kormann

graph_rewrite sort src

1

improves matching and almost solves symbolic equality

multi src alu uop

1

- [x] cstyle add mul - [x] llvm add mul - [ ] assembly add mul

Uop symbolics rewrite rules

5

- [x] test_lt - [x] test_ge - [x] test_lt_remove - [x] test_ge_remove - [x] test_ge_divides

keep track of linearizer times in process_replay

8

re: https://github.com/tinygrad/tinygrad/pull/5093#issuecomment-2183598699 will show if linearizer delay changes like this: ``` Perf: 3.27 -> 3.64 ns r_256_256 | 0/1 Perf: 1.34 -> 3.18 ns r_2560_2560 100%|█████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00

lstm training slo repro

4

minimal repro of lstm being slow on first forward pass. could be sped up with use of .realize but should also work without as i understand seems to correlate with...

whisper

3

- [x] fix jit error for long input - [x] stream input output - [x] timestamps - [x] long batch inference - [x] (speed up if necessary)

whisper long batch

1

couple conceptual cleanups for whisper inference + support for long batching working towards completely merging #5922

paper Switch 2 typo fix

1

I believe there has been a slight typo. not sure how to compile the pdf though