vllm-project--vllm-omni
wehub-resource-sync
47 行
1.5 KiB
YAML
47 行
1.5 KiB
YAML
# Covo-Audio-Chat deploy: fused_thinker_talker → code2wav on a single GPU.
|
|
# Verified on 1x A100-80G.
|
|
#
|
|
# Fields omitted from a stage fall back to StageDeployConfig defaults (see
|
|
# vllm_omni/config/stage_config.py). Non-default choices in this file:
|
|
# * async_chunk: false — pipeline defines no async-chunk processor funcs.
|
|
# * Stage 1 max_num_batched_tokens=2048, max_model_len=2048 — audio code
|
|
# sequences are short; the 32k default is unnecessary here.
|
|
# * Stage 1 dtype: float32 — BigVGAN vocoder weights are fp32; the default
|
|
# bf16 triggers mat1/mat2 dtype mismatch during warmup.
|
|
# * Stage 1 enforce_eager: true — the code2wav forward pass performs a
|
|
# GPU→CPU sync (int(code.flatten()[0]) in covo_audio.py) which is
|
|
# illegal during CUDA graph capture.
|
|
async_chunk: false
|
|
trust_remote_code: true
|
|
|
|
stages:
|
|
- stage_id: 0
|
|
max_num_seqs: 1
|
|
gpu_memory_utilization: 0.5
|
|
mm_processor_cache_gb: 0
|
|
devices: "0"
|
|
default_sampling_params:
|
|
temperature: 0.0
|
|
top_p: 1.0
|
|
top_k: -1
|
|
max_tokens: 2048
|
|
seed: 42
|
|
repetition_penalty: 1.1
|
|
|
|
- stage_id: 1
|
|
max_num_seqs: 1
|
|
gpu_memory_utilization: 0.2
|
|
enforce_eager: true
|
|
dtype: float32
|
|
max_num_batched_tokens: 2048
|
|
max_model_len: 2048
|
|
mm_processor_cache_gb: 0
|
|
devices: "0"
|
|
default_sampling_params:
|
|
temperature: 0.0
|
|
top_p: 1.0
|
|
top_k: -1
|
|
max_tokens: 2048
|
|
seed: 42
|
|
repetition_penalty: 1.1
|