mirror of
https://github.com/elder-plinius/OBLITERATUS.git
synced 2026-08-30 06:30:37 +02:00
88 lines
2.5 KiB
Python
88 lines
2.5 KiB
Python
"""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
|