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/status.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'app/backends/sglomni/status.py') diff --git a/app/backends/sglomni/status.py b/app/backends/sglomni/status.py index 099817e..2b77307 100644 --- a/app/backends/sglomni/status.py +++ b/app/backends/sglomni/status.py @@ -27,9 +27,9 @@ def build_spec(entry: ModelEntry) -> ServerSpec: Public because the run preparation (hub) and the CLI's managed-server bootstrap need to boot exactly the model their run selected, which can differ from detect()'s default. The config yaml is the vendored copy - from the catalog; models without one (Higgs, ZONOS2) run from - --model-path alone — except that a GPU the FP8 kernels cannot run - launches ZONOS2's vendored bf16 config instead (launch_config_path). + from the catalog; ZONOS2 alone runs from --model-path alone — except + that a GPU the FP8 kernels cannot run launches its vendored bf16 config + instead (launch_config_path). """ url = config.SGLOMNI_API_URL argv: List[str] = [ -- cgit v1.2.3