Measured notes on Arabic text handling and on measuring video, and the tools and upstream fixes that came out of them. GitHub
Generating a token streams the weights through the memory bus once, so tokens per second is bandwidth divided by bytes per token. On a base M4 mini that is 21.3 tok/s on a 7B, 77% of the theoretical bus. The same arithmetic pins a 128 GB box at 5-6 t/s on a dense 70B while it reaches 66 t/s on a mixture-of-experts model.
The check I wanted was melting geometry between frames. Real single-camera footage changes structure three times more than a literal cross-dissolve between two unrelated worlds, so there is no threshold in either direction. What survived is a narrower question: did the move you asked for actually happen?
SentencePiece’s default normalizer mapped two of the three Unicode directional marks to a space and left U+061C alone. Taku Kudo fixed it in three days, because the argument was consistency rather than frequency. The other Google repositories I checked came back clean, and that denominator is the point.
Arabic-Indic digits sat inside pypdf’s right-to-left range, so a page reading ١٢٣٤ extracted as ٤٣٢١. A two-line fix, merged — and a reversed number is more dangerous than reversed prose, because it still looks like a number.
Two of the four Arabic rendering bugs I filed on 8 September are fixed. The maintainer chose a version gate over deleting the call, for a reason that is nowhere in the source: his updater never bumps an already-installed package.
A reversed chart label is fixable. The same recipe run once before a file is written bakes the damage in permanently. 276 datasets readable, 119,517 text fields, one fully corrupted — and the first answer I got was wrong.
A code search finds the recipe in thousands of Python files, but whether it is a bug depends entirely on what draws the text. I read six of them: three were broken, three were correct, and one of those three does not even run.
Undoing reshape+bidi corruption round-trips exactly, until the text contains a lam-alef ligature. Then السلام comes back as السالم — a real word, a different word, one that survives a proofread.
Shaping support is decided when the wheel is built, not by the version number. The same Pillow renders Arabic correctly on one machine and broken on another, and the correct code differs between them.
The most-copied recipe for rendering Arabic in Python reverses your text on matplotlib 3.11, with no error. What happened when I filed it against Pillow and matplotlib.