master-1093-a59298e

fix: align CUDA extension architectures with ggml (#2347)

master-1092-5d78740

feat: support interleaved image prompts for vision LLMs (#2341)

master-1091-b9d7176

feat: unify text encoder selection across H3 variants (#2340)

master-1090-8b1e9df

feat: allow the P-H3-Image text encoder for MiniMax-H3-Image (#2337)

master-1088-32d516c

feat: add MiniMax-H3-Image support (#2324)

master-1087-f8417d0

perf(cuda): fuse RMSNorm and RoPE in Sol-H3-VAE (#2323)

master-1086-465adf3

fix: avoid structured binding capture in MiniMax H3 SOL attention

master-1081-e1ca5e8

perf(cuda): fuse bias into INT8 CUTLASS epilogues (#2320)

master-1080-a3f46b6

feat(server): make async job queue limits configurable (#2316)

master-1079-a3c49b1

refactor: unify ggml operation helpers around runner context (#2314)