# nemo-speech NMT-only server config. # # Load with: # nemo-speech serve --config config/nmt.example.yaml # # Requires a build with -DNEMO_SPEECH_BUILD_NMT=ON. The translation model is a # Riva-Translate GGUF; the language pair is selected per request with the # source_language and target_language fields. nmt: enabled: true # true | false | auto backend: gpu: 0 # GPU device index, -1 = CPU model: path: /models/riva-translate-4b-instruct-v2.q8_0.gguf # f16 also works n_ctx: 1024 # decode context length; sized for sentence-level # translation. Raise for document-length inputs # (KV cache grows ~0.13 MiB/token per context). generation: max_new_tokens: 256 # cap per input text pool: contexts: 1 # concurrent decode contexts; each adds one # n_ctx-sized KV cache. Raise for concurrency.