Size the whole-frame profile for a 6 GB card

TensorRT plans its memory for the profile's largest shape, and up to a whole 6D frame with Best's border (4608 x 6656) it asked for 4.9-5.9 GB and would not build on the RTX 3050. The profile now ends at 4608 x 3328 (15 MP), tuned for 4160 x 3248, and the tiler cuts a 6D frame into two such tiles: 27 MP of work for 20 MP kept, against 49 MP in 1408 tiles. The engine's directory names the profile, so a later range never loads an engine built for this one.
2026-10-06 23:14:10 -04:00
parent 3761281dd1
commit 1e1aa1442b
3 changed files with 30 additions and 8 deletions
@@ -619,6 +619,21 @@ mod tests {
        let squares = plan(3648, 5472, 256, Sizes::Square(1408)).unwrap();
        assert_eq!(squares.grid, (5, 7));
        assert!(one.work() * 2 < squares.work());
+        // The whole-frame engine's limit on a 6 GB card: two tiles, each
+        // within it, and still under half the work of the 1408 squares.
+        let halves = plan(
+            3648,
+            5472,
+            256,
+            Sizes::Any {
+                align: 16,
+                max: (4608, 3328),
+            },
+        )
+        .unwrap();
+        assert_eq!(halves.grid, (1, 2));
+        assert_eq!((halves.rows, halves.cols), (4160, 3248));
+        assert!(halves.work() * 2 < squares.work());
        // A limit no tile fits under.
        assert!(plan(
            3648,