fix(gpu): permit bounded CUDA context residue

obliteratus/benchmark_lifecycle.py

@@ -7,6 +7,12 @@ import gc
 from obliteratus.gpu_lifecycle import measure_torch_memory
 
 
+# PyTorch keeps a small allocator/driver context alive in a long-running UI
+# process after every model tensor has been freed.  The root supervisor remains
+# authoritative and independently refuses release above 2 GiB of process VRAM.
+MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES = 128 * 1024 * 1024
+
+
 def admit_benchmark(lifecycle, model_id: str) -> Exception | None:
     """Request admission before a benchmark worker may allocate CUDA memory."""
     try:
@@ -42,7 +48,10 @@ def release_benchmark_pipeline(
         torch_module.cuda.synchronize()
     device_module.empty_cache()
     memory = measure_torch_memory(torch_module)
-    if memory.allocated_bytes or memory.reserved_bytes:
+    if (
+        memory.allocated_bytes > MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES
+        or memory.reserved_bytes > MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES
+    ):
         lifecycle.resize(memory)
         raise RuntimeError(
             "benchmark CUDA allocations remain after cleanup; retaining GPU lease"

tests/test_app_benchmark_lifecycle.py

@@ -6,6 +6,7 @@ from types import SimpleNamespace
 import pytest
 
 from obliteratus.benchmark_lifecycle import (
+    MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES,
     admit_benchmark,
     mark_benchmark_ready,
     release_benchmark_pipeline,
@@ -99,9 +100,36 @@ def test_cleanup_releases_only_after_cuda_is_gone(monkeypatch):
     assert lifecycle.events == [("release", "benchmark_complete")]
 
 
+def test_cleanup_releases_with_bounded_cuda_context_residue(monkeypatch):
+    lifecycle = _LifecycleRecorder()
+    memory = MemoryUsage(
+        allocated_bytes=25_559_040,
+        reserved_bytes=62_914_560,
+        device_count=3,
+    )
+    assert memory.reserved_bytes < MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES
+    monkeypatch.setattr(
+        "obliteratus.benchmark_lifecycle.measure_torch_memory", lambda _torch: memory
+    )
+
+    release_benchmark_pipeline(
+        [SimpleNamespace(handle=SimpleNamespace(model=object(), tokenizer=object()))],
+        reason="benchmark_complete",
+        lifecycle=lifecycle,
+        torch_module=SimpleNamespace(cuda=SimpleNamespace(is_available=lambda: False)),
+        device_module=SimpleNamespace(empty_cache=lambda: None),
+    )
+
+    assert lifecycle.events == [("release", "benchmark_complete")]
+
+
 def test_cleanup_retains_lease_when_cuda_remains(monkeypatch):
     lifecycle = _LifecycleRecorder()
-    memory = MemoryUsage(allocated_bytes=1, reserved_bytes=2, device_count=1)
+    memory = MemoryUsage(
+        allocated_bytes=MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES + 1,
+        reserved_bytes=MAX_RELEASABLE_ALLOCATOR_RESIDUE_BYTES + 1,
+        device_count=1,
+    )
     monkeypatch.setattr(
         "obliteratus.benchmark_lifecycle.measure_torch_memory", lambda _torch: memory
     )