From 53de9ecd13fbf7e098fea88f73d1a9403a40d83b Mon Sep 17 00:00:00 2001 From: Void Agent Date: Sun, 2 Aug 2026 13:32:16 +0100 Subject: loss_reweight: pass Y to forward for full logits (Karpathy nanoGPT returns last-position logits when targets=None); regression test --- src/loss_reweight.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'src') diff --git a/src/loss_reweight.py b/src/loss_reweight.py index bcfdf07..1bf3124 100644 --- a/src/loss_reweight.py +++ b/src/loss_reweight.py @@ -108,8 +108,8 @@ def train(mode, seed, max_iters, batch_size): for _ in range(50): X, Y = get_batch('val') with torch.no_grad(): - logits = model(X)[0] - lv.append(F.cross_entropy(logits.view(-1, V), Y.view(-1)).item()) + _, loss = model(X, Y) + lv.append(loss.item()) v = np.mean(lv) model.train() if v < best_val: @@ -119,7 +119,7 @@ def train(mode, seed, max_iters, batch_size): if it % 1000 == 0: print(f" iter {it}: val={v:.4f}") X, Y = get_batch('train') - logits = model(X)[0] + logits = model(X, Y)[0] # pass Y: targets=None would give last-position logits only loss = _weighted_loss(logits, Y, mode, q_id, it, V, device) loss.backward() opt.step() -- cgit v1.2.3