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 #14076] qwen3-coder-next: Metal GPU crash with 'tensor buffer is nil' on Apple M4 Max (0.15.5-rc2) #55706

New Issue
Closed
opened 2026-04-29 09:36:52 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-29 09:36:52 -05:00
Owner
Copy Link

Originally created by @antonelli182 on GitHub (Feb 4, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14076

What is the issue?

Running qwen3-coder-next on Apple M4 Max with Metal GPU crashes during inference with:

ggml_metal_buffer_get_id: error: tensor ' (view)' buffer is nil

The model loads successfully into VRAM (~43 layers on GPU, ~6 on CPU), but crashes when attempting to generate tokens.

Workaround: Setting num_gpu: 0 forces CPU inference and works correctly, but is significantly slower.

OS

macOS 15.3 (Sequoia) - Apple M4 Max, 64GB unified memory

GPU

Apple M4 Max (Metal)

  • 51.8 GB available VRAM detected
  • MTLGPUFamilyApple9 / MTLGPUFamilyMetal4

CPU

Apple M4 Max (12 cores)

Ollama version

0.15.5-rc2 (pre-release)

Steps to reproduce

  1. Install Ollama 0.15.5-rc2 pre-release
  2. Pull the model: ollama pull qwen3-coder-next
  3. Attempt any inference:
curl http://localhost:11434/api/generate -d '{"model":"qwen3-coder-next","prompt":"hello","stream":false}'

Relevant server log

ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices
ggml_metal_device_init: GPU name:   Apple M4 Max
ggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
ggml_metal_device_init: has tensor            = false
...
time=2026-02-04T13:25:08.090-08:00 level=INFO source=server.go:1387 msg="llama runner started in 18.42 seconds"
ggml_metal_buffer_get_id: error: tensor ' (view)' buffer is nil
ggml_metal_buffer_get_id: error: tensor ' (view)' buffer is nil
[GIN] 2026/02/04 - 13:25:49 | 500 | 59.989722208s | 127.0.0.1 | POST "/api/generate"

Memory allocation (from logs)

  • System memory: 64.0 GiB total, 53.2 GiB free
  • GPU memory: 51.8 GiB available
  • Model weights (Metal): 42.8 GiB
  • Model weights (CPU): 5.4 GiB
  • KV cache (Metal): 7.7 GiB
  • KV cache (CPU): 788.4 MiB
  • Total: 57.4 GiB

Additional context

  • Issue may be related to qwen3next architecture handling in ggml Metal backend
  • The line tensor API disabled for pre-M5 and pre-A19 devices suggests M4 Max is being treated differently
  • CPU inference (num_gpu: 0) works correctly at ~15 tok/s
  • Similar CUDA issues reported in #14068 (Windows)
Originally created by @antonelli182 on GitHub (Feb 4, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14076 ### What is the issue? Running `qwen3-coder-next` on Apple M4 Max with Metal GPU crashes during inference with: ``` ggml_metal_buffer_get_id: error: tensor ' (view)' buffer is nil ``` The model loads successfully into VRAM (~43 layers on GPU, ~6 on CPU), but crashes when attempting to generate tokens. **Workaround:** Setting `num_gpu: 0` forces CPU inference and works correctly, but is significantly slower. ### OS macOS 15.3 (Sequoia) - Apple M4 Max, 64GB unified memory ### GPU Apple M4 Max (Metal) - 51.8 GB available VRAM detected - MTLGPUFamilyApple9 / MTLGPUFamilyMetal4 ### CPU Apple M4 Max (12 cores) ### Ollama version 0.15.5-rc2 (pre-release) ### Steps to reproduce 1. Install Ollama 0.15.5-rc2 pre-release 2. Pull the model: `ollama pull qwen3-coder-next` 3. Attempt any inference: ```bash curl http://localhost:11434/api/generate -d '{"model":"qwen3-coder-next","prompt":"hello","stream":false}' ``` ### Relevant server log ``` ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices ggml_metal_device_init: GPU name: Apple M4 Max ggml_metal_device_init: GPU family: MTLGPUFamilyApple9 (1009) ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003) ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002) ggml_metal_device_init: has tensor = false ... time=2026-02-04T13:25:08.090-08:00 level=INFO source=server.go:1387 msg="llama runner started in 18.42 seconds" ggml_metal_buffer_get_id: error: tensor ' (view)' buffer is nil ggml_metal_buffer_get_id: error: tensor ' (view)' buffer is nil [GIN] 2026/02/04 - 13:25:49 | 500 | 59.989722208s | 127.0.0.1 | POST "/api/generate" ``` ### Memory allocation (from logs) - System memory: 64.0 GiB total, 53.2 GiB free - GPU memory: 51.8 GiB available - Model weights (Metal): 42.8 GiB - Model weights (CPU): 5.4 GiB - KV cache (Metal): 7.7 GiB - KV cache (CPU): 788.4 MiB - Total: 57.4 GiB ### Additional context - Issue may be related to `qwen3next` architecture handling in ggml Metal backend - The line `tensor API disabled for pre-M5 and pre-A19 devices` suggests M4 Max is being treated differently - CPU inference (`num_gpu: 0`) works correctly at ~15 tok/s - Similar CUDA issues reported in #14068 (Windows)
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#55706
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?