fix: marker_mask bool + remove laya_common debug import

This commit is contained in:
dingbin.wang 2026-11-05 14:44:31 +08:00
parent 5b17c686bb
commit 562e9cb82b

View File

@ -64,7 +64,7 @@ def evaluate(model, loader, dev):
with torch.no_grad():
for b in loader:
logits, _ = model(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].bool().to(dev),
qtype=b["qtype"].to(dev))
# noul: markers = [false, true], true marker 是最后一个有效 marker
# logits shape: (batch, 8) — 取 true marker (index 1 in each pair... 简化: 取 max)
@ -107,7 +107,7 @@ def main():
target_modules=["Wqkv"], task_type=None)
# 手动对 encoder 加 LoRA (peft 对 bare nn.Module 用 get_peft_model)
dm.encoder = get_peft_model(dm.encoder, lora)
dm.print_trainable_parameters()
print(f'[jev] trainable: {sum(p.numel() for p in dm.parameters() if p.requires_grad):,} / {sum(p.numel() for p in dm.parameters()):,}')
dm.train()
tok = AutoTokenizer.from_pretrained(args.base)
@ -135,7 +135,7 @@ def main():
for ep in range(args.epochs):
for b in dl:
logits, act = dm(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].bool().to(dev),
qtype=b["qtype"].to(dev))
# noul markers = [false, true]: true marker 是每对中最后一个有效
# marker_mask (batch, 8): 每样本 2 个有效 marker
@ -158,7 +158,7 @@ def main():
with torch.no_grad():
for b in loader:
lg, _ = dm(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].bool().to(dev),
qtype=b["qtype"].to(dev))
p = torch.softmax(lg.float(), -1)[..., 1]
ps += p.tolist(); ls += b["label"].tolist()