Add Gemma 4 model family support (#1512)

* Add Gemma 4 model family support

Author: YunchaoYang

CHANGELOG.md

@@ -3,6 +3,11 @@ All notable changes to fairseq2 are documented in this file.
 
 The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
+## [0.8.1] - Unreleased
+- Gemma 4 model family (E4B, 31B, 26B-A4B) with base and instruction-tuned variants. Includes decoder with Per-Layer Embeddings (PLE), partial RoPE, KV sharing across sliding/global attention layers, Mixture-of-Experts (26B-A4B), QK/V-norm, logit soft-capping, audio tower (Conformer encoder for multimodal E4B), bidirectional HuggingFace state dict conversion, FSDP/activation checkpointing/tensor parallel support, and SFT recipe configs.
+- Bump transformers~=v5.5 and loosen huggingface_hub upper bound. (#1508)
+- Fixed typo in WerMetric: use `hyp_seqs` instead of `ref_seqs` for `hyp_seqs_list`. (#1506)
+
 ## [0.8.0] - March 25th, 2026
 - fsspec integration for remote filesystem support. Checkpoints can be saved to and loaded from S3 via `--checkpoint-dir s3://bucket/path/`. Requires `s3fs`. (#1126)
 - New `GlobalFileSystem` replaces `LocalFileSystem` as default, dispatching to the appropriate backend based on URI scheme. (#1126)

doc/source/_static/img/gemma4/gemma4_architecture.svg

@@ -67,6 +67,7 @@ other content generation tasks.
     reference/fairseq2.gang
     reference/fairseq2.model_checkpoint
     reference/fairseq2.models
+    reference/fairseq2.models.gemma4
     reference/fairseq2.models.hg
     reference/fairseq2.models.hub
     reference/fairseq2.models.llama