Change8

b8400

📦 llama-cppView on GitHub →
7 features🔧 6 symbols

Summary

This release introduces several new element-wise unary operations (neg, exp, sigmoid, softplus) and tensor manipulation ops (CONT, REPEAT) primarily to support Qwen 3.5's DeltaNet architecture on Hexagon hardware.

✨ New Features

  • Added element-wise unary operations (neg, exp, sigmoid, softplus) required by Qwen 3.5's DeltaNet linear attention layers.
  • Implemented 'neg' operation via scaling by -1.0.
  • Implemented 'exp' operation using existing hvx_exp_f32 HVX intrinsics.
  • Implemented 'sigmoid' operation using existing hvx_sigmoid_f32_aa HVX intrinsics.
  • Implemented 'softplus' operation with a scalar fallback for log(1 + exp(x)).
  • Implemented 'CONT' (contiguous) operation by reusing existing CPY infrastructure (same-type copy).
  • Implemented 'REPEAT' operation using tiled memory copy with multi-threaded execution via the worker pool, supporting f32 and f16 types.

Affected Symbols