.act tool invocation fails - KeyError: value (broken model config json)

[stndart]

Problem

When model tries to run a tool using .act, the call fails with KeyError: value (kv["value"])

Replication

Model

Any of
Qwen3-32B
QwQ-32B

Software

LM-Studio 0.3.17 (Build 11)
OS - Windows 10 Pro
CUDA llama.cpp v1.37.1
LMStudio Python SDK 1.4.0

Code

Tested with official documentation code: https://lmstudio.ai/docs/python/agent/act (scoped client variant over network)

Error:

Traceback (most recent call last):
  File "/root/radarange-orchestrator/examples/lmstudio_tool_docs.py", line 11, in <module>
    model.act(
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/root/radarange-orchestrator/.venv/lib/python3.10/site-packages/lmstudio/sync_api.py", line 1390, in act
    for event in prediction_stream._iter_events():
  File "/root/radarange-orchestrator/.venv/lib/python3.10/site-packages/lmstudio/sync_api.py", line 926, in _iter_events
    for event in endpoint.iter_message_events(contents):
  File "/root/radarange-orchestrator/.venv/lib/python3.10/site-packages/lmstudio/json_api.py", line 1299, in iter_message_events
    load_config=parse_llm_load_config(load_kvconfig),
  File "/root/radarange-orchestrator/.venv/lib/python3.10/site-packages/lmstudio/_kv_config.py", line 347, in parse_llm_load_config
    return LlmLoadModelConfig._from_any_api_dict(parse_server_config(server_config))
  File "/root/radarange-orchestrator/.venv/lib/python3.10/site-packages/lmstudio/_kv_config.py", line 340, in parse_server_config
    value = kv["value"]
KeyError: 'value'

Speculations

When model tries to call a tool, it passes it's config to the sdk (json-api.py:1298, then _kv_config.py:346)
The config looks like:

{
'loadModelConfig': {
        'fields': [
            {
                'key': 'load.gpuSplitConfig',
                'value': {
                    'customRatio': [],
                    'disabledGpus': [],
                    'priority': [],
                    'strategy': 'evenly',
                },
            },
            {'key': 'load.gpuStrictVramCap', 'value': False},
            {'key': 'llm.load.llama.cpuThreadPoolSize', 'value': 6},
            {'key': 'llm.load.contextLength', 'value': 80000},
            {'key': 'llm.load.llama.acceleration.offloadRatio'},
            {'key': 'llm.load.llama.flashAttention', 'value': True},
        ]
    }
}

It clearly lacks 'value' field at 'llm.load.llama.acceleration.offloadRatio' key.

[stndart]

After messing with LlmLoadModelConfig, passed to llm settings, i think i found the issue.
When passing config in dict[str, Any] format as config argument to lms.Llm(), the python SDK parser seem to be ok with GpuSetting.ratio = None, and the LM Studio seem also fine with it. However, when dumping None values back to json, LMStudio skips the 'value' field comletely, and there's where python SDK complains.
Regarding documentation example bug - i think i forgot to unload the model before trying to test the example, in which case LMStudio does nothing - even if the config differs.

[ncoghlan]

Good investigation, thanks. Do you know the original input config passed to the SDK that resulted in the malformed config stack?

I'm attempting to reproduce this for a test case, and am finding that omitting the ratio key or setting it to None results in the llm.load.llama.acceleration.offloadRatio config field being omitted entirely by the SDK client (as expected). That means there may be an additional server issue involved in order to produce the incorrectly formatted config stack in the first place.

[ncoghlan]

Received config parsing has been relaxed for 1.4.1