gemini-cli - 💡(How to fix) Fix BUG: Failed to create engine: OpenCL build error on RTX 2080 Ti despite latest drivers and DXC DLLs

What happened?

The Gemini CLI fails to initialize the local inference engine for gemma3-1b-gpu-custom on Windows 11 with an NVIDIA RTX 2080 Ti.

The LiteRT server starts successfully on port 9379, but any request (e.g., routing classification) triggers a 500 Internal Server Error: failed to create the engine: failed to create engine for model gemma3-1b-gpu-custom with backend gpu_artisan: failed to create engine.

The server logs reveal multiple OpenCL kernel compilation failures in cl_kernel.h, specifically related to type conversions: note: candidate function not viable: no known conversion from '__global half4' (vector of 4 'half' values) to 'float8' (vector of 8 'float' values) for 1st argument.

Despite manually installing the officially recommended dxil.dll and dxcompiler.dll (v1.8.2505.1) into the binary directory as per the documentation, the engine persists in attempting (and failing) to build an incompatible OpenCL kernel instead of utilizing the DirectX/WebGPU backend properly.

What did you expect to happen?

The local Gemma model should initialize correctly using the GPU acceleration (DirectXShaderCompiler backend) and handle inference requests without falling back to a failing OpenCL compilation.

Client information

> /about
About Gemini CLI

CLI Version                               0.40.1
Model                                     Auto (Gemini 3)
OS                                        win32
Auth Method                               Signed in with Google (REDACTED)

Login information

Signed in with Google Account (OAuth).

Anything else we need to know?

• GPU: NVIDIA GeForce RTX 2080 Ti (Turing)
• Driver Version: 591.86 (Released Jan 27, 2026)
• Environment: Windows 11
• Attempts made:

1. Verified latest NVIDIA drivers.
2. Manually placed dxil.dll and dxcompiler.dll from DXC release v1.8.2505.1 into %USERPROFILE%\.gemini\bin\litert\.
3. Restarted server with --debug flag.

• Reference: This issue seems identical to #26502, suggesting a regression or persistent compatibility bug with the gpu_artisan backend on certain NVIDIA architectures.