gpu offload does not work in lms

[FBTCAD]

when loading a new instance of the model with lms the gpu parameter is still an issue.

model = client.llm.load_new_instance(model_name,config={
"temperature": temperature,
"contextLength": min(tmp_max_context_length,max_tokens),
"gpu": {"ratio": model_gpuol, }
})

Model info: LlmInstanceInfo.from_dict({
"architecture": "qwen3",
"contextLength": 32768,
"displayName": "Qwen3 8B",
"format": "gguf",
"identifier": "qwen3-8b",
"instanceReference": "OdO6BqpmD+eu5898DBjJrOnA",
"maxContextLength": 32768,
"modelKey": "qwen3-8b",
"paramsString": "8B",
"path": "lmstudio-community/Qwen3-8B-GGUF/Qwen3-8B-Q8_0.gguf",
"sizeBytes": 8709518624,
"trainedForToolUse": true,
"type": "llm",
"vision": false
})
Found max context length: 32768
gpuoffload calculated: 1.0
ERROR:main:Error loading model 'qwen3-8b': 'mainGpu'

[ncoghlan]

Which version of the LM Studio app is running?

The GPU offload config fix in 1.3.1 has been tested against 0.3.16 build 7, but not against earlier versions:

>>> model = client.llm.load_new_instance(model_name, config={"gpu": {"ratio": 1.0}})
>>> model2 = client.llm.load_new_instance(model_name, config={"gpu": {"ratio": 0.5}})
>>> model3 = client.llm.load_new_instance(model_name, config={"gpu": {"ratio": 0.5, "mainGpu": 0, "splitStrategy": "favorMainGpu"}})
>>> model.get_load_config().gpu
GpuSetting(ratio=1.0, main_gpu=None, split_strategy='evenly', disabled_gpus=[])
>>> model2.get_load_config().gpu
GpuSetting(ratio=0.5, main_gpu=None, split_strategy='evenly', disabled_gpus=[])
>>> model3.get_load_config().gpu
GpuSetting(ratio=0.5, main_gpu=0, split_strategy='favorMainGpu', disabled_gpus=[])
>>>

[ncoghlan]

Closing on the assumption that this was covered by the fix in SDK 1.3.1