# Audex 30B-A3B thinker-only deploy: audio understanding (ASR / audio QA).
# Single-stage NemotronHAudexForConditionalGeneration on
# checkpoint_folder_full. Hybrid Mamba: no prefix caching. Pass this yaml
# explicitly (see audex_tts_30b.yaml).
pipeline: audex_thinker_only
async_chunk: false

stages:
  - stage_id: 0
    max_num_seqs: 8
    gpu_memory_utilization: 0.9
    trust_remote_code: true
    enable_prefix_caching: false
    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