#2894 [Bug] Qwen3.6-35B-A3B setup fails on Windows with prebuilt engine v0.1.45

本文

### Description

I'm trying to run Qwen3.6-35B-A3B in a separate Strata installation on Windows, while keeping my existing Strata installation running Qwen3.8-Flash-Next unchanged.

I downloaded the latest Strata release (v0.1.45) and ran the setup. The llama.cpp source, prebuilt Strata engine, and NVIDIA CUDA libraries all downloaded and installed successfully. However, setup stopped with the following error:

```text
[ok] llama.cpp 3cf0325 (gguf-py, ggml, mtmd)
[ok] ready-made engine 0.1.45 for sm_75, sm_86, sm_89, sm_120 (CUDA 13.0)
[ok] NVIDIA CUDA libraries (cuBLAS, CUDA runtime; ~0.4 GB) installed

[X] this engine (0.1.45) does not run Qwen3.6-35B-A3B yet
    update Strata, or compile the engine with --build, and run setup again
```

### System information

- **OS:** Windows
- **Strata version:** v0.1.45
- **GPU:** NVIDIA GeForce RTX 5070 Ti (Blackwell, compute capability 12.0 / sm_120)
- **Existing installation:** Strata running Qwen3.8-Flash-Next
- **New installation:** Separate Strata folder for Qwen3.6-35B-A3B
- **Python:** Using Strata's bundled virtual environment; Python is not installed system-wide

### Expected behavior

The setup should install a compatible prebuilt Windows engine and allow Qwen3.6-35B-A3B to run on the RTX 5070 Ti without requiring a manual source build.

### Actual behavior

The downloaded engine advertises support for `sm_120`, and all CUDA dependencies install successfully, but Strata's setup explicitly rejects the engine for Qwen3.6-35B-A3B.

### Questions

1. Is Qwen3.6-35B-A3B supported on Windows with the RTX 5070 Ti in v0.1.45?
2. Does the prebuilt engine need a newer llama.cpp version or specific changes to support this model architecture?
3. Could a compatible precompiled Windows engine be provided so users do not have to build from source?

I would prefer a ready-to-run solution because I do not want to compile the engine manually.

Thank you!