55 lines
2.3 KiB
YAML
55 lines
2.3 KiB
YAML
# Single-stage Qwen3-TTS: the Talker decodes every frame with the stateful
|
|
# streaming codec decoder and returns PCM itself (no Code2Wav stage).
|
|
# reset_prefix_cache(reset_running_requests=True) returns False while requests
|
|
# are running; wait for completion or abort them before resetting the cache.
|
|
pipeline: qwen3_tts_fused
|
|
model_runner: v2
|
|
async_chunk: true
|
|
|
|
stages:
|
|
- stage_id: 0
|
|
devices: "0"
|
|
max_num_seqs: 128
|
|
# Exact conditioned Talker inputs for preemption replay are allocated
|
|
# after KV profiling: 16 MiB per active request at 1.7B/max_model_len=4096
|
|
# (~2 GiB at 128 requests), in addition to codec state and output buffers.
|
|
# Keep headroom when increasing gpu_memory_utilization or capacity.
|
|
gpu_memory_utilization: 0.5
|
|
trust_remote_code: true
|
|
enable_prefix_caching: false
|
|
# Model options that multi-stage deployments put on the connector: a
|
|
# single stage has no connector edge.
|
|
additional_config:
|
|
talker_stream_decode: true
|
|
# Opt in to first-frame delivery on its own queue for lower first-audio
|
|
# latency; leave disabled when prioritizing peak throughput.
|
|
talker_stream_first_audio: false
|
|
# Match the standard two-stage profile's voice-clone context boundary.
|
|
# A shorter context (e.g. 25) can change timbre continuity.
|
|
ref_code_context_frames: 72
|
|
code_predictor_kv_cache: true
|
|
code_predictor_fused_sampling: true
|
|
code_predictor_fused: true
|
|
# Talker decode contexts are short (~100-700 tokens): FA3 split-KV plus
|
|
# its combine kernel costs more than one pass per head.
|
|
attention_config:
|
|
flash_attn_max_num_splits_for_cuda_graph: 1
|
|
async_scheduling: true
|
|
max_num_batched_tokens: 512
|
|
max_model_len: 4096
|
|
compilation_config:
|
|
cudagraph_mode: FULL
|
|
cudagraph_capture_sizes: [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 144, 160, 176, 192, 208, 224, 240, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512]
|
|
max_cudagraph_capture_size: 512
|
|
default_sampling_params:
|
|
temperature: 0.9
|
|
top_k: 50
|
|
max_tokens: 4096
|
|
min_tokens: 2
|
|
repetition_penalty: 1.05
|
|
subtalker_sampling_params:
|
|
do_sample: true
|
|
temperature: 0.9
|
|
top_k: 50
|
|
top_p: 1.0
|