v3.0.0 — IQ-quants, Diarize, Speculative Decoding, Thermal

whisperGF v3.0.0

تقليل استهلاك الذاكرة (RAM)

• IQ-quants (IQ2_XXS, IQ3_S, IQ4_XS, TQ1_0, TQ2_0) with imatrix
• KV Cache Quantization (INT8/INT4)
• Model Sharding (gguf-split) auto-handled
• K-quants auto-detect from GGUF

تسريع الاستدلال (Inference Speed)

• Speculative Decoding (draft + verify, 2x speedup)
• Qualcomm Hexagon HTP/QNN backend
• Thermal-Aware Threading
• Flash Attention enabled by default

تحسين تجربة المطور (DX)

• Tinydiarize (speaker diarization)
• Integrated Model Downloader (HuggingFace)
• Stateless VAD Reset
• RAII wrappers for JNI

البناء

• ABI: arm64-v8a only
• Size: 5.3 MB