Files
vllm-omni/vllm_omni/deploy/cosyvoice3_packed_streaming.yaml
T
2026-09-29 15:48:36 +00:00

19 lines
705 B
YAML

# Opt-in single-Hopper streaming profile: AR64 / codec16, Flow10.
# Request stream=true to measure first audible audio; non-stream latency is not TTFP.
# Keeps each request's HiFT mel/phase state on GPU and uses mainline GPU FP32 F0.
# Shared serving settings come from the full-response profile; only chunk
# boundaries, codec admission capacity, and stage environment differ.
base_config: cosyvoice3_full_response.yaml
connectors:
connector_of_shared_memory:
extra:
codec_chunk_frames: 25
stages:
- stage_id: 1
env:
COSYVOICE3_TRT: "0"
COSYVOICE3_BATCH_FLOW: "1"
COSYVOICE3_FULL_RESPONSE_OPTIMIZATIONS: "0"
COSYVOICE3_PACKED_STREAMING: "1"
max_num_seqs: 16