# VoxCPM2 deploy: single-stage AR pipeline with per-request state batching.
# Runs on large (H20 141GB) and small (24GB L4) cards alike: the KV cache is
# pinned to a right-sized budget instead of ballooning to fill the card.
#
# VoxCPM2-local runtime knobs are passed through hf_overrides.
async_chunk: false
dtype: bfloat16

stages:
  - stage_id: 0
    max_num_seqs: 8
    gpu_memory_utilization: 0.9
    # Right-size the KV cache instead of letting it fill gpu_memory_utilization.
    # max_num_seqs(8) * max_model_len(4096) ~= 32k tokens; 6 GiB leaves headroom
    # for longer live requests while keeping VoxCPM2 TTS decode batches small
    # enough to avoid the high ITL observed with large admission caps.
    # The remaining VRAM stays free for the VoxCPM2 diffusion side-path
    # (LocDiT / AudioVAE / code2wav + batch CUDA-graph pools), which allocates
    # outside vLLM's KV accounting during inference. Keeps peak ~13 GiB on any
    # card and avoids per-rebase OOM tuning on small CI GPUs.
    # Assumes prefix caching off and typical live decode concurrency <= ~8;
    # raise this cap if either grows or enable_prefix_caching is turned on.
    kv_cache_memory_bytes: 6442450944  # 6 GiB
    enforce_eager: true
    trust_remote_code: true
    enable_prefix_caching: false
    async_scheduling: true
    max_num_batched_tokens: 4096
    max_model_len: 4096
    devices: "0"
    engine_extras:
      hf_overrides:
        voxcpm2_runtime_config:
          audio_emit_every: 1
          vae_decode_every: 3
          enable_batched_cfm: true
          enable_batched_vae_decode: true
          coalesce_audio_d2h: true
          enable_loc_dit_fused_qkv: true
          enable_loc_dit_fused_mlp: true
          enable_loc_dit_zero_dt_cache: true
          enable_loc_dit_skip_qkv_contig: true
          enable_loc_dit_reduce_overhead_no_cg: true
          enable_unified_decode_graph: true
          unified_decode_graph_max_batch_size: 8
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 4096
      seed: 42
      repetition_penalty: 1.0