Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
Skip redundant initialization when loading checkpoints (#195)
Sophiefea4aa2f1fe1
diff --git a/laya/agent.py b/laya/agent.pyindex 15e2eb56..39553a74 100644--- a/laya/agent.py+++ b/laya/agent.py@@ -115,6 +115,10 @@ class Agent:"""from safetensors.torch import load_filefrom transformers import AutoTokenizer+ try:+ from transformers.initialization import no_init_weights+ except ImportError: # Transformers 4.x+ from transformers.modeling_utils import no_init_weightsmodel_dir = model_id_or_pathif not os.path.exists(model_dir):@@ -185,7 +189,10 @@ class Agent:self.tok = AutoTokenizer.from_pretrained(tok_dir if os.path.exists(tok_dir) else self.cfg.get("encoder"))enc_dir = os.path.join(model_dir, "encoder")- self.model = build_model(self.cfg, encoder_dir=enc_dir if os.path.exists(enc_dir) else None)+ # The checkpoint supplies every parameter; skip random/base-model weights.+ with no_init_weights():+ self.model = build_model(self.cfg, encoder_dir=enc_dir if os.path.exists(enc_dir) else None,+ pretrained=False)# Load weights and verify architectural compatibilityweights = load_file(weights_path)diff --git a/laya/common.py b/laya/common.pyindex fb900053..2d030002 100644--- a/laya/common.py+++ b/laya/common.py@@ -136,11 +136,11 @@ class DecisionModel(nn.Module):return logits, act_logits-def build_model(cfg: Dict, encoder_dir: Optional[str] = None) -> DecisionModel:+def build_model(cfg: Dict, encoder_dir: Optional[str] = None, pretrained: bool = True) -> DecisionModel:from transformers import AutoConfig, AutoModel- if encoder_dir and os.path.exists(encoder_dir):- ecfg = AutoConfig.from_pretrained(encoder_dir)+ if not pretrained or (encoder_dir and os.path.exists(encoder_dir)):+ ecfg = AutoConfig.from_pretrained(encoder_dir or cfg["encoder"])enc = AutoModel.from_config(ecfg, attn_implementation="sdpa")else:enc = AutoModel.from_pretrained(cfg["encoder"], attn_implementation="sdpa")diff --git a/tests/test_download.py b/tests/test_download.pyindex 34942eaf..4a34b3e7 100644--- a/tests/test_download.py+++ b/tests/test_download.py@@ -17,7 +17,7 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))import torch # noqa: E402from huggingface_hub.utils import filter_repo_objects # noqa: E402-from safetensors.torch import save_file # noqa: E402+from safetensors.torch import load_file, save_file # noqa: E402from tokenizers import Tokenizer # noqa: E402from tokenizers.models import WordLevel # noqa: E402from transformers import BertConfig, BertModel, PreTrainedTokenizerFast # noqa: E402@@ -113,6 +113,29 @@ class DownloadTests(unittest.TestCase):self.assertEqual(agent.predict("hello", self.questions), self.expected)download.assert_not_called()+ def test_loading_skips_initialization_and_preserves_weights(self):+ weights = load_file(self.repo / "model.safetensors")+ for bundled_encoder in (True, False):+ with self.subTest(bundled_encoder=bundled_encoder), tempfile.TemporaryDirectory() as destination:+ path = Path(destination)+ for filename in self.runtime_files:+ if not bundled_encoder and filename.startswith("encoder/"):+ continue+ target = path / filename+ target.parent.mkdir(parents=True, exist_ok=True)+ shutil.copyfile(self.repo / filename, target)+ cfg_path = path / "rl_agent_config.json"+ cfg = json.loads(cfg_path.read_text())+ cfg["encoder"] = str(self.repo / "encoder")+ cfg_path.write_text(json.dumps(cfg))+ rng = torch.get_rng_state()+ with patch("transformers.AutoModel.from_pretrained", side_effect=AssertionError("Unused base weights")):+ agent = load(str(path), device="cpu")+ self.assertTrue(torch.equal(torch.get_rng_state(), rng), "Loading initialized random weights")+ for name, value in agent.model.state_dict().items():+ torch.testing.assert_close(value, weights[name], rtol=0, atol=0)+ self.assertEqual(agent.predict("hello", self.questions), self.expected)+if __name__ == "__main__":torch.set_num_threads(1)