Spaces:
Sleeping
Sleeping
File size: 27,068 Bytes
3656ccf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 | //! Self-play generation core: cross-game batched MCTS over shakmaty boards.
//!
//! Python keeps only the batched NN forward; this crate owns boards, search
//! trees and game lifecycle for many concurrent games. Protocol per round:
//!
//! boards, player, castling, ep = engine.collect() # deduped NN-pending positions
//! logits, values = nn_forward(...) # (n, 64, 73) f32, (n,) f32
//! engine.apply(logits, values) # expand/backup, play ready moves
//! for game in engine.drain_finished(): ... # completed game records
//!
//! Search semantics mirror `Pos2MoveV2MctsBot` (PUCT + FPU, virtual loss,
//! negamax backup, terminal rules incl. in-search threefold repetition).
//! Encodings and tree primitives live in the shared `chess-core` crate —
//! see the parity test.
//!
//! Two expert-iteration options, both off by default (parity holds):
//!
//! * Playout cap randomization (KataGo): with `fast_sims > 0`, each move is a
//! full search (`num_sims`) with probability `full_frac`, else a fast one
//! (`fast_sims`). Every position is recorded with a `full` flag; only full
//! searches are strong enough to be policy targets, while all of them carry
//! the game outcome. Fast moves buy more games -- more independent value
//! labels -- per GPU hour.
//! * Root Dirichlet noise (AlphaZero): on full searches the root priors are
//! mixed `(1 - eps) * p + eps * Dir(alpha)`. Without it a move under ~0.5 %
//! prior is never visited, so the visit target can only re-confirm the prior.
//!
//! `SelfPlayEngine` (selfplay_engine.rs) runs self-play on the deployed bot's
//! own search instead, and `MatchEngine` (matches.rs) plays matches with it;
//! this engine's search is the older one, kept for comparison.
use std::collections::HashMap;
use chess_core::encoding::{
action_plane, encode_position, halfmove, move_coords, terminal_value, zobrist, NUM_PLANES,
};
use chess_core::search::sample_gamma;
use chess_core::tree::{backup_vl, puct_select, softmax_priors, Node};
use numpy::{IntoPyArray, PyArray1, PyArray2, PyArrayMethods, PyReadonlyArray1, PyReadonlyArray3};
use pyo3::prelude::*;
use pyo3::types::PyDict;
use rand::prelude::*;
use rand::rngs::StdRng;
use serde::Deserialize;
use shakmaty::fen::Fen;
use shakmaty::{CastlingMode, Chess, Color, Move, Position};
mod matches;
mod selfplay_engine;
// ── per-game state ───────────────────────────────────────────────────────
struct Rec {
tokens: [u8; 64],
player: u8,
castling: u8,
ep: u8,
halfmove: u16,
root_v: f32,
full: u8,
visit_idx: Vec<u16>, // from_sq * 73 + plane
visit_cnt: Vec<u16>,
}
struct Game {
id: usize,
start_fen: String,
pos: Chess,
history: Vec<u64>, // zobrist of every position incl. start
tree: Vec<Node>,
sims_done: u32,
inflight: u32,
move_sims: u32, // search budget of the current move
full: bool, // current move is a full search (policy target)
noise_applied: bool, // root noise already mixed into this move's root
records: Vec<Rec>,
moves_uci: Vec<String>,
resign_allowed: bool,
resign_streak: u32,
resign_sign: i32,
plies: u32,
active: bool,
}
struct Finished {
records: Vec<Rec>,
moves_uci: Vec<String>,
start_fen: String,
z_white: f32,
end: &'static str,
resign_allowed: bool,
id: usize,
}
struct NnItem {
key: u64,
tokens: [u8; 64],
player: u8,
castling: u8,
ep: u8,
move_fp: Vec<(usize, usize)>, // (from_sq, plane) per legal move, in move-list order
}
struct PendingLeaf {
game: usize,
node: usize,
path: Vec<(usize, usize)>, // (node_idx, child_idx)
key: u64,
moves: Vec<Move>,
}
#[derive(Deserialize)]
struct Config {
num_sims: u32,
#[serde(default = "d_games")] num_parallel_games: usize,
#[serde(default = "d_leaves")] leaves_per_wave: u32,
#[serde(default = "d_cpuct")] c_puct: f64,
#[serde(default = "d_fpu")] fpu: f64,
#[serde(default = "d_one")] prior_temp: f32,
#[serde(default = "d_one_f64")] move_temp: f64,
#[serde(default = "d_temp_plies")] temp_plies: u32,
#[serde(default = "d_resign")] resign_threshold: f32,
#[serde(default = "d_resign_plies")] resign_plies: u32,
#[serde(default = "d_noresign")] no_resign_frac: f64,
#[serde(default = "d_max_plies")] max_plies: u32,
#[serde(default = "d_tt_cap")] tt_cap: usize,
#[serde(default)] fast_sims: u32,
#[serde(default = "d_one_f64")] full_frac: f64,
#[serde(default)] dirichlet_alpha: f64,
#[serde(default = "d_dir_eps")] dirichlet_eps: f64,
}
fn d_dir_eps() -> f64 { 0.25 }
fn d_games() -> usize { 64 }
fn d_leaves() -> u32 { 4 }
fn d_cpuct() -> f64 { 1.0 }
fn d_fpu() -> f64 { 0.2 }
fn d_one() -> f32 { 1.0 }
fn d_one_f64() -> f64 { 1.0 }
fn d_temp_plies() -> u32 { 30 }
fn d_resign() -> f32 { 0.93 }
fn d_resign_plies() -> u32 { 6 }
fn d_noresign() -> f64 { 0.1 }
fn d_max_plies() -> u32 { 300 }
fn d_tt_cap() -> usize { 2_000_000 }
// move_temp is deserialized as f64 for sampling math convenience.
#[pyclass]
struct Engine {
cfg: Config,
openings: Vec<Chess>,
opening_fens: Vec<String>,
games: Vec<Game>,
pending: Vec<PendingLeaf>,
nn_items: Vec<NnItem>,
nn_index: HashMap<u64, usize>,
tt: HashMap<u64, (Vec<f32>, f32)>, // priors (legal-move order) + stm value
// The previous TT generation. A full TT is rotated here instead of cleared, so the
// subtrees of the searches in progress stay cached across the rotation.
tt_old: HashMap<u64, (Vec<f32>, f32)>,
finished: Vec<Finished>,
rng: StdRng,
started: usize,
completed: usize,
target: usize,
positions_recorded: usize,
}
impl Engine {
fn fresh_tree() -> Vec<Node> {
vec![Node::default()]
}
/// TT lookup across both generations; an old-generation hit is promoted.
fn tt_get(&mut self, key: u64) -> Option<(Vec<f32>, f32)> {
if let Some(e) = self.tt.get(&key) {
return Some(e.clone());
}
let e = self.tt_old.get(&key)?.clone();
self.tt.insert(key, e.clone());
Some(e)
}
/// (search budget, is_full) for the next move.
fn draw_budget(&mut self) -> (u32, bool) {
if self.cfg.fast_sims == 0 || self.rng.gen::<f64>() < self.cfg.full_frac {
(self.cfg.num_sims, true)
} else {
(self.cfg.fast_sims, false)
}
}
/// Mix Dirichlet noise into the root priors, once per full search, as soon as
/// the root is expanded. Never touches the TT, which keeps the clean priors.
fn maybe_noise_root(&mut self, gi: usize) {
let alpha = self.cfg.dirichlet_alpha;
let g = &self.games[gi];
if alpha <= 0.0 || !g.full || g.noise_applied || !g.tree[0].expanded {
return;
}
let k = g.tree[0].priors.len();
let noise: Vec<f64> = (0..k).map(|_| sample_gamma(&mut self.rng, alpha)).collect();
let total: f64 = noise.iter().sum();
let eps = self.cfg.dirichlet_eps;
let g = &mut self.games[gi];
if total > 0.0 {
for (p, n) in g.tree[0].priors.iter_mut().zip(&noise) {
*p = ((1.0 - eps) * (*p as f64) + eps * n / total) as f32;
}
}
g.noise_applied = true;
}
fn new_game(&mut self) -> Game {
let oi = self.rng.gen_range(0..self.openings.len());
let pos = self.openings[oi].clone();
let id = self.started;
self.started += 1;
let (move_sims, full) = self.draw_budget();
Game {
id,
start_fen: self.opening_fens[oi].clone(),
history: vec![zobrist(&pos)],
pos,
tree: Self::fresh_tree(),
sims_done: 0,
inflight: 0,
move_sims,
full,
noise_applied: false,
records: Vec::new(),
moves_uci: Vec::new(),
resign_allowed: self.rng.gen::<f64>() >= self.cfg.no_resign_frac,
resign_streak: 0,
resign_sign: 0,
plies: 0,
active: true,
}
}
/// Descend with virtual loss to an unexpanded node. Returns the leaf node
/// index, the path, the leaf position and the zobrist keys pushed en route.
fn descend(&mut self, gi: usize) -> (usize, Vec<(usize, usize)>, Chess, Vec<u64>) {
let mut node_idx = 0usize;
let mut pos = self.games[gi].pos.clone();
let mut path = Vec::new();
let mut keys = Vec::new();
loop {
if !self.games[gi].tree[node_idx].expanded {
return (node_idx, path, pos, keys);
}
let idx = puct_select(&self.games[gi].tree[node_idx], self.cfg.c_puct, self.cfg.fpu);
{
let node = &mut self.games[gi].tree[node_idx];
node.n[idx] += 1; // virtual loss
node.w[idx] += 1.0;
}
path.push((node_idx, idx));
let mv = self.games[gi].tree[node_idx].moves[idx].clone();
pos.play_unchecked(&mv);
keys.push(zobrist(&pos));
let child = self.games[gi].tree[node_idx].children[idx];
node_idx = if child >= 0 {
child as usize
} else {
self.games[gi].tree.push(Node::default());
let new_idx = self.games[gi].tree.len() - 1;
self.games[gi].tree[node_idx].children[idx] = new_idx as i32;
new_idx
};
}
}
fn resolve_leaf(game: &mut Game, node_idx: usize, moves: Vec<Move>, priors: &[f32],
value: f32, path: &[(usize, usize)]) {
if !game.tree[node_idx].expanded {
game.tree[node_idx].expand(moves, priors.to_vec());
}
backup_vl(&mut game.tree, path, value);
game.sims_done += 1;
}
/// Search budget reached: record the position, play the move, handle game
/// end / next-move reset.
fn finish_move(&mut self, gi: usize) {
let cfg_move_temp = self.cfg.move_temp;
let cfg_temp_plies = self.cfg.temp_plies;
let (tokens, player, castling, ep) = encode_position(&self.games[gi].pos);
let halfmove = halfmove(&self.games[gi].pos);
// Visit distribution + move selection from the root.
let (best_idx, root_v, visit_idx, visit_cnt) = {
let root = &self.games[gi].tree[0];
let mut visit_idx = Vec::new();
let mut visit_cnt = Vec::new();
for i in 0..root.moves.len() {
if root.n[i] > 0 {
let (f, t, p) = move_coords(&root.moves[i]);
visit_idx.push((f * NUM_PLANES + action_plane(f, t, p)) as u16);
visit_cnt.push(root.n[i].min(u16::MAX as u32) as u16);
}
}
let best = if cfg_move_temp > 0.0 && halfmove < cfg_temp_plies {
let weights: Vec<f64> = root
.n
.iter()
.map(|&n| (n as f64).powf(1.0 / cfg_move_temp))
.collect();
let total: f64 = weights.iter().sum();
if total > 0.0 {
let mut r = self.rng.gen::<f64>() * total;
let mut pick = root.n.len() - 1;
for (i, w) in weights.iter().enumerate() {
if r < *w {
pick = i;
break;
}
r -= w;
}
pick
} else {
0
}
} else {
(0..root.n.len()).max_by_key(|&i| root.n[i]).unwrap_or(0)
};
let rv = if root.n[best] > 0 {
(-root.w[best] / root.n[best] as f64) as f32
} else {
0.0
};
(best, rv, visit_idx, visit_cnt)
};
let full = self.games[gi].full as u8;
self.games[gi].records.push(Rec {
tokens, player, castling, ep,
halfmove: halfmove.min(u16::MAX as u32) as u16,
root_v,
full,
visit_idx, visit_cnt,
});
self.positions_recorded += 1;
let mv = self.games[gi].tree[0].moves[best_idx].clone();
let uci = mv.to_uci(CastlingMode::Standard).to_string();
let game = &mut self.games[gi];
game.moves_uci.push(uci);
game.pos.play_unchecked(&mv);
game.history.push(zobrist(&game.pos));
game.plies += 1;
// Resign bookkeeping (white-POV root value).
let v_white = if player == 1 { root_v } else { -root_v };
if v_white.abs() >= self.cfg.resign_threshold {
let sign = if v_white > 0.0 { 1 } else { -1 };
game.resign_streak = if sign == game.resign_sign {
game.resign_streak + 1
} else {
1
};
game.resign_sign = sign;
} else {
game.resign_streak = 0;
}
// Game over?
let cur_key = *game.history.last().unwrap();
let term = terminal_value(&game.pos, &game.history, cur_key);
let outcome: Option<(f32, &'static str)> = if let Some(t) = term {
let z = if t < 0.0 {
// side to move is mated; the side that just moved wins
if game.pos.turn() == Color::Black { 1.0 } else { -1.0 }
} else {
0.0
};
Some((z, "terminal"))
} else if game.resign_allowed && game.resign_streak >= self.cfg.resign_plies {
Some((game.resign_sign as f32, "resign"))
} else if game.plies >= self.cfg.max_plies {
Some((0.0, "max_plies"))
} else {
None
};
match outcome {
Some((z_white, end)) => {
let game = &mut self.games[gi];
let fin = Finished {
records: std::mem::take(&mut game.records),
moves_uci: std::mem::take(&mut game.moves_uci),
start_fen: game.start_fen.clone(),
z_white,
end,
resign_allowed: game.resign_allowed,
id: game.id,
};
self.finished.push(fin);
self.completed += 1;
if self.started < self.target {
self.games[gi] = self.new_game();
} else {
self.games[gi].active = false;
}
}
None => {
let (move_sims, full) = self.draw_budget();
let game = &mut self.games[gi];
game.tree = Self::fresh_tree();
game.sims_done = 0;
game.move_sims = move_sims;
game.full = full;
game.noise_applied = false;
}
}
}
}
#[pymethods]
impl Engine {
#[new]
fn new(config_json: &str, openings: Vec<String>, seed: u64, target: usize) -> PyResult<Self> {
let cfg: Config = serde_json::from_str(config_json)
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("config: {e}")))?;
let mut parsed = Vec::new();
for fen in &openings {
let f: Fen = fen
.parse()
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("fen {fen}: {e}")))?;
let pos: Chess = f
.into_position(CastlingMode::Standard)
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("fen {fen}: {e}")))?;
parsed.push(pos);
}
let n_games = cfg.num_parallel_games.min(target.max(1));
let mut engine = Engine {
cfg,
openings: parsed,
opening_fens: openings,
games: Vec::new(),
pending: Vec::new(),
nn_items: Vec::new(),
nn_index: HashMap::new(),
tt: HashMap::new(),
tt_old: HashMap::new(),
finished: Vec::new(),
rng: StdRng::seed_from_u64(seed),
started: 0,
completed: 0,
target,
positions_recorded: 0,
};
for _ in 0..n_games {
let g = engine.new_game();
engine.games.push(g);
}
Ok(engine)
}
/// Run descent waves over all active games; immediately resolve terminal
/// and TT-cached leaves. Returns encoded positions needing an NN forward
/// (deduped), as numpy arrays.
fn collect<'py>(
&mut self,
py: Python<'py>,
) -> (
Bound<'py, PyArray2<u8>>,
Bound<'py, PyArray1<u8>>,
Bound<'py, PyArray1<u8>>,
Bound<'py, PyArray1<u8>>,
) {
for gi in 0..self.games.len() {
if !self.games[gi].active {
continue;
}
for _ in 0..self.cfg.leaves_per_wave {
let g = &self.games[gi];
if g.sims_done + g.inflight >= g.move_sims {
break;
}
let root_unexpanded = !self.games[gi].tree[0].expanded;
if root_unexpanded && self.games[gi].inflight > 0 {
break; // root eval already queued
}
let (node_idx, path, pos, path_keys) = self.descend(gi);
let key = zobrist(&pos);
let mut all_keys = self.games[gi].history.clone();
all_keys.extend_from_slice(&path_keys);
if let Some(t) = terminal_value(&pos, &all_keys, key) {
let game = &mut self.games[gi];
backup_vl(&mut game.tree, &path, t);
game.sims_done += 1;
continue;
}
let moves: Vec<Move> = pos.legal_moves().to_vec();
if let Some((priors, value)) = self.tt_get(key) {
let game = &mut self.games[gi];
Self::resolve_leaf(game, node_idx, moves, &priors, value, &path);
if node_idx == 0 {
self.maybe_noise_root(gi);
}
continue;
}
self.nn_index.entry(key).or_insert_with(|| {
let (tokens, player, castling, ep) = encode_position(&pos);
let move_fp = moves
.iter()
.map(|m| {
let (f, t, p) = move_coords(m);
(f, action_plane(f, t, p))
})
.collect();
self.nn_items.push(NnItem { key, tokens, player, castling, ep, move_fp });
self.nn_items.len() - 1
});
self.pending.push(PendingLeaf { game: gi, node: node_idx, path, key, moves });
self.games[gi].inflight += 1;
if root_unexpanded {
break; // nothing else to search until the root is expanded
}
}
}
let n = self.nn_items.len();
let mut boards = Vec::with_capacity(n * 64);
let mut player = Vec::with_capacity(n);
let mut castling = Vec::with_capacity(n);
let mut ep = Vec::with_capacity(n);
for item in &self.nn_items {
boards.extend_from_slice(&item.tokens);
player.push(item.player);
castling.push(item.castling);
ep.push(item.ep);
}
let boards = PyArray1::from_vec(py, boards)
.reshape([n, 64])
.expect("reshape");
(
boards,
player.into_pyarray(py),
castling.into_pyarray(py),
ep.into_pyarray(py),
)
}
/// Consume NN results for the batch returned by the last collect():
/// expand + back up all pending leaves, then play moves for games whose
/// search budget is complete (recording positions / finishing games).
fn apply(
&mut self,
logits: PyReadonlyArray3<f32>,
values: PyReadonlyArray1<f32>,
) -> PyResult<()> {
let logits = logits.as_array();
let values = values.as_slice()?;
if values.len() != self.nn_items.len() {
return Err(pyo3::exceptions::PyValueError::new_err(format!(
"apply: got {} values for {} pending NN items",
values.len(),
self.nn_items.len()
)));
}
if self.tt.len() + self.nn_items.len() > self.cfg.tt_cap {
self.tt_old = std::mem::take(&mut self.tt);
}
for (i, item) in self.nn_items.iter().enumerate() {
let mut scores: Vec<f32> = item
.move_fp
.iter()
.map(|&(f, p)| logits[[i, f, p]])
.collect();
softmax_priors(&mut scores, self.cfg.prior_temp);
self.tt.insert(item.key, (scores, values[i]));
}
let pending = std::mem::take(&mut self.pending);
for leaf in pending {
let (priors, value) = self.tt.get(&leaf.key).cloned().expect("just inserted");
let game = &mut self.games[leaf.game];
Self::resolve_leaf(game, leaf.node, leaf.moves, &priors, value, &leaf.path);
game.inflight -= 1;
if leaf.node == 0 {
self.maybe_noise_root(leaf.game);
}
}
self.nn_items.clear();
self.nn_index.clear();
for gi in 0..self.games.len() {
if self.games[gi].active
&& self.games[gi].inflight == 0
&& self.games[gi].sims_done >= self.games[gi].move_sims
{
self.finish_move(gi);
}
}
Ok(())
}
/// Completed games since the last call, as dicts of numpy arrays.
fn drain_finished<'py>(&mut self, py: Python<'py>) -> PyResult<Vec<Bound<'py, PyDict>>> {
let finished = std::mem::take(&mut self.finished);
let mut out = Vec::with_capacity(finished.len());
for f in finished {
let p = f.records.len();
let mut boards = Vec::with_capacity(p * 64);
let (mut player, mut castling, mut ep) =
(Vec::with_capacity(p), Vec::with_capacity(p), Vec::with_capacity(p));
let mut halfmove = Vec::with_capacity(p);
let (mut root_v, mut z) = (Vec::with_capacity(p), Vec::with_capacity(p));
let mut full = Vec::with_capacity(p);
let (mut visit_idx, mut visit_cnt) = (Vec::new(), Vec::new());
let mut visit_ptr = Vec::with_capacity(p + 1);
visit_ptr.push(0i64);
for r in &f.records {
boards.extend_from_slice(&r.tokens);
player.push(r.player);
castling.push(r.castling);
ep.push(r.ep);
halfmove.push(r.halfmove);
root_v.push(r.root_v);
full.push(r.full);
z.push(if r.player == 1 { f.z_white } else { -f.z_white });
visit_idx.extend_from_slice(&r.visit_idx);
visit_cnt.extend_from_slice(&r.visit_cnt);
visit_ptr.push(visit_idx.len() as i64);
}
let d = PyDict::new(py);
d.set_item("boards", PyArray1::from_vec(py, boards).reshape([p, 64]).expect("reshape"))?;
d.set_item("player", player.into_pyarray(py))?;
d.set_item("castling", castling.into_pyarray(py))?;
d.set_item("ep", ep.into_pyarray(py))?;
d.set_item("halfmove", halfmove.into_pyarray(py))?;
d.set_item("root_v", root_v.into_pyarray(py))?;
d.set_item("full", full.into_pyarray(py))?;
d.set_item("z", z.into_pyarray(py))?;
d.set_item("visit_idx", visit_idx.into_pyarray(py))?;
d.set_item("visit_cnt", visit_cnt.into_pyarray(py))?;
d.set_item("visit_ptr", visit_ptr.into_pyarray(py))?;
d.set_item("moves", f.moves_uci)?;
d.set_item("fen", f.start_fen)?;
d.set_item("z_white", f.z_white)?;
d.set_item("end", f.end)?;
d.set_item("resign_allowed", f.resign_allowed)?;
d.set_item("id", f.id)?;
out.push(d);
}
Ok(out)
}
fn done(&self) -> bool {
self.completed >= self.target
}
/// (games started, games completed, positions recorded, TT entries)
fn stats(&self) -> (usize, usize, usize, usize) {
(self.started, self.completed, self.positions_recorded, self.tt.len() + self.tt_old.len())
}
}
// ── debug helpers for the Python parity test ────────────────────────────
#[pyfunction]
fn debug_encode(fen: &str) -> PyResult<(Vec<u8>, u8, u8, u8)> {
let f: Fen = fen
.parse()
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("{e}")))?;
let pos: Chess = f
.into_position(CastlingMode::Standard)
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("{e}")))?;
let (tokens, player, castling, ep) = encode_position(&pos);
Ok((tokens.to_vec(), player, castling, ep))
}
#[pyfunction]
fn debug_move_planes(fen: &str) -> PyResult<Vec<(String, u8, u8)>> {
let f: Fen = fen
.parse()
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("{e}")))?;
let pos: Chess = f
.into_position(CastlingMode::Standard)
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("{e}")))?;
let mut out = Vec::new();
for m in pos.legal_moves() {
let (from, to, promo) = move_coords(&m);
out.push((
m.to_uci(CastlingMode::Standard).to_string(),
from as u8,
action_plane(from, to, promo) as u8,
));
}
Ok(out)
}
#[pyfunction]
fn debug_zobrist(fen: &str) -> PyResult<u64> {
let f: Fen = fen
.parse()
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("{e}")))?;
let pos: Chess = f
.into_position(CastlingMode::Standard)
.map_err(|e| pyo3::exceptions::PyValueError::new_err(format!("{e}")))?;
Ok(zobrist(&pos))
}
#[pymodule]
fn selfplay_core(m: &Bound<'_, PyModule>) -> PyResult<()> {
m.add_class::<Engine>()?;
m.add_class::<matches::MatchEngine>()?;
m.add_class::<selfplay_engine::SelfPlayEngine>()?;
m.add_function(wrap_pyfunction!(debug_encode, m)?)?;
m.add_function(wrap_pyfunction!(debug_move_planes, m)?)?;
m.add_function(wrap_pyfunction!(debug_zobrist, m)?)?;
Ok(())
}
|