Commit Graph

10 Commits

Author SHA1 Message Date
George Hotz 3169cb386d
remove graph [pr] (#7085) 2024-10-16 11:40:07 +08:00
chenyu 7c8849010a
fix var_vals in MCTS (#6882)
tested with JITBEAM=100 llama
2024-10-04 08:19:35 -04:00
qazal c23d44c779
AST is UOp (#6030)
* most of the work from the uops2 branch

* schedule

* realize

* kernel

* lowerer

* search

* green

* merge uops with ops

* Revert "merge uops with ops"

This reverts commit 1408a59f12c97e3466679884266b247cf9df46bc.

* fix benchmark

* remove extra dedup
2024-08-16 22:09:00 +03:00
George Hotz 6c6d74d922
parallel mcts (#5626)
* start work on parallel mcts

* compile was linearizing twice

* typing + more early stopping

* fix compiler error
2024-07-21 14:53:23 -07:00
George Hotz ef179087a4
mcts exit condition wasn't right, also use it with BEAM>=100 (#5619)
* mcts exit condition wasn't right, also use it with BEAM>=100

* mcts touchups

* clean up sample
2024-07-21 10:16:47 -07:00
George Hotz 0f67ef4674
mcts graph and dedup support (#5618)
* mcts graph and dedup support

* usable graph

* mcts colors

* C=4 seems better

* C=3 even better

* sample_tree

* backprop is external function

* late expand to match algo
2024-07-20 23:29:14 -07:00
chenyu eddc5bcfd7
MCTS tweaks (#5616)
MCTS 500 is competitive with BEAM=8 on resnet on M1 Max.
- increment trial times even with compiled error and runtime error.
- use best time of children as the node value.
2024-07-20 19:45:59 -07:00
George Hotz 1113e47f96 print best in MCTS + light up the winner in hcopt 2024-07-20 09:39:36 -07:00
George Hotz ac99ecd94e
use statistics.median for timing (#5606) 2024-07-20 08:37:32 -07:00
George Hotz 06e336bccb
mcts search (#5598)
* mcts search

* mcts cleanups

* mcts cleanup

* random shuffle children order

* mcts in handcode_opt

* src and remove_node

* debug 3 to print ast

* print the type

* mcts in extra
2024-07-19 21:38:39 -07:00