Train manual impl - #8

Draft

rgodha24 wants to merge 59 commits into

mainfrom

Commits on Mar 26, 2026

init the lib
rgodha24
committed
2b7b36f
copy over siebren
rgodha24
committed
a61cf25
add nix flake & deps
rgodha24
committed
61904e1
wip: migrate
rgodha24
committed
05a4e52
im lazy as shit but almost done
rgodha24
committed
6bd967f

Commits on Mar 27, 2026

Refactor training: remove mcts/types, add python training module
rgodha24
committed
3f9dfe1
batch descent evaluations
rgodha24
committed
69db64e
add torch python dep lol
rgodha24
committed
b48a98e
add packed uint16 observations with triton decode
rgodha24
committed
ce7d37d
wire self-play samples into value training
rgodha24
committed
6d0076b
speed up self-play value inference
rgodha24
committed
de01f36
stream tree samples during self-play
rgodha24
committed
cefbc08
shorten training games for Descent
rgodha24
committed
0ad3544
shorten self-play games for training
rgodha24
committed
1b990ee
revert the weird ass game changes
rgodha24
committed
14b8e0b
stabilize value training targets
rgodha24
committed
92c7ecf
add wandb and update some constants
rgodha24
committed
f83ff32
instrument self-play and stop replaying reused tree samples
rgodha24
committed
dd487c7
push nodes we are pruning to replay buffer
rgodha24
committed
0143afa
cargo fmt
rgodha24
committed
36120b1
limit iteration count by gpu evaluations
rgodha24
committed
499fd97
const updates
rgodha24
committed
91e6060

Commits on Mar 28, 2026

more timing & bench work
rgodha24
committed
3c99bdc
increase self-play queue headroom
rgodha24
committed
26e4a16
try a mixer style model
rgodha24
committed
df13b0c
Revert "try a mixer style model"
rgodha24
committed
7342047
Reapply "try a mixer style model"
rgodha24
committed
6b27c11
pls stop fucking deadlocking why is it doing this im so confused
rgodha24
committed
1e4f428
Revert "Reapply "try a mixer style model""
rgodha24
committed
5491ad7
my fuck ass queue has so many fucking deadlocks
rgodha24
committed
8fcf64b
hopefully fix last bug
rgodha24
committed
13d53d9
fix weird ass bug in descent where node has no children
rgodha24
committed
c4e9b14
pretrain
rgodha24
committed
d1f2c57
resnet in engine
rgodha24
committed
e8c4ddc
optimize tf out of resnet cpu eval
rgodha24
committed
2a7d643
train a larger resnet
rgodha24
committed
c5848e9
wip: multigpu
rgodha24
committed
3d163dd
config for a multigpu
rgodha24
committed
6b3d24a
Revert "config for a multigpu"
rgodha24
committed
9936dbd
Revert "wip: multigpu"
rgodha24
committed
08ef34f

Commits on Mar 29, 2026

stabilize selfplay value targets
rgodha24
committed
12a8c21
cap search by total turn time
rgodha24
committed
f518c83
retune default 90k training config
rgodha24
committed
9014f50
some debug stuff
rgodha24
committed
86d2389
cleanup logging
rgodha24
committed
851374a

Commits on Mar 30, 2026

refine terminal value heuristic
rgodha24
committed
49abb9a
mix in some terminal states into replay buffer
rgodha24
committed
cd9c93c
fix bug in terminal state mixing
rgodha24
committed
5f6d7fc
simplify logging
rgodha24
committed
daf39f5
even larger resnet
rgodha24
committed
0f3ca54
multigpu!!
rgodha24
committed
db517fa
scuffed nnue training but like we r evaling nnue on gpu for some reason lol
rgodha24
committed
a06e645
try tanh nnue
rgodha24
committed
a06371c
fix weird bugs with training data not being bounded
rgodha24
committed
91f33a7

Commits on Mar 31, 2026