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 #14855] AMD Strix Halo (gfx1151) — ROCm Working Guide: 40 tok/s on 30B models #56095

New Issue
Closed
opened 2026-04-29 10:15:26 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-29 10:15:26 -05:00
Owner
Copy Link

Originally created by @queensone on GitHub (Mar 14, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14855

What is the issue?

AMD Strix Halo (gfx1151) — ROCm Working Guide: 40 tok/s on 30B models

Hardware

  • CPU/GPU: AMD Ryzen AI MAX+ (Strix Halo) — Radeon 8060S Graphics (gfx1151)
  • RAM: 128GB unified memory (~110GB available VRAM)
  • OS: Kubuntu 25.10, kernel 6.19
  • Ollama: v0.18.0

TL;DR

ROCm works natively on gfx1151 with HSA_OVERRIDE_GFX_VERSION=11.5.1. No Vulkan needed. Getting 40 tok/s on GLM-4.7-flash q8_0 (30B) and Qwen3.5 35B at full context. Way faster than Vulkan, which hangs or crashes on these models.

What doesn't work

  • Vulkan backend (OLLAMA_VULKAN=1 + HIP_VISIBLE_DEVICES=-1): Hangs on inference for Qwen3.5 models (known issues #14487, #14509). GLM loads but sometimes crashes. GPU shows 100% usage but produces no tokens.
  • ROCm without HSA override: gfx1151 not recognized by default in some ROCm builds.

What works — the override.conf

Create/edit /etc/systemd/system/ollama.service.d/override.conf:

[Service]
Environment="OLLAMA_KEEP_ALIVE=-1"
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="HSA_OVERRIDE_GFX_VERSION=11.5.1"

Key points:

  • Do NOT set OLLAMA_VULKAN=1
  • Do NOT set HIP_VISIBLE_DEVICES=-1 (this disables ROCm GPU access)
  • HSA_OVERRIDE_GFX_VERSION=11.5.1 tells ROCm to recognize gfx1151

Then reload and restart:

sudo systemctl daemon-reload
sudo systemctl restart ollama

Verification

Ollama logs should show:

inference compute id=0 library=ROCm compute=gfx1151 name=ROCm0 description="Radeon 8060S Graphics" total="111.5 GiB" available="110.5 GiB"

All layers offloaded to GPU:

offloaded 48/48 layers to GPU
model weights device=ROCm0 size="29.3 GiB"

Benchmarks

Model Quant Context tok/s VRAM Used
GLM-4.7-flash Q8_0 (30B) 4K ~40 tok/s ~30 GB
GLM-4.7-flash Q8_0 (30B) 202K ~40 tok/s ~50 GB
Qwen3.5 35B-a3b Q8_0 65K ~40 tok/s ~36 GB

VRAM Tips for 128GB Unified Memory

With unified memory, the GPU shares RAM with the system. Be careful with large context windows:

  • 200K+ context on a 30B model uses ~50-70GB, leaving less for your desktop/browser
  • If VRAM runs out, the amdgpu kernel driver spams Not enough memory for command submission! and can crash the entire system
  • Use Ollama's num_ctx parameter or your frontend's context setting to limit context size
  • A safe starting point: 65K-96K context for 30B models, which uses ~36-45GB

What I tried that failed

  1. Vulkan + HIP_VISIBLE_DEVICES=-1 — worked initially for some models but hung on Qwen3.5 and became unreliable
  2. ROCm without HSA_OVERRIDE_GFX_VERSION — GPU not detected
  3. 202K context with 30B model — works but risky on unified memory, can OOM and crash the system if other apps use too much RAM

Hope this helps other Strix Halo users! This took a full day of troubleshooting to figure out.

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @queensone on GitHub (Mar 14, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14855 ### What is the issue? # AMD Strix Halo (gfx1151) — ROCm Working Guide: 40 tok/s on 30B models ## Hardware - **CPU/GPU**: AMD Ryzen AI MAX+ (Strix Halo) — Radeon 8060S Graphics (gfx1151) - **RAM**: 128GB unified memory (~110GB available VRAM) - **OS**: Kubuntu 25.10, kernel 6.19 - **Ollama**: v0.18.0 ## TL;DR ROCm works natively on gfx1151 with `HSA_OVERRIDE_GFX_VERSION=11.5.1`. No Vulkan needed. Getting **40 tok/s** on GLM-4.7-flash q8_0 (30B) and Qwen3.5 35B at full context. Way faster than Vulkan, which hangs or crashes on these models. ## What doesn't work - **Vulkan backend** (`OLLAMA_VULKAN=1` + `HIP_VISIBLE_DEVICES=-1`): Hangs on inference for Qwen3.5 models (known issues #14487, #14509). GLM loads but sometimes crashes. GPU shows 100% usage but produces no tokens. - **ROCm without HSA override**: gfx1151 not recognized by default in some ROCm builds. ## What works — the override.conf Create/edit `/etc/systemd/system/ollama.service.d/override.conf`: ```ini [Service] Environment="OLLAMA_KEEP_ALIVE=-1" Environment="OLLAMA_FLASH_ATTENTION=1" Environment="HSA_OVERRIDE_GFX_VERSION=11.5.1" ``` **Key points:** - Do NOT set `OLLAMA_VULKAN=1` - Do NOT set `HIP_VISIBLE_DEVICES=-1` (this disables ROCm GPU access) - `HSA_OVERRIDE_GFX_VERSION=11.5.1` tells ROCm to recognize gfx1151 Then reload and restart: ```bash sudo systemctl daemon-reload sudo systemctl restart ollama ``` ## Verification Ollama logs should show: ``` inference compute id=0 library=ROCm compute=gfx1151 name=ROCm0 description="Radeon 8060S Graphics" total="111.5 GiB" available="110.5 GiB" ``` All layers offloaded to GPU: ``` offloaded 48/48 layers to GPU model weights device=ROCm0 size="29.3 GiB" ``` ## Benchmarks | Model | Quant | Context | tok/s | VRAM Used | |-------|-------|---------|-------|-----------| | GLM-4.7-flash | Q8_0 (30B) | 4K | ~40 tok/s | ~30 GB | | GLM-4.7-flash | Q8_0 (30B) | 202K | ~40 tok/s | ~50 GB | | Qwen3.5 | 35B-a3b Q8_0 | 65K | ~40 tok/s | ~36 GB | ## VRAM Tips for 128GB Unified Memory With unified memory, the GPU shares RAM with the system. Be careful with large context windows: - 200K+ context on a 30B model uses ~50-70GB, leaving less for your desktop/browser - If VRAM runs out, the amdgpu kernel driver spams `Not enough memory for command submission!` and can crash the entire system - Use Ollama's `num_ctx` parameter or your frontend's context setting to limit context size - A safe starting point: 65K-96K context for 30B models, which uses ~36-45GB ## What I tried that failed 1. Vulkan + `HIP_VISIBLE_DEVICES=-1` — worked initially for some models but hung on Qwen3.5 and became unreliable 2. ROCm without `HSA_OVERRIDE_GFX_VERSION` — GPU not detected 3. 202K context with 30B model — works but risky on unified memory, can OOM and crash the system if other apps use too much RAM Hope this helps other Strix Halo users! This took a full day of troubleshooting to figure out. ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the documentationamdbug labels 2026-04-29 10:15:27 -05:00
GiteaMirror commented 2026-04-29 10:15:31 -05:00
Author
Owner
Copy Link

@queensone commented on GitHub (Mar 15, 2026):

Update: Ollama 0.18 has native gfx1151 support!

Just tested — HSA_OVERRIDE_GFX_VERSION=11.5.1 is no longer needed on Ollama 0.18. ROCm detects the Radeon 8060S (gfx1151) natively
without any override.

Updated override.conf (simplified):
[Service]
Environment="OLLAMA_KEEP_ALIVE=-1"
Environment="OLLAMA_FLASH_ATTENTION=1"

That's it — two lines. If you're on Ollama 0.17 or earlier, you still need HSA_OVERRIDE_GFX_VERSION=11.5.1.

Even simpler now — two lines instead of three!

<!-- gh-comment-id:4062859961 --> @queensone commented on GitHub (Mar 15, 2026): Update: Ollama 0.18 has native gfx1151 support! Just tested — HSA_OVERRIDE_GFX_VERSION=11.5.1 is no longer needed on Ollama 0.18. ROCm detects the Radeon 8060S (gfx1151) natively without any override. Updated override.conf (simplified): [Service] Environment="OLLAMA_KEEP_ALIVE=-1" Environment="OLLAMA_FLASH_ATTENTION=1" That's it — two lines. If you're on Ollama 0.17 or earlier, you still need HSA_OVERRIDE_GFX_VERSION=11.5.1. --- Even simpler now — two lines instead of three!
GiteaMirror commented 2026-04-29 10:15:31 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 15, 2026):

That's it — two lines. If you're on Ollama 0.17 or earlier, you still need HSA_OVERRIDE_GFX_VERSION=11.5.1.

ollama has natively supported gfx1151 on linux since 0.5.13.

ollama  | time=2026-03-15T16:48:59.812Z level=INFO source=routes.go:1277 msg="Listening on [::]:11434 (version 0.5.13)"
ollama  | time=2026-03-15T16:48:59.812Z level=INFO source=gpu.go:217 msg="looking for compatible GPUs"
ollama  | time=2026-03-15T16:48:59.815Z level=INFO source=amd_linux.go:386 msg="amdgpu is supported" gpu=0 gpu_type=gfx1151
ollama  | time=2026-03-15T16:48:59.817Z level=INFO source=types.go:130 msg="inference compute" id=0 library=rocm variant="" compute=gfx1151 driver=6.12 name=1002:1586 total="96.0 GiB" available="95.9 GiB"
<!-- gh-comment-id:4063395872 --> @rick-github commented on GitHub (Mar 15, 2026): > That's it — two lines. If you're on Ollama 0.17 or earlier, you still need HSA_OVERRIDE_GFX_VERSION=11.5.1. ollama has natively supported gfx1151 on linux since 0.5.13. ``` ollama | time=2026-03-15T16:48:59.812Z level=INFO source=routes.go:1277 msg="Listening on [::]:11434 (version 0.5.13)" ollama | time=2026-03-15T16:48:59.812Z level=INFO source=gpu.go:217 msg="looking for compatible GPUs" ollama | time=2026-03-15T16:48:59.815Z level=INFO source=amd_linux.go:386 msg="amdgpu is supported" gpu=0 gpu_type=gfx1151 ollama | time=2026-03-15T16:48:59.817Z level=INFO source=types.go:130 msg="inference compute" id=0 library=rocm variant="" compute=gfx1151 driver=6.12 name=1002:1586 total="96.0 GiB" available="95.9 GiB" ```
GiteaMirror commented 2026-04-29 10:15:32 -05:00
Author
Owner
Copy Link

@queensone commented on GitHub (Mar 16, 2026):

Thanks for the clarification! The issue on my end was that I had OLLAMA_VULKAN=1 and HIP_VISIBLE_DEVICES=-1 in
my override.conf, which was hiding ROCm entirely. Once I removed those, ROCm detected gfx1151 natively.

So the real fix for anyone stuck on Vulkan with Strix Halo: just remove OLLAMA_VULKAN=1 and
HIP_VISIBLE_DEVICES=-1 from your override. ROCm works out of the box.

Performance went from hanging/unreliable on Vulkan to solid 40 tok/s on ROCm. Benchmarks and config details in
the original post above.

<!-- gh-comment-id:4064607069 --> @queensone commented on GitHub (Mar 16, 2026): Thanks for the clarification! The issue on my end was that I had OLLAMA_VULKAN=1 and HIP_VISIBLE_DEVICES=-1 in my override.conf, which was hiding ROCm entirely. Once I removed those, ROCm detected gfx1151 natively. So the real fix for anyone stuck on Vulkan with Strix Halo: just remove OLLAMA_VULKAN=1 and HIP_VISIBLE_DEVICES=-1 from your override. ROCm works out of the box. Performance went from hanging/unreliable on Vulkan to solid 40 tok/s on ROCm. Benchmarks and config details in the original post above.
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 documentation
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#56095
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?