# DGX-Spark-Variante von vllm_omni/deploy/qwen3_tts.yaml (vllm-omni 0.30.0):
#   * model_runner v1 statt v2 -- MRV2 importiert RoutedExpertsTensors, das im
#     vLLM-dev-Build des Images (vllm-node-030-omni) fehlt -> ImportError.
#   * gpu_memory_utilization 0.08 / 0.06 statt 2x 0.3 (Unified Memory ist geteilt).
#   * max_num_seqs 8 statt 64.
#
# Qwen3-TTS deploy: talker → code2wav via shared-memory chunk streaming.
# Verified on 1x H100.
#
# Runner selection:
#   Model Runner V2 (MRV2) is EXPERIMENTAL and is the default runner on CUDA
#   for this model. Set `model_runner: v1` to opt out. The `platforms:`
#   sections at the end of this file keep V1 on NPU, XPU, ROCm and MUSA.
#   See docs/configuration/stage_configs.md ("Qwen3-TTS with Model Runner V2").
#
# Fields omitted from a stage fall back to StageDeployConfig defaults (see
# vllm_omni/config/stage_config.py). Notable choices for this model:
#   * Stage 0 (talker) sets max_num_batched_tokens=512 for async-chunk
#     latency tuning (not correctness) — small per-step batches keep
#     first-chunk latency low, and the MRV2 profiles rely on that bound.
#   * Stage 1 (code2wav) sets max_num_batched_tokens=65536 for correctness:
#     codec prefill length (Q * num_frames) exceeds the 32k default.
#   * Stage 0 does not set enforce_eager — talker runs cudagraph by default.
#     Stage 1 defaults to false so Code2Wav can use its inner cudagraph;
#     set true to disable that path in eager mode.
model_runner: v1
async_chunk: true

connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      codec_streaming: true
      connector_get_sleep_s: 0.01
      connector_get_max_wait_first_chunk: 3000
      connector_get_max_wait: 300
      # Must match the decoder sliding attention window.
      codec_chunk_frames: 25
      codec_left_context_frames: 72
      # Emit only the first audio chunk early, then return to codec_chunk_frames.
      initial_codec_chunk_frames: 1
      # Chunk ramp (opt-in): gradual size increase for first N chunks to
      # eliminate the IC→steady cliff (1→25). Uncomment to enable.
      # Each entry is the size of that chunk index; indices past the table
      # fall back to codec_chunk_frames. When enabled, ramp overrides
      # initial_codec_chunk_frames (including the per-request API field).
      # Example:
      #   codec_chunk_ramp: [4, 4, 8, 16, 25]
      # Adaptive chunk ramp (opt-in, takes precedence over fixed codec_chunk_ramp):
      # Chunk 0 uses dynamic IC; chunk 1+ sizes from live buffer-depth feedback.
      # Self-tunes across hardware (NPU/NVIDIA) and load conditions.
      # codec_chunk_adaptive: true
      # codec_chunk_min_frames: 2
      # Code2Wav decode + transfer margin
      # codec_chunk_safety_margin_ms: 50
      # EWMA->delta divisor (delta=max(delta_min, int(ewma/divisor))); higher=gentler ramp
      # codec_chunk_ramp_divisor: 15
      # min chunk-size growth per emit
      # codec_chunk_ramp_delta_min: 2
      # Decoder CUDA Graph sizing:
      #   * async_chunk=true derives all stateful ICL/xvec shapes from
      #     codec_chunk_ramp when enabled, otherwise from
      #     initial_codec_chunk_frames, codec_chunk_frames, and sliding_window;
      #     decode_cudagraph_capture_sizes is ignored in this mode.
      #   * async_chunk=false captures stateless shapes 150 and
      #     decode_chunk_frames + decode_left_context_frames (325 by default).
      #     Extra stateless shapes may be added (union, not intersection) with:
      #       decode_cudagraph_capture_sizes: [97, 400]
      # Graph batch size defaults to 1. Add larger buckets only when needed:
      #   decode_cudagraph_batch_sizes: [1, 2, 4]
      # Limit eager batches to reduce temporary stateful cache allocations; CUDA Graph batching uses the buckets above.
      decode_batch_max_size: 4

stages:
  - stage_id: 0
    max_num_seqs: 8
    gpu_memory_utilization: 0.08
    trust_remote_code: true
    # Omni prefix cache (reuse of talker codec rows on shared prefixes). Off
    # by default; set true when text repeats across requests (same voice or
    # shared ref_audio). The talker opts out of hidden-state caching and
    # defers codes.audio writes, so per-step cost is small. Stage 1 (codec
    # decoder) is not a supported prefix-cache stage: keep it false there.
    # See docs/design/feature/prefix_caching.md for the stage matrix.
    enable_prefix_caching: false
    async_scheduling: true
    # Bound per-step prefill work (validated MRV2 basic profile). Small
    # per-step batches keep first-chunk latency low.
    max_num_batched_tokens: 512
    max_model_len: 4096
    devices: "0"
    output_connectors:
      to_stage_1: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.9
      top_k: 50
      max_tokens: 4096
      # Official qwen-tts generates the talker with min_new_tokens=2; without
      # it, checkpoints that sample codec EOS on the first frame return empty
      # audio (HTTP 400). See #4962.
      min_tokens: 2
      repetition_penalty: 1.05
    # Suppress codebook-0 silence tokens for the first N decode frames of a
    # request (#4966). Applies to every request on this stage, so intended
    # for x-vector-only (Base voice-clone) deploys. 0 = off.
    silence_ban_frames: 0
    subtalker_sampling_params:
      do_sample: true
      temperature: 0.9
      top_k: 50
      top_p: 1.0

  - stage_id: 1
    # Decoder compute dtype on GPU/NPU; 310P requires FP16 and overrides this.
    dtype: bfloat16
    # Tuned for Base voice clone; CustomVoice / VoiceDesign are TTFA-optimal at 1.
    max_num_seqs: 8
    gpu_memory_utilization: 0.06
    enforce_eager: false
    trust_remote_code: true
    enable_prefix_caching: false
    async_scheduling: true
    # Must be divisible by num_code_groups and cover (left_context + chunk).
    # Prefill length is Q * num_frames (e.g. 16 * 2148 = 34368); keep
    # headroom past 32k.
    max_num_batched_tokens: 65536
    # async_chunk appends windows per step; max_model_len must cover the
    # accumulated flat codec stream.
    max_model_len: 65536
    devices: "0"
    input_connectors:
      from_stage_0: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 65536
      repetition_penalty: 1.0

platforms:
  rocm:
    # MRV2 is not qualified on ROCm; keep the legacy V1 runner.
    model_runner: v1
    stages:
      - stage_id: 1
        # MIOpen Conv1d cannot be invoked while vLLM captures the outer
        # Code2Wav model graph. Keep CUDA's graph-enabled default, but run
        # Stage 1 eager on ROCm so decoder convolutions execute normally.
        enforce_eager: true
  npu:
    model_runner: v1
    stages:
      - stage_id: 0
        enforce_eager: false
        compilation_config:
          cudagraph_mode: PIECEWISE
      - stage_id: 1
        compilation_config:
          cudagraph_mode: PIECEWISE
  xpu:
    model_runner: v1
  musa:
    model_runner: v1