* Fix model download in test workflow

* Use hf CLI in test workflow

* Use hf CLI name in CI and docs

* Reference PR in changelog

abetlen#2150)

* fix(ci): use supported macos runner label

* fix(ci): add apple silicon macos test coverage

* fix(ci): run standard macos tests on apple silicon

* fix(ci): simplify apple silicon macos install

* fix(ci): disable ggml native on apple silicon runner

* docs: update changelog for macos ci runner fix

* Add Ruff formatting and safe lint baseline

* Update changelog for Ruff setup

* Update llama.cpp and sync bindings

* Clean up binding compatibility shims

* Remove flash attention property shim

* Remove mtmd verbosity shim

* Add docstrings for new bindings

* Format Ruff files and add changelog entry

* ci: add riscv64 wheel builds to release workflow

Add a build_wheels_riscv64 job mirroring the existing arm64 QEMU-based
build. Uses cibuildwheel with QEMU emulation for linux/riscv64, targeting
CPython 3.10-3.14 on manylinux.

Closes abetlen#2138

* ci: use cibuildwheel 3.1.2 for riscv64 wheels

* docs: update changelog for riscv64 wheel PR

---------

Co-authored-by: abetlen <abetlen@gmail.com>

* fix: handle Qwen 3.5 hybrid prefix reuse

* test: fix Qwen runtime unit mocks

* test: drop Qwen runtime unit tests

* docs: credit Qwen fix contributors in changelog

* docs/tests: update default Qwen model to 3.5 0.8B

* test: rebaseline Qwen 3.5 outputs

* test: stabilize low-level Qwen sampling check

* test: tighten Qwen 3.5 completion prompts

* fix(ci): harden release wheel workflow

* fix(ci): document and pin release wheel baselines

* fix(ci): speed up release arch builds

* fix(ci): split riscv64 by python version

* fix(ci): sanitize riscv64 artifact names

* fix(ci): harden cuda wheel workflow

* fix(ci): pin cuda toolkit versions accurately

* fix(ci): resolve exact cuda toolkit installs

* fix(ci): align cuda toolkit roots and tags

* fix(ci): pin cuda packages to nvidia label

* fix(ci): allow cuda solver to mix non-cuda deps

* fix(ci): harden docker build workflow

* docs: update changelog for ci workflows

* feat: expose attention_type parameter in Llama.__init__

* docs: preserve attention_type in pickled state

* docs: update changelog for attention_type

---------

Co-authored-by: Victor Biederbeck <victor@moria.hiddencove.xyz>
Co-authored-by: abetlen <abetlen@gmail.com>

…ent arches and one PTX target for forward compatibility (abetlen#2158)

* fix(ci): shrink CUDA wheel fatbins

* docs: update changelog for cuda wheel size fix

* Fix embedding models without KV memory

* Add changelog entry for embedding memory fix

* Update llama.cpp to c0159f9c1

* Add changelog entry for llama.cpp update

…#2165)

* fix(ci): publish distinct manylinux and musllinux cpu wheels

* docs: add changelog entry for linux wheel repair fix

* ci: publish CPU wheels as py3-none

* docs: add changelog entry for py3-none wheel tags