mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-09-27 21:46:57 +02:00
Walk the binding offset back until the distance to the tensor is a whole number of blocks, so block quantized views get a valid element offset in the shader.