On-device LLM inference engine
- Completed the QNN Embedding model export pipeline. Routed Embedding and standard LLMs through their own export flows by model type, filled in the required configuration and artifact mapping, and made Embedding model export and deployment reliable.
- Fixed an AWQ quantization initialization crash in multimodal models. Straightened out how calibration data is passed into visual position encoding, preventing null values from interrupting quantization initialization, and added multimodal regression validation.
- Fixed state leakage across successive Gemma 4 Omni text requests. Handled text and multimodal paths separately so a new request no longer reuses cache state left over from the previous decoding round, making multi-turn results stable and reproducible.
- Fixed attention mask selection in the Omni generation stage. Let generation and embedding modes each match their correct mask implementation dynamically, preventing hybrid-attention models from failing on the first forward pass and improving on-device multimodal inference stability.
