Merge 8f5184fbab into 43b7241e67

utils.py

@@ -1,18 +1,139 @@
 import os
-from typing import Dict, Tuple, Union, Optional
+from typing import Dict, Tuple, Union, Optional, List
 
+import torch
 from torch.nn import Module
 from transformers import AutoModel, AutoTokenizer
 from transformers.tokenization_utils import PreTrainedTokenizer
 
 
-def auto_configure_device_map(num_gpus: int) -> Dict[str, int]:
+def calculate_per_gpu_layers(gpu_list: List[int], total_layers: int) -> Dict[int, int]:
+    """
+    Calculate the number of layers to be allocated to each GPU based on the memory ratio.
+
+    Args:
+        gpu_list (List[int]): A list of GPU indices.
+        total_layers (int): The total number of layers in the model.
+
+    Returns:
+        Dict[int, int]: A dictionary mapping GPU indices to the number of layers assigned to each GPU.
+
+    >>> from unittest import mock
+    >>> import torch
+    >>> mock_get_device_properties = mock.Mock()
+    >>> fake_device_properties = lambda gpu: type('', (), {'total_memory': (gpu + 1) * 1024})()
+    >>> mock_get_device_properties.side_effect = fake_device_properties
+    >>> torch.cuda.get_device_properties = mock_get_device_properties
+    >>> calculate_per_gpu_layers([0, 1, 2], 30)
+    {0: 5, 1: 10, 2: 15}
+    """
+
+    # 根据每个GPU的显存大小，计算每个GPU应分配的层数
+    # 获取每个gpu的显存大小
+    gpu_memory_map = {
+        gpu: torch.cuda.get_device_properties(gpu).total_memory
+        for gpu in gpu_list
+    }
+
+    # 计算总显存大小
+    total_memory = sum(gpu_memory_map.values())
+
+    # 计算每个GPU的显存比例
+    gpu_memory_ratios = {
+        gpu: memory / total_memory
+        for gpu, memory in gpu_memory_map.items()
+    }
+
+    # 计算每个 GPU 应分配的层数
+    per_gpu_layers = {
+        gpu: int(round(total_layers * ratio))
+        for gpu, ratio in gpu_memory_ratios.items()
+    }
+
+    # 修正分配误差，确保总层数为total_layers
+    while True:
+        diff = total_layers - sum(per_gpu_layers.values())
+        if diff > 0:
+            gpu_with_max_memory = max(gpu_memory_ratios, key=gpu_memory_ratios.get)
+            per_gpu_layers[gpu_with_max_memory] += diff
+        elif diff < 0:
+            gpu_with_min_memory = min(gpu_memory_ratios, key=gpu_memory_ratios.get)
+            per_gpu_layers[gpu_with_min_memory] -= -diff
+        else:
+            break
+
+    return per_gpu_layers
+
+
+def auto_configure_device_map(num_gpus: int = 2, gpu_list: Optional[List[int]] = None) -> Dict[str, int]:
+    """
+    Automatically configure the device map for model parallelism based on the number of GPUs and their memory ratios.
+
+    Args:
+        num_gpus (int): The number of GPUs to be used.
+        gpu_list (Optional[List[int]]): An optional list of GPU indices. Defaults to None.
+
+    Returns:
+        Dict[str, int]: A dictionary representing the device map for model parallelism.
+
+    >>> from unittest import mock
+    >>> import torch
+    >>> # mock torch.cuda.get_device_properties
+    >>> mock_get_device_properties = mock.Mock()
+    >>> fake_device_properties = lambda gpu: type('', (), {'total_memory': (gpu + 1) * 1024})()
+    >>> mock_get_device_properties.side_effect = fake_device_properties
+    >>> torch.cuda.get_device_properties = mock_get_device_properties
+    >>> # mock torch.cuda.device_count
+    >>> mock_device_count = mock.Mock()
+    >>> mock_device_count.return_value = 3
+    >>> torch.cuda.device_count = mock_device_count
+    >>> for k, v in auto_configure_device_map(3).items():
+    ...     print(f"{k}: {v}")
+    transformer.word_embeddings: 0
+    transformer.final_layernorm: 0
+    lm_head: 0
+    transformer.layers.0: 0
+    transformer.layers.1: 0
+    transformer.layers.2: 0
+    transformer.layers.3: 1
+    transformer.layers.4: 1
+    transformer.layers.5: 1
+    transformer.layers.6: 1
+    transformer.layers.7: 1
+    transformer.layers.8: 1
+    transformer.layers.9: 1
+    transformer.layers.10: 1
+    transformer.layers.11: 1
+    transformer.layers.12: 1
+    transformer.layers.13: 2
+    transformer.layers.14: 2
+    transformer.layers.15: 2
+    transformer.layers.16: 2
+    transformer.layers.17: 2
+    transformer.layers.18: 2
+    transformer.layers.19: 2
+    transformer.layers.20: 2
+    transformer.layers.21: 2
+    transformer.layers.22: 2
+    transformer.layers.23: 2
+    transformer.layers.24: 2
+    transformer.layers.25: 2
+    transformer.layers.26: 2
+    transformer.layers.27: 2
+    """
+
     # transformer.word_embeddings 占用1层
     # transformer.final_layernorm 和 lm_head 占用1层
     # transformer.layers 占用 28 层
     # 总共30层分配到num_gpus张卡上
     num_trans_layers = 28
-    per_gpu_layers = 30 / num_gpus
+
+    if gpu_list is None:
+        gpu_list = list(range(num_gpus))
+    assert len(gpu_list) <= torch.cuda.device_count(), "分配的GPU数量超过了实际可用的GPU数量"
+
+    # 获取每个gpu的承载的层数
+    per_gpu_layer_dict = calculate_per_gpu_layers(gpu_list, total_layers=num_trans_layers + 2)
 
     # bugfix: 在linux中调用torch.embedding传入的weight,input不在同一device上,导致RuntimeError
     # windows下 model.device 会被设置成 transformer.word_embeddings.device
@@ -20,33 +141,58 @@ def auto_configure_device_map(num_gpus: int) -> Dict[str, int]:
     # 在调用chat或者stream_chat时,input_ids会被放到model.device上
     # 如果transformer.word_embeddings.device和model.device不同,则会导致RuntimeError
     # 因此这里将transformer.word_embeddings,transformer.final_layernorm,lm_head都放到第一张卡上
-    device_map = {'transformer.word_embeddings': 0,
-                  'transformer.final_layernorm': 0, 'lm_head': 0}
+    current_gpu_index = 0
+    current_gpu = gpu_list[current_gpu_index]
+    device_map = {
+        'transformer.word_embeddings': current_gpu,
+        'transformer.final_layernorm': current_gpu,
+        'lm_head': current_gpu
+    }
 
     used = 2
-    gpu_target = 0
-    for i in range(num_trans_layers):
-        if used >= per_gpu_layers:
-            gpu_target += 1
-            used = 0
-        assert gpu_target < num_gpus
-        device_map[f'transformer.layers.{i}'] = gpu_target
-        used += 1
 
+    # 分配剩余的层数
+    for i in range(num_trans_layers):
+        if used < per_gpu_layer_dict[current_gpu]:
+            used += 1
+        else:
+            # 当前 GPU 的层数已分配完，切换到下一个 GPU
+            current_gpu_index += 1
+            current_gpu = gpu_list[current_gpu_index]
+            used = 1
+
+        device_map[f"transformer.layers.{i}"] = current_gpu
     return device_map
 
 
 def load_model_on_gpus(checkpoint_path: Union[str, os.PathLike], num_gpus: int = 2,
+                       gpu_list: Optional[List[int]] = None,
                        multi_gpu_model_cache_dir: Union[str, os.PathLike] = "./temp_model_dir",
                        device_map: Optional[Dict[str, int]] = None,
                        tokenizer: Optional[PreTrainedTokenizer] = None, **kwargs) -> Module:
+    """
+    Load a pretrained model on multiple GPUs.
+
+    Args:
+        checkpoint_path (Union[str, os.PathLike]): The path to the checkpoint or model directory.
+        num_gpus (int, optional): The number of GPUs to use. Defaults to 2.
+        gpu_list (Optional[List[int]], optional): A list of GPU indices. Defaults to None.
+        multi_gpu_model_cache_dir (Union[str, os.PathLike], optional): A directory to cache the multi-GPU model.
+        device_map (Optional[Dict[str, int]], optional): A dictionary representing the device map for model parallelism.
+        tokenizer (Optional[PreTrainedTokenizer], optional): The tokenizer to be used with the model. Defaults to None.
+        **kwargs: Additional keyword arguments for loading the model.
+
+    Returns:
+        Module: The pretrained model on multiple GPUs.
+    """
+
     from accelerate import load_checkpoint_and_dispatch
 
     model = AutoModel.from_pretrained(checkpoint_path, trust_remote_code=True, **kwargs)
     model = model.eval()
 
     if device_map is None:
-        device_map = auto_configure_device_map(num_gpus)
+        device_map = auto_configure_device_map(num_gpus, gpu_list)
     try:
         model = load_checkpoint_and_dispatch(
             model, checkpoint_path, device_map=device_map, offload_folder="offload", offload_state_dict=True).half()
@@ -69,13 +215,28 @@ def load_model_on_gpus(checkpoint_path: Union[str, os.PathLike], num_gpus: int =
 
 def load_model_and_tokenizer(checkpoint_path: Union[str, os.PathLike], num_gpus: int = 1,
                              multi_gpu_model_cache_dir: Union[str, os.PathLike] = "./temp_model_dir",
+                             gpu_list: Optional[List[int]] = None,
                              **kwargs) -> Tuple[Module, PreTrainedTokenizer]:
+    """
+    Load a pretrained model and its tokenizer.
+
+    Args:
+        checkpoint_path (Union[str, os.PathLike]): The path to the checkpoint or model directory.
+        num_gpus (int, optional): The number of GPUs to use. Defaults to 1.
+        multi_gpu_model_cache_dir (Union[str, os.PathLike], optional): A directory to cache the multi-GPU model.
+        gpu_list (Optional[List[int]], optional): A list of GPU indices. Defaults to None.
+        **kwargs: Additional keyword arguments for loading the model and tokenizer.
+
+    Returns:
+        Tuple[Module, PreTrainedTokenizer]: A tuple containing the loaded model and tokenizer.
+    """
+
     tokenizer = AutoTokenizer.from_pretrained(checkpoint_path, trust_remote_code=True, **kwargs)
     if num_gpus < 2:
         model = AutoModel.from_pretrained(checkpoint_path, trust_remote_code=True, **kwargs).half().cuda()
         model = model.eval()
     else:
-        model = load_model_on_gpus(checkpoint_path, num_gpus=num_gpus,
+        model = load_model_on_gpus(checkpoint_path, num_gpus=num_gpus, gpu_list=gpu_list,
                                    multi_gpu_model_cache_dir=multi_gpu_model_cache_dir,
                                    tokenizer=tokenizer, **kwargs)
     return model, tokenizer