Merge pull request #1349 from 545999961/master

545999961 · web-flow · commit f9f673e4ff15 · 2025-01-22T16:43:55.000+08:00
Fix Bug: OOM
diff --git a/FlagEmbedding/abc/inference/AbsEmbedder.py b/FlagEmbedding/abc/inference/AbsEmbedder.py
@@ -36,6 +36,7 @@ class AbsEmbedder(ABC):
         passage_max_length (int, optional): Maximum length for passage. Defaults to :data:`512`.
         convert_to_numpy (bool, optional): If True, the output embedding will be a Numpy array. Otherwise, it will be a Torch Tensor. 
             Defaults to :data:`True`.
+        multi_GPU_type (str): The type of multi-GPU inference. Defaults to :data:`"dp"`. You can choose ['dp', 'multi_process'].
         kwargs (Dict[Any], optional): Additional parameters for HuggingFace Transformers config or children classes.
     """
 
@@ -52,6 +53,7 @@ def __init__(
         query_max_length: int = 512,
         passage_max_length: int = 512,
         convert_to_numpy: bool = True,
+        multi_GPU_type: str = 'dp',
         **kwargs: Any,
     ):
         query_instruction_format = query_instruction_format.replace('\\n', '\n')
@@ -66,6 +68,8 @@ def __init__(
         self.query_max_length = query_max_length
         self.passage_max_length = passage_max_length
         self.convert_to_numpy = convert_to_numpy
+        self._multi_GPU_type = multi_GPU_type
+        self._dp_set = False
 
         for k in kwargs:
             setattr(self, k, kwargs[k])
@@ -77,6 +81,21 @@ def __init__(
         self.model = None
         self.pool = None
     
+    def start_dp(self):
+        if self._multi_GPU_type == 'dp' and \
+            (isinstance(self.target_devices, list) and len(self.target_devices) > 1) and \
+            (isinstance(self.target_devices[0], int) or 'cuda' in self.target_devices[0]) and \
+            self._dp_set == False:
+
+            if self.use_fp16: self.model.half()
+            self.model = self.model.to(torch.device("cuda"))
+            if isinstance(self.target_devices[0], int):
+                self.model = torch.nn.DataParallel(self.model, device_ids = self.target_devices)
+            else:
+                devices = [int(e.split(':')[-1].strip()) for e in self.target_devices]
+                self.model = torch.nn.DataParallel(self.model, device_ids = devices)
+            self._dp_set = True
+
     def stop_self_pool(self):
         if self.pool is not None:
             self.stop_multi_process_pool(self.pool)
@@ -107,7 +126,10 @@ def get_target_devices(devices: Union[str, int, List[str], List[int]]) -> List[s
             elif is_torch_npu_available():
                 return [f"npu:{i}" for i in range(torch.npu.device_count())]
             elif torch.backends.mps.is_available():
-                return [f"mps:{i}" for i in range(torch.mps.device_count())]
+                try:
+                    return [f"mps:{i}" for i in range(torch.mps.device_count())]
+                except:
+                    return ["mps"]
             else:
                 return ["cpu"]
         elif isinstance(devices, str):
@@ -253,6 +275,15 @@ def encode(
                 device=self.target_devices[0],
                 **kwargs
             )
+        
+        if self._multi_GPU_type == 'dp':
+            return self.encode_only(
+                sentences,
+                batch_size=batch_size,
+                max_length=max_length,
+                convert_to_numpy=convert_to_numpy,
+                **kwargs
+            )
 
         if self.pool is None:
             self.pool = self.start_multi_process_pool(AbsEmbedder._encode_multi_process_worker)
@@ -262,6 +293,7 @@ def encode(
             batch_size=batch_size,
             max_length=max_length,
             convert_to_numpy=convert_to_numpy,
+            device=torch.device("cuda"),
             **kwargs
         )
         return embeddings
@@ -284,6 +316,20 @@ def encode_single_device(
         """
         pass
 
+    def encode_only(
+        self,
+        sentences: Union[List[str], str],
+        batch_size: int = 256,
+        max_length: int = 512,
+        convert_to_numpy: bool = True,
+        device: Any = None,
+        **kwargs: Any,
+    ):
+        """
+        This method should encode sentences and return embeddings on a single device.
+        """
+        pass
+
     # adapted from https://github.com/UKPLab/sentence-transformers/blob/1802076d4eae42ff0a5629e1b04e75785d4e193b/sentence_transformers/SentenceTransformer.py#L807
     def start_multi_process_pool(
         self,
diff --git a/FlagEmbedding/inference/embedder/decoder_only/base.py b/FlagEmbedding/inference/embedder/decoder_only/base.py
@@ -257,7 +257,7 @@ def encode_single_device(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         # encode
diff --git a/FlagEmbedding/inference/embedder/decoder_only/icl.py b/FlagEmbedding/inference/embedder/decoder_only/icl.py
@@ -409,7 +409,7 @@ def encode_queries_single_device(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         # encode
@@ -519,7 +519,7 @@ def encode_single_device(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         # encode
diff --git a/FlagEmbedding/inference/embedder/encoder_only/base.py b/FlagEmbedding/inference/embedder/encoder_only/base.py
@@ -196,6 +196,38 @@ def encode_single_device(
         if self.use_fp16: self.model.half()
 
         self.model.to(device)
+        
+        return self.encode_only(
+            sentences=sentences,
+            batch_size=batch_size,
+            max_length=max_length,
+            convert_to_numpy=convert_to_numpy,
+            device=device,
+            **kwargs
+        )
+    
+    @torch.no_grad()
+    def encode_only(
+        self,
+        sentences: Union[List[str], str],
+        batch_size: int = 256,
+        max_length: int = 512,
+        convert_to_numpy: bool = True,
+        device: Any = None,
+        **kwargs: Any
+    ):
+        """Encode input sentences.
+
+        Args:
+            sentences (Union[List[str], str]): Input sentences to encode.
+            batch_size (int, optional): Number of sentences for each iter. Defaults to :data:`256`.
+            max_length (int, optional): Maximum length of tokens. Defaults to :data:`512`.
+            convert_to_numpy (bool, optional): If True, the output embedding will be a Numpy array. Otherwise, it will 
+                be a Torch Tensor. Defaults to :data:`True`.
+
+        Returns:
+            Union[torch.Tensor, np.ndarray]: return the embedding vectors in a numpy array or tensor.
+        """
         self.model.eval()
 
         input_was_string = False
@@ -238,7 +270,7 @@ def encode_single_device(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         # encode
diff --git a/FlagEmbedding/inference/embedder/encoder_only/m3.py b/FlagEmbedding/inference/embedder/encoder_only/m3.py
@@ -406,7 +406,7 @@ def _process_colbert_vecs(colbert_vecs: np.ndarray, attention_mask: list):
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         # encode
diff --git a/FlagEmbedding/inference/reranker/decoder_only/base.py b/FlagEmbedding/inference/reranker/decoder_only/base.py
@@ -412,7 +412,7 @@ def compute_score_single_gpu(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         dataset, dataloader = None, None
diff --git a/FlagEmbedding/inference/reranker/decoder_only/layerwise.py b/FlagEmbedding/inference/reranker/decoder_only/layerwise.py
@@ -282,7 +282,7 @@ def compute_score_single_gpu(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         dataset, dataloader = None, None
diff --git a/FlagEmbedding/inference/reranker/decoder_only/lightweight.py b/FlagEmbedding/inference/reranker/decoder_only/lightweight.py
@@ -368,7 +368,7 @@ def compute_score_single_gpu(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         all_scores = []
diff --git a/FlagEmbedding/inference/reranker/encoder_only/base.py b/FlagEmbedding/inference/reranker/encoder_only/base.py
@@ -169,7 +169,7 @@ def compute_score_single_gpu(
                 flag = True
             except RuntimeError as e:
                 batch_size = batch_size * 3 // 4
-            except torch.OutOfMemoryError as e:
+            except torch.cuda.OutOfMemoryError as e:
                 batch_size = batch_size * 3 // 4
 
         all_scores = []
