v0.3.16-cu124-AVX2-win-20250913

feat: Sync llama: use FA + max. GPU layers by default
feat: Sync sampling : optimize samplers by reusing bucket sort
feat: Sync aLoRA Support
feat: Update Submodule vendor/llama.cpp to 40be511

v0.3.16-cu124-AVX2-linux-20250913

feat: Sync llama: use FA + max. GPU layers by default
feat: Sync sampling : optimize samplers by reusing bucket sort
feat: Sync aLoRA Support
feat: Update Submodule vendor/llama.cpp to 40be511

v0.3.16-cu128-AVX2-win-20250831

feat: Update Submodule vendor/llama.cpp 6c442f4..bbbf5ec
feat: Sync llama : remove KV cache defragmentation logic
feat: Sync model : jina-embeddings-v3 support
feat: Sync llama: use FA + max. GPU layers by default, the flash_attn parameter in context_params has been deleted and replaced by flash_attn_type as the default auto initialization parameter.

v0.3.16-cu128-AVX2-linux-20250831

feat: Update Submodule vendor/llama.cpp 6c442f4..bbbf5ec
feat: Sync llama : remove KV cache defragmentation logic
feat: Sync model : jina-embeddings-v3 support
feat: Sync llama: use FA + max. GPU layers by default, the flash_attn parameter in context_params has been deleted and replaced by flash_attn_type as the default auto initialization parameter.

v0.3.16-cu126-AVX2-win-20250831

feat: Update Submodule vendor/llama.cpp 6c442f4..bbbf5ec
feat: Sync llama : remove KV cache defragmentation logic
feat: Sync model : jina-embeddings-v3 support
feat: Sync llama: use FA + max. GPU layers by default, the flash_attn parameter in context_params has been deleted and replaced by flash_attn_type as the default auto initialization parameter.

v0.3.16-cu126-AVX2-linux-20250831

feat: Update Submodule vendor/llama.cpp 6c442f4..bbbf5ec
feat: Sync llama : remove KV cache defragmentation logic
feat: Sync model : jina-embeddings-v3 support
feat: Sync llama: use FA + max. GPU layers by default, the flash_attn parameter in context_params has been deleted and replaced by flash_attn_type as the default auto initialization parameter.

v0.3.16-cu124-AVX2-win-20250831

feat: Update Submodule vendor/llama.cpp 6c442f4..bbbf5ec
feat: Sync llama : remove KV cache defragmentation logic
feat: Sync model : jina-embeddings-v3 support
feat: Sync llama: use FA + max. GPU layers by default, the flash_attn parameter in context_params has been deleted and replaced by flash_attn_type as the default auto initialization parameter.

v0.3.16-cu124-AVX2-linux-20250831

feat: Update Submodule vendor/llama.cpp 6c442f4..bbbf5ec
feat: Sync llama : remove KV cache defragmentation logic
feat: Sync model : jina-embeddings-v3 support
feat: Sync llama: use FA + max. GPU layers by default, the flash_attn parameter in context_params has been deleted and replaced by flash_attn_type as the default auto initialization parameter.

v0.3.16-cu128-AVX2-win-20250822

• feat: Sync llama.cpp API 20250822
• feat: Try to add python3.13 version compilation
• feat: Add some context API funstion
• fix: Correct some function return data type to c_size_t
• fix: Improve the Qwen25VLChatHandler's ability to handle multiple image inputs to avoid the illusion of multiple image inputs

v0.3.16-cu128-AVX2-linux-20250821

• feat: Sync llama.cpp API 20250822
• feat: Try to add python3.13 version compilation
• feat: Add some context API funstion
• fix: Correct some function return data type to c_size_t
• fix: Improve the Qwen25VLChatHandler's ability to handle multiple image inputs to avoid the illusion of multiple image inputs