async_chunk: false
trust_remote_code: true
pipeline: indextts2_5
connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      codec_streaming: false
      connector_get_sleep_s: 0.01
      connector_get_max_wait_first_chunk: 5000
      connector_get_max_wait: 500
stages:
  - stage_id: 0
    max_num_seqs: 4
    # Triton matches the native GPT-2 logits for this non-RoPE model.
    # FlashInfer produces materially different semantic-token logits.
    attention_backend: TRITON_ATTN
    gpu_memory_utilization: 0.4
    enable_prefix_caching: false
    enable_chunked_prefill: false
    skip_tokenizer_init: true
    async_scheduling: true
    max_num_batched_tokens: 2560
    max_model_len: 2560
    devices: "0"
    output_connectors:
      to_stage_1: connector_of_shared_memory
    hf_overrides:
      use_gpt_latent: false
    # Stage 0 uses plain vLLM sampling, not the upstream default
    # num_beams=3 beam search. Use upstream num_beams=1 for parity comparisons.
    default_sampling_params:
      temperature: 0.8
      top_p: 0.8
      top_k: 30
      max_tokens: 1500
      repetition_penalty: 10.0
      stop_token_ids: [8193]
      detokenize: false
  - stage_id: 1
    max_num_seqs: 4
    gpu_memory_utilization: 0.4
    enable_prefix_caching: false
    skip_tokenizer_init: true
    enforce_eager: true
    async_scheduling: true
    max_num_batched_tokens: 8192
    max_model_len: 32768
    devices: "0"
    input_connectors:
      from_stage_0: connector_of_shared_memory
    hf_overrides:
      use_gpt_latent: false
      diffusion_steps: 25
      s2mel_dit_bf16: true
      s2mel_cfm_batch_size: 4
      inference_cfg_rate: 0.7
      s2mel_dit_cuda_graph: false
      s2mel_dit_torch_compile: true
      s2mel_dit_torch_compile_mode: default
      s2mel_vocoder_torch_compile: false
    default_sampling_params:
      temperature: 0.0
      max_tokens: 32767