**Root Cause:** 
The requested context length (`ctx=262144`) requires a KV cache that vastly exceeds the available GPU VRAM (`free=1024 MiB`). KV cache memory scales linearly with context length, batch size, and model dimensions. Allocating a 262k-token context window on a GPU with only 1 GB free is physically impossible without offloading or quantization.

**Next Two Actions:**
1. **Reduce `--ctx-size`** (or `-c`) to a value that fits within the available VRAM (typically 4096–8192 tokens for most models on consumer GPUs).
2. **Adjust GPU offloading or enable KV cache quantization**: Lower `--n-gpu-layers` (`-ngl`) to reserve VRAM for the KV cache, or add `--cache-type-k q8_0 --cache-type-v q8_0` (if supported by your `llama.cpp` build) to compress the KV cache and fit it into the remaining memory.