Data-dependent bidirectional mask (Gemma-4 vision-block) forces standard Attention over GQA — split prefill/decode decoder graphs?

BFloat16 logits returned as garbage — Logits::Get() only casts Float16 to float32

OGA export generating an invalid model when a lm_head unquantized model is exported

Copilot design of Pipeline-as-Config

Feature Request: Support for Google Gemma 4 model family (PLE architecture, variable head dims, KV cache sharing)

cannot build Qwen/Qwen3-14B on machine with 64GB RAM

[Java] `onnxruntime-genai-cuda.dll` missing from Java JAR and extraction logic in CUDA builds

Issue with OGA granite-4.0-1b model response

Feature Request: Support for Qwen3.5

FP16 model supported for CPU