Starting OVMS 2026.2 with two models, each on a dedicated GPU, leads to abort

[avk1im]

Describe the bug
Starting OVMS 2026.2 with two models on two different GPUs leads to abort.

System: Intel(R) Core(TM) Ultra 7 265K, 128 GB RAM
GPU.1: Arc Pro B60
GPU.2: Arc B580

Model1: phi-4-int8-ov
Model2: phi-4-int4-ov

Models are tested by a simple chat completion script addressing rest_port 8000, attached.

Each model runs fine on either GPU (even if Model1 is very slow on GPU.2 since it doesn't fit in GPU RAM -- yet still functional). Both models also run together fine on the single GPU.1, barely fitting in the GPU RAM. I haven't tested both together on GPU.2 since that would take a very long time.

Yet when I try to run together
-Model1 on GPU.1 AND
-Model2 on GPU.2
OR
-Model1 on GPU.2 AND
-Model2 on GPU.1
the startup process of OVMS auto-aborts after ca. 20 sec with a short message: "Abort was called at 308 line in file:" (no file name is given) -- s. console output attached.

To Reproduce
Steps to reproduce the behavior:

1. Steps to prepare models repository:
   Model1 pull:
   ovms --pull --source_model "OpenVINO/phi-4-int8-ov" --model_repository_path "C:\ovms-models" --model_name phi-4-int8-ov --task text_generation
   Model1 run on GPU.1, works fine:
   ovms --source_model "OpenVINO/phi-4-int8-ov" --model_repository_path "C:\ovms-models" --model_name phi-4-int8-ov --target_device GPU.1 --task text_generation --rest_port 8000

Model2 pull:
ovms --pull --source_model "OpenVINO/phi-4-int4-ov" --model_repository_path "C:\ovms-models" --model_name phi-4-int4-ov --task text_generation
Model2 run on GPU.2, works fine:
ovms --source_model "OpenVINO/phi-4-int4-ov" --model_repository_path "C:\ovms-models" --model_name phi-4-int4-ov --target_device GPU.2 --task text_generation --rest_port 8000

Create JSON config:
ovms --add_to_config --config_path c:\ovms-models\config.json --model_name phi-4-int8-ov --model_path C:\ovms-models\OpenVINO\phi-4-int8-ov
ovms --add_to_config --config_path c:\ovms-models\config.json --model_name phi-4-int4-ov --model_path C:\ovms-models\OpenVINO\phi-4-int4-ov

3. OVMS launch command '....'
   ovms.exe --config_path "C:\ovms-models\config.json" --rest_port 8000 --log_level DEBUG --log_path Model1OnB60_Model2OnB580_LeadsToABORT.txt

4. Client command (additionally client code if not using official client or demo) '....'
   A simple chat script addressing rest_port 8000 (attached). In the script, I ask "Hi! Please explain C# generics, with code examples".

5. See errorExpected behaviorA clear and concise description of what you expected to happen.
   When functional, each model gives a pretty good explanation.

Logs
Logs from OVMS, ideally with --log_level DEBUG. Logs from client.

Configuration

1. OVMS version: 2026.2
   OpenVINO Model Server 2026.2.5e9dcfc46
   OpenVINO backend 2026.2.0-21902-90214e5be05-releases/2026/2
   OpenVINO GenAI backend 2026.2.0.0-3121-adf73e80e66
   Bazel build flags: --config=win_mp_on_py_off

2. OVMS config.json file
   {
   "model_config_list": [
   {
   "config": {
   "name": "phi-4-int8-ov",
   "base_path": "C:\ovms-models\OpenVINO\phi-4-int8-ov"
   }
   },
   {
   "config": {
   "name": "phi-4-int4-ov",
   "base_path": "C:\ovms-models\OpenVINO\phi-4-int4-ov"
   }
   }
   ]
   }

3. CPU, accelerator's versions if applicable

Intel® Arc™ Pro B60 Graphics
Driver Details
Up to date
Provider
Intel Corporation
Version
32.0.101.8801
Date
2026-12-05
Device Details
Adapter Compatibility
Intel Corporation
Video Processor
Intel® Arc™ Pro B60 Graphics Family
Adapter DAC Type
Internal
Availability
Offline
Status
This device is working properly.
Location
PCI bus 4, device 0, function 0
Device Id
PCI\VEN_8086&DEV_E211&SUBSYS_0300172F&REV_00\6&2895C387&0&00080030

Intel® Arc™ B580 Graphics
Driver Details
Up to date
Provider
Intel Corporation
Version
32.0.101.8801
Date
2026-12-05
Device Details
Adapter Compatibility
Intel Corporation
Video Processor
Intel® Arc™ B580 Graphics Family
Adapter DAC Type
Internal
Availability
Offline
Status
This device is working properly.
Location
PCI bus 133, device 0, function 0
Device Id
PCI\VEN_8086&DEV_E20B&SUBSYS_11008086&REV_00\6&8497E81&0&000800E4

4. Model repository directory structure
   C:\ovms-models\OpenVINO\phi-4-int8-ov
   C:\ovms-models\OpenVINO\phi-4-int4-ov

5. Model or publicly available similar model that reproduces the issueAdditional context
   Attaching:

• Debug logs of 3 cases: Both models on GPU.1 (ok), Model1 on GPU1 and Model2 on GPU2 (abort), Model1 on GPU2 and Model2 on GPU1 (abort)
• Console output for 2 bad cases: Model1 on GPU1 and Model2 on GPU2 (abort), Model1 on GPU2 and Model2 on GPU1 (abort)
• config.json
• pbtxt files for both models

[avk1im]

BothModelsOnB60_OK.txt
Model1OnB60_Model2OnB580_LeadsToABORT.txt
Model1OnB60_Model2OnB580_ConsoleOutput.txt
Model1OnB580_Model2OnB60_LeadsToABORT.txt
Model1OnB580_Model2OnB60_ConsoleOutput.txt

chat2.CCHANGE_to_ps1_to_run.txt

config.json

graph_Model2.(pb).txt

graph_Model1.(pb).txt