The llama.cpp b10666 release expands the test-save-load-state suite to run across all architectures and resolves several critical bugs in state management and graph handling. Key changes include fixing dangling references in minimax-01, aligning chunk copying for on-device sequences, and correcting indexer head counts for deepseek4.
- The test-save-load-state suite now runs against every *.gguf model via a new --models DIR mode, replacing the previous single-model limitation.
- On-device state sequence copy logic was relaxed to handle non-contiguous sources by using a byte cursor instead of requiring exact chunk alignment.
- The dummy DSA indexer parameters were updated (key_length=128, head_count=64) to ensure the fused Lightning Indexer op runs on the GPU rather than falling back to CPU.
- deepseek4's indexer head count was set to 64 to match the fixed kernel requirements, preventing device-mismatch warnings.
- A dangling hparams reference in minimax-01's LA graph input was replaced with a copy to prevent stack memory corruption during graph reuse.
These updates improve test coverage and stability for complex architectures like deepseek4, gemma2, and minimax-01, ensuring state serialization works correctly across diverse hardware backends.