Files
vllm-omni/vllm_omni/deploy/cosyvoice3_full_response.yaml
T
2026-09-29 15:48:36 +00:00

63 lines
1.8 KiB
YAML

# Opt-in single-Hopper full-response throughput profile (not streaming TTFP).
# Stage 1 selects the packed backend through its environment below.
# The 4096-frame internal boundary exceeds this profile's 2048-token output cap.
async_chunk: true
connectors:
connector_of_shared_memory:
extra:
codec_chunk_frames: 4096
codec_left_context_frames: 25
codec_streaming: true
codec_vocab_size: 6561
connector_get_max_wait: 300
connector_get_max_wait_first_chunk: 3000
connector_get_sleep_s: 0.01
name: SharedMemoryConnector
cuda_mps: true
pipeline: cosyvoice3
stages:
- default_sampling_params:
max_tokens: 2048
repetition_penalty: 1.0001
temperature: 1.0
top_k: 25
top_p: 0.8
devices: '0'
disable_hybrid_kv_cache_manager: true
# Prompt embedding rearrangement requires the complete prefill block.
enable_chunked_prefill: false
enable_prefix_caching: false
enforce_eager: false
gpu_memory_utilization: 0.4
max_num_batched_tokens: 32768
max_num_seqs: 64
mm_processor_cache_gb: 0
# The talker runs on Model Runner V2 (both RAS and standard sampling);
# set model_runner: v1 to fall back. The codec stage stays on V1.
model_runner: v2
output_connectors:
to_stage_1: connector_of_shared_memory
skip_mm_profiling: true
stage_id: 0
trust_remote_code: true
- default_sampling_params:
max_tokens: 2048
devices: '0'
disable_hybrid_kv_cache_manager: true
enable_prefix_caching: false
enforce_eager: true
gpu_memory_utilization: 0.25
input_connectors:
from_stage_0: connector_of_shared_memory
max_model_len: 32768
max_num_batched_tokens: 32768
max_num_seqs: 8
skip_mm_profiling: true
stage_id: 1
env:
COSYVOICE3_TRT: "0"
COSYVOICE3_BATCH_FLOW: "1"
COSYVOICE3_FULL_RESPONSE_OPTIMIZATIONS: "1"
COSYVOICE3_PACKED_STREAMING: "0"
trust_remote_code: true