Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
59 commits
Select commit Hold shift + click to select a range
2b7b36f
init the lib
rgodha24 Mar 26, 2026
a61cf25
copy over siebren
rgodha24 Mar 26, 2026
61904e1
add nix flake & deps
rgodha24 Mar 26, 2026
05a4e52
wip: migrate
rgodha24 Mar 26, 2026
6bd967f
im lazy as shit but almost done
rgodha24 Mar 26, 2026
3f9dfe1
Refactor training: remove mcts/types, add python training module
rgodha24 Mar 27, 2026
69db64e
batch descent evaluations
rgodha24 Mar 27, 2026
b48a98e
add torch python dep lol
rgodha24 Mar 27, 2026
ce7d37d
add packed uint16 observations with triton decode
rgodha24 Mar 27, 2026
6d0076b
wire self-play samples into value training
rgodha24 Mar 27, 2026
de01f36
speed up self-play value inference
rgodha24 Mar 27, 2026
cefbc08
stream tree samples during self-play
rgodha24 Mar 27, 2026
0ad3544
shorten training games for Descent
rgodha24 Mar 27, 2026
1b990ee
shorten self-play games for training
rgodha24 Mar 27, 2026
14b8e0b
revert the weird ass game changes
rgodha24 Mar 27, 2026
92c7ecf
stabilize value training targets
rgodha24 Mar 27, 2026
f83ff32
add wandb and update some constants
rgodha24 Mar 27, 2026
dd487c7
instrument self-play and stop replaying reused tree samples
rgodha24 Mar 27, 2026
0143afa
push nodes we are pruning to replay buffer
rgodha24 Mar 27, 2026
36120b1
cargo fmt
rgodha24 Mar 27, 2026
499fd97
limit iteration count by gpu evaluations
rgodha24 Mar 27, 2026
91e6060
const updates
rgodha24 Mar 27, 2026
3c99bdc
more timing & bench work
rgodha24 Mar 28, 2026
26e4a16
increase self-play queue headroom
rgodha24 Mar 28, 2026
df13b0c
try a mixer style model
rgodha24 Mar 28, 2026
7342047
Revert "try a mixer style model"
rgodha24 Mar 28, 2026
6b27c11
Reapply "try a mixer style model"
rgodha24 Mar 28, 2026
1e4f428
pls stop fucking deadlocking why is it doing this im so confused
rgodha24 Mar 28, 2026
5491ad7
Revert "Reapply "try a mixer style model""
rgodha24 Mar 28, 2026
8fcf64b
my fuck ass queue has so many fucking deadlocks
rgodha24 Mar 28, 2026
13d53d9
hopefully fix last bug
rgodha24 Mar 28, 2026
c4e9b14
fix weird ass bug in descent where node has no children
rgodha24 Mar 28, 2026
d1f2c57
pretrain
rgodha24 Mar 28, 2026
e8c4ddc
resnet in engine
rgodha24 Mar 28, 2026
2a7d643
optimize tf out of resnet cpu eval
rgodha24 Mar 28, 2026
c5848e9
train a larger resnet
rgodha24 Mar 28, 2026
3d163dd
wip: multigpu
rgodha24 Mar 28, 2026
6b3d24a
config for a multigpu
rgodha24 Mar 28, 2026
9936dbd
Revert "config for a multigpu"
rgodha24 Mar 28, 2026
08ef34f
Revert "wip: multigpu"
rgodha24 Mar 28, 2026
12a8c21
stabilize selfplay value targets
rgodha24 Mar 29, 2026
f518c83
cap search by total turn time
rgodha24 Mar 29, 2026
9014f50
retune default 90k training config
rgodha24 Mar 29, 2026
86d2389
some debug stuff
rgodha24 Mar 29, 2026
851374a
cleanup logging
rgodha24 Mar 29, 2026
49abb9a
refine terminal value heuristic
rgodha24 Mar 30, 2026
cd9c93c
mix in some terminal states into replay buffer
rgodha24 Mar 30, 2026
5f6d7fc
fix bug in terminal state mixing
rgodha24 Mar 30, 2026
daf39f5
simplify logging
rgodha24 Mar 30, 2026
0f3ca54
even larger resnet
rgodha24 Mar 30, 2026
db517fa
multigpu!!
rgodha24 Mar 30, 2026
a06e645
scuffed nnue training but like we r evaling nnue on gpu for some reas…
rgodha24 Mar 30, 2026
a06371c
try tanh nnue
rgodha24 Mar 30, 2026
91f33a7
fix weird bugs with training data not being bounded
rgodha24 Mar 30, 2026
616397f
port over killshots
rgodha24 Mar 31, 2026
72a959b
include player location in nnue
rgodha24 Mar 31, 2026
7d5fd63
switch packed format decode to custom triton kernel
rgodha24 Mar 31, 2026
702691a
Revert "switch packed format decode to custom triton kernel"
rgodha24 Mar 31, 2026
8a4b6dd
fix stack overflow issue for DoubleArray::splat
rgodha24 Mar 31, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -6,3 +6,4 @@ build/

.venv/
.idea/
runs/
8 changes: 7 additions & 1 deletion AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,4 +9,10 @@ our search impl will also be in rust.

check out /mnt/github/TheConverseEngineer/AlphaSnake/ for last years impl

`just test` runs tests
`just test` runs tests. u should use this rather than doign it yourself bc this auto installs the package

- this just tests the rust board impl. do not use it for eg training code changes

`nix develop` GETS YOU CUDA
`maturin develop --release` to bring python bindings in. you can `--manifest-path training/Cargo.toml` etc to choose the package to reinstall
`PYTHONPATH=python` is uesful sometimes
Loading