Skip to content

Open-Source Frontier Voice AI

Python

fix: auto-tune per-worker env vars in DP mode

Jianwei Yue6b65abb9bbc
diff --git a/vllm_plugin/scripts/start_server.py b/vllm_plugin/scripts/start_server.py
index 375c29e..19e70d0 100644
--- a/vllm_plugin/scripts/start_server.py
+++ b/vllm_plugin/scripts/start_server.py
@@ -214,6 +214,14 @@ def start_dp_server(model_path: str, frontend_port: int,
f"but only {num_gpus} available"
)
+ # Auto-tune per-worker env vars based on dp size
+ ffmpeg_concurrency = max(
+ 64, int(os.environ.get("VIBEVOICE_FFMPEG_MAX_CONCURRENCY", "64"))
+ )
+ media_threads = max(
+ 8, int(os.environ.get("VLLM_MEDIA_LOADING_THREAD_COUNT", "8"))
+ )
+
_install_nginx()
# Assign internal ports: frontend_port + 100, +101, ...
@@ -228,6 +236,8 @@ def start_dp_server(model_path: str, frontend_port: int,
print(f" GPUs per replica: {gpus_per_replica}")
print(f" Max Num Seqs: {max_num_seqs}")
print(f" Max Model Len: {max_model_len}")
+ print(f" FFmpeg concurrency (per worker): {ffmpeg_concurrency}")
+ print(f" Media loading threads (per worker): {media_threads}")
print(f"{'='*60}\n")
# Write nginx config
@@ -242,6 +252,8 @@ def start_dp_server(model_path: str, frontend_port: int,
env = os.environ.copy()
env["CUDA_VISIBLE_DEVICES"] = gpu_ids
+ env["VIBEVOICE_FFMPEG_MAX_CONCURRENCY"] = str(ffmpeg_concurrency)
+ env["VLLM_MEDIA_LOADING_THREAD_COUNT"] = str(media_threads)
vllm_cmd = _build_vllm_cmd(
model_path, port,