[sebcrozet]

v0.5.0

Added

• New ml feature enabling the vortx::ml module of machine-learning operators: activations (incl. tanh/elu backward), unary ops, softmax/log-softmax, RMS norm, layer norm, RoPE, SiLU, fused attention, axis reductions, concat, gather, select, conv2d/conv-transpose-2d, im2col, pooling, window partitioning, the Adam optimizer, PPO gradients, and quantized GEMV (Q4/Q5/Q8 and K-quants, with CPU-side quantization blocks). The optional rand feature adds Distribution impls for sampling random quantized blocks.
• cuda-oxide feature to compile the CUDA kernels with cuda-oxide instead of rust-cuda. It is commented out in the published crates: it needs khal from GitHub through [patch].
• vortx_shaders::linalg::{contiguous, op_assign}::MAX_NUM_THREADS are now public.

Changed

• Update to khal/khal-std/khal-builder 0.4.
• gemm_tiled accumulates with Vec4 FMAs.

Fixed

• Contiguous and OpAssign now clamp their dispatch size to MAX_NUM_THREADS, fixing tensors too large for the workgroup-count limit.