fix(ui): show VRAM for every visible GPU

tests/test_app_vram.py

@@ -0,0 +1,87 @@
+"""Deterministic contracts for the Gradio accelerator-memory display."""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+
+
+def test_vram_html_covers_cpu_single_and_multi_accelerator_topologies():
+    """Exercise app state away from Gradio's import-time worker sockets."""
+    script = r'''
+from types import SimpleNamespace
+
+import app
+
+
+def memory(used, reserved, total, name):
+    return SimpleNamespace(
+        used_gb=used,
+        reserved_gb=reserved,
+        total_gb=total,
+        device_name=name,
+    )
+
+
+# Zero accelerators preserves the CPU-only message.
+app.dev.is_gpu_available = lambda: False
+html = app._get_vram_html()
+assert "CPU ONLY — NO GPU DETECTED" in html
+assert "data-device-index" not in html
+
+# A single non-CUDA accelerator preserves one unified-memory row.
+app.dev.is_gpu_available = lambda: True
+app.dev.is_cuda = lambda: False
+app.dev.get_memory_info = lambda _index=0: memory(4, 0, 16, "Apple M3 (MPS)")
+html = app._get_vram_html()
+assert html.count("data-device-index=") == 1
+assert 'data-device-index="single"' in html
+assert "Apple M3 (MPS)" in html
+assert "4.0 / 16.0 GB (25%)" in html
+assert "Automatic sharding" not in html
+
+# Homogeneous CUDA devices each receive an indexed row and query.
+memories = [
+    memory(4, 5, 80, "NVIDIA A100"),
+    memory(16, 20, 80, "NVIDIA A100"),
+    memory(72, 74, 80, "NVIDIA A100"),
+]
+queried = []
+app.dev.is_cuda = lambda: True
+app.dev.device_count = lambda: len(memories)
+
+def memory_info(index):
+    queried.append(index)
+    return memories[index]
+
+app.dev.get_memory_info = memory_info
+html = app._get_vram_html()
+assert queried == [0, 1, 2]
+assert html.count("data-device-index=") == 3
+assert "GPU 0 · NVIDIA A100" in html
+assert "GPU 1 · NVIDIA A100" in html
+assert "GPU 2 · NVIDIA A100" in html
+assert "Automatic sharding uses additional GPUs" in html
+
+# Heterogeneous device names stay aligned with their CUDA indices.
+memories = [
+    memory(2, 3, 24, "NVIDIA RTX 4090"),
+    memory(8, 9, 80, "NVIDIA A100"),
+]
+app.dev.device_count = lambda: len(memories)
+app.dev.get_memory_info = memories.__getitem__
+html = app._get_vram_html()
+assert "GPU 0 · NVIDIA RTX 4090" in html
+assert "GPU 1 · NVIDIA A100" in html
+assert html.index("GPU 0 · NVIDIA RTX 4090") < html.index("GPU 1 · NVIDIA A100")
+'''
+
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        timeout=120,
+        check=False,
+    )
+
+    assert result.returncode == 0, result.stdout + result.stderr

tests/test_device_boundaries.py

@@ -45,9 +45,11 @@ def test_explicit_device_validation(monkeypatch):
 
 def test_names_and_device_counts(monkeypatch):
     monkeypatch.setattr(device, "is_cuda", lambda: True)
-    monkeypatch.setattr(device.torch.cuda, "get_device_name", lambda _index: "Test GPU")
+    names = ["Test GPU 0", "Test GPU 1", "Test GPU 2", "Test GPU 3"]
+    monkeypatch.setattr(device.torch.cuda, "get_device_name", names.__getitem__)
     monkeypatch.setattr(device.torch.cuda, "device_count", lambda: 4)
-    assert device.get_device_name() == "Test GPU"
+    assert device.get_device_name() == "Test GPU 0"
+    assert device.get_device_name(3) == "Test GPU 3"
     assert device.device_count() == 4
 
     monkeypatch.setattr(device, "is_cuda", lambda: False)
@@ -88,11 +90,18 @@ def test_system_memory_sources_and_fallback(monkeypatch):
 def test_memory_info_for_cuda_and_cuda_fallback(monkeypatch):
     gib = 1024**3
     monkeypatch.setattr(device, "is_cuda", lambda: True)
-    monkeypatch.setattr(device, "get_device_name", lambda: "GPU")
+    queried_names = []
+
+    def device_name(index=0):
+        queried_names.append(index)
+        return f"GPU {index}"
+
+    monkeypatch.setattr(device, "get_device_name", device_name)
     monkeypatch.setattr(device.torch.cuda, "mem_get_info", lambda _index: (6 * gib, 8 * gib))
     monkeypatch.setattr(device.torch.cuda, "memory_allocated", lambda _index: 1 * gib)
     monkeypatch.setattr(device.torch.cuda, "memory_reserved", lambda _index: 2 * gib)
-    assert device.get_memory_info(2) == device.MemoryInfo(1, 2, 8, 6, "GPU")
+    assert device.get_memory_info(2) == device.MemoryInfo(1, 2, 8, 6, "GPU 2")
+    assert queried_names == [2]
 
     monkeypatch.setattr(device.torch.cuda, "mem_get_info", Mock(side_effect=RuntimeError("unsupported")))
     monkeypatch.setattr(
@@ -100,7 +109,11 @@ def test_memory_info_for_cuda_and_cuda_fallback(monkeypatch):
         "get_device_properties",
         lambda _index: SimpleNamespace(total_memory=10 * gib),
     )
-    assert device.get_memory_info(2) == device.MemoryInfo(total_gb=10, free_gb=10, device_name="GPU")
+    assert device.get_memory_info(2) == device.MemoryInfo(
+        total_gb=10,
+        free_gb=10,
+        device_name="GPU 2",
+    )
 
 
 def test_memory_info_for_mps_cpu_and_total_free(monkeypatch):