boatmanandYueqian Lin
4f9b65074d
Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com> Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
100 lines
3.4 KiB
YAML
100 lines
3.4 KiB
YAML
# Qwen3-TTS high-concurrency deploy profile.
|
|
#
|
|
# This is intentionally separate from qwen3_tts.yaml. Use it only when running
|
|
# sustained high-concurrency serving on two GPUs, for example:
|
|
#
|
|
# vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-Base --omni \
|
|
# --deploy-config vllm_omni/deploy/qwen3_tts_high_concurrency.yaml
|
|
#
|
|
# Profile validated for the c64 / PROMPTS=512 performance experiments:
|
|
# Stage 0 talker on GPU 0 with S0=64, Stage 1 Code2Wav on GPU 1 with S1=10.
|
|
#
|
|
async_chunk: true
|
|
|
|
connectors:
|
|
connector_of_shared_memory:
|
|
name: SharedMemoryConnector
|
|
extra:
|
|
codec_streaming: true
|
|
connector_get_sleep_s: 0.01
|
|
connector_get_max_wait_first_chunk: 3000
|
|
connector_get_max_wait: 300
|
|
codec_chunk_frames: 25
|
|
codec_left_context_frames: 72
|
|
# Stage0 code-predictor prefix CUDA graphs for the c64 hot path.
|
|
# These keys are consumed by Qwen3-TTS talker and ignored by Code2Wav.
|
|
code_predictor_prefix_graphs: true
|
|
code_predictor_prefix_graph_buckets: [64]
|
|
code_predictor_prefix_graph_seq_lens: [2, 3, 4, 5, 6, 7, 8]
|
|
# Keep voice-clone reference context bounded so Stage1 chunk lengths are
|
|
# stable across different reference-audio durations.
|
|
ref_code_context_frames: 72
|
|
# Emit only the first audio chunk early, then return to codec_chunk_frames.
|
|
initial_codec_chunk_frames: 1
|
|
# Chunk ramp (opt-in): gradual size increase for first N chunks to
|
|
# eliminate the IC→steady cliff (1→25). Uncomment to enable.
|
|
# Each entry is the size of that chunk index; indices past the table
|
|
# fall back to codec_chunk_frames. When enabled, ramp overrides
|
|
# initial_codec_chunk_frames (including the per-request API field).
|
|
# Example:
|
|
# codec_chunk_ramp: [4, 4, 8, 16, 25]
|
|
# async_chunk derives stateful ICL/xvec graph shapes from the ramp when
|
|
# enabled, otherwise from the fixed chunk settings above, and ignores
|
|
# decode_cudagraph_capture_sizes. Keep B>1
|
|
# captures opt-in; c64 e2e validation did not show a stable win.
|
|
decode_cudagraph_batch_sizes: [1]
|
|
decode_batch_max_size: 1
|
|
decode_enable_tf32: false
|
|
|
|
stages:
|
|
- stage_id: 0
|
|
max_num_seqs: 64
|
|
gpu_memory_utilization: 0.3
|
|
trust_remote_code: true
|
|
enable_prefix_caching: false
|
|
async_scheduling: true
|
|
max_num_batched_tokens: 512
|
|
max_model_len: 4096
|
|
devices: "0"
|
|
output_connectors:
|
|
to_stage_1: connector_of_shared_memory
|
|
default_sampling_params:
|
|
temperature: 0.9
|
|
top_k: 50
|
|
max_tokens: 4096
|
|
# Official qwen-tts generates the talker with min_new_tokens=2; without
|
|
# it, checkpoints that sample codec EOS on the first frame return empty
|
|
# audio (HTTP 400). See #4962.
|
|
min_tokens: 2
|
|
repetition_penalty: 1.05
|
|
subtalker_sampling_params:
|
|
do_sample: true
|
|
temperature: 0.9
|
|
top_k: 50
|
|
top_p: 1.0
|
|
|
|
- stage_id: 1
|
|
max_num_seqs: 10
|
|
gpu_memory_utilization: 0.3
|
|
enforce_eager: true
|
|
trust_remote_code: true
|
|
enable_prefix_caching: false
|
|
async_scheduling: true
|
|
max_num_batched_tokens: 65536
|
|
max_model_len: 65536
|
|
devices: "1"
|
|
input_connectors:
|
|
from_stage_0: connector_of_shared_memory
|
|
default_sampling_params:
|
|
temperature: 0.0
|
|
top_p: 1.0
|
|
top_k: -1
|
|
max_tokens: 65536
|
|
repetition_penalty: 1.0
|
|
|
|
platforms:
|
|
npu:
|
|
stages:
|
|
- stage_id: 0
|
|
enforce_eager: true
|