**Likely Root Cause:** The requested context length (`ctx=262144`) requires a KV cache that vastly exceeds the available 1024 MiB of GPU VRAM. KV cache memory scales linearly with context length, and 262k tokens typically demands several gigabytes of VRAM even with aggressive quantization. The allocator correctly refuses to proceed when the required memory > free memory.

**Next Two Actions:**
1. Reduce the `--ctx-size` (or `-c`) parameter to a value that fits within ~1 GB of VRAM (e.g., `--ctx-size 8192` or lower, depending on model dimension and precision).
2. Lower `--n-gpu-layers` (or `-ngl`) to offload fewer transformer layers to the GPU, freeing VRAM for the KV cache, or switch to a more heavily quantized model (e.g., Q4_K_M → Q2_K) to reduce baseline VRAM consumption.