# NemotronVoiceChat-11B STREAMING speech-to-speech deployment (async-chunk).
#
# Same 3 stages as nemotron_labs_voicechat.yaml, but the stages overlap:
# the thinker streams its frame-locked text timeline to the talker per decode
# step (cumulative ids.all payloads), and the talker streams prompt-trimmed
# cumulative code stacks to code2wav every `codec_chunk_frames` frames. The
# codec re-decodes the full prefix per chunk and emits only the new samples
# (NeMo decode_one_audio_step semantics), so the concatenated stream is
# seam-free. First audio arrives after ~codec_chunk_frames thinker steps
# instead of after the whole pipeline completes.
#
# The non-streaming yaml remains the parity-acceptance path (bit-exact WAV);
# streamed audio is validated by ASR/WER instead (prefix decodes are not
# bit-identical to the one-shot decode of the final stack).
pipeline: nemotron_voicechat
async_chunk: true

connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      shm_threshold_bytes: 65536
      # Streaming cadence for talker -> code2wav: emit a chunk once this many
      # NEW frames (80 ms each) have accumulated; ~1 s of audio per chunk.
      codec_chunk_frames: 13

stages:
  - stage_id: 0  # thinker: speech -> frame-locked text (final text output)
    max_num_seqs: 1
    gpu_memory_utilization: 0.62
    enforce_eager: true
    enable_prefix_caching: false
    async_scheduling: false
    max_num_batched_tokens: 8192
    max_model_len: 8192
    dtype: float32
    mamba_ssm_cache_dtype: float32
    devices: "0"
    output_connectors:
      to_stage_1: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      ignore_eos: true
      seed: 0

  - stage_id: 1  # talker: text timeline -> 31-quantizer code stacks
    max_num_seqs: 1
    gpu_memory_utilization: 0.12
    enforce_eager: true
    enable_prefix_caching: false
    async_scheduling: false
    skip_tokenizer_init: true
    max_num_batched_tokens: 4096
    max_model_len: 16384
    dtype: float32
    devices: "0"
    input_connectors:
      from_stage_0: connector_of_shared_memory
    output_connectors:
      to_stage_2: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      max_tokens: 16383
      stop_token_ids: [1]
      detokenize: false
      seed: 0

  - stage_id: 2  # code2wav: codes -> 22.05 kHz PCM (fp32 codec)
    max_num_seqs: 1
    gpu_memory_utilization: 0.06
    enforce_eager: true
    enable_prefix_caching: false
    async_scheduling: false
    skip_tokenizer_init: true
    max_num_batched_tokens: 65536
    max_model_len: 65536
    devices: "0"
    input_connectors:
      from_stage_1: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      max_tokens: 1
      detokenize: false
      seed: 0