Fork of https://github.com/xenova/microgpt.js

Optimize Zig port: remove @constCast, f64 inference fast path, readability improvements

- Eliminate all @constCast by changing [][]const *Value to [][]*Value - Add f64 fast-path for inference (doLinearF64, doSoftmaxF64, doRmsnormF64, doGptF64) - Break up dense Mersenne Twister expressions into readable intermediates - Use std.mem.swap in mtShuffle - Inline sum in doLinear to avoid temp allocation per row - Pre-size ArrayLists (tokens, losses) where lengths are known - Shrink Value.gen from u64 to u32


Author moderation Date Commit Parent e21de2f3
+764
1 changed file
+764
1 changed file