llama.cpp b10273 Release Fixes Sampler Initialization in Backend Sampling
The llama.cpp b10273 release removes "full-context windows" from history-based samplers to resolve initialization order issues. Instead of resolving to the full context length (previously represented by -1), the default value is now resolved to 1024, and history-based samplers share a default value of 64. This release addresses a critical initialization conflict introduced by backend-sampling, where samplers are constructed before the full llama context is established. Resolving this ensures smoother execution and prevents crashes or incorrect behavior when running inference with hardware-accelerated sampling. The change removes the `context_size` parameter from history-based samplers because the system cannot infer the final context length at the time of sampler construction. This minor build release also includes pre-built binaries for various platforms, though KleidiAI support for macOS Apple Silicon remains disabled.
## BACKGROUND
In LLM inference, samplers determine how the model selects the next token from probability distributions. "Backend-sampling" in llama.cpp integrates this sampling step directly into the computation graph on backends like CUDA to reduce CPU-GPU data transfer overhead, but this requires initializing samplers before the complete context object is fully created.