v0.13.1

v0.13.1 (2026-04-23)

This release is published under the Apache-2.0 License.

Bug Fixes

• Correct GQA/MQA head expansion logic for Q and K (62b73bd)

• Introduce RoutingResult and fix probability renormalization (a1dabe3)

• Remove GPU-CPU sync that was happening during bincount() call in MoE layer (534facc)

• Remove MoE statistics interface from qwen3 decoder layer and module (99e0379)

---

Detailed Changes: v0.13.0...v0.13.1

v0.13.0

v0.13.0 (2026-04-20)

This release is published under the Apache-2.0 License.

Features

• Migrate to Flash Attention 4 (19f8483)

---

Detailed Changes: v0.12.0...v0.13.0

v0.12.0

v0.12.0 (2026-04-14)

This release is published under the Apache-2.0 License.

Features

• Add partial RoPE to GQA (4dcfe7e)

---

Detailed Changes: v0.11.0...v0.12.0

v0.11.0

v0.11.0 (2026-04-13)

This release is published under the Apache-2.0 License.

Features

• Add Gated DeltaNet layer that use fla-core kernels (e0e466f)

• Add shared expert to MoE block (0e93f45)

---

Detailed Changes: v0.10.0...v0.11.0

v0.10.0

v0.10.0 (2026-04-13)

This release is published under the Apache-2.0 License.

Features

• Add custom RMSNorm kernel supporting zero-centered weight (365ad1f)

• Add optional output gating for grouped query attention (af3e411)

---

Detailed Changes: v0.9.0...v0.10.0

v0.9.0

v0.9.0 (2026-04-08)

This release is published under the Apache-2.0 License.

Bug Fixes

• Allow using MoE parallelization function with [EP]=1 (fdc7d6d)

• Fix huggingface mappers for classification models (caa2e63)

Features

• Add model state mappers that convert HuggingFace model states into d9d ones and vice versa (47bae88)

• Make HF model state mappers for Qwen3 MoE support new fused experts format from transformers v5 (cdecc29)

---

Detailed Changes: v0.8.0...v0.9.0

v0.8.0

v0.8.0 (2026-03-22)

This release is published under the Apache-2.0 License.

Features

• Add training hooks system based on event bus along with a set of predefined events (13a76e2)

---

Detailed Changes: v0.7.0...v0.8.0

v0.7.0

v0.7.0 (2026-03-18)

This release is published under the Apache-2.0 License.

Features

• Add Multi-Latent Attention implementation along with support for different RoPE layouts (56684bc)

• Add Qwen3 Dense model (08047bd)

---

Detailed Changes: v0.6.0...v0.7.0

v0.6.0

v0.6.0 (2026-03-11)

This release is published under the Apache-2.0 License.

Bug Fixes

• Render multi-process model state loading progress using multiple progress bars (fixes #1) (defea97)

Features

• Implement composition-based classificaiton metrics API (7d7476c)

---

Detailed Changes: v0.5.4...v0.6.0

v0.5.4

v0.5.4 (2026-03-10)

This release is published under the Apache-2.0 License.

Bug Fixes

• Reorder contents in README (973c7cb)

Documentation

• Migrate from mkdocs-shadcn to zensical (7c131c4)

---

Detailed Changes: v0.5.3...v0.5.4