# VoxCPM2 deploy: single-stage AR pipeline with per-request state batching. # Verified on 1x H20 141GB. # # VoxCPM2-local runtime knobs are passed through hf_overrides. async_chunk: false dtype: bfloat16 stages: - stage_id: 0 max_num_seqs: 32 gpu_memory_utilization: 0.9 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 default_sampling_params: temperature: 0.0 top_p: 1.0 top_k: -1 max_tokens: 4096 seed: 42 repetition_penalty: 1.0