Notable changes: April 9–25, 2026

[justinchuby]

Summary

Over the past three weeks (April 9–25), mobius merged 50 PRs from 5 contributors (plus 2 bots). The headline additions are Gemma 4 multimodal support (text + vision + audio with MoE), significant improvements to CUDA/GPU execution provider compatibility, and a push to bring L1–L3 test coverage above 95%.

---

🧠 New Model Support

• Gemma 4 (any-to-any multimodal) — Full text, vision, and audio support with Mixture-of-Experts routing and KV-cache sharing (feat: add Gemma 4 support (text, vision, audio, MoE) #137, feat: add input_features_mask to Gemma4 audio encoder #207) Investigation coordinated improvements in onnxruntime, genai and ort extentions
• NemotronH — Added MoE layer support for Nemotron-3 30B/120B hybrid architectures (Add MoE layer support for NemotronH (Nemotron-3 30B/120B) #188)
• Pixtral / Ministral3 — Vision-language model support with dynamic vision encoding, FP8 dequantization, and multi-head attention (Add Pixtral/Ministral3 VLM support: dynamic vision, FP8 dequant, PatchMerger fix, MHA #130)
• Mamba2 — Reimplemented using the new LinearAttention component for cleaner architecture (Use LinearAttention to implement Mamba2 and other cleanup #163)
• Qwen3.5-MoE — Fixed weight preprocessing to correctly strip language_model prefix and drop unsupported keys (fix: Qwen3.5-MoE preprocess_weights strips language_model prefix and drops mtp/visual keys #195)

⚡ CUDA / GPU Inference

• GroupQueryAttention for CUDA EP — Emit com.microsoft.GroupQueryAttention directly from the Attention component when the target execution provider supports it, including sliding-window attention for Gemma 4 (feat: emit GroupQueryAttention directly in Attention when EP supports it #134, Emit com.microsoft.GroupQueryAttention for Gemma4 with sliding window #190)
• Gemma 4 CUDA accuracy — Fixed CUDA EP correctness issues and audio encoder numerical accuracy (Fix Gemma4 CUDA EP support and audio encoder accuracy #172)
• WebGPU support — Enabled Shape op support for WebGPU EP and removed workaround rewrite rules (feat: Enable WebGPU Shape op support, remove EliminateShape rewrite rules #132)
• MXFP4 quantization — Fixed GPT-OSS MXFP4 dequantization and Gemma 3 VL auto-routing for quantized models (fix: GPT-OSS MXFP4 dequant + Gemma3 VL auto-routing #194)

�� ORT GenAI Integration

• Gemma 4 end-to-end export — Fixed genai_config generation including decoder path, vision inputs, image_processor.json, processor_config.json, and spatial_merge_size (fix: Gemma4 genai config - vision inputs, decoder input_ids, processor #199, fix: genai config gaps - decoder path, image_processor.json, spatial_merge_size #204, Add Gemma4 image_processor.json generation with onnxruntime-extension… #209)
• Chat template propagation — Export now copies chat_template.jinja to the ORT GenAI output folder (fix: copy chat_template.jinja to ORT GenAI export folder #198)
• Config-mode fixes — Corrected model_type and token ID propagation when using --config mode, plus tokenizer copy (fix: --config mode model_type and token ID propagation #152, fix: genai_config defaults and tokenizer copy for --config mode #153)
• Tied weight deduplication — Shared tied weights are now emitted as a single ONNX initializer instead of duplicated copies, reducing model size (fix: share tied weights as single ONNX initializer instead of copying #154)

✅ Quality & Testing

• L1–L3 coverage at 95%+ — Fixed L4/L5 test failures and expanded coverage with golden data and seq2seq generation tests (Fix L4/L5 test failures and improve L1-L3 coverage to 95%+ #178, Expand L4/L5 test coverage: golden data, seq2seq generation, ci_skip_reason #183)
• GPU CI overhaul — Split L4 and L5 GPU workflows, extracted reusable workflow templates, wired affected-model detection to skip unchanged models, and added MOBIUS_TEST_DEVICE for CUDA (ci: Split L5 GPU workflows from L4 #164, CI: Fix GPU workflow dependencies and increase timeouts #179, ci: add MOBIUS_TEST_DEVICE environment variable for CUDA in workflows #180, CI: Narrow affected-models detection by resolving __init__.py re-exports #182, CI: Extract L4/L5 GPU jobs into reusable workflows and wire detect-affected #185)
• Smart change detection — CI now resolves __init__.py re-exports to accurately detect which models changed, avoiding unnecessary GPU test runs (ci: Fix model change detection script and registry parsing logic #169, CI: Narrow affected-models detection by resolving __init__.py re-exports #182)
• Codecov integration — Added coverage upload to all CI pipelines (ci: add codecov upload step to all CI pipelines #135)
• Bug fix: silent NoPE-model corruption — Made RoPE config nullable so models without positional embeddings are no longer silently corrupted (Make RoPE config nullable to fix silent NoPE-model corruption #177)
• Cross-platform — Fixed Windows PermissionError in test_patch_merger (Fix Windows PermissionError in test_patch_merger #160)

🛠️ Developer Experience & Code Quality

• CLI docs consolidation — Merged scattered CLI documentation into a single cli_reference.md and clarified --ep vs --optimize flags (docs: clarify --ep vs --optimize in getting-started guide #201, docs: consolidate CLI documentation into cli_reference.md #203)
• Agent skills restructuring — Moved skills to .agents/skills/, applied progressive disclosure, and added proven Gemma 4 patterns (docs: update skills with proven Gemma 4 patterns (MoE, Any-to-Any, KV sharing) #140, Restructure skills: move to .agents/skills/, apply progressive disclosure, merge overlapping skills #181)
• Quality checklist — Added a definition-of-done checklist for new model additions covering L1–L5 testing, ORT GenAI validation, and code review (Add quality checklist (definition of done) for new model additions #139)
• onnxscript 0.7.0 adoption — Refactored ir.Function builders to use the new build_function API (Refactor ir.Function builders to use build_function from onnxscript 0.7.0 #189)
• Cleaner ops — Replaced op.Constant(value_float=...) with Python float literals for dtype-safe auto-casting (Replace op.Constant(value_float=...) with Python float literals, use autocast #166), replaced FusedMatMul with standard Transpose+MatMul (refactor: replace FusedMatMul with Transpose+MatMul, add initializer-folding passes #129)
• Sphinx docs timestamps — Pages now show accurate git-based creation and last-update dates (Add git-based page creation and last-update timestamps to Sphinx docs #145, Fix docs pages showing identical created/updated dates #149)

---

Justin's notes:

• Olive pass support: using mobius to acquire an onnx model as an Olive pass

---

🔭 What's Next

• Open source release
• E2E Olive recipes: Expand e2e Olive recipes for a batch of models.
• Expanding GPU validation — Continue broadening L4/L5 GPU test coverage across more model families
• Additional multimodal architectures — Audio-to-audio, TTS, etc.
• Support generating models from Olive quantized PyTorch weights (gptq etc.).-
• Quantization support — Improve MXFP4/INT4 weight handling across more architectures / from gguf models
• ORT GenAI runtime parity — Close remaining gaps in config generation for newer models
• Performance testing by integrating with ep cert testing infra