Backend supervisor: TypeError: unsupported operand type(s) for /: 'WindowsPath' and 'NoneType'

[zhanglt]

Before you start

• I have read the FAQ and my problem is not answered there.
• I have read the Roadmap and this is not already planned there.
• I have searched existing issues and found no duplicate.
• I have restarted the Desktop app to pick up the latest update and the problem still happens.

What happened

[14:13:19] cmd/INFO ft serve --model E:\FreeToken\models\gpt-oss-20b --port 1919 --moe-backend auto --max-running-requests 4 --memory-ratio 0.85 --host 127.0.0.1 --cors-origins tauri://localhost,http://tauri.localhost,http://localhost:1420
[14:13:20] stdout/INFO serve started (pid=33276 model=E:\FreeToken\models\gpt-oss-20b port=1919)
[14:13:31] stdout/WARN [2026-09-23|14:13:31] WARNING --moe-backend is deprecated; use --moe-strategy
[14:13:31] stdout/INFO [2026-09-23|14:13:31] INFO Parsed arguments:
[14:13:31] stdout/INFO ServerArgs(model_path='E:\FreeToken\models\gpt-oss-20b', tp_info=DistributedInfo(rank=0, size=1), dtype=torch.bfloat16, max_running_req=4, attention_backend='auto', moe_strategy='auto', quant_backend=None, ple_backend='disk', expert_load='auto', moe_cache_size=0, moe_cache_rate=None, moe_cache_auto=False, kv_reserve_tokens=8192, moe_cache_policy='lru', moe_prefill_overlap=True, moe_prefill_hit_d2d=False, moe_collect_stats=False, moe_cpu_threads=0, moe_cpu_layers=None, moe_hybrid_max_fetch=-1, cuda_graph_bs=None, cuda_graph_max_bs=None, page_size=1, memory_ratio=0.85, linear_state_cache_ratio=2.0, swa_full_tokens_ratio=0.2, swa_num_pages_override=None, distributed_timeout=60.0, use_dummy_weight=False, use_pynccl=True, max_seq_len_override=None, num_page_override=None, num_token_override=None, mm=MultimodalConfig(disabled_encoders=frozenset(), embed_cache_device='cpu', encoder_weights='host', image_min_tokens=None, image_max_tokens=None, processor_kwargs={}), max_extend_tokens=8192, cache_type='radix', offline_mode=False, decode_log_interval=40, special_token_ckpt=False, _unique_suffix='.pid=7516', _ipc_base_port=18315, server_host='127.0.0.1', server_port=1919, num_tokenizer=0, silent_output=False, shell_mode=False, served_model_name='gpt-oss-20b', tool_call_parser='gpt_oss', reasoning_parser='gpt_oss', sampling_defaults='model', max_output_tokens=None, enable_cache_report=False, allowed_media_domains='', allowed_local_media_path='', cors_origins='tauri://localhost,http://tauri.localhost,http://localhost:1420', gpu=(), gpu_assigned=None)
[14:13:31] stdout/INFO [2026-09-23|14:13:31|FrontendAPI] INFO Default sampling config (source=framework): temperature=0.0, top_k=-1, top_p=1.0
[14:13:31] stdout/INFO INFO: Started server process [7516]
[14:13:31] stdout/INFO INFO: Waiting for application startup.
[14:13:31] stdout/INFO INFO: Application startup complete.
[14:13:31] stdout/INFO INFO: Uvicorn running on http://127.0.0.1:1919 (Press CTRL+C to quit)
[14:13:39] stdout/INFO C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\server\launch.py:81: FutureWarning: torch.cuda._set_allocator_settings is deprecated. Use torch._C._accelerator_setAllocatorSettings instead.
[14:13:39] stdout/INFO scheduler = Scheduler(args)
[14:13:39] stdout/INFO [2026-09-23|14:13:39|core|rank=0] INFO Enabled expandable_segments (override via PYTORCH_ALLOC_CONF)
[14:13:39] stdout/INFO [2026-09-23|14:13:39|core|rank=0] INFO Auto-selected attention backend: triton
[14:13:39] stdout/INFO [2026-09-23|14:13:39|core|rank=0] INFO Auto-selected MoE strategy: offload
[14:13:39] stdout/INFO [2026-09-23|14:13:39|core|rank=0] INFO No MoE cache sizing flag given; defaulting to --moe-cache-auto for auto-selected strategy 'offload'
[14:13:39] stdout/INFO [2026-09-23|14:13:39|core|rank=0] INFO Resolved config: moe_strategy='offload', attention_backend='triton', cache_type='swa_radix', page_size=1
[14:13:39] stdout/INFO [2026-09-23|14:13:39|core|rank=0] INFO Free memory before loading model: 6.93 GiB
[14:13:39] stdout/INFO C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\torch\utils_device.py:116: UserWarning: expandable_segments not supported on this platform (Triggered internally at C:\actions-runner_work\pytorch\pytorch\pytorch\c10/cuda/CUDAAllocatorConfig.h:39.)
[14:13:39] stdout/INFO return func(*args, **kwargs)
[14:13:40] stdout/INFO [2026-09-23|14:13:40|core|rank=0] INFO kernel triton_gptoss selected; skipped triton: standard MXFP4 kernel reads the concatenated gate|up row order (×24)
[14:13:43] stdout/INFO [2026-09-23|14:13:43|core|rank=0] INFO MoE experts: mxfp4 via triton_gptoss
[14:13:43] stdout/INFO [2026-09-23|14:13:43|core|rank=0] INFO expert banks: slow path (serial build)
[14:13:54] stdout/INFO [2026-09-23|14:13:54|core|rank=0] INFO --moe-cache-auto resolved moe_cache_size=93 num_pages=8275 (prefill_overlap=True)
[14:13:54] stdout/INFO [2026-09-23|14:13:54|core|rank=0] INFO Allocating 8275 tokens for KV cache, K + V = 0.26 GiB
[14:13:54] stdout/INFO [2026-09-23|14:13:54|core|rank=0] INFO Free memory after initialization: 1.90 GiB
[14:13:54] stdout/INFO [2026-09-23|14:13:54|core|rank=0] INFO Start capturing CUDA graphs with sizes: [1, 2, 4]
[14:13:54] stdout/INFO [2026-09-23|14:13:54|core|rank=0] INFO Free GPU memory before capturing CUDA graphs: 1.90 GiB
[14:13:54] stdout/INFO Process freetoken-TP0-scheduler:
[14:13:54] stdout/ERROR [2026-09-23|14:13:54|FrontendAPI] ERROR Backend supervisor: TypeError: unsupported operand type(s) for /: 'WindowsPath' and 'NoneType'
[14:13:54] stdout/INFO Traceback (most recent call last):
[14:13:54] stdout/INFO File "D:\Python312\Lib\multiprocessing\process.py", line 314, in _bootstrap
[14:13:54] stdout/INFO self.run()
[14:13:54] stdout/INFO File "D:\Python312\Lib\multiprocessing\process.py", line 108, in run
[14:13:54] stdout/INFO self._target(*self._args, **self._kwargs)
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\server\launch.py", line 81, in _run_scheduler
[14:13:54] stdout/INFO scheduler = Scheduler(args)
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "python/freetoken/scheduler/scheduler.py", line 68, in freetoken.scheduler.scheduler.Scheduler.init
[14:13:54] stdout/INFO File "python/freetoken/engine/engine.py", line 521, in freetoken.engine.engine.Engine.init
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\engine\graph.py", line 134, in init
[14:13:54] stdout/INFO self._capture_graphs(max_seq_len, vocab_size, model)
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\engine\graph.py", line 164, in _capture_graphs
[14:13:54] stdout/INFO self._reset_moe_offload_cache()
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\engine\graph.py", line 138, in _reset_moe_offload_cache
[14:13:54] stdout/INFO self.moe_offload_cache.reset()
[14:13:54] stdout/INFO File "python/freetoken/moe/offload_cache.py", line 886, in freetoken.moe.offload_cache.OffloadMoeCache.reset
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\moe\offload_kernels.py", line 90, in reset_cache
[14:13:54] stdout/INFO _reset_cache_gpu(cache)
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\moe\offload_kernels.py", line 212, in _reset_cache_gpu
[14:13:54] stdout/INFO _reset_cache_kernel[grid](
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\jit.py", line 370, in
[14:13:54] stdout/INFO return lambda *args, **kwargs: self.run(grid=grid, warmup=False, *args, **kwargs)
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\jit.py", line 700, in run
[14:13:54] stdout/INFO device = driver.active.get_current_device()
[14:13:54] stdout/INFO ^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\driver.py", line 28, in active
[14:13:54] stdout/INFO self._active = self.default
[14:13:54] stdout/INFO ^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\driver.py", line 22, in default
[14:13:54] stdout/INFO self._default = _create_driver()
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\driver.py", line 10, in _create_driver
[14:13:54] stdout/INFO return active_drivers0
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\backends\nvidia\driver.py", line 759, in init
[14:13:54] stdout/INFO self.utils = CudaUtils() # TODO: make static
[14:13:54] stdout/INFO ^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\backends\nvidia\driver.py", line 73, in init
[14:13:54] stdout/INFO mod = compile_module_from_src(
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\build.py", line 186, in compile_module_from_src
[14:13:54] stdout/INFO so = _build(name, src_path, tmpdir, library_dirs or [], include_dirs or [], libraries or [], ccflags or [])
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\build.py", line 104, in _build
[14:13:54] stdout/INFO cc = get_cc()
[14:13:54] stdout/INFO ^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\runtime\build.py", line 34, in get_cc
[14:13:54] stdout/INFO cc, _, _ = find_msvc_winsdk(env_only=True)
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\windows_utils.py", line 279, in find_msvc_winsdk
[14:13:54] stdout/INFO msvc_bin_path, msvc_inc_dirs, msvc_lib_dirs = find_msvc(env_only)
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\windows_utils.py", line 165, in find_msvc
[14:13:54] stdout/INFO msvc_base_path, version = f()
[14:13:54] stdout/INFO ^^^
[14:13:54] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\windows_utils.py", line 72, in find_msvc_env
[14:13:54] stdout/INFO if not check_msvc(msvc_base_path, version):
[14:13:54] stdout/INFO ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[14:13:55] stdout/INFO File "C:\Users\Administrator\AppData\Local\FreeToken\venv\Lib\site-packages\triton\windows_utils.py", line 59, in check_msvc
[14:13:55] stdout/INFO msvc_base_path / version / "bin" / "Hostx64" / "x64" / "cl.exe",
[14:13:55] stdout/INFO ~~~~~~~~~~~~~~~^~~~~~~~~
[14:13:55] stdout/INFO TypeError: unsupported operand type(s) for /: 'WindowsPath' and 'NoneType'
[14:14:02] health/ERROR TypeError: unsupported operand type(s) for /: 'WindowsPath' and 'NoneType' (×4)

Desktop app version

v0.2.0-beta.21

OS

Windows 10

OS details

No response

GPU and driver

NVIDIA GeForce RTX 3080 Laptop GPU (8 GB), Intel(R) UHD Graphics(128 MB)

CPU and system RAM

Intel(R) Core(TM)i7-10870H CPU @2.20GHz 32G RAM

Checkpoint

gpt-0ss-20b MXFP4

Model settings

No response

Engine log

Anything else

No response

[jason-fxz]

Thanks for the detailed log, it pins this down exactly.

What is happening

This is a bug in triton-windows (the Triton build FreeToken uses on Windows), triggered by an environment variable on your machine.

Your system has VCINSTALLDIR set permanently, but not its companion VCToolsVersion. Triton sees VCINSTALLDIR, assumes it is running inside a Visual Studio developer shell, and tries to build the MSVC path from VCToolsVersion without checking that it exists. Instead of falling back to its bundled compiler it crashes with the WindowsPath / NoneType error, which takes the engine process down.

Neither variable is meant to be permanent: the Visual Studio vcvars scripts set both of them, together, for the lifetime of one shell. A lone VCINSTALLDIR in the system environment is a leftover from a manual edit or an older toolchain. The model, the checkpoint and your GPU are fine; the log shows loading completed and the crash happens at the first Triton kernel launch.

Workaround

Remove the stray variable, then restart the Desktop app. Open System Properties → Environment Variables and delete VCINSTALLDIR from both the user and system lists, or in an admin PowerShell:

[Environment]::SetEnvironmentVariable("VCINSTALLDIR", $null, "User")
[Environment]::SetEnvironmentVariable("VCINSTALLDIR", $null, "Machine")

After restarting, loading the model should get past Start capturing CUDA graphs without the error.