Merge pull request #156 from elder-plinius/fix/155-gpu-memory-refresh

fix(ui): refresh GPU memory gauge periodically

app.py

@@ -59,7 +59,12 @@ import gradio as gr
 import torch
 from obliteratus import device as dev
 from obliteratus.credential_sources import resolve_first, resolve_secret, secret_available
-from obliteratus.ui_vram import render_vram_html
+from obliteratus.ui_vram import (
+    DEFAULT_VRAM_REFRESH_INTERVAL,
+    VRAM_REFRESH_CHOICES,
+    render_vram_html,
+    resolve_vram_refresh_interval,
+)
 from obliteratus.model_load_settings import (
     DTYPE_CHOICES,
     QUANTIZATION_CHOICES,
@@ -3957,6 +3962,28 @@ html:not(.dark) body::after { display: none; }
     margin-bottom: 8px;
 }
 
+/* ---- GPU MEMORY MONITOR ---- */
+.vram-monitor-row { align-items: flex-end !important; }
+.vram-display { flex-grow: 1 !important; }
+.vram-refresh-control {
+    max-width: 108px !important;
+    margin-left: auto !important;
+    opacity: 0.72;
+}
+.vram-refresh-control:hover,
+.vram-refresh-control:focus-within { opacity: 1; }
+.vram-refresh-control label span {
+    font-size: 0.62rem !important;
+    letter-spacing: 0.06em !important;
+}
+.vram-refresh-control .wrap { min-height: 30px !important; }
+@media (max-width: 640px) {
+    .vram-refresh-control {
+        max-width: 100px !important;
+        margin-left: auto !important;
+    }
+}
+
 /* ---- TAB STYLING ---- */
 .tabs { border-bottom: 1px solid #1a1f2e !important; }
 button.tab-nav {
@@ -4207,8 +4234,24 @@ with gr.Blocks(title="OBLITERATUS", fill_height=True) as demo:
         </div>
     """)
 
-    # GPU VRAM monitor — refreshed on page load and after key operations
-    vram_display = gr.HTML(value=_get_vram_html())
+    # GPU VRAM monitor — periodically refreshed and updated after key operations
+    with gr.Row(elem_classes=["vram-monitor-row"]):
+        vram_display = gr.HTML(
+            value=_get_vram_html(),
+            scale=12,
+            elem_classes=["vram-display"],
+        )
+        vram_refresh_interval = gr.Dropdown(
+            choices=VRAM_REFRESH_CHOICES,
+            value=DEFAULT_VRAM_REFRESH_INTERVAL,
+            label="GPU refresh",
+            filterable=False,
+            allow_custom_value=False,
+            scale=1,
+            min_width=96,
+            elem_classes=["vram-refresh-control"],
+        )
+    vram_refresh_timer = gr.Timer(value=DEFAULT_VRAM_REFRESH_INTERVAL, active=True)
 
     # ZeroGPU info — only shown when running on HF Spaces with ZeroGPU
     if _ZEROGPU_AVAILABLE:
@@ -5497,6 +5540,25 @@ Built on the shoulders of:
     # Refresh VRAM on page load
     demo.load(fn=_get_vram_html, outputs=[vram_display])
 
+    # Poll independently of long-running queued model work. Updating the one
+    # timer's interval avoids accumulating polling loops when the selector changes.
+    vram_refresh_timer.tick(
+        fn=_get_vram_html,
+        outputs=[vram_display],
+        queue=False,
+        concurrency_limit=1,
+        concurrency_id="vram-refresh",
+        api_visibility="private",
+    )
+    vram_refresh_interval.change(
+        fn=resolve_vram_refresh_interval,
+        inputs=[vram_refresh_interval],
+        outputs=[vram_refresh_timer],
+        queue=False,
+        show_progress="hidden",
+        api_visibility="private",
+    )
+
 
 # ---------------------------------------------------------------------------
 # Launch

obliteratus/ui_vram.py

@@ -5,14 +5,24 @@ from __future__ import annotations
 from typing import Any
 
 
+VRAM_REFRESH_CHOICES = (("10 sec", 10.0), ("30 sec", 30.0), ("1 min", 60.0))
+DEFAULT_VRAM_REFRESH_INTERVAL = 30.0
+
+
+def resolve_vram_refresh_interval(value: float) -> float:
+    """Return a supported VRAM polling interval, falling back to the default."""
+    allowed = {interval for _label, interval in VRAM_REFRESH_CHOICES}
+    return value if value in allowed else DEFAULT_VRAM_REFRESH_INTERVAL
+
+
 def render_vram_html(device: Any) -> str:
     """Return per-device GPU/accelerator memory usage as styled bars."""
-    if not device.is_gpu_available():
-        return (
-            '<div style="text-align:center;color:#4a5568;font-size:0.72rem;'
-            'letter-spacing:1px;margin-top:6px;">CPU ONLY — NO GPU DETECTED</div>'
-        )
     try:
+        if not device.is_gpu_available():
+            return (
+                '<div style="text-align:center;color:#4a5568;font-size:0.72rem;'
+                'letter-spacing:1px;margin-top:6px;">CPU ONLY — NO GPU DETECTED</div>'
+            )
         cuda_count = device.device_count() if device.is_cuda() else 0
         device_indices = range(cuda_count) if cuda_count else (None,)
         rows = []

tests/conditional/test_operator_ui.py

@@ -31,3 +31,72 @@ def test_gradio_application_constructs_without_public_listener():
     assert probe.returncode == 0, probe.stderr
     assert probe.stdout.strip() == "operator UI constructed without a listener"
     assert probe.stderr == ""
+
+
+def test_gpu_memory_refresh_controls_have_one_configurable_polling_loop():
+    """Exercise the rendered Gradio graph rather than relying on source shape."""
+    pytest.importorskip("gradio", reason="install the locked spaces extra")
+    probe = subprocess.run(
+        [
+            sys.executable,
+            "-c",
+            r'''
+import asyncio
+import app
+
+config = app.demo.get_config_file()
+components = config["components"]
+timers = [item for item in components if item["type"] == "timer"]
+assert len(timers) == 1
+timer = timers[0]
+assert timer["props"]["value"] == 30.0
+assert timer["props"]["active"] is True
+
+selectors = [
+    item for item in components
+    if "vram-refresh-control" in item["props"].get("elem_classes", [])
+]
+assert len(selectors) == 1
+selector = selectors[0]
+assert selector["type"] == "dropdown"
+assert selector["props"]["label"] == "GPU refresh"
+assert selector["props"]["choices"] == [
+    ("10 sec", 10.0), ("30 sec", 30.0), ("1 min", 60.0)
+]
+assert selector["props"]["value"] == 30.0
+assert selector["props"]["filterable"] is False
+assert selector["props"]["allow_custom_value"] is False
+assert selector["props"]["min_width"] == 96
+
+timer_ticks = [
+    dependency for dependency in config["dependencies"]
+    if (timer["id"], "tick") in dependency["targets"]
+]
+assert len(timer_ticks) == 1
+assert timer_ticks[0]["queue"] is False
+assert timer_ticks[0]["outputs"] == [app.vram_display._id]
+
+interval_changes = [
+    dependency for dependency in config["dependencies"]
+    if (selector["id"], "change") in dependency["targets"]
+    and dependency["outputs"] == [timer["id"]]
+]
+assert len(interval_changes) == 1
+assert interval_changes[0]["queue"] is False
+
+async def verify_interval_changes():
+    for seconds in (10.0, 30.0, 60.0):
+        result = await app.demo.process_api(
+            interval_changes[0]["id"], [seconds], state=None
+        )
+        assert result["data"] == [seconds]
+
+asyncio.run(verify_interval_changes())
+''',
+        ],
+        capture_output=True,
+        text=True,
+        timeout=120,
+        check=False,
+    )
+    assert probe.returncode == 0, probe.stdout + probe.stderr

tests/test_app_vram.py

@@ -4,7 +4,12 @@ from __future__ import annotations
 
 from types import SimpleNamespace
 
-from obliteratus.ui_vram import render_vram_html
+from obliteratus.ui_vram import (
+    DEFAULT_VRAM_REFRESH_INTERVAL,
+    VRAM_REFRESH_CHOICES,
+    render_vram_html,
+    resolve_vram_refresh_interval,
+)
 
 
 def _memory(used, reserved, total, name):
@@ -79,3 +84,19 @@ def test_vram_html_handles_memory_errors():
         get_memory_info=lambda _index: (_ for _ in ()).throw(RuntimeError("unavailable")),
     )
     assert "Memory: unavailable" in render_vram_html(device)
+
+    device.is_gpu_available = lambda: (_ for _ in ()).throw(RuntimeError("unavailable"))
+    assert "Memory: unavailable" in render_vram_html(device)
+
+
+def test_vram_refresh_interval_contract():
+    assert VRAM_REFRESH_CHOICES == (
+        ("10 sec", 10.0),
+        ("30 sec", 30.0),
+        ("1 min", 60.0),
+    )
+    assert DEFAULT_VRAM_REFRESH_INTERVAL == 30.0
+    assert resolve_vram_refresh_interval(10.0) == 10.0
+    assert resolve_vram_refresh_interval(30.0) == 30.0
+    assert resolve_vram_refresh_interval(60.0) == 60.0
+    assert resolve_vram_refresh_interval(15.0) == 30.0