38 lines
1.2 KiB
YAML
38 lines
1.2 KiB
YAML
# Stage config for Step-Audio2 CI tests (offline inference, thinker-only).
|
|
# Single Thinker stage (audio/text understanding → text). ASR/text-output
|
|
# tests don't need Token2Wav; a two-stage config would deadlock on stage-1 skip.
|
|
|
|
# The following config has been verified on 2x 24GB GPU (L4/RTX3090/RTX4090).
|
|
stage_args:
|
|
- stage_id: 0
|
|
runtime:
|
|
process: true
|
|
devices: "0"
|
|
max_batch_size: 1
|
|
engine_args:
|
|
model_stage: thinker
|
|
model_arch: StepAudio2ThinkerForConditionalGeneration
|
|
tensor_parallel_size: 1
|
|
worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
|
|
scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
|
|
max_model_len: 4096
|
|
max_num_batched_tokens: 4096
|
|
max_num_seqs: 1
|
|
gpu_memory_utilization: 0.8
|
|
skip_mm_profiling: true
|
|
enforce_eager: true
|
|
trust_remote_code: true
|
|
enable_prefix_caching: false
|
|
engine_output_type: text # Output text + audio tokens
|
|
is_comprehension: true
|
|
final_output: true
|
|
final_output_type: text
|
|
default_sampling_params:
|
|
temperature: 0.7
|
|
top_p: 0.9
|
|
top_k: -1
|
|
max_tokens: 256
|
|
seed: 42
|
|
detokenize: true
|
|
repetition_penalty: 1.05
|