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 #15548] Gemma 4 E2B fails to load on Apple M5 — Metal library compilation error #71994

New Issue
Closed
opened 2026-05-05 03:15:56 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-05-05 03:15:56 -05:00
Owner
Copy Link

Originally created by @Chloeyz13 on GitHub (Apr 13, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15548

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

Chip: Apple M5
Ollama model: gemma4:e2b
Running ollama run gemma4:e2b results in a 500 Internal Server Error. The model never loads.

❯ ollama run gemma3:4b
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

Relevant log output

ggml_metal_device_init: simdgroup reduction   = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = true
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  = 19069.67 MB
ggml_metal_init: allocating
ggml_metal_init: picking default device: Apple M5
ggml_metal_init: the device does not have a precompiled Metal library - this is unexpected
ggml_metal_init: will try to compile it on the fly
ggml_metal_library_init: using embedded metal library
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]
ggml_metal_init: error: failed to initialize the Metal library
2   ollama                              0x0000000105bfe544 ggml_backend_get_default_buffer_type + 76
3   ollama                              0x0000000105b7a838 _cgo_c81fd19bee02_Cfunc_ggml_backend_get_default_buffer_type + 36
SIGABRT: abort
github.com/ollama/ollama/ml/backend/ggml._Cfunc_ggml_backend_get_default_buffer_type(0x0)
fault   0x1836b95b0

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.20.6

Originally created by @Chloeyz13 on GitHub (Apr 13, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15548 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? Chip: Apple M5 Ollama model: gemma4:e2b Running ollama run gemma4:e2b results in a 500 Internal Server Error. The model never loads. ❯ ollama run gemma3:4b 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 ### Relevant log output ```shell ggml_metal_device_init: simdgroup reduction = true ggml_metal_device_init: simdgroup matrix mul. = true ggml_metal_device_init: has unified memory = true ggml_metal_device_init: has bfloat = true ggml_metal_device_init: has tensor = true ggml_metal_device_init: use residency sets = true ggml_metal_device_init: use shared buffers = true ggml_metal_device_init: recommendedMaxWorkingSetSize = 19069.67 MB ggml_metal_init: allocating ggml_metal_init: picking default device: Apple M5 ggml_metal_init: the device does not have a precompiled Metal library - this is unexpected ggml_metal_init: will try to compile it on the fly ggml_metal_library_init: using embedded metal library 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] ggml_metal_init: error: failed to initialize the Metal library 2 ollama 0x0000000105bfe544 ggml_backend_get_default_buffer_type + 76 3 ollama 0x0000000105b7a838 _cgo_c81fd19bee02_Cfunc_ggml_backend_get_default_buffer_type + 36 SIGABRT: abort github.com/ollama/ollama/ml/backend/ggml._Cfunc_ggml_backend_get_default_buffer_type(0x0) fault 0x1836b95b0 ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.20.6
GiteaMirror added the bug label 2026-05-05 03:15:56 -05:00
GiteaMirror commented 2026-05-05 03:16:02 -05:00
Author
Owner
Copy Link

@Horacehxw commented on GitHub (Apr 13, 2026):

I can reproduce this on a local Apple M5 Pro / macOS 26.4 / Ollama 0.20.6 setup with gemma4:26b (and local aliases based on it).

Two additional data points that may help narrow this down:

  1. GGML_METAL_TENSOR_DISABLE=1 is an effective workaround for the raw runner crash on this machine.

    • With that env set, the runner reports:
      • ggml_metal_device_init: has tensor = false
      • offloaded 31/31 layers to GPU
    • Without it, the runner reports has tensor = true, then tries to compile the Metal library on the fly and aborts with the MetalPerformancePrimitives bfloat / half static_assert mismatch.

  2. The Ollama macOS app appears not to pass this env through to the ollama serve child process it launches.

    • launchctl print gui/$(id -u)/com.ollama.ollama shows GGML_METAL_TENSOR_DISABLE=1 in the app's inherited environment.
    • But ps eww -p <serve_pid> for /Applications/Ollama.app/Contents/Resources/ollama serve does not include GGML_METAL_TENSOR_DISABLE, OLLAMA_KV_CACHE_TYPE, or OLLAMA_FLASH_ATTENTION.
    • In that app-launched server path, requests still fail with 500 and logs show:
      • ggml_metal_device_init: has tensor = true
      • ggml_metal_init: the device does not have a precompiled Metal library - this is unexpected
      • ggml_metal_init: will try to compile it on the fly
      • MPPTensorOpsMatMul2dImpl.h ... static_assert failed due to requirement '__is_same_v<bfloat, half>'
      • SIGABRT: abort

I also observed a second app-specific issue: if I manually start ollama serve with the workaround env and then open the macOS app UI, the app sends concurrent initialization requests and I can hit context canceled during model load, which turns a generate request into 500 even though the underlying Metal workaround is active.

A stable successful request on this machine was:

GGML_METAL_TENSOR_DISABLE=1 \
OLLAMA_KV_CACHE_TYPE=q8_0 \
OLLAMA_FLASH_ATTENTION=1 \
/Applications/Ollama.app/Contents/Resources/ollama serve

Then:

curl -s http://127.0.0.1:11434/api/chat \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gemma4-agent",
    "stream": false,
    "think": false,
    "messages": [{"role": "user", "content": "Reply with exactly one word: pong"}],
    "options": {"num_predict": 32}
  }'

Response:

{"model":"gemma4-agent","message":{"role":"assistant","content":"pong"},"done":true,"done_reason":"stop"}

So at least on this machine, there seem to be two separate bugs:

  • Metal tensor path compilation/regression on Apple M5 + macOS 26.x
  • macOS app process chain not propagating critical runtime env vars to ollama serve
<!-- gh-comment-id:4238011621 --> @Horacehxw commented on GitHub (Apr 13, 2026): I can reproduce this on a local Apple M5 Pro / macOS 26.4 / Ollama 0.20.6 setup with `gemma4:26b` (and local aliases based on it). Two additional data points that may help narrow this down: 1. `GGML_METAL_TENSOR_DISABLE=1` is an effective workaround for the raw runner crash on this machine. - With that env set, the runner reports: - `ggml_metal_device_init: has tensor = false` - `offloaded 31/31 layers to GPU` - Without it, the runner reports `has tensor = true`, then tries to compile the Metal library on the fly and aborts with the `MetalPerformancePrimitives` `bfloat` / `half` static_assert mismatch. 2. The Ollama macOS app appears not to pass this env through to the `ollama serve` child process it launches. - `launchctl print gui/$(id -u)/com.ollama.ollama` shows `GGML_METAL_TENSOR_DISABLE=1` in the app's inherited environment. - But `ps eww -p <serve_pid>` for `/Applications/Ollama.app/Contents/Resources/ollama serve` does **not** include `GGML_METAL_TENSOR_DISABLE`, `OLLAMA_KV_CACHE_TYPE`, or `OLLAMA_FLASH_ATTENTION`. - In that app-launched server path, requests still fail with `500` and logs show: - `ggml_metal_device_init: has tensor = true` - `ggml_metal_init: the device does not have a precompiled Metal library - this is unexpected` - `ggml_metal_init: will try to compile it on the fly` - `MPPTensorOpsMatMul2dImpl.h ... static_assert failed due to requirement '__is_same_v<bfloat, half>'` - `SIGABRT: abort` I also observed a second app-specific issue: if I manually start `ollama serve` with the workaround env and then open the macOS app UI, the app sends concurrent initialization requests and I can hit `context canceled` during model load, which turns a generate request into `500` even though the underlying Metal workaround is active. A stable successful request on this machine was: ```bash GGML_METAL_TENSOR_DISABLE=1 \ OLLAMA_KV_CACHE_TYPE=q8_0 \ OLLAMA_FLASH_ATTENTION=1 \ /Applications/Ollama.app/Contents/Resources/ollama serve ``` Then: ```bash curl -s http://127.0.0.1:11434/api/chat \ -H 'Content-Type: application/json' \ -d '{ "model": "gemma4-agent", "stream": false, "think": false, "messages": [{"role": "user", "content": "Reply with exactly one word: pong"}], "options": {"num_predict": 32} }' ``` Response: ```json {"model":"gemma4-agent","message":{"role":"assistant","content":"pong"},"done":true,"done_reason":"stop"} ``` So at least on this machine, there seem to be two separate bugs: - Metal tensor path compilation/regression on Apple M5 + macOS 26.x - macOS app process chain not propagating critical runtime env vars to `ollama serve`
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#71994
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?