Consistent errors on the ComfyUI-Zluda package related to thread causing non-unwinding panic.

[UnsexSituation]

What happened?

Right after a clean install of Stability Matrix, HIP SDK, Adrenalin 25.9.1 on my RX 7600 XT, system running on Windows 11, I decided to use ComfyUI-Zluda because I wanted to use it in tandem with Inference. However, right after installing, I ran the package, letting it compile, and it spat out the error provided. I've tried re-installing drivers, HIP SDK, switching to PRO drivers, yet the issue persists.

Note: I CANNOT reproduce the error on a regular installation (i.e. without Stability Matrix) of ComfyUI-Zluda.

Steps to reproduce

1. Open Stability Matrix
2. Install and Run ComfyUI-Zluda.
3. Wait for it to compile, afterwards watch as it produces an error without being able to generate an image.

Relevant logs

Adding extra search path checkpoints C:\Users\***\Documents\Models\StableDiffusion
Adding extra search path diffusers C:\Users\***\Documents\Models\Diffusers
Adding extra search path loras C:\Users\***\Documents\Models\Lora
Adding extra search path loras C:\Users\***\Documents\Models\LyCORIS
Adding extra search path clip C:\Users\***\Documents\Models\TextEncoders
Adding extra search path clip_vision C:\Users\***\Documents\Models\ClipVision
Adding extra search path embeddings C:\Users\***\Documents\Models\Embeddings
Adding extra search path vae C:\Users\***\Documents\Models\VAE
Adding extra search path vae_approx C:\Users\***\Documents\Models\ApproxVAE
Adding extra search path controlnet C:\Users\***\Documents\Models\ControlNet
Adding extra search path controlnet C:\Users\***\Documents\Models\T2IAdapter
Adding extra search path gligen C:\Users\***\Documents\Models\GLIGEN
Adding extra search path upscale_models C:\Users\***\Documents\Models\ESRGAN
Adding extra search path upscale_models C:\Users\***\Documents\Models\RealESRGAN
Adding extra search path upscale_models C:\Users\***\Documents\Models\SwinIR
Adding extra search path hypernetworks C:\Users\***\Documents\Models\Hypernetwork
Adding extra search path ipadapter C:\Users\***\Documents\Models\IpAdapter
Adding extra search path ipadapter C:\Users\***\Documents\Models\IpAdapters15
Adding extra search path ipadapter C:\Users\***\Documents\Models\IpAdaptersXl
Adding extra search path prompt_expansion C:\Users\***\Documents\Models\PromptExpansion
Adding extra search path ultralytics C:\Users\***\Documents\Models\Ultralytics
Adding extra search path ultralytics_bbox C:\Users\***\Documents\Models\Ultralytics\bbox
Adding extra search path ultralytics_segm C:\Users\***\Documents\Models\Ultralytics\segm
Adding extra search path sams C:\Users\***\Documents\Models\Sams
Adding extra search path diffusion_models C:\Users\***\Documents\Models\DiffusionModels
Checkpoint files will always be loaded safely.
Total VRAM 16368 MB, total RAM 32635 MB
pytorch version: 2.7.0+cu118
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 7600 XT [ZLUDA] : native
Using pytorch attention
Python version: 3.11.13 (main, Jul 23 2025, 00:29:09) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.3.57
ComfyUI frontend version: 1.25.11
[Prompt Server] web root: C:\Users\***\Documents\StabilityMatrix-win-x64\Data\Packages\ComfyUI-Zluda\venv\Lib\site-packages\comfyui_frontend_package\static

Import times for custom nodes:
   0.0 seconds: C:\Users\***\Documents\StabilityMatrix-win-x64\Data\Packages\ComfyUI-Zluda\custom_nodes\websocket_image_save.py

Context impl SQLiteImpl.
Will assume non-transactional DDL.
No target revision found.
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
model weight dtype torch.float16, manual cast: None
model_type EPS
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load SDXLClipModel
loaded completely 9.5367431640625e+25 1560.802734375 True
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
Requested to load SDXL
loaded completely 12836.39580078125 4897.0483474731445 True

  ::  Checking package versions...
Found pydantic: 2.11.7, pydantic-settings: 2.10.1
  ::  Pydantic packages are compatible, skipping reinstall
Installed version of comfyui-frontend-package: 1.25.11
Installed version of comfyui-workflow-templates: 0.1.75
Installed version of av: 15.1.0
Installed version of comfyui-embedded-docs: 0.2.6
  ::  Package version check complete.

  ::  ------------------------ ZLUDA -----------------------  ::  
  ::  Auto-detecting AMD GPU architecture for Triton...
  ::  Detected GPU via Windows registry: AMD Radeon RX 7600 XT
  ::  Set TRITON_OVERRIDE_ARCH=gfx1102
  ::  Triton not installed
  ::  ONNX Runtime not installed — skipping patch.
  ::  Using ZLUDA with device: AMD Radeon RX 7600 XT [ZLUDA]
  ::  Applying core ZLUDA patches...
  ::  Triton optimizations skipped (not available)
  ::  Configuring PyTorch backends...
  ::  Disabled CUDA flash attention
  ::  Enabled math attention fallback
  ::  ZLUDA initialization complete
  ::  ------------------------ ZLUDA -----------------------  ::  


thread '<unnamed>' panicked at zluda_dnn\src\lib.rs:195:14:
[ZLUDA] Unknown data type: 10
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

thread '<unnamed>' panicked at library\core\src\panicking.rs:218:5:
panic in a function that cannot unwind
stack backtrace:
   0:     0x7ff8d25e3cf1 - cudnnConvolutionBackwardData
   1:     0x7ff8d25f117a - cudnnConvolutionBackwardData
   2:     0x7ff8d25e23b7 - cudnnConvolutionBackwardData
   3:     0x7ff8d25e3b35 - cudnnConvolutionBackwardData
   4:     0x7ff8d25e4fe2 - cudnnConvolutionBackwardData
   5:     0x7ff8d25e4dd3 - cudnnConvolutionBackwardData
   6:     0x7ff8d25e5acf - cudnnConvolutionBackwardData
   7:     0x7ff8d25e5922 - cudnnConvolutionBackwardData
   8:     0x7ff8d25e442f - cudnnConvolutionBackwardData
   9:     0x7ff8d25e555e - cudnnConvolutionBackwardData
  10:     0x7ff8d25f9c05 - cudnnConvolutionBackwardData
  11:     0x7ff8d25f9cb3 - cudnnConvolutionBackwardData
  12:     0x7ff8d25f9d95 - cudnnConvolutionBackwardData
  13:     0x7ff8d25dd503 - cudnnBackendSetAttribute
  14:     0x7ff8ef600430 - _CxxFrameHandler3
  15:     0x7ff8ef5f342d - is_exception_typeof
  16:     0x7ff903ea6006 - RtlCaptureContext2
  17:     0x7ff8d25dd4e0 - cudnnBackendSetAttribute
  18:     0x7ffe0e7908bd - at::native::cudnn_convolution_transpose
  19:     0x7ffe0e7e75ce - at::native::_use_cudnn_ctc_loss_tensor
  20:     0x7ffe0e7e9baf - at::native::_use_cudnn_ctc_loss_tensor
  21:     0x7ffe0e7ea17a - at::native::_use_cudnn_ctc_loss_tensor
  22:     0x7ffe0e7ea348 - at::native::_use_cudnn_ctc_loss_tensor
  23:     0x7ffe0e7d2272 - at::native::_use_cudnn_ctc_loss_tensor
  24:     0x7ffe0e7cc841 - at::native::_use_cudnn_ctc_loss_tensor
  25:     0x7ffe0e8256ac - at::native::_use_cudnn_ctc_loss_tensor
  26:     0x7ffe101ae8a8 - at::native::_cudnn_attention_forward
  27:     0x7ffe104fc10e - at::cuda::where_outf
  28:     0x7ffe10438e2a - at::cuda::bucketize_outf
  29:     0x7fff0ffc7f46 - at::_ops::zeros_like_out::redispatch
  30:     0x7fff1001368f - at::_ops::_cudnn_attention_forward::call
  31:     0x7ffe101b1058 - at::native::_fused_sdp_choice_cuda
  32:     0x7ffe105070d8 - at::cuda::where_outf
  33:     0x7ffe10438eeb - at::cuda::bucketize_outf
  34:     0x7fff0fdaba00 - at::_ops::_scaled_dot_product_cudnn_attention::call
  35:     0x7fff0f999252 - at::native::scaled_dot_product_attention
  36:     0x7fff104b253c - at::compositeimplicitautograd::where
  37:     0x7fff1049126f - at::compositeimplicitautograd::broadcast_to_symint
  38:     0x7fff0fe02851 - at::_ops::scaled_dot_product_attention::call
  39:     0x7ff8130d812f - torch::PythonArgs::toSymIntOptional
  40:     0x7ff813107d15 - torch::PythonArgs::toSymIntOptional
  41:     0x7ff87436f47e - PyArg_ParseTuple
  42:     0x7ff8743f0449 - PyObject_Call
  43:     0x7ff8743f006d - PyObject_CallObject
  44:     0x7ff87439f87c - PyEval_EvalFrameDefault
  45:     0x7ff874366721 - PyFunction_Vectorcall
  46:     0x7ff874367b2f - PyFunction_Vectorcall
  47:     0x7ff8743effdd - PyObject_CallObject
  48:     0x7ff87439f87c - PyEval_EvalFrameDefault
  49:     0x7ff874366721 - PyFunction_Vectorcall
  50:     0x7ff874367b2f - PyFunction_Vectorcall
  51:     0x7ff8743effdd - PyObject_CallObject
  52:     0x7ff87439f87c - PyEval_EvalFrameDefault
  53:     0x7ff874366721 - PyFunction_Vectorcall
  54:     0x7ff8743f88bc - PyObject_FastCallDictTstate
  55:     0x7ff87448d3a7 - PyObject_Call_Prepend
  56:     0x7ff87448d2d4 - PyLong_FromBytes
  57:     0x7ff87436cf0b - PyObject_MakeTpCall
  58:     0x7ff87439966f - PyObject_Vectorcall
  59:     0x7ff87439ac6f - PyEval_EvalFrameDefault
  60:     0x7ff874366721 - PyFunction_Vectorcall
  61:     0x7ff874367b2f - PyFunction_Vectorcall
  62:     0x7ff8743effdd - PyObject_CallObject
  63:     0x7ff87439f87c - PyEval_EvalFrameDefault
  64:     0x7ff874366721 - PyFunction_Vectorcall
  65:     0x7ff874367b2f - PyFunction_Vectorcall
  66:     0x7ff8743effdd - PyObject_CallObject
  67:     0x7ff87439f87c - PyEval_EvalFrameDefault
  68:     0x7ff874366721 - PyFunction_Vectorcall
  69:     0x7ff8743f88bc - PyObject_FastCallDictTstate
  70:     0x7ff87448d3a7 - PyObject_Call_Prepend
  71:     0x7ff87448d2d4 - PyLong_FromBytes
  72:     0x7ff87436cf0b - PyObject_MakeTpCall
  73:     0x7ff87439966f - PyObject_Vectorcall
  74:     0x7ff87439ac6f - PyEval_EvalFrameDefault
  75:     0x7ff8743679d1 - PyFunction_Vectorcall
  76:     0x7ff8743f0085 - PyObject_CallObject
  77:     0x7ff87439f87c - PyEval_EvalFrameDefault
  78:     0x7ff8743679d1 - PyFunction_Vectorcall
  79:     0x7ff8743f0085 - PyObject_CallObject
  80:     0x7ff87439f87c - PyEval_EvalFrameDefault
  81:     0x7ff874366721 - PyFunction_Vectorcall
  82:     0x7ff8743f8847 - PyObject_FastCallDictTstate
  83:     0x7ff87448d3a7 - PyObject_Call_Prepend
  84:     0x7ff87448d2d4 - PyLong_FromBytes
  85:     0x7ff87436cf0b - PyObject_MakeTpCall
  86:     0x7ff87439966f - PyObject_Vectorcall
  87:     0x7ff87439ac6f - PyEval_EvalFrameDefault
  88:     0x7ff8743679d1 - PyFunction_Vectorcall
  89:     0x7ff8743f0085 - PyObject_CallObject
  90:     0x7ff87439f87c - PyEval_EvalFrameDefault
  91:     0x7ff8743679d1 - PyFunction_Vectorcall
  92:     0x7ff8743f0085 - PyObject_CallObject
  93:     0x7ff87439f87c - PyEval_EvalFrameDefault
  94:     0x7ff874366721 - PyFunction_Vectorcall
  95:     0x7ff874367b2f - PyFunction_Vectorcall
  96:     0x7ff8743effdd - PyObject_CallObject
  97:     0x7ff87439f87c - PyEval_EvalFrameDefault
  98:     0x7ff874366721 - PyFunction_Vectorcall
  99:     0x7ff874367b2f - PyFunction_Vectorcall
 100:     0x7ff8743effdd - PyObject_CallObject
 101:     0x7ff87439f87c - PyEval_EvalFrameDefault
 102:     0x7ff874366721 - PyFunction_Vectorcall
 103:     0x7ff8743f88bc - PyObject_FastCallDictTstate
 104:     0x7ff87448d3a7 - PyObject_Call_Prepend
 105:     0x7ff87448d2d4 - PyLong_FromBytes
 106:     0x7ff8743f00b0 - PyObject_CallObject
 107:     0x7ff87439f87c - PyEval_EvalFrameDefault
 108:     0x7ff874366721 - PyFunction_Vectorcall
 109:     0x7ff874367b2f - PyFunction_Vectorcall
 110:     0x7ff8743effdd - PyObject_CallObject
 111:     0x7ff87439f87c - PyEval_EvalFrameDefault
 112:     0x7ff874366721 - PyFunction_Vectorcall
 113:     0x7ff874367b2f - PyFunction_Vectorcall
 114:     0x7ff8743effdd - PyObject_CallObject
 115:     0x7ff87439f87c - PyEval_EvalFrameDefault
 116:     0x7ff874366721 - PyFunction_Vectorcall
 117:     0x7ff874367b2f - PyFunction_Vectorcall
 118:     0x7ff8743effdd - PyObject_CallObject
 119:     0x7ff87439f87c - PyEval_EvalFrameDefault
 120:     0x7ff874366721 - PyFunction_Vectorcall
 121:     0x7ff8743f0085 - PyObject_CallObject
 122:     0x7ff87439f87c - PyEval_EvalFrameDefault
 123:     0x7ff8743679d1 - PyFunction_Vectorcall
 124:     0x7ff8743f0085 - PyObject_CallObject
 125:     0x7ff87439f87c - PyEval_EvalFrameDefault
 126:     0x7ff874366721 - PyFunction_Vectorcall
 127:     0x7ff874367b2f - PyFunction_Vectorcall
 128:     0x7ff8743effdd - PyObject_CallObject
 129:     0x7ff87439f87c - PyEval_EvalFrameDefault
 130:     0x7ff874366721 - PyFunction_Vectorcall
 131:     0x7ff8743f88bc - PyObject_FastCallDictTstate
 132:     0x7ff87448d3a7 - PyObject_Call_Prepend
 133:     0x7ff87448d2d4 - PyLong_FromBytes
 134:     0x7ff87436cf0b - PyObject_MakeTpCall
 135:     0x7ff87439966f - PyObject_Vectorcall
 136:     0x7ff87439ac6f - PyEval_EvalFrameDefault
 137:     0x7ff874366721 - PyFunction_Vectorcall
 138:     0x7ff8743f88bc - PyObject_FastCallDictTstate
 139:     0x7ff87448d3a7 - PyObject_Call_Prepend
 140:     0x7ff87448d2d4 - PyLong_FromBytes
 141:     0x7ff8743f00b0 - PyObject_CallObject
 142:     0x7ff87439f87c - PyEval_EvalFrameDefault
 143:     0x7ff874366721 - PyFunction_Vectorcall
 144:     0x7ff8743effdd - PyObject_CallObject
 145:     0x7ff87439f87c - PyEval_EvalFrameDefault
 146:     0x7ff874366721 - PyFunction_Vectorcall
 147:     0x7ff8743effdd - PyObject_CallObject
 148:     0x7ff87439f87c - PyEval_EvalFrameDefault
 149:     0x7ff8743679d1 - PyFunction_Vectorcall
 150:     0x7ff8743f0085 - PyObject_CallObject
 151:     0x7ff87439f87c - PyEval_EvalFrameDefault
 152:     0x7ff8743679d1 - PyFunction_Vectorcall
 153:     0x7ff8743f0085 - PyObject_CallObject
 154:     0x7ff87439f87c - PyEval_EvalFrameDefault
 155:     0x7ff874366721 - PyFunction_Vectorcall
 156:     0x7ff874367b2f - PyFunction_Vectorcall
 157:     0x7ff8743effdd - PyObject_CallObject
 158:     0x7ff87439f87c - PyEval_EvalFrameDefault
 159:     0x7ff8743608de - PyDict_MergeFromSeq2
 160:     0x7ff8744a914c - PyArg_NoPositional
 161:     0x7ff87439e452 - PyEval_EvalFrameDefault
 162:     0x7ff8743608de - PyDict_MergeFromSeq2
 163:     0x7ff8744a914c - PyArg_NoPositional
 164:     0x7ff87439e452 - PyEval_EvalFrameDefault
 165:     0x7ff8743608de - PyDict_MergeFromSeq2
 166:     0x7ff8744a914c - PyArg_NoPositional
 167:     0x7ff87439e452 - PyEval_EvalFrameDefault
 168:     0x7ff8743608de - PyDict_MergeFromSeq2
 169:     0x7ff8744a914c - PyArg_NoPositional
 170:     0x7ff87439e452 - PyEval_EvalFrameDefault
 171:     0x7ff8743608de - PyDict_MergeFromSeq2
 172:     0x7ff8744a914c - PyArg_NoPositional
 173:     0x7ff8d6e159e0 - PyInit__asyncio
 174:     0x7ff8d6e15853 - PyInit__asyncio
 175:     0x7ff87436cf0b - PyObject_MakeTpCall
 176:     0x7ff874596b2c - PyContext_NewHamtForTests
 177:     0x7ff874596dd9 - PyContext_NewHamtForTests
 178:     0x7ff8743c2d8c - PyArg_CheckPositional
 179:     0x7ff8743f03ff - PyObject_Call
 180:     0x7ff8743f006d - PyObject_CallObject
 181:     0x7ff87439f87c - PyEval_EvalFrameDefault
 182:     0x7ff874366721 - PyFunction_Vectorcall
 183:     0x7ff8743f0085 - PyObject_CallObject
 184:     0x7ff87439f87c - PyEval_EvalFrameDefault
 185:     0x7ff874366721 - PyFunction_Vectorcall
 186:     0x7ff874367b89 - PyFunction_Vectorcall
 187:     0x7ff8743f03ff - PyObject_Call
 188:     0x7ff87434881c - PyInterpreterState_Delete
 189:     0x7ff874348782 - PyInterpreterState_Delete
 190:     0x7ff9016137b0 - wcsrchr
 191:     0x7ff90306e8d7 - BaseThreadInitThunk
 192:     0x7ff903dbc34c - RtlUserThreadStart
thread caused non-unwinding panic. aborting.
  0%|          | 0/1 [00:00<?, ?it/s]
Stopped process at 9/9/2025 6:20:36 PM +03:00

Version

v2.15.0

What Operating System are you using?

Windows

[InMyDarkMind]

Same issue ComfyUI-Zluda from Stability Matrix
GPU: 7900xtx
OS: Win11

Used to work before but reinstalled package and no longer works.

[AyelLynx]

Same issue ComfyUI-Zluda from Stability Matrix GPU: 7900xtx OS: Win11

Used to work before but reinstalled package and no longer works.

have you found a fix?

[UnsexSituation]

Same issue ComfyUI-Zluda from Stability Matrix GPU: 7900xtx OS: Win11
Used to work before but reinstalled package and no longer works.

have you found a fix?

There's a 'fix' that worked for me, which was setting the --force-fp32 flag. It tanks the speed too much though

[mitchy030]

Getting the same issue. My guess is pytorch. mine was working when i was using version 2.4.0. now it does not work on 2.7.0.

[AyelLynx]

Getting the same issue. My guess is pytorch. mine was working when i was using version 2.4.0. now it does not work on 2.7.0.

Mine got it fix by uninstalling StabilityMatrix and just using the pure ComfyUI-Zluda. It need some work around to get it working but there's instructions you can follow.

[kndol]

The same issue ComfyUI-Zluda from Stability Matrix
GPU: 6600xt
OS: Win11

[Bro3x]

For me the laucnh option --use-quad-cross-attention instead of --use-pytorch-cross-attention works with fp16. My be the bug ist related to comfyanonymous/ComfyUI#7016.