GTX 1060 CUDA not working for Qwen3-4B-Q4_K_M.gguf with llama-cpp-python 0.3.23-cu125 on Windows #2234

Title:
GTX 1060 (sm_61) Windows 10: ggml-cuda.cu:102 CUDA error with Qwen3-4B-Q4_K_M.gguf

Body:
Environment:

• OS: Windows 10 64-bit
• GPU: NVIDIA GeForce GTX 1060 6GB (compute capability 6.1)
• Driver: 53x.xx (supports CUDA 12.5)
• llama-cpp-python: 0.3.23-cu125 (prebuilt wheel)
• Model: Qwen3-4B-Q4_K_M.gguf
• Python: 3.10

Error messages:

1. llama_context: n_ctx_seq (512) < n_ctx_train (40960)
2. ggml-cuda.cu:102: CUDA error

Symptoms:

• With default n_ctx=512 and n_gpu_layers=35, CUDA error occurs immediately.
• After changing n_ctx=2048 and n_gpu_layers=20, the warning remains, and CUDA error still happens.
• GPU usage jumps briefly then drops to 0%; model falls back to CPU.

Suspected cause:
The prebuilt cu125 wheel does NOT include sm_61 (Pascal/GTX 1060) kernels, leading to CUDA error when offloading layers.

Questions:

1. Does the official cu125 wheel support sm_61?
2. Is there a Windows 10 prebuilt wheel with sm_61 included?
3. What is the recommended n_ctx / n_gpu_layers for Qwen3-4B-Q4_K_M.gguf on GTX 1060 6GB?