# Step-Audio2 online CI deploy: thinker only (audio/text -> text).
#
# The two-stage topology can deadlock when an audio-to-text request skips the
# Token2Wav stage, so this config explicitly selects the ASR pipeline.
# Keeps the legacy 1x 24GB CI sizing; validated here on 1x H100-80GB.
pipeline: step_audio_2_asr
async_chunk: false
trust_remote_code: true
enable_prefix_caching: false

stages:
  - stage_id: 0
    devices: "0"
    tensor_parallel_size: 1
    max_model_len: 1024
    max_num_batched_tokens: 1024
    max_num_seqs: 1
    gpu_memory_utilization: 0.7
    skip_mm_profiling: true
    enforce_eager: true
    async_scheduling: false
    hf_overrides:
      architectures: ["StepAudio2ThinkerForConditionalGeneration"]
    default_sampling_params:
      temperature: 0.7
      top_p: 0.9
      top_k: -1
      max_tokens: 256
      seed: 42
      detokenize: true
      repetition_penalty: 1.05