0
Fork 0
mirror of https://github.com/zml/zml.git synced 2026-11-16 19:11:09 +00:00

[LLMD] llmd oneapi image: startup panic when an Intel integrated GPU is exposed next to an Arc card #1081

Open
opened 2026-11-13 20:14:23 +00:00 by davlgd · 2 comments
davlgd commented 2026-11-13 20:14:23 +00:00 (Migrated from github.com)

Summary

On a machine with an Intel integrated GPU, the README command for Intel GPUs (--device=/dev/dri) exposes both the iGPU and the Arc card. llmd then panics during startup, before loading the weights, whatever the model. The integrated GPU alone fails to compile the model; restricting llmd to the Arc card works.

Configuration:

  • Model: any (tested with Qwen/Qwen3-0.6B and Qwen/Qwen2.5-1.5B-Instruct)
  • OS: Debian 13, kernel 7.2.12
  • Platform: oneapi
  • Accelerator: Intel Arc Pro B60 24 GB
  • Driver version: bundled in the image (Level Zero GPU driver 26.31.39395.14); kernel driver xe
  • Version: docker image sha256:5a4af1e7545249f8bff581745f362f5a23625f4c9737246796242d89d32f7281 (tag 20261112.0-oneapi)
  • Command line: docker run -p 8000:8000 --device=/dev/dri zmlai/llmd:oneapi --model=hf://Qwen/Qwen3-0.6B

Environment

  • Image: zmlai/llmd:oneapi (tag 20261112.0-oneapi, digest sha256:5a4af1e7545249f8bff581745f362f5a23625f4c9737246796242d89d32f7281)
  • Discrete GPU: Intel Arc Pro B60 24 GB (ASRock, PCI 8086:e211), xe driver, GuC 70.72.1
  • Integrated GPU: Intel Arrow Lake-S graphics (PCI 8086:7d67), i915 driver, CPU Core Ultra 7 270K Plus
  • Host: Debian 13, kernel 7.2.12 (trixie-backports), Docker 26.1.5
  • No Intel compute runtime installed on the host: the container only gets the device nodes

Steps to reproduce

docker run -p 8000:8000 --device=/dev/dri zmlai/llmd:oneapi --model=hf://Qwen/Qwen3-0.6B

Actual result

info(llmd): Devices:
info(llmd): 	- Intel(R) Arc(TM) Pro B60 Graphics (oneapi:0)
info(llmd): 	- Intel(R) Graphics (oneapi:1)
thread 6 panic: PJRT_LoadedExecutable_Execute failed with: error.InvalidArgument

The container exits with code 134. Same result with hf://Qwen/Qwen2.5-1.5B-Instruct.

What I tested

All with --device=/dev/dri and hf://Qwen/Qwen3-0.6B:

Variation Devices listed by llmd Result
none (README command) Arc Pro B60 + Intel(R) Graphics panic at startup
--data-parallel=1 Arc Pro B60 + Intel(R) Graphics panic at startup
-e ZE_AFFINITY_MASK=0 Arc Pro B60 works
-e ONEAPI_DEVICE_SELECTOR=level_zero:0 Arc Pro B60 works
-e ZE_AFFINITY_MASK=1 Intel(R) Graphics fails at startup, exit code 1
only the Arc device nodes (--device=/dev/dri/card1 --device=/dev/dri/renderD129) Arc Pro B60 works

With the integrated GPU alone (ZE_AFFINITY_MASK=1), the compile step fails six times with:

warning(pjrt): [PJRT_Client_Compile] InvalidArgument: GPU compute capability of serialized executable doesn't match target device capability. (serialized: Unknown, target: Xe1)
error(zml/Compiler): Pjrt failed to compile the following MLIR (error.InvalidArgument):

So the integrated GPU is not usable by this build, and its presence breaks startup even though a usable device is also available.

llmd --help lists no option to choose a device.

Expected result

llmd skips devices it cannot compile for (or lets the user pick one) and starts on the Arc card, or fails with a message naming the unsupported device.

Workaround

docker run -p 8000:8000 --device=/dev/dri -e ZE_AFFINITY_MASK=0 \
  zmlai/llmd:oneapi --model=hf://Qwen/Qwen3-0.6B

The index depends on the enumeration order: here the Arc card is device 0.

Scope

Tested on one machine, with one integrated GPU (Arrow Lake-S) and one discrete card (Arc Pro B60).

### Summary On a machine with an Intel integrated GPU, the README command for Intel GPUs (`--device=/dev/dri`) exposes both the iGPU and the Arc card. llmd then panics during startup, before loading the weights, whatever the model. The integrated GPU alone fails to compile the model; restricting llmd to the Arc card works. **Configuration:** - Model: any (tested with `Qwen/Qwen3-0.6B` and `Qwen/Qwen2.5-1.5B-Instruct`) - OS: Debian 13, kernel 7.2.12 - Platform: `oneapi` - Accelerator: Intel Arc Pro B60 24 GB - Driver version: bundled in the image (Level Zero GPU driver 26.31.39395.14); kernel driver `xe` - Version: docker image `sha256:5a4af1e7545249f8bff581745f362f5a23625f4c9737246796242d89d32f7281` (tag `20261112.0-oneapi`) - Command line: `docker run -p 8000:8000 --device=/dev/dri zmlai/llmd:oneapi --model=hf://Qwen/Qwen3-0.6B` ### Environment - Image: `zmlai/llmd:oneapi` (tag `20261112.0-oneapi`, digest `sha256:5a4af1e7545249f8bff581745f362f5a23625f4c9737246796242d89d32f7281`) - Discrete GPU: Intel Arc Pro B60 24 GB (ASRock, PCI `8086:e211`), `xe` driver, GuC 70.72.1 - Integrated GPU: Intel Arrow Lake-S graphics (PCI `8086:7d67`), `i915` driver, CPU Core Ultra 7 270K Plus - Host: Debian 13, kernel 7.2.12 (trixie-backports), Docker 26.1.5 - No Intel compute runtime installed on the host: the container only gets the device nodes ### Steps to reproduce ```bash docker run -p 8000:8000 --device=/dev/dri zmlai/llmd:oneapi --model=hf://Qwen/Qwen3-0.6B ``` ### Actual result ``` info(llmd): Devices: info(llmd): - Intel(R) Arc(TM) Pro B60 Graphics (oneapi:0) info(llmd): - Intel(R) Graphics (oneapi:1) thread 6 panic: PJRT_LoadedExecutable_Execute failed with: error.InvalidArgument ``` The container exits with code 134. Same result with `hf://Qwen/Qwen2.5-1.5B-Instruct`. ### What I tested All with `--device=/dev/dri` and `hf://Qwen/Qwen3-0.6B`: | Variation | Devices listed by llmd | Result | |---|---|---| | none (README command) | Arc Pro B60 + Intel(R) Graphics | panic at startup | | `--data-parallel=1` | Arc Pro B60 + Intel(R) Graphics | panic at startup | | `-e ZE_AFFINITY_MASK=0` | Arc Pro B60 | works | | `-e ONEAPI_DEVICE_SELECTOR=level_zero:0` | Arc Pro B60 | works | | `-e ZE_AFFINITY_MASK=1` | Intel(R) Graphics | fails at startup, exit code 1 | | only the Arc device nodes (`--device=/dev/dri/card1 --device=/dev/dri/renderD129`) | Arc Pro B60 | works | With the integrated GPU alone (`ZE_AFFINITY_MASK=1`), the compile step fails six times with: ``` warning(pjrt): [PJRT_Client_Compile] InvalidArgument: GPU compute capability of serialized executable doesn't match target device capability. (serialized: Unknown, target: Xe1) error(zml/Compiler): Pjrt failed to compile the following MLIR (error.InvalidArgument): ``` So the integrated GPU is not usable by this build, and its presence breaks startup even though a usable device is also available. `llmd --help` lists no option to choose a device. ### Expected result llmd skips devices it cannot compile for (or lets the user pick one) and starts on the Arc card, or fails with a message naming the unsupported device. ### Workaround ```bash docker run -p 8000:8000 --device=/dev/dri -e ZE_AFFINITY_MASK=0 \ zmlai/llmd:oneapi --model=hf://Qwen/Qwen3-0.6B ``` The index depends on the enumeration order: here the Arc card is device 0. ### Scope Tested on one machine, with one integrated GPU (Arrow Lake-S) and one discrete card (Arc Pro B60).
steeve commented 2026-11-13 20:25:56 +00:00 (Migrated from github.com)

I think we should be able to support the Arrow Lake-S GPU as it's Xe based (only missing XMX).

zml/llmd assumes all GPUs discovered are compatible for collectives (oneCCL in this case). Maybe? If not, we'll let the user decide which one they want to expose to zml/llmd.

would that satisfy your use case @davlgd ?

I think we should be able to support the Arrow Lake-S GPU as it's Xe based (only missing XMX). zml/llmd assumes all GPUs discovered are compatible for collectives (oneCCL in this case). Maybe? If not, we'll let the user decide which one they want to expose to zml/llmd. would that satisfy your use case @davlgd ?
davlgd commented 2026-11-13 20:32:22 +00:00 (Migrated from github.com)

Thanks for the quick answer!

Using the integrated GPU wasn't my initial goal: I simply ran the README command, and --device=/dev/dri exposed it along with the Arc card. But if it can be made usable, I'll take it.

Thanks for the quick answer! Using the integrated GPU wasn't my initial goal: I simply ran the README command, and `--device=/dev/dri` exposed it along with the Arc card. But if it can be made usable, I'll take it.
Sign in to join this conversation.
No milestone
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
DKB/zml#1081
No description provided.