 ROCR_VISIBLE_DEVICES:GPU-3ae513de0b7ca695]" error="failed to finish discovery before timeout" time=2026-03-31T00:00:56.727+02:00 level=TRACE source=runner.go:467 msg="runner enumerated devices" OLLAMA_LIBRARY_PATH="[/usr/lib/ollama /usr/lib/ollama/rocm]" devices=[] time=2026-03-31T00:00:56.727+02:00 level=DEBUG source=runner.go:437 msg="bootstrap discovery took" duration=30.001268652s OLLAMA_LIBRARY_PATH="[/usr/lib/ollama /usr/lib/ollama/rocm]" extra_envs="map[GGML_CUDA_INIT:1 ROCR_VISIBLE_DEVICES:GPU-3ae513de0b7ca695]" time=2026-03-31T00:00:56.727+02:00 level=DEBUG source=runner.go:153 msg="filtering device which didn't fully initialize" id=GPU-3ae513de0b7ca695 libdir=/usr/lib/ollama/rocm pci_id=0000:03:00.0 library=ROCm time=2026-03-31T00:00:56.727+02:00 level=TRACE source=runner.go:174 msg="supported GPU library combinations before filtering" supported=map[] time=2026-03-31T00:00:56.727+02:00 level=TRACE source=runner.go:183 msg="removing unsupported or overlapping GPU combination" libDir=/usr/lib/ollama/rocm description="AMD Radeon RX 9060 XT" compute=gfx1200 pci_id=0000:03:00.0 time=2026-03-31T00:00:56.727+02:00 level=DEBUG source=runner.go:40 msg="GPU bootstrap discovery took" duration=30.150257536s time=2026-03-31T00:00:56.727+02:00 level=INFO source=types.go:60 msg="inference compute" id=cpu library=cpu compute="" name=cpu description=cpu libdirs=ollama driver="" pci_id="" type="" total="15.6 GiB" available="15.0 GiB" time=2026-03-31T00:00:56.727+02:00 level=INFO source=routes.go:1850 msg="vram-based default context" total_vram="0 B" default_num_ctx=4096 ``` ### OS WSL2 ### GPU AMD ### CPU AMD ### Ollama version 0.19.0
GiteaMirror added the bugneeds more info labels 2026-05-10 05:28:01 -05:00
GiteaMirror commented 2026-05-10 05:28:04 -05:00
Author
Owner
Copy Link

@mspservices commented on GitHub (Mar 30, 2026):

Additional data point — NVIDIA RTX 5090 on WSL2 (not ROCm)

Same symptom (GPU not detected, 0 layers offloaded) but on NVIDIA/CUDA, not AMD/ROCm.

Environment:

  • OS: Ubuntu 24.04 on WSL2 (kernel 6.6.87.2-microsoft-standard-WSL2)
  • GPU: NVIDIA GeForce RTX 5090 (32 GB VRAM, Blackwell)
  • Driver: 595.97, CUDA 13.2
  • Ollama: 0.19.0 (regression from 0.18.1 which worked fine)

nvidia-smi works:

NVIDIA-SMI 595.58.02    Driver Version: 595.97    CUDA Version: 13.2
GPU 0: NVIDIA GeForce RTX 5090, 2394MiB / 32607MiB

Ollama debug log shows 0 GPUs discovered:

level=INFO  source=runner.go:67  msg="discovering available GPUs..."
level=DEBUG source=runner.go:437 msg="bootstrap discovery took" duration=53.569471ms
level=DEBUG source=runner.go:124 msg="evaluating which, if any, devices to filter out" initial_count=0
level=INFO  source=ggml.go:494  msg="offloaded 0/65 layers to GPU"
level=INFO  source=device.go:245 msg="model weights" device=CPU size="16.2 GiB"

Key detail: initial_count=0 — the CUDA discovery subprocess finds no devices at all, despite nvidia-smi seeing the GPU fine. The 0.18.1 tarball includes both cuda_v12 and cuda_v13 lib dirs; haven't checked if 0.19.0 dropped cuda_v13 or if the detection logic changed.

Workaround: Downgrading to 0.18.1 (testing now, will confirm).

<!-- gh-comment-id:4158561800 --> @mspservices commented on GitHub (Mar 30, 2026): **Additional data point — NVIDIA RTX 5090 on WSL2 (not ROCm)** Same symptom (GPU not detected, 0 layers offloaded) but on NVIDIA/CUDA, not AMD/ROCm. **Environment:** - OS: Ubuntu 24.04 on WSL2 (kernel 6.6.87.2-microsoft-standard-WSL2) - GPU: NVIDIA GeForce RTX 5090 (32 GB VRAM, Blackwell) - Driver: 595.97, CUDA 13.2 - Ollama: 0.19.0 (regression from 0.18.1 which worked fine) **`nvidia-smi` works:** ``` NVIDIA-SMI 595.58.02 Driver Version: 595.97 CUDA Version: 13.2 GPU 0: NVIDIA GeForce RTX 5090, 2394MiB / 32607MiB ``` **Ollama debug log shows 0 GPUs discovered:** ``` level=INFO source=runner.go:67 msg="discovering available GPUs..." level=DEBUG source=runner.go:437 msg="bootstrap discovery took" duration=53.569471ms level=DEBUG source=runner.go:124 msg="evaluating which, if any, devices to filter out" initial_count=0 level=INFO source=ggml.go:494 msg="offloaded 0/65 layers to GPU" level=INFO source=device.go:245 msg="model weights" device=CPU size="16.2 GiB" ``` **Key detail:** `initial_count=0` — the CUDA discovery subprocess finds no devices at all, despite `nvidia-smi` seeing the GPU fine. The 0.18.1 tarball includes both `cuda_v12` and `cuda_v13` lib dirs; haven't checked if 0.19.0 dropped `cuda_v13` or if the detection logic changed. **Workaround:** Downgrading to 0.18.1 (testing now, will confirm).
GiteaMirror commented 2026-05-10 05:28:05 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 30, 2026):

@nexxyrus

pid:588 tid:0x7f3c0edfb6c0 [hsakmt_hsa_signal_store_screlease] hsa_signal_store_screlease not found - /opt/rocm-7.2.1/lib/librocdxg.so: undefined symbol: hsa_signal_store_screlease

Might be a library incompatibility, works fine on a Strix Halo machine running rocm-7.2.0.

@mspservices

level=DEBUG source=runner.go:124 msg="evaluating which, if any, devices to filter out" initial_count=0

Set OLLAMA_DEBUG=2, restart the ollama server and post the full log.

<!-- gh-comment-id:4158708696 --> @rick-github commented on GitHub (Mar 30, 2026): @nexxyrus ``` pid:588 tid:0x7f3c0edfb6c0 [hsakmt_hsa_signal_store_screlease] hsa_signal_store_screlease not found - /opt/rocm-7.2.1/lib/librocdxg.so: undefined symbol: hsa_signal_store_screlease ``` Might be a library incompatibility, works fine on a Strix Halo machine running rocm-7.2.0. @mspservices ``` level=DEBUG source=runner.go:124 msg="evaluating which, if any, devices to filter out" initial_count=0 ``` Set `OLLAMA_DEBUG=2`, restart the ollama server and post the full log.
GiteaMirror commented 2026-05-10 05:28:05 -05:00
Author
Owner
Copy Link

@mspservices commented on GitHub (Mar 31, 2026):

@rick-github Thanks for looking at this. Here's the full log with OLLAMA_DEBUG=2 on Ollama 0.19.0:

Note: We've downgraded to 0.18.1 which works fine with the same GPU/driver. I'll need to temporarily reinstall 0.19.0 to capture the debug=2 log. Will do that and post the full output shortly.

In the meantime, key facts:

  • nvidia-smi sees the GPU fine (RTX 5090, driver 595.97, CUDA 13.2)
  • 0.18.1 binary from the same install path works perfectly (28.9 GB in VRAM, all 65 layers offloaded)
  • 0.19.0 finds initial_count=0 in 53ms — the CUDA discovery subprocess doesn't find any device
  • The 0.18.1 tarball includes both cuda_v12 and cuda_v13 lib dirs; we confirmed both are present in /usr/local/lib/ollama/
  • WSL2 provides libcuda.so at /usr/lib/wsl/lib/ (driver-level), no CUDA toolkit installed in WSL itself

Will capture OLLAMA_DEBUG=2 logs with 0.19.0 and post here.

<!-- gh-comment-id:4162153630 --> @mspservices commented on GitHub (Mar 31, 2026): @rick-github Thanks for looking at this. Here's the full log with `OLLAMA_DEBUG=2` on Ollama 0.19.0: **Note:** We've downgraded to 0.18.1 which works fine with the same GPU/driver. I'll need to temporarily reinstall 0.19.0 to capture the debug=2 log. Will do that and post the full output shortly. In the meantime, key facts: - `nvidia-smi` sees the GPU fine (RTX 5090, driver 595.97, CUDA 13.2) - 0.18.1 binary from the same install path works perfectly (28.9 GB in VRAM, all 65 layers offloaded) - 0.19.0 finds `initial_count=0` in 53ms — the CUDA discovery subprocess doesn't find any device - The 0.18.1 tarball includes both `cuda_v12` and `cuda_v13` lib dirs; we confirmed both are present in `/usr/local/lib/ollama/` - WSL2 provides `libcuda.so` at `/usr/lib/wsl/lib/` (driver-level), no CUDA toolkit installed in WSL itself Will capture `OLLAMA_DEBUG=2` logs with 0.19.0 and post here.
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 needs more info
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#87390
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?