Merge pull request #216 from Modalities/fix_num_params_hybrid_sharding

fix: Computation of  Number of Parameters in HYBRID Sharding

Author: mali-git

src/modalities/util.py

@@ -76,6 +76,16 @@ def get_total_number_of_trainable_parameters(model: FSDP) -> Number:
     num_params_tensor = torch.tensor(num_params).cuda()
     dist.all_reduce(num_params_tensor, op=dist.ReduceOp.SUM)
     total_num_params = num_params_tensor.item()
+    # For HYBRID sharding, divide by sharding factor to get the correct number of parameters
+    # TODO: Define constant instead of hardcoding string
+    if model.sharding_strategy.name == "HYBRID_SHARD":
+        # Assumes that CUDA is available and each node has the same number of GPUs
+        # Note: Per default FSDP constructs process groups for the user to shard intra-node and replicate inter-node.
+        # However, users can also provide their own sharding process groups (currently not supported in Modalities)
+        # which would require to adapt the code.
+        sharding_factor_hybrid_sharding = dist.get_world_size() // torch.cuda.device_count()
+        total_num_params = total_num_params // sharding_factor_hybrid_sharding
+
     return total_num_params
 
 

tests/test_utils.py

@@ -1,3 +1,5 @@
+from enum import Enum
+
 import torch
 
 import modalities
@@ -44,12 +46,19 @@ def test_get_total_number_of_trainable_parameters():
     model = torch.nn.Sequential(torch.nn.Linear(10, 5), torch.nn.ReLU(), torch.nn.Linear(5, 2))
 
     # Calculate the expected number of trainable parameters
-    expected_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    expected_params = 10*5 + 5 + 5*2 + 2  # weights_1 + bias_1 + weights_2 + bias_2 = 67
+    world_size = 8
+    num_gpus_per_node = 4
 
     # Create a mock FSDP model
     class MockFSDP:
+        class ShardingStrategy(Enum):
+            FULL_SHARD = "FULL_SHARD"
+            HYBRID_SHARD = "HYBRID_SHARD"
+
         def __init__(self, model):
             self.model = model
+            self.sharding_strategy = self.ShardingStrategy.FULL_SHARD
 
     fsdp_model = MockFSDP(model)
 
@@ -61,11 +70,29 @@ def mock_all_reduce(tensor, op):
     def mock_cuda(tensor):
         return tensor
 
+    def mock_world_size():
+        return world_size
+
+    def mock_device_count():
+        return num_gpus_per_node
+
     def mock_get_local_number_of_trainable_parameters(model: MockFSDP):
-        return get_local_number_of_trainable_parameters(model.model)
+        if model.sharding_strategy == MockFSDP.ShardingStrategy.FULL_SHARD:
+            return get_local_number_of_trainable_parameters(model.model)
+        elif model.sharding_strategy == MockFSDP.ShardingStrategy.HYBRID_SHARD:
+            sharding_factor = world_size // num_gpus_per_node
+            return sharding_factor * get_local_number_of_trainable_parameters(model.model)
+        else:
+            raise ValueError(f"Sharding strategy {model.sharding_strategy} not supported.")
 
     modalities.util.get_local_number_of_trainable_parameters = mock_get_local_number_of_trainable_parameters
     torch.distributed.all_reduce = mock_all_reduce
+    torch.distributed.get_world_size = mock_world_size
+    torch.cuda.device_count = mock_device_count
     torch.Tensor.cuda = mock_cuda
 
     assert get_total_number_of_trainable_parameters(fsdp_model) == expected_params
+
+    fsdp_model.sharding_strategy = MockFSDP.ShardingStrategy.HYBRID_SHARD
+    modalities.util.get_local_number_of_trainable_parameters = mock_get_local_number_of_trainable_parameters
+    assert get_total_number_of_trainable_parameters(fsdp_model) == expected_params