# Audex (Nemotron-Labs-Audex-2B) thinker-only deploy: audio understanding
# (ASR / audio QA) -> text. Single-stage configuration of the audio-capable
# full checkpoint (checkpoint_folder_full): NV-Whisper encoder + relu2
# projector + 2B dense LM.
#
# The official audioqa script decodes ASR at temperature 0; max_tokens 256
# covers transcription and short QA answers (raise for long-form answers).
pipeline: audex_thinker_only
async_chunk: false

stages:
  - stage_id: 0
    max_num_seqs: 8
    gpu_memory_utilization: 0.7
    trust_remote_code: true
    enable_prefix_caching: true
    max_num_batched_tokens: 8192
    # 30 clips x 750 sound embeddings (900 s cap) + text headroom.
    max_model_len: 32768
    dtype: bfloat16
    devices: "0"
    default_sampling_params:
      temperature: 0.0
      max_tokens: 256
      seed: 0