AI Software Just Got Smarter on Your Phone
This update makes AI run faster on your devices without needing the cloud
Deep Dive
Llama.cpp’s latest pre-release fixes a bug in conv_transpose_2d for multiple batches. The ggml implementation previously computed only the first batch, leaving later batches as zero — now the loop covers all batches and the work buffer scales accordingly. The Metal kernel had the same issue, only handling batch 0; it now covers all batches by decoding the batch from the grid and adjusting input and output offsets. A multi-batch test was added, and the release fixes ggml-org/ggml#1448.
Key Points
- Free AI tool Llama.cpp fixed a bug so AI runs faster on your devices like phones and laptops
- No internet needed anymore — AI works offline, saving time and data costs
- The update is free and works on most devices, from old laptops to new phones
Why It Matters
Your AI helpers just got faster, cheaper, and more private — all without buying new gadgets