github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-10 23:48:01 -05:00
Code Issues Packages Projects Releases Wiki Activity
11,682 Commits 203 Branches 13 Tags
master
Commit Graph

116 Commits

Author SHA1 Message Date
George Hotz
e0201922e3 Q network for pruning BEAM / uops deduping / BEAM_ESTIMATE (#2142) 
* stable diffusion < 324ms

* revert swap action

* fix tests due to more sum splitting

* REDUCEOP_SPLIT_THRESHOLD env var

* added from unaligned np test (#2134)

* align cpu buffer before copy into cl buffer (#2135)

* remove shelve from handcode_resnet50_opt.py (#2139)

* Add dictionary keys to reduce db size (#2131)

* work

* ignore beam cache

* dictionary keys are generic

* minor db cleanups

* fix baseline and extract dataset

* fix training

* log likelihood

* more lin to feats

* sts

* training policynet

* net sort of works

* dedup

* refactor, stupid new actions

* fix uops deduping

* BEAM_ESTIMATE

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
Co-authored-by: imaolo <56898718+imaolo@users.noreply.github.com>
 2023-10-27 10:53:06 -10:00
chenyu
0ca0e9ee5e exclude ast with variables from beam search (#2140) 
* exclude ast with variables from beam search

* test that

* add to CI
 2023-10-25 16:35:29 -04:00
George Hotz
cea2bc7964 Add dictionary keys to reduce db size (#2131) 
* work

* ignore beam cache

* dictionary keys are generic

* minor db cleanups

* fix baseline and extract dataset

* fix training

* log likelihood
 2023-10-24 10:49:22 -04:00
George Hotz
abeba8f1fc optimization: get actions in CI (#2125) 
* get actions in CI

* actually run the test

* pythonpath
 2023-10-20 12:22:01 -07:00
George Hotz
c36d306606 KOPT is over, BEAM is upstream (#2071) 
* create cache for q learning

* make linter happy

* global beam

* where it belongs

* bugfix

* ditch the kopt, use the beam

* faster lin and DEBUG=2 okay

* remove kopt, move search to features
 2023-10-16 09:46:03 -07:00
George Hotz
49bcfec383 0s in the action space (#2070) 
* 0s in the action space

* simpler

* skip duplicate actions
 2023-10-14 11:22:48 -07:00
George Hotz
4124cf1df5 cleanup tensor cores, expose exclude local upcast (#2064) 
* expose exclude_local_upcast

* convert apply tensor cores to ops

* update comment

* put LOCAL back to what it was, BEAM is better than way
 2023-10-14 09:21:03 -07:00
George Hotz
90c777d815 remove apply_auto_opt (#2063) 2023-10-13 07:44:14 -07:00
George Hotz
6f1810af2d with unroll, the action space goes from 161 -> 127 (#2060) 
* with unroll, the action space goes from 161 -> 127

* more reliable instrumentation

* beam search is so op

* beam bugfix
 2023-10-12 20:52:23 -07:00
George Hotz
c5edb3c374 train value net, improve API, add BCE (#2047) 
* api cleanups, BCE losses

* valuenet

* fixup examples

* learning okay

* add valuenet runner

* net improvements

* net improvements

* 40% win rate
 2023-10-12 07:56:38 -07:00
George Hotz
0ba629c7b9 add world dataset (#2045) 2023-10-11 15:54:30 -07:00
George Hotz
0c3b6f13a8 Latest opt (#2044) 
* split out actions

* rl algorithm
 2023-10-11 15:46:14 -07:00
George Hotz
41bfeb2c1e start work on auto opt (#2034) 
* start work on auto opt

* lin failure

* not beating hcopt

* greedy

* timing is fast

* codegen.search

* greedy search in handcode_opt

* track running gflops

* clean up those files

* no failure
 2023-10-11 12:54:53 -07:00
chenyu
1c980517c5 s/var_vals_from_ast/vars_from_ast (#2038) 2023-10-10 20:21:55 -07:00
George Hotz
f139060103 Rewrite hand coded opt with action space (#2030) 
* tests passing

* hand coded opt with new abstractions

* simpler opts

* split out tensor cores
 2023-10-10 07:38:38 -07:00
George Hotz
16ca8410f8 op logger + replay (#2021) 
* logops

* fix dtype printing

* needs inf

* ops dataset

* minor improvements

* 12k kernels

* opt can compile

* graph flops
 2023-10-08 15:10:18 -07:00