Files
vllm-omni/vllm_omni/deploy/cosyvoice3_full_response_high_concurrency_standard.yaml
2026-09-29 15:48:36 +00:00

16 lines
533 B
YAML

# Standard (non-RAS) sampling variant of the C64 throughput profile, used for
# SGLang-parity validation. Launch with
# --hf-overrides '{"cosyvoice3_sampling_mode": "standard"}'.
# Standard sampling needs a real repetition penalty (RAS handles repetition
# itself); these defaults match the SGLang reference.
base_config: cosyvoice3_full_response_high_concurrency.yaml
stages:
- stage_id: 0
default_sampling_params:
max_tokens: 2048
repetition_penalty: 1.21
temperature: 0.7
top_k: 20
top_p: 0.8