fix layer grad_norm with pp

internlm/core/context/parallel_context.py

@@ -150,7 +150,7 @@ class ParallelContext(metaclass=SingletonMeta):
         self.virtual_pipeline_parallel_size = None
         self.virtual_pipeline_parallel_rank = None
         self._expert_parallel_group_names = []
-        self.layer_names = {"unknown", "embedding", "norm", "head"}
+        self.layer_names = ["unknown"]
 
     @property
     def config(self):

internlm/model/modeling_internlm.py

@@ -24,6 +24,7 @@ from internlm.solver.pipeline_utils import partition_uniform
 from internlm.utils.checkpoint import activation_checkpoint
 from internlm.utils.common import filter_kwargs
 from internlm.utils.logger import get_logger
+from internlm.utils.parallel import set_model_params_layer_name
 from internlm.utils.registry import MODEL_INITIALIZER
 
 MODEL_TYPE = "INTERNLM"
@@ -418,6 +419,21 @@ def _build_generic_model_1d(num_layers, num_chunks, device=torch.device("cuda"),
     if gpc.is_rank_for_log():
         logger.info(f"The layer sharding is {all_parts}.")
 
+    # config gpc.layer_name
+    # get names of first and last layers
+    kwargs["num_layers"] = 1
+    kwargs["device"] = device
+    kwargs["first"] = True
+    kwargs["last"] = True
+    kwargs["start_layer_idx"] = 0
+    tmp_chunk = PackedFlashInternLm1D(**filter_kwargs(PackedFlashInternLm1D.__init__, kwargs)).cpu()
+    # get names of middle layers
+    for idx in range(num_layers):
+        layer_name = f"{PackedFlashBaseLayer1D.__name__}.{idx}"
+        gpc.layer_names.append(layer_name)
+    set_model_params_layer_name(tmp_chunk)
+    torch.cuda.empty_cache()
+
     models = []
 
     for start, end in parts:

internlm/model/modeling_moe.py

@@ -26,6 +26,7 @@ from internlm.solver.pipeline_utils import partition_uniform
 from internlm.utils.checkpoint import activation_checkpoint
 from internlm.utils.common import filter_kwargs
 from internlm.utils.logger import get_logger
+from internlm.utils.parallel import set_model_params_layer_name
 from internlm.utils.registry import MODEL_INITIALIZER
 
 MODEL_TYPE = "INTERNLM_MoE"
@@ -516,6 +517,21 @@ def _build_generic_model_1d(num_layers, num_chunks, device=torch.device("cuda"),
     if gpc.is_rank_for_log():
         logger.info(f"The layer sharding is {all_parts}.")
 
+    # config gpc.layer_name
+    # get names of first and last layers
+    kwargs["num_layers"] = 1
+    kwargs["device"] = device
+    kwargs["first"] = True
+    kwargs["last"] = True
+    kwargs["start_layer_idx"] = 0
+    tmp_chunk = PackedFlashInternLm1D(**filter_kwargs(PackedFlashInternLm1D.__init__, kwargs)).cpu()
+    # get names of middle layers
+    for idx in range(num_layers):
+        layer_name = f"{PackedFlashBaseLayer1D.__name__}.{idx}"
+        gpc.layer_names.append(layer_name)
+    set_model_params_layer_name(tmp_chunk)
+    torch.cuda.empty_cache()
+
     models = []
 
     for start, end in parts:

internlm/utils/parallel.py

@@ -78,14 +78,16 @@ def set_model_params_layer_name(model):
         if isinstance(_chunk, NaiveAMPModel):
             _chunk = _chunk.model
         # Create a unique layer name based on the block's class name and index
-        for name, children in _chunk.named_children():
+        for _, children in _chunk.named_children():
             if isinstance(children, nn.ModuleList):
                 for idx, block in enumerate(children):
                     for param in block.parameters():
                         layer_name = f"{block.__class__.__name__}.{idx}"
-                        gpc.layer_names.add(layer_name)
+                        gpc.layer_names.append(layer_name)
                         param.__setattr__("layer_name", layer_name)
             else:
                 for param in children.parameters():
-                    gpc.layer_names.add(name)
-                    param.__setattr__("layer_name", name)
+                    layer_name = f"{children.__class__.__name__}"
+                    gpc.layer_names.append(layer_name)
+                    param.__setattr__("layer_name", layer_name)
+    gpc.layer_names = sorted(set(gpc.layer_names))