fix(server): keep memory free for full-size LightOnOCR-3-4B pages

packages/sie_server/models/lightonai__LightOnOCR-3-4B.yaml

@@ -43,10 +43,15 @@ tasks:
 # two trained prompts accepted. The `transformers` profile runs the same contract in-process in the
 # transformers5 bundle.
 #
-# Throughput: up to 128 pages run in the engine at once. The radix cache is off: every page is a
-# new image, so it saves nothing, and with it on SGLang turns overlap scheduling off for this
-# hybrid (Gated DeltaNet) model. FlashInfer serves the attention layers. Sampling is unchanged and
-# outputs stay within bf16 run-to-run variance.
+# Throughput: up to 128 pages run in the engine at once; on an L4 at this memory reservation SGLang
+# caps it at 64. The radix cache is off: every page is a new image, so it saves nothing, and with it
+# on SGLang turns overlap scheduling off for this hybrid (Gated DeltaNet) model. FlashInfer serves
+# the attention layers. Sampling is unchanged and outputs stay within bf16 run-to-run variance.
+#
+# Memory: SGLang keeps (1 - mem_fraction_static) of the memory free at launch as runtime headroom.
+# One page at the 16 MP limit peaks 4.8 GiB above the memory in use after load, in the vision
+# encoder. On an L4, 0.75 and 0.85 run out of memory there and the engine exits, which fails every
+# page in flight. 0.70 leaves 0.7 GiB free at that peak.
 max_sequence_length: 24576
 profiles:
   default:
@@ -55,7 +60,7 @@ profiles:
     adapter_path: sie_server.adapters.sglang_vision_extract.adapter:SGLangVisionExtractAdapter
     adapter_options:
       loadtime:
-        mem_fraction_static: 0.85
+        mem_fraction_static: 0.70
         served_model_name: lightonai/LightOnOCR-3-4B
         trust_remote_code: false
         max_concurrent_requests: 128
@@ -78,7 +83,7 @@ profiles:
     adapter_options:
       # Profile inheritance replaces loadtime, so retain every engine/sampling option.
       loadtime:
-        mem_fraction_static: 0.85
+        mem_fraction_static: 0.70
         served_model_name: lightonai/LightOnOCR-3-4B
         trust_remote_code: false
         max_concurrent_requests: 128

packages/sie_server/tests/config/test_lightonocr3_profiles.py

@@ -127,6 +127,14 @@ def test_markdown_profile_retains_default_generation_context_and_engine_options(
     assert adapter._extra_launch_args == ["--disable-radix-cache"]
 
 
+def test_sglang_profiles_keep_their_l4_calibrated_reservation() -> None:
+    profiles = yaml.safe_load(_CONFIG.read_text())["profiles"]
+    for name in ("default", "markdown"):
+        loadtime = profiles[name]["adapter_options"]["loadtime"]
+        assert loadtime["mem_fraction_static"] == 0.70
+        assert loadtime["max_concurrent_requests"] == 128
+
+
 def test_default_routes_to_sglang_vision_extract_bundle_only() -> None:
     assert _ADAPTER in yaml.safe_load((_BUNDLES / "transformers5.yaml").read_text())["adapters"]
     for bundle in sorted(_BUNDLES.glob("*.yaml")):

packages/sie_server/tests/config/test_sglang_memory_profiles.py

@@ -128,9 +128,7 @@ def test_extending_profile_forwards_its_resolved_kv_budget():
     assert adapter._memory_budget.kv_budget_tokens == 65536
 
 
-@pytest.mark.parametrize(
-    "name", ["Qwen__Qwen3.5-4B", "lightonai__LightOnOCR-3-4B", "Qwen__Qwen3-VL-8B-Instruct", "zai-org__GLM-OCR"]
-)
+@pytest.mark.parametrize("name", ["Qwen__Qwen3.5-4B", "Qwen__Qwen3-VL-8B-Instruct", "zai-org__GLM-OCR"])
 def test_specialized_profiles_retain_calibrated_reservation(name):
     config = load_model_config(MODELS / f"{name}.yaml")
     assert config.resolve_profile("default").loadtime["mem_fraction_static"] == 0.85