v7.2.0

7.2.0 (2026-10-21)

🍕 Features

• Load Qwen3 embedding models from GGUF files in the Candle backend (#325) (287fa9d) - By @dhkts1, @anush008

🚀 Performance

• Three Qwen3 Metal speedups: one q/k norm-and-rope kernel, attention output in o_proj's order, and the last layer for the pooled position (#324) (73035d3) - By @dhkts1
• Hand Qwen3's q, k and v to the fused SDPA as strided views (#321) (1d5f16e) - By @dhkts1
• Run Qwen3's RMSNorm and RoPE through Candle's fused kernels (#320) (dc44316) - By @dhkts1
• Run padded Qwen3 batches through the fused SDPA kernel on Metal (#319) (8c1b457) - By @dhkts1
• Run Qwen3 attention through the fused SDPA kernel on Metal when nothing is padded (#318) (9c80de7) - By @dhkts1
• Use the fused softmax in Qwen3 attention, nearly halving peak memory on long texts. (#317) (e669409) - By @dhkts1