💡 All-in-one AI framework for semantic search, LLM orchestration and language model workflows
Python agentsaiai-agentsembeddingsinformation-retrievallanguage-modellarge-language-modelsllmnlppythonragretrieval-augmented-generationsearchsearch-enginesemantic-searchsentence-embeddingstransformerstxtaivector-databasevector-search
Various training pipeline fixes for v5, closes #1088
davidmezzetti109a4d3c66bb
diff --git a/src/python/txtai/data/sequences.py b/src/python/txtai/data/sequences.pyindex 5a9274d1..a34596a3 100644--- a/src/python/txtai/data/sequences.py+++ b/src/python/txtai/data/sequences.py@@ -38,9 +38,8 @@ class Sequences(Data):source = [self.prefix + x if self.prefix else x for x in data[source]]inputs = self.tokenizer(source, max_length=self.maxlength, padding=False, truncation=True)- # Tokenize target- with self.tokenizer.as_target_tokenizer():- targets = self.tokenizer(data[target], max_length=self.maxlength, padding=False, truncation=True)+ # Tokenize targets+ targets = self.tokenizer(data[target], max_length=self.maxlength, padding=False, truncation=True)# Combine inputsinputs["labels"] = targets["input_ids"]diff --git a/src/python/txtai/models/onnx.py b/src/python/txtai/models/onnx.pyindex 4abba63a..77a0b1f7 100644--- a/src/python/txtai/models/onnx.py+++ b/src/python/txtai/models/onnx.py@@ -122,7 +122,9 @@ class OnnxModel(PreTrainedModel):value = value.cpu().numpy()# Cast to numpy array if not already one- features[key] = np.asarray(value)+ # Only add valid model inputs+ if any(x.name == key for x in self.model.get_inputs()):+ features[key] = np.asarray(value)return featuresdiff --git a/src/python/txtai/models/tokendetection.py b/src/python/txtai/models/tokendetection.pyindex 49cfbea0..d3448b96 100644--- a/src/python/txtai/models/tokendetection.py+++ b/src/python/txtai/models/tokendetection.py@@ -98,6 +98,7 @@ class TokenDetection(PreTrainedModel):loss = goutputs[0] + self.weight * doutputs[0]return loss, goutputs[1], doutputs[1], dlabels+ # pylint: disable=W0613def save_pretrained(self, output, state_dict=None, **kwargs):"""Saves current model to output directory.@@ -108,9 +109,6 @@ class TokenDetection(PreTrainedModel):kwargs: additional keyword arguments"""- # Save combined model to support training from checkpoints- super().save_pretrained(output, state_dict, **kwargs)-# Save generator tokenizer and modelgpath = os.path.join(output, "generator")self.tokenizer.save_pretrained(gpath)diff --git a/src/python/txtai/pipeline/train/hfonnx.py b/src/python/txtai/pipeline/train/hfonnx.pyindex 956e7f0e..37428964 100644--- a/src/python/txtai/pipeline/train/hfonnx.py+++ b/src/python/txtai/pipeline/train/hfonnx.py@@ -61,7 +61,7 @@ class HFOnnx(Tensors):# Export model to ONNXexport(- model,+ ModelWrapper(model),(dummy,),output,opset_version=opset,@@ -138,7 +138,7 @@ class HFOnnx(Tensors):config = {"default": (OrderedDict({"last_hidden_state": {0: "batch", 1: "sequence"}}), AutoModel.from_pretrained),- "pooling": (OrderedDict({"embeddings": {0: "batch", 1: "sequence"}}), lambda x: PoolingOnnx(x, -1)),+ "pooling": (OrderedDict({"embeddings": {0: "batch", 1: "sequence"}}), lambda x: PoolingFactory.create({"path": x, "device": -1})),"question-answering": (OrderedDict({@@ -157,29 +157,27 @@ class HFOnnx(Tensors):return (inputs,) + config[task]-class PoolingOnnx(nn.Module):+class ModelWrapper(nn.Module):"""- Extends Pooling methods to name inputs to model, which is required to export to ONNX.+ Wraps an model to control named inputs for export."""- def __init__(self, path, device):+ def __init__(self, model):"""- Creates a new PoolingOnnx instance.+ Creates a new ModelWrapper.Args:- path: path to model, accepts Hugging Face model hub id or local path- device: tensor device id+ model: model to export"""super().__init__()- # Create pooling method based on configuration- self.model = PoolingFactory.create({"path": path, "device": device})+ self.model = model.eval()# pylint: disable=W0221def forward(self, input_ids=None, attention_mask=None, token_type_ids=None):"""- Runs inputs through pooling model and returns outputs.+ Runs inputs through model and returns outputs.Args:inputs: model inputsdiff --git a/src/python/txtai/pipeline/train/hftrainer.py b/src/python/txtai/pipeline/train/hftrainer.pyindex 64a9b1e8..711053e0 100644--- a/src/python/txtai/pipeline/train/hftrainer.py+++ b/src/python/txtai/pipeline/train/hftrainer.py@@ -273,8 +273,8 @@ class HFTrainer(Tensors):return AutoModelForSeq2SeqLM.from_pretrained(base, config=config, quantization_config=quantization)if task == "token-detection":return TokenDetection(- AutoModelForMaskedLM.from_pretrained(base, config=config, quantization_config=quantization),- AutoModelForPreTraining.from_pretrained(base, config=config, quantization_config=quantization),+ AutoModelForMaskedLM.from_pretrained(base, config=config, quantization_config=quantization, attn_implementation="eager"),+ AutoModelForPreTraining.from_pretrained(base, config=config, quantization_config=quantization, attn_implementation="eager"),tokenizer,)diff --git a/test/python/testpipeline/testtrain/testtrainer.py b/test/python/testpipeline/testtrain/testtrainer.pyindex 7e204d2e..ccc418a1 100644--- a/test/python/testpipeline/testtrain/testtrainer.py+++ b/test/python/testpipeline/testtrain/testtrainer.py@@ -6,8 +6,6 @@ import osimport unittestimport tempfile-from unittest.mock import patch-import torchfrom transformers import AutoTokenizer, AutoModelForSequenceClassification@@ -220,15 +218,11 @@ class TestTrainer(unittest.TestCase):labels = Labels((model, tokenizer), dynamic=False)self.assertEqual(labels("cat")[0][0], 1)- @patch("importlib.util.find_spec")- def testPEFT(self, spec):+ def testPEFT(self):"""Test training a model with causal language modeling and PEFT"""- # Disable triton- spec.return_value = None-trainer = HFTrainer()model, _ = trainer("hf-internal-testing/tiny-random-gpt2",@@ -297,7 +291,7 @@ class TestTrainer(unittest.TestCase):output = os.path.join(tempfile.gettempdir(), "trainer.rtd")trainer = HFTrainer()- model, _ = trainer("hf-internal-testing/tiny-random-electra", self.data, task="token-detection", save_safetensors=False, output_dir=output)+ model, _ = trainer("hf-internal-testing/tiny-random-electra", self.data, task="token-detection", output_dir=output)# Test model completed successfullyself.assertIsNotNone(model)