ggml/llama.cpp releases: b11009

TL;DR

The ggml/llama.cpp project released a new version addressing issues with split states and granularity for fused QKV gemm operations, crucial for models like Qwen35. This update ensures correct handling of attention layers when using specific configurations, enhancing the model's performance and accuracy.

Detailed Summary

The ggml/llama.cpp project released updates addressing issues with split states and granularity in the fused QKV gemma4 and qwen35 models. These changes are crucial for handling different embedding dimensions, ensuring compatibility with --fuse-qkv flag. The broader impact includes improved model performance and flexibility, particularly for large-scale language models like Qwen35.

Key Points

• • fix split state and granularity for fused QKV gemma4, qwen35
• • calculate split states for attn_qkv from n_head * n_embd_head_k
• • handle fused full attention layers for qwen35/qwen35