5014fea4a0
Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com> Signed-off-by: Yueqian Lin <linyueqian@outlook.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Yueqian Lin <linyueqian@outlook.com> Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
70 lines
2.4 KiB
YAML
70 lines
2.4 KiB
YAML
# Audex (Nemotron-Labs-Audex-2B) text-to-audio deploy: TTA thinker → XCodec1.
|
|
#
|
|
# Notable choices:
|
|
# * Sync full-payload (async_chunk: false): XCodec1 is a CNN codec the
|
|
# official flow decodes over the full sequence; there is no streaming
|
|
# reference and chunked decode risks boundary artifacts.
|
|
# * Stage 0 sampling defaults follow the official TTA script (temperature
|
|
# 1.0, top_k 80; stop on <audiogen_end>; codec cap 4000 tokens = 1000
|
|
# frames generated, decode truncates at 500 frames ≈ 10 s).
|
|
# * Classifier-free guidance is effectively mandatory for TTA quality; the
|
|
# official default is cfg_scale 3.0. The RVQ phase mask is required for
|
|
# validity — both logits processors are registered here and driven per
|
|
# request via extra_args (cfg_* and tta_rvq).
|
|
# * Stage 1's model is the EXTERNAL hf-audio XCodec1 checkpoint (override
|
|
# with a local path here or via XCODEC1_PATH); its tokenizer is borrowed
|
|
# from the thinker checkpoint automatically.
|
|
pipeline: audex_tta
|
|
async_chunk: false
|
|
|
|
connectors:
|
|
connector_of_shared_memory:
|
|
name: SharedMemoryConnector
|
|
extra:
|
|
shm_threshold_bytes: 65536
|
|
connector_get_sleep_s: 0.01
|
|
connector_get_max_wait_first_chunk: 3000
|
|
connector_get_max_wait: 300
|
|
audiocodec_token_offset: 196613
|
|
audiocodec_vocab_size: 8192
|
|
|
|
stages:
|
|
- stage_id: 0
|
|
max_num_seqs: 8
|
|
gpu_memory_utilization: 0.4
|
|
trust_remote_code: true
|
|
enable_prefix_caching: true
|
|
async_scheduling: true
|
|
max_num_batched_tokens: 512
|
|
max_model_len: 8192
|
|
dtype: bfloat16
|
|
devices: "0"
|
|
logits_processors:
|
|
- "vllm_omni.model_executor.models.audex.cfg:AudexCFGLogitsProcessor"
|
|
- "vllm_omni.model_executor.models.audex.tta:TTARVQPhaseMaskLogitsProcessor"
|
|
output_connectors:
|
|
to_stage_1: connector_of_shared_memory
|
|
default_sampling_params:
|
|
temperature: 1.0
|
|
top_k: 80
|
|
max_tokens: 4200
|
|
seed: 0
|
|
|
|
- stage_id: 1
|
|
model: hf-audio/xcodec-hubert-general-balanced
|
|
max_num_seqs: 8
|
|
gpu_memory_utilization: 0.25
|
|
enforce_eager: true
|
|
trust_remote_code: true
|
|
enable_prefix_caching: false
|
|
async_scheduling: false
|
|
max_num_batched_tokens: 8192
|
|
max_model_len: 8192
|
|
devices: "0"
|
|
input_connectors:
|
|
from_stage_0: connector_of_shared_memory
|
|
default_sampling_params:
|
|
temperature: 0.0
|
|
max_tokens: 8192
|
|
seed: 0
|