feat(backend): zerogpu duration estimator (clamped 60-180s)

backend.py

@@ -0,0 +1,44 @@
+"""ZImageStudioBackend — wraps the DiffSynth pipeline; applies @spaces.GPU on HF Spaces."""
+from __future__ import annotations
+
+import os
+from typing import Any
+
+# Spaces import is optional — running locally we don't have it.
+try:
+    import spaces  # type: ignore
+except ImportError:
+    spaces = None  # type: ignore[assignment]
+
+
+_BASE_DURATION_S: dict[str, int] = {
+    "t2i":        20,   # fixed setup + decode
+    "controlnet": 30,   # + preprocessor + control patch
+    "upscale":    50,   # + realesrgan pixel-space step
+}
+_PER_STEP_S: dict[tuple[str, str], float] = {
+    ("t2i", "Base"):  2.4,
+    ("t2i", "Turbo"): 1.6,
+    ("controlnet", "Turbo"): 2.0,
+    ("upscale", "Turbo"):    1.6,
+}
+
+
+def duration_for(
+    mode: str,
+    params: dict[str, Any],
+    multiplier: float = 1.0,
+) -> int:
+    """Estimate ZeroGPU duration for a request. Pure function; clamped to [60, 180]."""
+    model = params.get("model", "Turbo")
+    steps = int(params.get("steps") or params.get("refine_steps") or 8)
+    width = int(params.get("width", 1024))
+    height = int(params.get("height", 1024))
+
+    base = _BASE_DURATION_S.get(mode, 30)
+    per_step = _PER_STEP_S.get((mode, model), _PER_STEP_S.get((mode, "Turbo"), 1.6))
+    size_factor = (width * height) / (1024 * 1024)
+    cold_buffer = 15  # CPU→GPU copy on first call after a quiet period
+
+    est = (base + per_step * steps + cold_buffer) * size_factor * multiplier
+    return max(60, min(int(est), 180))

tests/test_backend.py

@@ -0,0 +1,34 @@
+import backend
+
+
+def test_duration_t2i_turbo_is_short():
+    d = backend.duration_for(mode="t2i", params=dict(model="Turbo", steps=8, width=1024, height=1024))
+    assert 60 <= d <= 90
+
+
+def test_duration_t2i_base_is_longer():
+    d = backend.duration_for(mode="t2i", params=dict(model="Base", steps=25, width=1024, height=1024))
+    assert d > 60
+
+
+def test_duration_clamps_at_180():
+    d = backend.duration_for(mode="t2i", params=dict(model="Base", steps=200, width=2048, height=2048))
+    assert d == 180
+
+
+def test_duration_clamps_at_60():
+    d = backend.duration_for(mode="t2i", params=dict(model="Turbo", steps=1, width=256, height=256))
+    assert d == 60
+
+
+def test_duration_multiplier_scales_up():
+    base = backend.duration_for(mode="t2i", params=dict(model="Turbo", steps=8, width=1024, height=1024))
+    retry = backend.duration_for(mode="t2i", params=dict(model="Turbo", steps=8, width=1024, height=1024),
+                                  multiplier=2.0)
+    assert retry > base
+
+
+def test_duration_upscale_has_realesrgan_overhead():
+    t2i = backend.duration_for(mode="t2i", params=dict(model="Turbo", steps=8, width=1024, height=1024))
+    upsc = backend.duration_for(mode="upscale", params=dict(refine_steps=5, width=1024, height=1024))
+    assert upsc > t2i