From 7a7dca313750ee75e0f8a2a5442ca5d78e743294 Mon Sep 17 00:00:00 2001 From: historia Date: Wed, 2 Sep 2026 20:46:42 -0400 Subject: feat: reserve 20% vram when running higgs with sglang-omni --- app/backends/sglomni/constants.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'app/backends/sglomni/constants.py') diff --git a/app/backends/sglomni/constants.py b/app/backends/sglomni/constants.py index 97a4f3d..94a07bd 100644 --- a/app/backends/sglomni/constants.py +++ b/app/backends/sglomni/constants.py @@ -15,8 +15,9 @@ UV_PIP_PKG = "uv" # launching Python's, see pythonenv.prepare_env. # The vendored copies of upstream examples/configs/*.yaml (the `sgl-omni -# serve --config` argument). Two catalog models run without one (Higgs, -# ZONOS2) and are launched from --model-path alone. +# serve --config` argument). ZONOS2 is the one catalog model without a +# config (it runs from --model-path alone, with a GPU-conditional bf16 +# fallback config in its catalog entry). CONFIGS_DIR = Path(__file__).resolve().parent / "configs" # The managed server's port when the configured URL names none. -- cgit v1.2.3