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 #15448] Apple M5 + macOS 26.3.1: 500 Internal Server Error, Metal compiler failed to build request even with OLLAMA_NUM_GPU=0 #9873

New Issue
Open
opened 2026-04-12 22:44:02 -05:00 by GiteaMirror · 15 comments
GiteaMirror commented 2026-04-12 22:44:02 -05:00
Owner
Copy Link

Originally created by @frankcsliu on GitHub (Apr 9, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15448

What is the issue?

Describe the bug

On a MacBook Pro with Apple M5 chip and macOS 26.3.1, any attempt to run a model fails with:

Error: 500 Internal Server Error: llama runner process has terminated: %!w(<nil>)

This happens even when forcing CPU-only mode with OLLAMA_NUM_GPU=0. The system logs show repeated Metal shader compilation failures right before the runner terminates.

Environment

  • Device: MacBook Pro (Apple M5, 24 GB RAM)
  • OS: macOS 26.3.1 (from sw_vers):
    • ProductVersion: 26.3.1
    • ProductVersionExtra: (a)
    • BuildVersion: 25D771280a
  • Ollama version: ollama --version → 0.20.4
  • Install method: Homebrew / official install script / .dmg (please specify)

Steps to reproduce

On this machine:

  1. Install Ollama 0.20.4
  2. Pull a model, e.g.: 
    ollama pull llama3.1:8b
  3. Try to run the model in CPU-only mode: 
    OLLAMA_NUM_GPU=0 ollama run llama3.1:8b

Expected behavior

The model should start and respond (especially in CPU-only mode with OLLAMA_NUM_GPU=0).

Actual behavior

The command immediately fails with:

Error: 500 Internal Server Error: llama runner process has terminated: %!w(<nil>)

No additional error details are printed in the terminal.

Relevant logs

From:

log stream --predicate 'process == "ollama"' --info | grep -iE 'Error|Metal|fault|assert|panic'

I see (excerpt):

2026-04-09 17:49:22.708860+0800 0xa1b5     Activity    0x68f85              1940   0    ollama: (Metal) Metal Compiling Shader
2026-04-09 17:49:23.640608+0800 0xa1b5     Error       0x68f85              1940   0    ollama: (Metal) Compiler failed to build request

2026-04-09 17:49:23.799570+0800 0xa1cc     Error       0x0                  1998   0    ollama: (libCoreFSCache.dylib) flock failed to lock list file (<private>): errno = 35
2026-04-09 17:49:23.799940+0800 0xa1cc     Error       0x0                  1998   0    ollama: (libCoreFSCache.dylib) flock failed to lock list file (<private>): errno = 35

# more lines with MTLCompilerService and Metal Compiling Shader ...

The 500 error (llama runner process has terminated: %!w(<nil>)) appears every time right after the Metal compiler failure.

What I have already tried

  • Fully reinstalling Ollama (removing ~/.ollama, caches, logs) and installing 0.20.4 again
  • Rebooting the machine
  • Forcing CPU-only mode: 
    OLLAMA_NUM_GPU=0 ollama run llama3.1:8b
  • Deleting and re-pulling the model: 
    ollama rm llama3.1:8b
ollama pull llama3.1:8b

All of the above still result in the same 500 error and the Metal compiler failure in the logs.

Questions

  • Is this a known regression for Apple M5 + macOS 26.3.1 in the current Metal / GGML backend?
  • Is there any workaround to run models on this machine (even CPU-only) until a fix is available?

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @frankcsliu on GitHub (Apr 9, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15448 ### What is the issue? **Describe the bug** On a MacBook Pro with Apple M5 chip and macOS 26.3.1, any attempt to run a model fails with: `Error: 500 Internal Server Error: llama runner process has terminated: %!w(<nil>)` This happens even when forcing CPU-only mode with `OLLAMA_NUM_GPU=0`. The system logs show repeated Metal shader compilation failures right before the runner terminates. --- **Environment** - Device: MacBook Pro (Apple M5, 24 GB RAM) - OS: macOS 26.3.1 (from `sw_vers`): - `ProductVersion`: `26.3.1` - `ProductVersionExtra`: `(a)` - `BuildVersion`: `25D771280a` - Ollama version: `ollama --version` → `0.20.4` - Install method: Homebrew / official install script / .dmg (please specify) --- **Steps to reproduce** On this machine: 1. Install Ollama 0.20.4 2. Pull a model, e.g.: ```bash ollama pull llama3.1:8b ``` 3. Try to run the model in CPU-only mode: ```bash OLLAMA_NUM_GPU=0 ollama run llama3.1:8b ``` **Expected behavior** The model should start and respond (especially in CPU-only mode with `OLLAMA_NUM_GPU=0`). **Actual behavior** The command immediately fails with: ```text Error: 500 Internal Server Error: llama runner process has terminated: %!w(<nil>) ``` No additional error details are printed in the terminal. --- **Relevant logs** From: ```bash log stream --predicate 'process == "ollama"' --info | grep -iE 'Error|Metal|fault|assert|panic' ``` I see (excerpt): ```text 2026-04-09 17:49:22.708860+0800 0xa1b5 Activity 0x68f85 1940 0 ollama: (Metal) Metal Compiling Shader 2026-04-09 17:49:23.640608+0800 0xa1b5 Error 0x68f85 1940 0 ollama: (Metal) Compiler failed to build request 2026-04-09 17:49:23.799570+0800 0xa1cc Error 0x0 1998 0 ollama: (libCoreFSCache.dylib) flock failed to lock list file (<private>): errno = 35 2026-04-09 17:49:23.799940+0800 0xa1cc Error 0x0 1998 0 ollama: (libCoreFSCache.dylib) flock failed to lock list file (<private>): errno = 35 # more lines with MTLCompilerService and Metal Compiling Shader ... ``` The 500 error (`llama runner process has terminated: %!w(<nil>)`) appears every time right after the Metal compiler failure. --- **What I have already tried** - Fully reinstalling Ollama (removing `~/.ollama`, caches, logs) and installing 0.20.4 again - Rebooting the machine - Forcing CPU-only mode: ```bash OLLAMA_NUM_GPU=0 ollama run llama3.1:8b ``` - Deleting and re-pulling the model: ```bash ollama rm llama3.1:8b ollama pull llama3.1:8b ``` All of the above still result in the same 500 error and the Metal compiler failure in the logs. --- **Questions** - Is this a known regression for Apple M5 + macOS 26.3.1 in the current Metal / GGML backend? - Is there any workaround to run models on this machine (even CPU-only) until a fix is available? ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-12 22:44:02 -05:00
GiteaMirror commented 2026-04-12 22:44:03 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 9, 2026):

OLLAMA_NUM_GPU is not an ollama configuration variable.

Server logs will aid in debugging.

<!-- gh-comment-id:4213379819 --> @rick-github commented on GitHub (Apr 9, 2026): `OLLAMA_NUM_GPU` is not an ollama configuration variable. [Server logs](https://docs.ollama.com/troubleshooting) will aid in debugging.
GiteaMirror commented 2026-04-12 22:44:04 -05:00
Author
Owner
Copy Link

@egdavid commented on GitHub (Apr 9, 2026):

Same here on the M5. Downgrading to 0.20.3 seems to be the only stable fix for now, couldn't find a workaround for the current build.

<!-- gh-comment-id:4213434094 --> @egdavid commented on GitHub (Apr 9, 2026): Same here on the M5. Downgrading to 0.20.3 seems to be the only stable fix for now, couldn't find a workaround for the current build.
GiteaMirror commented 2026-04-12 22:44:04 -05:00
Author
Owner
Copy Link

@NadavM-ux commented on GitHub (Apr 9, 2026):

Same here, using Macbook pro M5 , Downgrading to 0.20.3 worked for me.

<!-- gh-comment-id:4213457966 --> @NadavM-ux commented on GitHub (Apr 9, 2026): Same here, using Macbook pro M5 , Downgrading to 0.20.3 worked for me.
GiteaMirror commented 2026-04-12 22:44:04 -05:00
Author
Owner
Copy Link

@poreporec commented on GitHub (Apr 9, 2026):

Same

<!-- gh-comment-id:4213502464 --> @poreporec commented on GitHub (Apr 9, 2026): Same
GiteaMirror commented 2026-04-12 22:44:05 -05:00
Author
Owner
Copy Link

@Alex4386 commented on GitHub (Apr 9, 2026):

It seems they shipped the ollama with libmlxc binary for x86_64 on arm64 platforms on ollama 0.20.4.

Error: error loading model: 500 Internal Server Error: mlx runner failed: Error: failed to initialize MLX: MLX: Failed to load /Applications/Ollama.app/Contents/Resources/libmlxc.dylib: dlopen(/Applications/Ollama.app/Contents/Resources/libmlxc.dylib, 0x0009): tried: '/Applications/Ollama.app/Contents/Resources/libmlxc.dylib' (mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64e' or 'arm64e.v1' or 'arm64' or 'arm64')), '/System/Volumes/Preboot/Cryptexes/OS/Applications/Ollama.app/Contents/Resources/libmlxc.dylib' (no such file), '/Applications/Ollama.app/Contents/Resources/libmlxc.dylib (exit: exit status 1)
<!-- gh-comment-id:4214084462 --> @Alex4386 commented on GitHub (Apr 9, 2026): It seems they shipped the ollama with libmlxc binary for `x86_64` on `arm64` platforms on ollama 0.20.4. ``` Error: error loading model: 500 Internal Server Error: mlx runner failed: Error: failed to initialize MLX: MLX: Failed to load /Applications/Ollama.app/Contents/Resources/libmlxc.dylib: dlopen(/Applications/Ollama.app/Contents/Resources/libmlxc.dylib, 0x0009): tried: '/Applications/Ollama.app/Contents/Resources/libmlxc.dylib' (mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64e' or 'arm64e.v1' or 'arm64' or 'arm64')), '/System/Volumes/Preboot/Cryptexes/OS/Applications/Ollama.app/Contents/Resources/libmlxc.dylib' (no such file), '/Applications/Ollama.app/Contents/Resources/libmlxc.dylib (exit: exit status 1) ```
GiteaMirror commented 2026-04-12 22:44:05 -05:00
Author
Owner
Copy Link

@yoon-jeong-ho15 commented on GitHub (Apr 9, 2026):

downgrading to 0.20.3 also worked for me.

<!-- gh-comment-id:4214255241 --> @yoon-jeong-ho15 commented on GitHub (Apr 9, 2026): downgrading to 0.20.3 also worked for me.
GiteaMirror commented 2026-04-12 22:44:06 -05:00
Author
Owner
Copy Link

@ThomasHorn1967 commented on GitHub (Apr 9, 2026):

same here, downgrading to 0.20.3 also worked for me.

<!-- gh-comment-id:4215043071 --> @ThomasHorn1967 commented on GitHub (Apr 9, 2026): same here, downgrading to 0.20.3 also worked for me.
GiteaMirror commented 2026-04-12 22:44:07 -05:00
Author
Owner
Copy Link

@fairyorbital commented on GitHub (Apr 9, 2026):

Same here.
Disabling Flash Attention with OLLAMA_FLASH_ATTENTION=0 did NOT resolve the issue. Only downgrading to 0.20.3 worked for me.

My Environment:

  • Ollama Version: v0.20.4 (Regression)
  • Hardware: Apple M5 Pro, 64GB RAM
  • OS: macOS

Error Log Snippet (from my server logs):

ggml_metal_library_init: error: Error Domain=MTLLibraryErrorDomain Code=3 "program_source:7131:28: warning: variable 'theta_base' is used uninitialized..."
...
 /System/Library/Frameworks/MetalPerformancePrimitives.framework/Headers/__impl/MPPTensorOpsMatMul2dImpl.h:3266:5: error: static_assert failed due to requirement
'__tensor_ops_detail::__is_same_v<bfloat, half>' "Input types must match cooperative tensor types"
...
ggml_metal_init: error: failed to create library
ggml_backend_metal_device_init: error: failed to allocate context
ggml-backend.cpp:258: GGML_ASSERT(backend) failed
SIGABRT: abort
<!-- gh-comment-id:4217440885 --> @fairyorbital commented on GitHub (Apr 9, 2026): Same here. Disabling Flash Attention with `OLLAMA_FLASH_ATTENTION=0` did NOT resolve the issue. Only downgrading to 0.20.3 worked for me. My Environment: - Ollama Version: v0.20.4 (Regression) - Hardware: Apple M5 Pro, 64GB RAM - OS: macOS Error Log Snippet (from my server logs): ``` ggml_metal_library_init: error: Error Domain=MTLLibraryErrorDomain Code=3 "program_source:7131:28: warning: variable 'theta_base' is used uninitialized..." ... /System/Library/Frameworks/MetalPerformancePrimitives.framework/Headers/__impl/MPPTensorOpsMatMul2dImpl.h:3266:5: error: static_assert failed due to requirement '__tensor_ops_detail::__is_same_v<bfloat, half>' "Input types must match cooperative tensor types" ... ggml_metal_init: error: failed to create library ggml_backend_metal_device_init: error: failed to allocate context ggml-backend.cpp:258: GGML_ASSERT(backend) failed SIGABRT: abort ```
GiteaMirror commented 2026-04-12 22:44:07 -05:00
Author
Owner
Copy Link

@frankli0324 commented on GitHub (Apr 9, 2026):

I this duplicates/closely relates to https://github.com/ollama/ollama/issues/14432

<!-- gh-comment-id:4218443579 --> @frankli0324 commented on GitHub (Apr 9, 2026): I this duplicates/closely relates to https://github.com/ollama/ollama/issues/14432
GiteaMirror commented 2026-04-12 22:44:09 -05:00
Author
Owner
Copy Link

@code-retriever commented on GitHub (Apr 10, 2026):

Confirming on Apple M5 Max (128GB, 40 GPU cores) — macOS 26.3.1

Version bisection results

Version Metal 4 GPU detection Model loading Notes
v0.19.0 ✅ Metal — Apple M5 Max, 107.5 GiB ✅ Works gemma4 architecture not supported
v0.20.0 ✅ Metal ✅ Works gemma4 not yet supported
v0.20.2 ✅ Metal ✅ Works gemma4:e4b ✅, gemma4:26b partial (see note below)
v0.20.3 ✅ Metal ✅ Works All models pass — this is the last working version for M5 Max
v0.20.4 ❌ CPU only ❌ All fail Regression introduced here
v0.20.5 ❌ CPU only ❌ All fail Same as v0.20.4

Note: On M5 Max 128GB, v0.20.2 and v0.20.3 both work. Other reporters with M5 24GB see failures on v0.20.2 — this may be a GPU core count / memory difference.

Root cause: kernel_rope_multi shader

On v0.20.5 with verbose logging, the Metal library compilation error is visible:

ggml_metal_library_init: error: Error Domain=MTLLibraryErrorDomain Code=3
  "program_source:7131:28: warning: variable 'theta_base' is used uninitialized
   whenever 'if' condition is false [-Wsometimes-uninitialized]
   } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t

The Metal 4 compiler (MTLGPUFamilyMetal4 / MTLGPUFamilyApple10) treats this uninitialized variable as a fatal error, whereas Metal 3 was more lenient.

What doesn't work

  • OLLAMA_FLASH_ATTENTION=0 — no effect (crash happens before FA is relevant)
  • OLLAMA_KV_CACHE_TYPE="" — no effect
  • OLLAMA_LLM_LIBRARY=cpu — runner still initializes Metal library and crashes
  • OLLAMA_NUM_GPU=0 — same, Metal init crash kills the process before CPU fallback
  • Clean environment (env -i) — same result

Workaround

Downgrade to v0.20.3 (confirmed working).

pkill -f ollama
gh release download v0.20.3 --repo ollama/ollama --pattern "Ollama-darwin.zip" --dir /tmp
rm -rf /Applications/Ollama.app
unzip -q /tmp/Ollama-darwin.zip -d /Applications/

Environment: Apple M5 Max, 128GB unified memory, 40 GPU cores, macOS 26.3.1 (Build 25D2128), Xcode CLT 26.4.

<!-- gh-comment-id:4223596689 --> @code-retriever commented on GitHub (Apr 10, 2026): ## Confirming on Apple M5 Max (128GB, 40 GPU cores) — macOS 26.3.1 ### Version bisection results | Version | Metal 4 GPU detection | Model loading | Notes | |---------|----------------------|---------------|-------| | **v0.19.0** | ✅ Metal — Apple M5 Max, 107.5 GiB | ✅ Works | gemma4 architecture not supported | | **v0.20.0** | ✅ Metal | ✅ Works | gemma4 not yet supported | | **v0.20.2** | ✅ Metal | ✅ Works | gemma4:e4b ✅, gemma4:26b partial (see note below) | | **v0.20.3** | ✅ Metal | ✅ Works | All models pass — **this is the last working version for M5 Max** | | **v0.20.4** | ❌ CPU only | ❌ All fail | Regression introduced here | | **v0.20.5** | ❌ CPU only | ❌ All fail | Same as v0.20.4 | > Note: On M5 Max 128GB, v0.20.2 and v0.20.3 both work. Other reporters with M5 24GB see failures on v0.20.2 — this may be a GPU core count / memory difference. ### Root cause: `kernel_rope_multi` shader On v0.20.5 with verbose logging, the Metal library compilation error is visible: ``` ggml_metal_library_init: error: Error Domain=MTLLibraryErrorDomain Code=3 "program_source:7131:28: warning: variable 'theta_base' is used uninitialized whenever 'if' condition is false [-Wsometimes-uninitialized] } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t ``` The Metal 4 compiler (MTLGPUFamilyMetal4 / MTLGPUFamilyApple10) treats this uninitialized variable as a fatal error, whereas Metal 3 was more lenient. ### What doesn't work - `OLLAMA_FLASH_ATTENTION=0` — no effect (crash happens before FA is relevant) - `OLLAMA_KV_CACHE_TYPE=""` — no effect - `OLLAMA_LLM_LIBRARY=cpu` — runner still initializes Metal library and crashes - `OLLAMA_NUM_GPU=0` — same, Metal init crash kills the process before CPU fallback - Clean environment (`env -i`) — same result ### Workaround Downgrade to **v0.20.3** (confirmed working). ```bash pkill -f ollama gh release download v0.20.3 --repo ollama/ollama --pattern "Ollama-darwin.zip" --dir /tmp rm -rf /Applications/Ollama.app unzip -q /tmp/Ollama-darwin.zip -d /Applications/ ``` Environment: Apple M5 Max, 128GB unified memory, 40 GPU cores, macOS 26.3.1 (Build 25D2128), Xcode CLT 26.4.
GiteaMirror commented 2026-04-12 22:44:09 -05:00
Author
Owner
Copy Link

@executeautomation commented on GitHub (Apr 11, 2026):

I am having the same issue as well with MacBook Pro M5 Max 128GB

Image
<!-- gh-comment-id:4228079775 --> @executeautomation commented on GitHub (Apr 11, 2026): I am having the same issue as well with MacBook Pro M5 Max 128GB <img width="1453" height="1063" alt="Image" src="https://github.com/user-attachments/assets/21c27356-5764-4d88-ac8a-7cc414335a35" />
GiteaMirror commented 2026-04-12 22:44:10 -05:00
Author
Owner
Copy Link

@chenghongm commented on GitHub (Apr 12, 2026):

ollama --version
ollama version is 0.20.5

same issue: "Error: 500 Internal Server Error: model failed to load, this may be due to resource limitations or an internal error, check ollama server logs for details", no log at all

<!-- gh-comment-id:4230499686 --> @chenghongm commented on GitHub (Apr 12, 2026): ollama --version ollama version is 0.20.5 same issue: "Error: 500 Internal Server Error: model failed to load, this may be due to resource limitations or an internal error, check ollama server logs for details", no log at all
GiteaMirror commented 2026-04-12 22:44:10 -05:00
Author
Owner
Copy Link

@reawake97 commented on GitHub (Apr 12, 2026):

Downgrading to 0.20.3 worked for me.

<!-- gh-comment-id:4230613051 --> @reawake97 commented on GitHub (Apr 12, 2026): Downgrading to 0.20.3 worked for me.
GiteaMirror commented 2026-04-12 22:44:11 -05:00
Author
Owner
Copy Link

@boite2chocolat commented on GitHub (Apr 12, 2026):

Hello folks
Same here, my macbook pro m5 with 24gb of ram is getting the same error when install models.
Downgrading to 0.20.3 worked for me as well.

<!-- gh-comment-id:4231578293 --> @boite2chocolat commented on GitHub (Apr 12, 2026): Hello folks Same here, my macbook pro m5 with 24gb of ram is getting the same error when install models. Downgrading to 0.20.3 worked for me as well.
GiteaMirror commented 2026-04-12 22:44:12 -05:00
Author
Owner
Copy Link

@exoad commented on GitHub (Apr 12, 2026):

downgrading to 0.20.3 worked for me as well on my m5 air!

<!-- gh-comment-id:4232275102 --> @exoad commented on GitHub (Apr 12, 2026): downgrading to 0.20.3 worked for me as well on my m5 air!
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#9873
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?