nnue_model.py

import torch
import torch.nn as nn
import onnx
import sys
sys.stdout.reconfigure(encoding='utf-8')

NUM_FEATURES = 10 * 64 * 64     # NNUE feature space
PAD_IDX = NUM_FEATURES

class NNUE(nn.Module):
    def __init__(self):
        super().__init__()
        self.embed = nn.Embedding(NUM_FEATURES + 1, 256, padding_idx=PAD_IDX)
        self.fc1 = nn.Linear(256 + 1, 256)
        self.fc2 = nn.Linear(256, 64)
        self.fc3 = nn.Linear(64, 1)

    def forward(self, feats, stm):
        # feats: [B, N] long
        x = self.embed(feats).sum(dim=1)

        # stm: [B] -> [-1, +1]
        stm = stm.float().unsqueeze(1) * 2 - 1
        x = torch.cat([x, stm], dim=1)

        x = torch.relu(self.fc1(x))
        x = torch.relu(self.fc2(x))
        return self.fc3(x).squeeze(1)


# ------------------------------
# Create model
# ------------------------------
torch_model = NNUE()
torch_model.eval()

# ------------------------------
# Correct example inputs
# ------------------------------
BATCH = 1
ACTIVE_FEATURES = 32  # typical NNUE sparse feature count

# Feature indices (LONG!)
feats = torch.randint(
    low=0,
    high=NUM_FEATURES,
    size=(BATCH, ACTIVE_FEATURES),
    dtype=torch.long
)

# Side to move: 0 = black, 1 = white
stm = torch.randint(0, 2, (BATCH,), dtype=torch.long)

example_inputs = (feats, stm)