GUI-Decisions Space

app.py

@@ -20,6 +20,13 @@ BASE = os.environ.get("BASE", "google/gemma-4-31B-it")
 ADAPTER = os.environ.get("ADAPTER", "infinitylogesh/GUI-Decisions-31B-LoRA")
 SLOT, VANILLA = (37, 99, 235), (234, 88, 12)  # blue: slot reads; orange: the base model writing JSON
 MAX_PX = 3840 * 2160
+WARM = set()  # functions that have run once in this process: the first call includes CUDA warm-up
+
+
+def warm_note(name):
+    first = name not in WARM
+    WARM.add(name)
+    return " The first request after a restart includes GPU warm-up; run it again for a warm time." if first else ""
 
 m = LocalSlotModel.from_pretrained(BASE, adapter=ADAPTER, processor=ADAPTER, device_map="cpu")
 m.model.to("cuda")
@@ -93,7 +100,7 @@ def decide(image, instruction, previous, compare):
 
     md = (f"### Slots: `{describe(a, d, text)}`\n\n"
           f"One forward pass: **{ms:.0f} ms** on this GPU (bf16, transformers"
-          + (", plus the base model writing the text" if text is not None else "") + ").")
+          + (", plus the base model writing the text" if text is not None else "") + ")." + warm_note("decide"))
     if compare:
         va, vpt, vargs, vtext, raw, vms = m.vanilla_decide(im, instruction, user_prev)
         if vpt is not None:
@@ -123,7 +130,7 @@ def box(image, description):
     w = unwarp(float(np.argmax(probs(lp["w"])) + 0.5) / m.arm.bins, "log")
     md = (f"### Box: `[{', '.join(f'{v:.3f}' for v in b)}]`\n\n"
           f"Centre and log-scaled size, four slots read in one forward pass: **{ms:.0f} ms** on this GPU. "
-          f"The heatmap is the centre's x × y distribution (most likely width {w:.1%} of the image).")
+          f"The heatmap is the centre's x × y distribution (most likely width {w:.1%} of the image)." + warm_note("box"))
     return out.convert("RGB"), md