WSL2 Multi-GPU: Prefill failing to CPU-only, token generation uses GPUs normally

### What is the issue? **Environment:** Host OS: Windows 11 Pro WSL2 Distribution: Ubuntu 24.04 GPU: 2x NVIDIA GeForce RTX 3060 (12GB) NVIDIA Driver Version: 576.80 (on Windows Host) Model: gemma-3-4b-it-qat.gguf **What happened?** In this WSL2 environment, Ollama consistently uses the CPU for the initial prefill phase of prompt processing. For large prompts, this CPU-bound prefill can take several minutes to complete, during which nvidia-smi reports 0% GPU utilization (although the full model can be seen sharded onto the gpus). Once this phase is over, token generation correctly utilizes the GPUs and achieves expected performance. I've only seen this with Ollama running in the WSL2 environment, it is not reproducible when running Ollama natively on the Windows host. The behavior persists even when num_ctx is reduced from 128k to 8k. The issue also occurs in v0.9.3 (what I started with, going down to 0.9.2 was a troubleshooting step) Several environment and system configurations were also tested with no change in the prefill behavior, including: - Using numactl to set the memory interleave policy. - Explicitly setting CUDA_VISIBLE_DEVICES=0,1. - Forcing the KV cache type with OLLAMA_KV_CACHE_TYPE=f16 in an attempt to address a "kv cache type not supported" warning found in logs. - Attempting to resolve potential library path issues via symbolic links (ln -s) and by explicitly setting LD_LIBRARY_PATH. The core issue appears to be a failure within the Ollama runner process to initialize a GPU-compatible backend for the prefill task, forcing a fallback to the CPU, despite being able to use the GPU for the decoding task. **What did you expect to happen?** The prefill computation for all prompts should be executed on the GPU, leveraging the available VRAM and compute capabilities. CPU usage should remain low during prefill, and nvidia-smi should show high GPU utilization, resulting in fast prompt processing times, consistent with the performance of native Windows Ollama.…