From 5f78c3431b091ad599ceb225a4a235488770d926 Mon Sep 17 00:00:00 2001 From: sls_0x Date: Mon, 21 Sep 2026 08:47:35 +0300 Subject: [PATCH] Declare the served-rig stub's real (weightless) memory footprint --- tests/_reader_served_rig.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/tests/_reader_served_rig.py b/tests/_reader_served_rig.py index 852b1011..13220691 100644 --- a/tests/_reader_served_rig.py +++ b/tests/_reader_served_rig.py @@ -404,7 +404,12 @@ def register_provider(self) -> str: "'timeout_seconds': 120, 'health_timeout_seconds': 5, 'temperature': 0.0, " "'supports_json_mode': False, 'supports_json_schema': False, 'context_window': 8192, " "'native_ollama_chat': True},\n" - f" metadata={{'input_modalities': {(['text', 'image'] if self.vision else ['text'])!r}}},\n" + f" metadata={{'input_modalities': {(['text', 'image'] if self.vision else ['text'])!r}, " + # The stub carries no weights: without a declared footprint the name-based estimator + # budgets it like an 8B model (7.5 GB), which the router's enforce_hardware_fit pass + # then excludes wholesale on smaller-RAM hosts (measured: the macOS CI runner lost + # every pinned stub lane to model_exceeds_hardware_budget). Declare the truth. + "'ram_budget_gb': 0.5, 'vram_budget_gb': 0.5},\n" " enabled=True,\n" "))\n" "print('registered')\n"