# GLM-TTS deploy: 2-stage AR -> DiT flow-matching.
#
# Default mode is async-chunk streaming through SharedMemoryConnector.
# Stage 0 (AR LLM) streams speech token chunks to Stage 1 (DiT + vocoder)
# via the shared-memory connector. Stage 1 runs as LLM_GENERATION type
# to support chunk_transfer_adapter consumption.
#
# Stage 1 keeps vLLM eager scheduling but enables the GLM-TTS bucketed DiT
# CUDA graph wrapper via hf_overrides.use_dit_cuda_graphs. Graph capture falls
# back to eager if a dynamic shape exceeds the captured buckets.
#
# Verified on 1x A40 48GB (both stages share GPU 0).
# Memory footprint ~16.6 GiB peak; fits 24GB cards with the settings below.
async_chunk: true
trust_remote_code: true

connectors:
  glm_tts_shm:
    name: SharedMemoryConnector
    extra:
      codec_streaming: true
      connector_get_sleep_s: 0.002
      connector_get_max_wait_first_chunk: 3000
      connector_get_max_wait: 300
      # Progressive chunk sizes matching original GLM-TTS defaults.
      codec_chunk_frames: [25, 50, 200]
      codec_left_context_frames: 25
      crossfade_sec: 0.1

stages:
  - stage_id: 0
    max_num_seqs: 1
    gpu_memory_utilization: 0.55
    # BF16 for AR decode: memory-bandwidth bound on single-batch decode,
    # bf16 halves HBM reads per step (~1.7x throughput on L4/A40).
    dtype: bfloat16
    enforce_eager: false
    async_scheduling: true
    enable_prefix_caching: false
    max_num_batched_tokens: 512
    max_model_len: 2048
    devices: "0"
    output_connectors:
      to_stage_1: glm_tts_shm
    default_sampling_params:
      temperature: 1.0
      top_k: 25
      top_p: 0.8
      max_tokens: 2048
      min_tokens: 1
      seed: 42
      detokenize: false
    hf_overrides:
      sample_method: ras
      ras_top_p: 0.8
      ras_top_k: 25
      ras_win_size: 10
      ras_tau_r: 0.1
      min_token_text_ratio: 2.0
      max_token_text_ratio: 20.0
    skip_mm_profiling: true

  - stage_id: 1
    max_num_seqs: 1
    gpu_memory_utilization: 0.25
    enforce_eager: true
    max_model_len: 8192
    max_num_batched_tokens: 2048
    # Force float32 for the flow-matching DiT stage.  The Euler ODE solver
    # (10 steps × 2 forward passes each) accumulates rounding errors in half
    # precision that manifest as audible hiss on H20/H100.  float32 matches
    # the reference implementation and eliminates the noise floor.
    dtype: float32
    devices: "0"
    enable_prefix_caching: false
    input_connectors:
      from_stage_0: glm_tts_shm
    default_sampling_params:
      max_tokens: 2048
    hf_overrides:
      use_dit_cuda_graphs: true
    disable_hybrid_kv_cache_manager: true
    skip_mm_profiling: true