32 lines
873 B
Python
Executable File
32 lines
873 B
Python
Executable File
#!/usr/bin/env python3
|
|
"""Start the SGLang CLI only after the H3 request-reset patch is installed."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import multiprocessing
|
|
import os
|
|
|
|
from sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.minimax_h3.stages.denoising import (
|
|
MiniMaxH3DenoisingStage,
|
|
)
|
|
from sitecustomize import _install_patch
|
|
|
|
|
|
_install_patch()
|
|
if not getattr(MiniMaxH3DenoisingStage, "_ymrl_cache_reset_patch", False):
|
|
raise RuntimeError("MiniMax-H3 Cache-DiT per-request reset patch did not install")
|
|
|
|
|
|
def launch() -> int:
|
|
if os.getenv("SGLANG_CACHE_PATCH_VERIFY_ONLY") == "1":
|
|
print("MiniMax-H3 Cache-DiT request-reset patch: installed")
|
|
return 0
|
|
from sglang.cli.main import main
|
|
|
|
return int(main() or 0)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
multiprocessing.freeze_support()
|
|
raise SystemExit(launch())
|