blob: d2c42f48b3fe4db0f7372e1214141d4ebc528c83 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
|
config_cls: DotsTTSPipelineConfig
model_path: dots-studio/dots.tts-mf
# The solver settings feed two consumers: the latent AR engine solves with
# them, and preprocessing sizes its generation schedule from them. The
# selector writes one value into both stages; an explicit per-stage entry
# under stages: would override it.
shared:
- select:
stages: [preprocessing, latent_engine]
factory:
num_steps: 4
max_generate_length: 500
stages:
reference_encode:
factory:
max_concurrency: 8
max_batch_size: 1
max_batch_wait_ms: 4
latent_engine:
factory:
optimize: true
engine:
mem_fraction_static: 0.20
max_running_requests: 16
# note (luojiaxuan): backbone decode runs through the SGLang CUDA graph
# with the model-owned feedback buffer; delete these two lines (or set
# disable_cuda_graph: true) to fall back to eager backbone decode.
disable_cuda_graph: false
cuda_graph_max_bs: 16
vocoder:
factory:
optimize: true
max_batch_size: 4
max_batch_wait_ms: 2
|
