Merge branch 'develop' of github.com:greycooker/PaddleNLP into develop

merge

Author: greycooker

paddlenlp/peft/lora/lora_model.py

@@ -195,8 +195,7 @@ def _merge_trainable_tensor_parallel(self, trainable_state_dict):
             if key in trainable_name_action_mappings:
                 ret = distributed_gather(tensor, group=mp_group, offload=True)
                 action = trainable_name_action_mappings[key]
-                is_collumn = self.lora_split_mapping[key]
-                if "_scale" in key and not is_collumn and is_dst:
+                if key in self.lora_split_mapping and not self.lora_split_mapping[key] and "_scale" in key and is_dst:
                     ret = paddle.to_tensor(ret)
                     tensor = paddle.max(ret, axis=0)
                 else:

paddlenlp/transformers/model_utils.py

@@ -41,7 +41,10 @@
 )
 from huggingface_hub.utils import EntryNotFoundError
 from paddle import Tensor
-from paddle.distributed.fleet.meta_parallel.parallel_layers import SharedLayerDesc
+from paddle.distributed.fleet.meta_parallel.parallel_layers import (
+    PipelineLayer,
+    SharedLayerDesc,
+)
 from paddle.nn import Embedding, Layer
 
 # TODO(fangzeyang) Temporary fix and replace by paddle framework downloader later
@@ -933,6 +936,18 @@ def _post_init(self, original_init, *args, **kwargs):
         ):
             self.init_weights()
 
+        # Note:
+        # 1. PipelineLayer will create parameters for each layer and
+        # call `_synchronize_shared_weights()` to synchronize the shared parameters.
+        # 2. When setting the model `state_dict`, `_synchronize_shared_weights` will be called to
+        # synchronize the shared parameters.
+        # However, `self._init_weights` will re-initialize the parameters without
+        # synchronizing the shared parameters. If the following step does not load a checkpoint,
+        # the shared parameters will be different.
+
+        if isinstance(self, PipelineLayer):
+            self._synchronize_shared_weights()
+
     def _init_weights(self, layer):
         """
         Initialize the weights. This method should be overridden by derived class.

pipelines/pipelines/document_stores/faiss.py

@@ -391,7 +391,7 @@ def update_embeddings(
 
                 vector_id_map = {}
                 for doc in document_batch:
-                    vector_id_map[str(doc.id)] = str(vector_id)
+                    vector_id_map[str(doc.id)] = str(vector_id) + "_" + index
                     vector_id += 1
                 self.update_vector_ids(vector_id_map, index=index)
                 progress_bar.set_description_str("Documents Processed")
@@ -443,7 +443,6 @@ def get_all_documents_generator(
         )
         if return_embedding is None:
             return_embedding = self.return_embedding
-
         for doc in documents:
             if return_embedding:
                 if doc.meta and doc.meta.get("vector_id") is not None:
@@ -588,7 +587,6 @@ def query_by_embedding(
 
         if filters:
             logger.warning("Query filters are not implemented for the FAISSDocumentStore.")
-
         index = index or self.index
         if not self.faiss_indexes.get(index):
             raise Exception(f"Index named '{index}' does not exists. Use 'update_embeddings()' to create an index.")
@@ -599,11 +597,9 @@ def query_by_embedding(
         query_emb = query_emb.reshape(1, -1).astype(np.float32)
         if self.similarity == "cosine":
             self.normalize_embedding(query_emb)
-
         score_matrix, vector_id_matrix = self.faiss_indexes[index].search(query_emb, top_k)
         vector_ids_for_query = [str(vector_id) + "_" + index for vector_id in vector_id_matrix[0] if vector_id != -1]
         documents = self.get_documents_by_vector_ids(vector_ids_for_query, index=index)
-
         # assign query score to each document
         scores_for_vector_ids: Dict[str, float] = {
             str(v_id): s for v_id, s in zip(vector_id_matrix[0], score_matrix[0])

pipelines/pipelines/document_stores/sql.py

@@ -216,15 +216,13 @@ def get_documents_by_vector_ids(
     ):
         """Fetch documents by specifying a list of text vector id strings"""
         index = index or self.index
-
         documents = []
         for i in range(0, len(vector_ids), batch_size):
             query = self.session.query(DocumentORM).filter(
                 DocumentORM.vector_id.in_(vector_ids[i : i + batch_size]), DocumentORM.index == index
             )
             for row in query.all():
                 documents.append(self._convert_sql_row_to_document(row))
-
         sorted_documents = sorted(documents, key=lambda doc: vector_ids.index(doc.meta["vector_id"]))
         return sorted_documents
 
@@ -405,7 +403,6 @@ def write_documents(
             document_objects = [Document.from_dict(d) if isinstance(d, dict) else d for d in documents]
         else:
             document_objects = documents
-
         document_objects = self._handle_duplicate_documents(
             documents=document_objects, index=index, duplicate_documents=duplicate_documents
         )

tests/transformers/load_subfolder/test_config.py

@@ -15,9 +15,11 @@
 
 from paddlenlp.transformers import AutoConfig, BertConfig, CLIPConfig, T5Config
 from paddlenlp.utils.log import logger
+from tests.testing_utils import slow
 
 
 class ConfigLoadTester(unittest.TestCase):
+    @slow
     def test_bert_config_load(self):
         logger.info("Download Bert Config from PaddleNLP BOS")
         bert_config = BertConfig.from_pretrained("bert-base-uncased", from_hf_hub=False)
@@ -43,6 +45,7 @@ def test_bert_config_load(self):
         bert_config = BertConfig.from_pretrained("aistudio/bert-base-uncased", from_aistudio=True)
         bert_config = AutoConfig.from_pretrained("aistudio/bert-base-uncased", from_aistudio=True)
 
+    @slow
     def test_clip_config_load(self):
         logger.info("Download CLIP Config from PaddleNLP BOS")
         clip_config = CLIPConfig.from_pretrained("openai/clip-vit-base-patch32", from_hf_hub=False)
@@ -68,6 +71,7 @@ def test_clip_config_load(self):
         clip_config = CLIPConfig.from_pretrained("aistudio/clip-vit-base-patch32", from_aistudio=True)
         clip_config = AutoConfig.from_pretrained("aistudio/clip-vit-base-patch32", from_aistudio=True)
 
+    @slow
     def test_t5_config_load(self):
         logger.info("Download T5 Config from PaddleNLP BOS")
         t5_config = T5Config.from_pretrained("t5-small", from_hf_hub=False)

tests/transformers/load_subfolder/test_image_processor.py

@@ -16,9 +16,11 @@
 
 from paddlenlp.transformers import AutoImageProcessor, CLIPImageProcessor
 from paddlenlp.utils.log import logger
+from tests.testing_utils import slow
 
 
 class ImageProcessorLoadTester(unittest.TestCase):
+    @slow
     def test_clip_load(self):
         logger.info("Download model from PaddleNLP BOS")
         clip_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32", from_hf_hub=False)

tests/transformers/load_subfolder/test_model.py

@@ -20,6 +20,7 @@
 
 from paddlenlp.transformers import AutoModel, BertModel, CLIPTextModel, T5Model
 from paddlenlp.utils.log import logger
+from tests.testing_utils import slow
 
 
 class ModelLoadTester(unittest.TestCase):
@@ -58,6 +59,7 @@ def test_cache_dir(
                 else:
                     assert any(".pdparams" in f for f in file_list), "*.pdparams not in cache_dir"
 
+    @slow
     def test_bert_load(self):
         # BOS
         logger.info("Download model from PaddleNLP BOS")
@@ -194,6 +196,7 @@ def test_bert_load(self):
             use_safetensors=False,
         )
 
+    @slow
     def test_bert_load_safe(self):
         # BOS
         logger.info("Download model from PaddleNLP BOS")
@@ -320,6 +323,7 @@ def test_bert_load_safe(self):
             use_safetensors=True,
         )
 
+    @slow
     def test_clip_load(self):
         # BOS
         logger.info("Download model from PaddleNLP BOS")
@@ -466,6 +470,7 @@ def test_clip_load(self):
             use_safetensors=False,
         )
 
+    @slow
     def test_clip_load_safe(self):
         # BOS
         logger.info("Download model from PaddleNLP BOS")
@@ -608,6 +613,7 @@ def test_clip_load_safe(self):
             use_safetensors=True,
         )
 
+    @slow
     def test_t5_load(self):
         # BOS
         logger.info("Download model from PaddleNLP BOS")
@@ -726,6 +732,7 @@ def test_t5_load(self):
             use_safetensors=False,
         )
 
+    @slow
     def test_t5_load_safe(self):
         # BOS
         logger.info("Download model from PaddleNLP BOS")

tests/transformers/load_subfolder/test_processor.py

@@ -17,9 +17,11 @@
 
 from paddlenlp.transformers import AutoProcessor, CLIPProcessor
 from paddlenlp.utils.log import logger
+from tests.testing_utils import slow
 
 
 class ProcessorLoadTester(unittest.TestCase):
+    @slow
     def test_clip_load(self):
         logger.info("Download model from PaddleNLP BOS")
         clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32", from_hf_hub=False)

tests/transformers/load_subfolder/test_tokenizer.py

@@ -22,9 +22,11 @@
     T5Tokenizer,
 )
 from paddlenlp.utils.log import logger
+from tests.testing_utils import slow
 
 
 class TokenizerLoadTester(unittest.TestCase):
+    @slow
     def test_bert_load(self):
         logger.info("Download model from PaddleNLP BOS")
         bert_tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", from_hf_hub=False)
@@ -57,6 +59,7 @@ def test_bert_load(self):
             "aistudio/paddlenlp-test-model", subfolder="bert-base-uncased", from_aistudio=True
         )
 
+    @slow
     def test_clip_load(self):
         logger.info("Download model from PaddleNLP BOS")
         clip_tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32", from_hf_hub=False)
@@ -89,6 +92,7 @@ def test_clip_load(self):
             "aistudio/paddlenlp-test-model", subfolder="clip-vit-base-patch32", from_aistudio=True
         )
 
+    @slow
     def test_t5_load(self):
         logger.info("Download model from PaddleNLP BOS")
         t5_tokenizer = T5Tokenizer.from_pretrained("t5-small", from_hf_hub=False)