github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 08:59:13 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #14562] Vulkan backend fails to detect AMD Radeon 860M iGPU (gfx1152/RDNA 3.5) on Windows #35205

New Issue
Open
opened 2026-04-22 19:34:51 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-22 19:34:51 -05:00
Owner
Copy Link

Originally created by @vicduch on GitHub (Mar 2, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14562

System Info

  • OS: Windows 11 Build 26200
  • CPU: AMD Ryzen AI 7 350 (Zen 5)
  • GPU: AMD Radeon 860M (RDNA 3.5, gfx1152) — integrated GPU
  • VRAM: 8 GB allocated (shared from 32 GB system RAM)
  • Driver: AMD Adrenalin 24.30.62.28
  • Ollama: v0.17.4 and v0.17.5 (both affected)

Vulkan system status

vulkaninfo --summary correctly detects the GPU:

GPU0: AMD Radeon(TM) 860M Graphics
apiVersion = 1.3.302
deviceType = PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
vendorID = 0x1002
deviceID = 0x1114
driverID = DRIVER_ID_AMD_PROPRIETARY

Problem

With OLLAMA_VULKAN=1, the Vulkan discovery runner finds zero devices, even though Vulkan is fully functional at the system level. The runner falls back to CPU with total_vram="0 B".

Debug logs

With OLLAMA_DEBUG=true and OLLAMA_NEW_ENGINE=true:

1. Vulkan runner — no device found

bootstrap discovery took 690ms
OLLAMA_LIBRARY_PATH=[...\lib\ollama ...\lib\ollama\vulkan]
extra_envs=map[]

No GPU reported. Falls through to CPU.

2. ROCm runner — device found but rejected

verifying if device is supported
  library=...\lib\ollama\rocm
  description="AMD Radeon(TM) 860M Graphics"
  compute=gfx1152
  id=0
  pci_id=0000:63:00.0

filtering device which didn't fully initialize
  id=0
  libdir=...\lib\ollama\rocm
  pci_id=0000:63:00.0
  library=ROCm

Device is detected by ROCm but fails verification (gfx1152 not in supported list). Setting HSA_OVERRIDE_GFX_VERSION=11.0.0 does not help — device still gets filtered.

Final result

inference compute id=cpu library=cpu total="23.8 GiB"
vram-based default context total_vram="0 B"

Root cause hypothesis

Ollama bundles vulkan-1.dll v1.4.321.1 in lib/ollama/vulkan/. Replacing it with the system's vulkan-1.dll v1.3.300.0 (which vulkaninfo uses successfully) did not fix the issue. The problem appears to be in ggml-vulkan.dll's device enumeration — it may be filtering out integrated GPUs with UMA (unified memory architecture) or devices reporting 0 bytes of dedicated VRAM.

Proof that Vulkan works on this GPU

KoboldCpp v1.108.2 using its own build of ggml-vulkan detects the GPU immediately:

ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon(TM) 860M Graphics (AMD proprietary driver)
  | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64
  | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat

llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon(TM) 860M Graphics)
load_tensors: offloaded 33/33 layers to GPU

Performance: Prefill 750 tok/s, Generation 8.6 tok/s — fully functional.

Expected behavior

OLLAMA_VULKAN=1 should detect and use the AMD Radeon 860M via Vulkan, as KoboldCpp does successfully with the same hardware, driver, and Vulkan API.

Workaround

Use KoboldCpp with --usevulkan --gpulayers 99 pointing to the GGUF blob in ~/.ollama/models/blobs/.

Originally created by @vicduch on GitHub (Mar 2, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14562 ## System Info - **OS:** Windows 11 Build 26200 - **CPU:** AMD Ryzen AI 7 350 (Zen 5) - **GPU:** AMD Radeon 860M (RDNA 3.5, gfx1152) — integrated GPU - **VRAM:** 8 GB allocated (shared from 32 GB system RAM) - **Driver:** AMD Adrenalin 24.30.62.28 - **Ollama:** v0.17.4 and v0.17.5 (both affected) ## Vulkan system status `vulkaninfo --summary` correctly detects the GPU: ``` GPU0: AMD Radeon(TM) 860M Graphics apiVersion = 1.3.302 deviceType = PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU vendorID = 0x1002 deviceID = 0x1114 driverID = DRIVER_ID_AMD_PROPRIETARY ``` ## Problem With `OLLAMA_VULKAN=1`, the Vulkan discovery runner finds **zero devices**, even though Vulkan is fully functional at the system level. The runner falls back to CPU with `total_vram="0 B"`. ## Debug logs With `OLLAMA_DEBUG=true` and `OLLAMA_NEW_ENGINE=true`: ### 1. Vulkan runner — no device found ``` bootstrap discovery took 690ms OLLAMA_LIBRARY_PATH=[...\lib\ollama ...\lib\ollama\vulkan] extra_envs=map[] ``` No GPU reported. Falls through to CPU. ### 2. ROCm runner — device found but rejected ``` verifying if device is supported library=...\lib\ollama\rocm description="AMD Radeon(TM) 860M Graphics" compute=gfx1152 id=0 pci_id=0000:63:00.0 filtering device which didn't fully initialize id=0 libdir=...\lib\ollama\rocm pci_id=0000:63:00.0 library=ROCm ``` Device is detected by ROCm but fails verification (gfx1152 not in supported list). Setting `HSA_OVERRIDE_GFX_VERSION=11.0.0` does not help — device still gets filtered. ### Final result ``` inference compute id=cpu library=cpu total="23.8 GiB" vram-based default context total_vram="0 B" ``` ## Root cause hypothesis Ollama bundles `vulkan-1.dll` v1.4.321.1 in `lib/ollama/vulkan/`. Replacing it with the system's `vulkan-1.dll` v1.3.300.0 (which `vulkaninfo` uses successfully) did **not** fix the issue. The problem appears to be in `ggml-vulkan.dll`'s device enumeration — it may be filtering out integrated GPUs with UMA (unified memory architecture) or devices reporting 0 bytes of dedicated VRAM. ## Proof that Vulkan works on this GPU **KoboldCpp v1.108.2** using its own build of `ggml-vulkan` detects the GPU immediately: ``` ggml_vulkan: Found 1 Vulkan devices: ggml_vulkan: 0 = AMD Radeon(TM) 860M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon(TM) 860M Graphics) load_tensors: offloaded 33/33 layers to GPU ``` Performance: Prefill 750 tok/s, Generation 8.6 tok/s — fully functional. ## Expected behavior `OLLAMA_VULKAN=1` should detect and use the AMD Radeon 860M via Vulkan, as KoboldCpp does successfully with the same hardware, driver, and Vulkan API. ## Workaround Use KoboldCpp with `--usevulkan --gpulayers 99` pointing to the GGUF blob in `~/.ollama/models/blobs/`.
GiteaMirror commented 2026-04-22 19:34:52 -05:00
Author
Owner
Copy Link

@Jasdfgh commented on GitHub (Mar 12, 2026):

nice writeup. the koboldcpp comparison is really the key piece — their ggml-vulkan picks up the 860M fine, so it's not a driver or system-level thing, it's somewhere in how ollama's vulkan build does device enumeration. probably filtering out devices with 0 dedicated VRAM or something UMA-related?

the vulkan backend is actually maintained upstream at ggml-org/llama.cpp, not in ollama itself — so might be worth cross-filing there. your logs from koboldcpp showing it working on the exact same hardware would make it pretty easy for them to narrow down.

<!-- gh-comment-id:4045392118 --> @Jasdfgh commented on GitHub (Mar 12, 2026): nice writeup. the koboldcpp comparison is really the key piece — their ggml-vulkan picks up the 860M fine, so it's not a driver or system-level thing, it's somewhere in how ollama's vulkan build does device enumeration. probably filtering out devices with 0 dedicated VRAM or something UMA-related? the vulkan backend is actually maintained upstream at ggml-org/llama.cpp, not in ollama itself — so might be worth cross-filing there. your logs from koboldcpp showing it working on the exact same hardware would make it pretty easy for them to narrow down.
GiteaMirror commented 2026-04-22 19:34:53 -05:00
Author
Owner
Copy Link

@hidp123 commented on GitHub (Apr 2, 2026):

Any solution after to get Ollama to use the GPU on AMD Ryzen AI 7 350 w/ Radeon 860M?

<!-- gh-comment-id:4179154941 --> @hidp123 commented on GitHub (Apr 2, 2026): Any solution after to get Ollama to use the GPU on AMD Ryzen AI 7 350 w/ Radeon 860M?
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
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#35205
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?