feat(optimizer): add layer norm to tensorboard (#429)

* add layer norm to tensorboard

* test moe layer norm

* add function: reduce grads

internlm/solver/optimizer/hybrid_zero_optim.py

@@ -34,7 +34,7 @@ from internlm.utils.megatron_timers import megatron_timer as timer
 from internlm.utils.timeout import llm_timeout
 
 from .base_optimizer import BaseOptimizer
-from .utils import compute_norm
+from .utils import compute_layer_norm, compute_norm, compute_param_norm
 
 inf = math.inf
 logger = get_logger(__file__)
@@ -520,6 +520,29 @@ class HybridZeroOptimizer(BaseOptimizer):
 
         return norm
 
+    def _compute_param_norm_stage(
+        self, group_id: int = 0, last_bucket: bool = False, last_stage: bool = False, previous_param_norms=None
+    ):
+        # compute norm for gradients that have been reduced
+        params, grads = self._param_store.get_reduced_param_for_compute_norm(group_id=group_id, last_bucket=last_bucket)
+
+        total_param_norms = {}
+        if len(params) == 0:
+            dtype = self.param_groups[group_id]["dtype"]
+            grads = [self.padding_grad.to(dtype)]
+            params = [self.padding_tensor.to(dtype)]
+
+        if self._clip_grad_norm > 0:
+            total_param_norms = compute_param_norm(
+                grads,
+                params,
+                last_stage=last_stage,
+                previous_param_norms=previous_param_norms,
+                zero_mode=self._broadcast_parallel_mode[group_id],
+                is_moe_group=self._is_moe_group(self.optim.param_groups[group_id]),
+            )
+        return total_param_norms
+
     @llm_timeout(func_name="optim_step")
     def step(self, closure=None):
         """Performs a single optimization step.
@@ -547,8 +570,11 @@ class HybridZeroOptimizer(BaseOptimizer):
 
         # compute norm for gradients in the before bucket
         groups_norms = []
+        groups_param_norms = []
         for group_id in range(self.num_param_groups):
             groups_norms.append(self._compute_norm_with_stage(group_id=group_id))
+            if gpc.config.get("grad_norm_profiling", False):
+                groups_param_norms.append(self._compute_param_norm_stage(group_id=group_id))
 
         # clear reduced grads
         # grads in the last bucket is reduced
@@ -561,6 +587,8 @@ class HybridZeroOptimizer(BaseOptimizer):
 
         # compute norm for gradients in the last bucket
         total_norms = {}
+        total_param_norms = {}
+        total_layer_norms = {}
         for group_id in range(self.num_param_groups):
             group_name = self.param_groups[group_id]["name"] if "name" in self.param_groups[group_id] else "default"
             group_name = f"{group_id}_{group_name}"
@@ -570,6 +598,16 @@ class HybridZeroOptimizer(BaseOptimizer):
                 last_stage=True,
                 previous_norm=groups_norms[group_id],
             )
+            if gpc.config.get("grad_norm_profiling", False):
+                param_norms = self._compute_param_norm_stage(
+                    group_id=group_id,
+                    last_bucket=True,
+                    last_stage=True,
+                    previous_param_norms=groups_param_norms[group_id],
+                )
+                total_layer_norms[group_name], total_param_norms[group_name] = compute_layer_norm(
+                    param_norms=param_norms, loss_scale=self.loss_scale.item()
+                )
 
             # Need to allreduce(avg) the norms across different ranks because moe params will not be synced
             # during allreduce
@@ -585,7 +623,12 @@ class HybridZeroOptimizer(BaseOptimizer):
         self._sync_grad()
         timer("sync_grad").stop()
 
-        return self._step(closure=closure, norms=total_norms)
+        state, global_norms = self._step(closure=closure, norms=total_norms)
+        if gpc.config.get("grad_norm_profiling", False):
+            global_norms["layer_norms"] = total_layer_norms
+            global_norms["param_norms"] = total_param_norms
+
+        return state, global_norms
 
     def _step(self, closure=None, norms=None):
         assert closure is None, "closure is not supported by step()"

internlm/solver/optimizer/utils.py

@@ -15,7 +15,7 @@ from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
 from internlm.core.context import ParallelMode
 from internlm.core.context import global_context as gpc
 from internlm.core.naive_amp import NaiveAMPModel
-from internlm.utils.common import get_tensor_norm, move_norm_to_cuda
+from internlm.utils.common import get_current_device, get_tensor_norm, move_norm_to_cuda
 from internlm.utils.logger import get_logger
 from internlm.utils.parallel import is_model_parallel_parameter
 
@@ -209,6 +209,49 @@ def calc_lp(grads, norm_type):
     return norm
 
 
+def reduce_grads(gradients, parameters, fine_grained=False):
+    parallel_grads = []
+    if fine_grained:
+        parallel_grads = {}
+
+    def append_grad(g, p):
+        if fine_grained:
+            param_name = p.param_name if hasattr(p, "param_name") else "unknown-padding"
+            if param_name not in parallel_grads:
+                parallel_grads[param_name] = []
+            parallel_grads[param_name].append(g.data.float())
+        else:
+            parallel_grads.append(g.data.float())
+
+    for g, p in zip(gradients, parameters):
+        # TODO: consider the pipeline shared parameter
+        if (
+            gpc.is_initialized(ParallelMode.PIPELINE)
+            and hasattr(p, "pipeline_shared_module_pg")
+            and dist.get_rank(p.pipeline_shared_module_pg) == 0
+        ):  # if shared between different pipe, only count o
+            append_grad(g, p)
+        elif (
+            gpc.is_initialized(ParallelMode.PIPELINE)
+            and hasattr(p, "pipeline_shared_module_pg")
+            and dist.get_rank(p.pipeline_shared_module_pg) != 0
+        ):
+            continue
+        elif (
+            gpc.is_initialized(ParallelMode.TENSOR)
+            and not is_model_parallel_parameter(p)
+            and gpc.get_local_rank(ParallelMode.TENSOR) == 0
+        ):  # if not used in each chunk, such as layernorm
+            append_grad(g, p)
+        elif is_model_parallel_parameter(p):
+            append_grad(g, p)
+        elif gpc.get_local_rank(ParallelMode.TENSOR) != 0:
+            continue
+        else:
+            raise RuntimeError("Should not arrive here")
+    return parallel_grads
+
+
 def compute_norm(
     gradients, parameters, last_stage=False, previous_norm=None, norm_type=2, zero_mode=ParallelMode.ZERO1
 ):
@@ -247,33 +290,7 @@ def compute_norm(
             )
         total_norm = total_norm_cuda[0].item()
     else:
-        tensor_parallel_grads = []
-        for g, p in zip(gradients, parameters):
-            # TODO: consider the pipeline shared parameter
-            if (
-                gpc.is_initialized(ParallelMode.PIPELINE)
-                and hasattr(p, "pipeline_shared_module_pg")
-                and dist.get_rank(p.pipeline_shared_module_pg) == 0
-            ):  # if shared between different pipe, only count o
-                tensor_parallel_grads.append(g.data.float())
-            elif (
-                gpc.is_initialized(ParallelMode.PIPELINE)
-                and hasattr(p, "pipeline_shared_module_pg")
-                and dist.get_rank(p.pipeline_shared_module_pg) != 0
-            ):
-                continue
-            elif (
-                gpc.is_initialized(ParallelMode.TENSOR)
-                and not is_model_parallel_parameter(p)
-                and gpc.get_local_rank(ParallelMode.TENSOR) == 0
-            ):  # if not used in each chunk, such as layernorm
-                tensor_parallel_grads.append(g.data.float())
-            elif is_model_parallel_parameter(p):
-                tensor_parallel_grads.append(g.data.float())
-            elif gpc.get_local_rank(ParallelMode.TENSOR) != 0:
-                continue
-            else:
-                raise RuntimeError("Should not arrive here")
+        tensor_parallel_grads = reduce_grads(gradients, parameters)
 
         if norm_type == 2.0 and enable_cuda_kernels:
             tensor_parallel_norm = calc_l2_norm(tensor_parallel_grads) ** norm_type
@@ -319,6 +336,124 @@ def compute_norm(
     return total_norm
 
 
+def compute_param_norm(
+    gradients,
+    parameters,
+    last_stage=False,
+    previous_param_norms=None,
+    norm_type=2,
+    zero_mode=ParallelMode.ZERO1,
+    is_moe_group=False,
+):
+    """Get the norm of params
+    Arguments:
+        gradients (Iterable[Tensor]): The gradient value.
+        parameters (Iterable[Tensor]): The parameter each gradient corresponds to.
+        norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for
+            infinity norm.
+
+    Returns:
+        The norm of the parameters.
+    """
+    enable_cuda_kernels = gradients[0].device.type == "cuda"
+    # Norm parameters.
+    norm_type = float(norm_type)
+    total_param_norms = {}
+
+    param_grads = reduce_grads(gradients, parameters, fine_grained=True)
+
+    param_norms = {}
+    for param_name, grads in param_grads.items():
+        if norm_type == inf:
+            param_norm = max(g.data.abs().max() for g in grads)
+        elif norm_type == 2.0 and enable_cuda_kernels:
+            param_norm = calc_l2_norm(grads) ** norm_type
+        else:
+            param_norm = calc_lp(grads, norm_type)
+        param_norms[param_name] = param_norm.item() if torch.is_tensor(param_norm) else param_norm
+
+    if last_stage is False:
+        return param_norms
+
+    if previous_param_norms is not None:
+        for key, value in previous_param_norms.items():
+            if key not in param_norms:
+                param_norms[key] = value
+                continue
+
+            if norm_type == inf:
+                param_norms[key] = max(param_norms[key], value)
+            else:
+                param_norms[key] += value
+
+    # model parallel
+    model_parallel_param_norms = {}
+    if gpc.is_initialized(ParallelMode.MODEL):
+        parallel_param_norms = [None for _ in range(gpc.get_world_size(ParallelMode.MODEL))]
+        dist.all_gather_object(parallel_param_norms, param_norms, group=gpc.get_group(ParallelMode.MODEL))
+        for local_param_norm in parallel_param_norms:
+            for param_name, param_norm in local_param_norm.items():
+                if param_name not in model_parallel_param_norms:
+                    model_parallel_param_norms[param_name] = 0.0
+                if norm_type == inf:
+                    model_parallel_param_norms[param_name] = max(model_parallel_param_norms[param_name], param_norm)
+                else:
+                    model_parallel_param_norms[param_name] += param_norm
+
+    # zero parallel
+    zero_param_norms = [None for _ in range(gpc.get_world_size(zero_mode))]
+    dist.all_gather_object(zero_param_norms, model_parallel_param_norms, group=gpc.get_group(zero_mode))
+    for local_param_norm in zero_param_norms:
+        for param_name, param_norm in local_param_norm.items():
+            if param_name not in total_param_norms:
+                total_param_norms[param_name] = 0.0
+            if norm_type == inf:
+                total_param_norms[param_name] = max(total_param_norms[param_name], param_norm)
+            else:
+                total_param_norms[param_name] += param_norm
+
+    # moe
+    if is_moe_group:
+        pg = gpc.get_group(ParallelMode.EXPERT)
+        scaled_param_norm = torch.cuda.FloatTensor(list(total_param_norms.values()), device=get_current_device())
+        scaled_param_norm = scaled_param_norm / float(gpc.get_world_size(ParallelMode.EXPERT))
+        dist.all_reduce(scaled_param_norm, group=pg)
+        for i, param_name in enumerate(total_param_norms.keys()):
+            total_param_norms[param_name] = scaled_param_norm[i].item()
+
+    # scale
+    for param_name, param_norm in total_param_norms.items():
+        if param_norm in (inf, -inf):
+            total_param_norms[param_name] = -1
+        elif math.isnan(param_norm):
+            total_param_norms[param_name] = -2
+
+    return total_param_norms
+
+
+def compute_layer_norm(param_norms, loss_scale):
+    """
+    compute layer norm by parameter norms
+    """
+    param_norms_groupby_layer = {}
+    layer_norms = {}
+
+    for param_name, param_norm in param_norms.items():
+        layer_name, param_key = param_name.split("-")
+        if layer_name not in param_norms_groupby_layer:
+            param_norms_groupby_layer[layer_name] = {}
+        if layer_name not in layer_norms:
+            layer_norms[layer_name] = 0.0
+
+        if param_norm not in (-1, -2):
+            param_norm = param_norm**0.5 / loss_scale
+
+        param_norms_groupby_layer[layer_name][param_key] = param_norm
+        layer_norms[layer_name] += param_norm
+
+    return layer_norms, param_norms_groupby_layer
+
+
 class BaseGradScaler(ABC):
     """A base class for the gradient scaler.
 

internlm/train/training_internlm.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python
 # -*- encoding: utf-8 -*-
 
+import copy
 import functools
 import time
 from functools import partial
@@ -52,7 +53,11 @@ from internlm.train.utils import create_param_groups
 from internlm.utils.common import DummyProfile
 from internlm.utils.logger import get_logger
 from internlm.utils.megatron_timers import megatron_timer as timer
-from internlm.utils.parallel import sync_model_param, sync_model_param_within_tp
+from internlm.utils.parallel import (
+    set_model_params_layer_name,
+    sync_model_param,
+    sync_model_param_within_tp,
+)
 from internlm.utils.registry import MODEL_INITIALIZER
 from internlm.utils.timeout import llm_timeout
 
@@ -154,6 +159,10 @@ def initialize_optimizer(model: Union[nn.Module, nn.ModuleList]):
     Returns:
         A tuple of (optimizer, beta2_scheduler, lr_scheduler).
     """
+    if gpc.config.get("grad_norm_profiling", False):
+        # set the layer name as an attribute of the model parameters
+        set_model_params_layer_name(model)
+
     if gpc.config.hybrid_zero_optimizer.overlap_sync_param:
         param_bcast_sync_handler = ParamBcastSyncHandler(model)
     else:
@@ -518,6 +527,21 @@ def record_current_batch_training_metrics(
         for key, value in acc_perplex.items():
             infos[key] = value
 
+        if gpc.config.get("grad_norm_profiling", False):
+            layer_norms = copy.deepcopy(grad_norm["layer_norms"])
+            param_norms = copy.deepcopy(grad_norm["param_norms"])
+            for group_name, value in layer_norms.items():
+                if value:
+                    title = f"laye_norm_group_{group_name}"
+                    writer.add_scalars(key=title, value=value, step=train_state.step_count)
+            for group_name, layer_group in param_norms.items():
+                if layer_group:
+                    for layer_name, param_group in layer_group.items():
+                        title = f"param_norm_{layer_name}_{group_name}"
+                        writer.add_scalars(key=title, value=param_group, step=train_state.step_count)
+            del grad_norm["layer_norms"]
+            del grad_norm["param_norms"]
+
         line = ""
         for key, value in infos.items():
             line += f"{key}={value} "

internlm/utils/parallel.py

@@ -2,9 +2,11 @@
 # -*- encoding: utf-8 -*-
 
 import torch.distributed as dist
+from torch import nn
 
 from internlm.core.context import IS_TENSOR_PARALLEL, ParallelMode
 from internlm.core.context import global_context as gpc
+from internlm.core.naive_amp import NaiveAMPModel
 
 
 def is_model_parallel_parameter(p):
@@ -61,3 +63,31 @@ def get_parallel_log_file_name():
         f"tp={gpc.get_local_rank(ParallelMode.TENSOR)}_pp={gpc.get_local_rank(ParallelMode.PIPELINE)}"
     )
     return log_file_name
+
+
+def set_model_params_layer_name(model):
+    r"""Set the layer name as an attribute of the model parameters.
+    Args:
+        model (:class:`torch.nn.Module`): A pyTorch model on whose parameters you check the consistency.
+    """
+    if not isinstance(model, nn.ModuleList):
+        model = [model]
+
+    for _chunk in model:
+        if isinstance(_chunk, NaiveAMPModel):
+            _chunk = _chunk.model
+        # Create a unique layer name based on the block's class name and index
+        for _, children in _chunk.named_children():
+            if isinstance(children, nn.ModuleList):
+                for idx, block in enumerate(children):
+                    for param_name, param in block.named_parameters():
+                        layer_name = f"{block.__class__.__name__}Block{idx}"
+                        layer_param_name = f"{layer_name}-{param_name}"
+                        param.__setattr__("layer_name", layer_name)
+                        param.__setattr__("param_name", layer_param_name)
+            else:
+                for param_name, param in children.named_parameters():
+                    layer_name = f"{children.__class__.__name__}"
+                    layer_param_name = f"{layer_name}-{param_name}"
+                    param.__setattr__("layer_name", layer_name)
+                    param.__setattr__("param_name", f"{layer_name}-{param_name}")