Failed to load qwen3.5-9b – com.microsoft:CausalConvWithState not registered

[suki-lqh]

Bug Report: Failed to load qwen3.5-9b-generic-gpu:2 – com.microsoft:CausalConvWithState not registered

1. Environment Info

• OS: Windows 11 25H2
• .NET Version: .NET 8.0
• NuGet Packages:
  • Microsoft.AI.Foundry.Local.WinML 1.1.0
  • Microsoft.ML.OnnxRuntime.Gpu.Linux 1.25.1
  • OpenAI 2.10.0
• Affected Model: qwen3.5-9b-generic-gpu:2
• Model Path: D:\foundry_models\Microsoft\qwen3.5-9b-generic-gpu-2

2. Issue Description

When loading the ONNX model qwen3.5-9b-generic-gpu:2 via Microsoft.AI.Foundry.Local, the application crashes with an unhandled exception.
ONNX Runtime GenAI fails to load text.onnx because the custom operator com.microsoft:CausalConvWithState is not registered.

3. Full Exception Stack Trace

Unhandled exception. Microsoft.AI.Foundry.Local.FoundryLocalException: Error loading model qwen3.5-9b-generic-gpu:2: Error: Microsoft.ML.OnnxRuntimeGenAI.OnnxRuntimeGenAIException: Load model from D:\foundry_models\Microsoft\qwen3.5-9b-generic-gpu-2\v2\text.onnx failed:Fatal error: com.microsoft:CausalConvWithState(-1) is not a registered function/op
   at Microsoft.ML.OnnxRuntimeGenAI.Result.VerifySuccess(IntPtr) + 0x47
   at Microsoft.ML.OnnxRuntimeGenAI.Model..ctor(Config) + 0x26
   at Microsoft.Neutron.OpenAI.Provider.OnnxLoadedModel..ctor(String, Config, GenAIConfig, InferenceModel, OnnxEP) + 0x56
   at Microsoft.AI.Foundry.Local.ModelManager.<LoadModelAsync>d__9.MoveNext() + 0xc8c
--- End of stack trace from previous location ---
   at Microsoft.AI.Foundry.Local.NativeInterop.<>c__DisplayClass10_0.<<ExecuteCommandManaged>b__5>d.MoveNext() + 0xa0
--- End of stack trace from previous location ---
   at Microsoft.AI.Foundry.Local.NativeInterop.<ExecuteWithTracker>d__9.MoveNext() + 0xb8

4. Root Cause Analysis

1. Platform mismatch: Referenced Microsoft.ML.OnnxRuntime.Gpu.Linux Linux build running on Windows, causing custom operator binaries missing.
2. The ONNX model qwen3.5-9b-generic-gpu relies on custom operator com.microsoft:CausalConvWithState, which is not registered in current OnnxRuntimeGenAI runtime.
3. Version incompatibility between Foundry.Local.WinML 1.1.0 and OnnxRuntime 1.25.1.

5. Suggested Resolution

1. Replace Microsoft.ML.OnnxRuntime.Gpu.Linux with Windows-compatible Microsoft.ML.OnnxRuntime.Gpu or Microsoft.ML.OnnxRuntime.DirectML.
2. Upgrade Microsoft.AI.Foundry.Local.WinML to a newer version that supports Qwen3.5 series ONNX models.
3. Ensure runtime extensions and custom operator libraries for CausalConvWithState are properly registered in OnnxRuntime GenAI.