v0.1.3

Fixes

• Progressive audio degradation past ~10 s (#7): the codec transformers are trained with 10 s sliding-window causal attention (causal_transformer_context_duration); the GGML port used an unbounded causal mask, so quality collapsed progressively after the 10 s mark (crackling, robotic voice, falling volume) on every backend. Both the decoder and the encoder (voice-cloning references > 10 s) are fixed.
• Audio repetition penalty now matches the reference: each token in the history is penalized once (torch.unique semantics) instead of once per occurrence, which compounded to penalty^k at the server's default penalty of 1.1.

Docs

• README build instructions updated for the vendored third_party/llama.cpp submodule (-DLLAMA_CPP_DIR is gone, as reported in #7); Windows output paths corrected; --n-batch / --n-ctx / --aux-cpu documented.
• scripts/convert_hf_to_gguf.py --llama-cpp-dir now defaults to the bundled submodule.

v0.1.2

• /v1/audio/speech: enforce the documented 0.25 lower bound for 'speed'
• README refresh: prebuilt binaries, error semantics

v0.1.1

• Bundle CUDA runtime (libcudart/libcublas/libcublasLt) in the CUDA release tarball instead of requiring a system CUDA toolkit; add Blackwell (sm_120) kernel + PTX fallback
• Windows: build with GGML_OPENMP=OFF (TheRock's Windows ROCm dist doesn't ship the LLVM OpenMP runtime; ggml's own thread pool covers it)

v0.1.0

Prebuilt backend binaries for Lemonade.