gpu(titan-24): reduce FLUX peak memory

dockerfiles/hermes-local-image-server.py

@@ -33,6 +33,7 @@ MODEL_REVISION = os.environ.get(
     "HERMES_LOCAL_IMAGE_REVISION", "e7b7dc27f91deacad38e78976d1f2b499d76a294"
 )
 MODEL_CACHE = os.environ.get("HF_HOME", "/models/huggingface")
+OFFLOAD_MODE = os.environ.get("HERMES_LOCAL_IMAGE_OFFLOAD_MODE", "sequential")
 LEASE_NAMESPACE = os.environ.get("LEASE_NAMESPACE", "hermes")
 LEASE_NAME = os.environ.get("LEASE_NAME", "titan-24-gpu-owner")
 LEASE_IDLE_OWNER = os.environ.get("LEASE_IDLE_OWNER", "hermes")
@@ -279,7 +280,17 @@ def _render_in_process(payload: dict[str, Any]) -> dict[str, Any]:
             torch_dtype=torch.bfloat16,
             cache_dir=MODEL_CACHE,
         )
-        pipe.enable_model_cpu_offload()
+        if OFFLOAD_MODE == "sequential":
+            # The shared 10 GiB card keeps the desktop and idle Wolf daemon
+            # resident. Layer-level offload preserves 1024px output quality
+            # without requiring those low-memory services to be torn down.
+            pipe.enable_sequential_cpu_offload()
+        elif OFFLOAD_MODE == "model":
+            pipe.enable_model_cpu_offload()
+        else:
+            raise ValueError("HERMES_LOCAL_IMAGE_OFFLOAD_MODE must be sequential or model")
+        pipe.enable_vae_slicing()
+        pipe.enable_vae_tiling()
         _set_state(phase="rendering")
         arguments: dict[str, Any] = {
             "prompt": prompt,

services/hermes/local-image-deployment.yaml

@@ -21,7 +21,7 @@ spec:
       annotations:
         ai.bstein.dev/model: black-forest-labs/FLUX.2-klein-4B
         ai.bstein.dev/gpu: titan-24 lease-shared image, desktop, and Wolf lane
-        ai.bstein.dev/config-rev: "20260811-flux2-klein-process-isolation"
+        ai.bstein.dev/config-rev: "20260811-flux2-klein-low-vram"
     spec:
       serviceAccountName: hermes-gpu-runtime
       runtimeClassName: nvidia
@@ -43,7 +43,7 @@ spec:
             sizeLimit: 2Gi
       containers:
         - name: local-image
-          image: registry.bstein.dev/bstein/hermes-local-image@sha256:b158e33adab1305f93e3857786ca577b9cf3a4540d556c2a847ba9eb411de0dc
+          image: registry.bstein.dev/bstein/hermes-local-image@sha256:a937f02fe1e85df9d86c0edb1972de6c020fd8ef4f768809624cbebdc521bca1
           imagePullPolicy: IfNotPresent
           ports:
             - name: local-image
@@ -80,6 +80,8 @@ spec:
               value: "1"
             - name: HERMES_LOCAL_IMAGE_RENDER_TIMEOUT
               value: "1200"
+            - name: HERMES_LOCAL_IMAGE_OFFLOAD_MODE
+              value: sequential
             - name: PYTORCH_CUDA_ALLOC_CONF
               value: expandable_segments:True
             - name: HF_HOME

testing/tests/test_hermes_chat_quality.py

@@ -625,6 +625,7 @@ def test_local_flux_runtime_and_gpu_handoff_are_flux_managed():
     assert model_env["HERMES_LOCAL_IMAGE_GPU_MAX_EXTERNAL_MEMORY_BYTES"] == (
         "3221225472"
     )
+    assert model_env["HERMES_LOCAL_IMAGE_OFFLOAD_MODE"] == "sequential"
     assert "nvidia-process-exporter-local.monitoring.svc.cluster.local" in model_env[
         "HERMES_LOCAL_IMAGE_GPU_ACTIVITY_URL"
     ]
@@ -760,6 +761,17 @@ def test_local_flux_renderer_uses_a_disposable_cuda_worker(monkeypatch):
     assert options["check"] is False
 
 
+def test_local_flux_uses_low_vram_offload_without_reducing_resolution():
+    """The shared 3080 lane must trade time, not image size, for headroom."""
+    source = (ROOT / "dockerfiles" / "hermes-local-image-server.py").read_text()
+    assert 'OFFLOAD_MODE = os.environ.get(' in source
+    assert '"HERMES_LOCAL_IMAGE_OFFLOAD_MODE", "sequential"' in source
+    assert "pipe.enable_sequential_cpu_offload()" in source
+    assert "pipe.enable_vae_slicing()" in source
+    assert "pipe.enable_vae_tiling()" in source
+    assert '"square": (1024, 1024)' in source
+
+
 def test_chat_auth_file_mount_survives_atomic_provider_refresh():
     statefulset = _documents(HERMES / "chat-statefulset.yaml")[0]
     containers = statefulset["spec"]["template"]["spec"]["containers"]