github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-16 21:25:46 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #15813] Metal backend crash on Apple M5 - bfloat/half type mismatch in MPPTensorOpsMatMul2d (v0.21.2) #87764

New Issue
Open
opened 2026-05-10 06:20:24 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-05-10 06:20:24 -05:00
Owner
Copy Link

Originally created by @egalarza007 on GitHub (Apr 25, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15813

Description

Ollama fails to run any model on Apple M5 hardware. The Metal shader compilation
fails with a bfloat/half type mismatch in MetalPerformancePrimitives.

Environment

  • Ollama version: 0.21.2
  • macOS: Sequoia 15.3+ (Apple M5)
  • GPU: Apple M5 (MTLGPUFamilyApple10)
  • VRAM: ~19GB available

Error

From ~/.ollama/logs/server.log:

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: error: Error Domain=MTLLibraryErrorDomain Code=3
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 initialize the Metal library
ggml_backend_metal_device_init: error: failed to allocate context
llama_init_from_model: failed to initialize the context: failed to initialize Metal backend
panic: unable to create llama context

The server reports the GPU as MTLGPUFamilyApple10 (1010) which is M5-specific.
The embedded Metal library does not have a precompiled library for this GPU family,
and on-the-fly compilation fails due to a bfloat16/half type mismatch in the
MetalPerformancePrimitives framework.

Impact

All models fail with: 500 Internal Server Error: llama runner process has terminated
OLLAMA_NO_METAL=1 is silently ignored in v0.21.2.

Workaround

Setting GGML_METAL_BF16_DISABLE=1 before starting the server allows models to
load and run (CPU fallback, no GPU acceleration):

GGML_METAL_BF16_DISABLE=1 ollama serve

Expected Behavior

Ollama should either:

  1. Ship a precompiled Metal library for MTLGPUFamilyApple10 (M5), or
  2. Correctly fall back to CPU when Metal compilation fails, or
  3. Honor OLLAMA_NO_METAL=1 to explicitly disable Metal
Originally created by @egalarza007 on GitHub (Apr 25, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15813 ## Description Ollama fails to run any model on Apple M5 hardware. The Metal shader compilation fails with a bfloat/half type mismatch in MetalPerformancePrimitives. ## Environment - Ollama version: 0.21.2 - macOS: Sequoia 15.3+ (Apple M5) - GPU: Apple M5 (MTLGPUFamilyApple10) - VRAM: ~19GB available ## Error From ~/.ollama/logs/server.log: 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: error: Error Domain=MTLLibraryErrorDomain Code=3 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 initialize the Metal library ggml_backend_metal_device_init: error: failed to allocate context llama_init_from_model: failed to initialize the context: failed to initialize Metal backend panic: unable to create llama context The server reports the GPU as MTLGPUFamilyApple10 (1010) which is M5-specific. The embedded Metal library does not have a precompiled library for this GPU family, and on-the-fly compilation fails due to a bfloat16/half type mismatch in the MetalPerformancePrimitives framework. ## Impact All models fail with: 500 Internal Server Error: llama runner process has terminated OLLAMA_NO_METAL=1 is silently ignored in v0.21.2. ## Workaround Setting GGML_METAL_BF16_DISABLE=1 before starting the server allows models to load and run (CPU fallback, no GPU acceleration): GGML_METAL_BF16_DISABLE=1 ollama serve ## Expected Behavior Ollama should either: 1. Ship a precompiled Metal library for MTLGPUFamilyApple10 (M5), or 2. Correctly fall back to CPU when Metal compilation fails, or 3. Honor OLLAMA_NO_METAL=1 to explicitly disable Metal
GiteaMirror commented 2026-05-10 06:20:25 -05:00
Author
Owner
Copy Link

@asqararslonov commented on GitHub (May 3, 2026):

model failed to load on Apple M5 (Metal 4) — Metal shader half/bfloat type mismatch

Environment:

  • Apple M5 chip (Metal 4)
  • macOS Tahoe (latest)
  • Ollama 0.22.1 (stable / Ollama.app)
  • Any model (gemma4, qwen3, llama3, etc.)

Symptom

Every model fails immediately with:

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

This is not a RAM issue. It affects all models regardless of size.

Root Cause

The embedded Metal shader in Ollama's ggml backend fails to compile on Metal 4 (M5). Apple's MetalPerformancePrimitives framework on Metal 4 enforces strict type matching for cooperative tensor matmul — but the shader mixes half and bfloat types:

static_assert failed: "Input types must match cooperative tensor types" (half vs bfloat)
ggml_metal_library_init: error: Error Domain=MTLLibraryErrorDomain Code=3
ggml_backend_metal_device_init: error: failed to allocate context
GGML_ASSERT(backend) failed → exit status 2

The following workarounds do not help — Metal is always initialized regardless:

  • OLLAMA_METAL=0
  • OLLAMA_LLM_LIBRARY=cpu
  • OLLAMA_NEW_ENGINE=true

Fix

Install the HEAD build via Homebrew, which includes the Metal 4 shader patch:

# Quit Ollama.app first if it's running
brew unlink ollama        # only if stable was already installed via brew
brew install --HEAD ollama
brew services start ollama

The HEAD brew service automatically sets OLLAMA_FLASH_ATTENTION=1 and
OLLAMA_KV_CACHE_TYPE=q8_0 in the launchd plist.

Also remove Ollama.app from Login Items (System Settings → General → Login Items) to prevent conflicts with the brew service on reboot.

Verification

After the fix, all models load and run on GPU:

$ ollama ps
NAME             ID              SIZE     PROCESSOR    CONTEXT
gemma4:latest    c6eb396dbd59    10 GB    100% GPU     4096

Tested with gemma4:latest (9.6 GB) and qwen3:8b on Apple M5.

<!-- gh-comment-id:4366508665 --> @asqararslonov commented on GitHub (May 3, 2026): ## `model failed to load` on Apple M5 (Metal 4) — Metal shader `half`/`bfloat` type mismatch **Environment:** - Apple M5 chip (Metal 4) - macOS Tahoe (latest) - Ollama 0.22.1 (stable / Ollama.app) - Any model (gemma4, qwen3, llama3, etc.) --- ### Symptom Every model fails immediately with: ``` 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 ``` This is **not** a RAM issue. It affects all models regardless of size. --- ### Root Cause The embedded Metal shader in Ollama's ggml backend fails to compile on **Metal 4 (M5)**. Apple's `MetalPerformancePrimitives` framework on Metal 4 enforces strict type matching for cooperative tensor matmul — but the shader mixes `half` and `bfloat` types: ``` static_assert failed: "Input types must match cooperative tensor types" (half vs bfloat) ggml_metal_library_init: error: Error Domain=MTLLibraryErrorDomain Code=3 ggml_backend_metal_device_init: error: failed to allocate context GGML_ASSERT(backend) failed → exit status 2 ``` The following workarounds do **not** help — Metal is always initialized regardless: - `OLLAMA_METAL=0` - `OLLAMA_LLM_LIBRARY=cpu` - `OLLAMA_NEW_ENGINE=true` --- ### Fix Install the HEAD build via Homebrew, which includes the Metal 4 shader patch: ```bash # Quit Ollama.app first if it's running brew unlink ollama # only if stable was already installed via brew brew install --HEAD ollama brew services start ollama ``` > The HEAD brew service automatically sets `OLLAMA_FLASH_ATTENTION=1` and > `OLLAMA_KV_CACHE_TYPE=q8_0` in the launchd plist. Also remove **Ollama.app** from Login Items (**System Settings → General → Login Items**) to prevent conflicts with the brew service on reboot. --- ### Verification After the fix, all models load and run on GPU: ``` $ ollama ps NAME ID SIZE PROCESSOR CONTEXT gemma4:latest c6eb396dbd59 10 GB 100% GPU 4096 ``` Tested with `gemma4:latest` (9.6 GB) and `qwen3:8b` on Apple M5.
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#87764
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?