Norms registry (mosaicml#1080)

README.md

@@ -306,7 +306,7 @@ dependencies = [
     "llm-foundry",
 ]
 
-[project.entry-points."llm_foundry.loggers"]
+[project.entry-points."llmfoundry_loggers"]
 my_logger = "foundry_registry.loggers:MyLogger"
 ```
 

llmfoundry/layers_registry.py

@@ -0,0 +1,20 @@
+# Copyright 2024 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Type
+
+import torch
+
+from llmfoundry.utils.registry_utils import create_registry
+
+# Layers
+_norm_description = """The norms registry is used to register classes that implement normalization layers."""
+norms = create_registry('llmfoundry',
+                        'norms',
+                        generic_type=Type[torch.nn.Module],
+                        entry_points=True,
+                        description=_norm_description)
+
+__all__ = [
+    'norms',
+]

llmfoundry/models/layers/__init__.py

@@ -9,7 +9,7 @@
 from llmfoundry.models.layers.custom_embedding import SharedEmbedding
 from llmfoundry.models.layers.fc import FC_CLASS_REGISTRY
 from llmfoundry.models.layers.ffn import FFN_CLASS_REGISTRY, MPTMLP, build_ffn
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY, LPLayerNorm
+from llmfoundry.models.layers.norm import LPLayerNorm
 
 __all__ = [
     'scaled_multihead_dot_product_attention',
@@ -23,7 +23,6 @@
     'ATTN_CLASS_REGISTRY',
     'MPTMLP',
     'MPTBlock',
-    'NORM_CLASS_REGISTRY',
     'LPLayerNorm',
     'FC_CLASS_REGISTRY',
     'SharedEmbedding',

llmfoundry/models/layers/attention.py

@@ -15,7 +15,7 @@
 from torch import nn
 
 from llmfoundry.models.layers.fc import FC_CLASS_REGISTRY
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
+from llmfoundry.models.layers.layer_builders import build_norm
 
 
 def is_flash_v2_installed(v2_version: str = '2.0.0'):
@@ -419,12 +419,19 @@ def __init__(
         self.Wqkv._fused = (0, fuse_splits)
 
         if self.qk_ln or self.qk_gn:
-            norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
             norm_size = self.head_dim if qk_gn else d_model
-            self.q_ln = norm_class(norm_size, device=device)
+            self.q_ln = build_norm(
+                name=norm_type.lower(),
+                normalized_shape=norm_size,
+                device=device,
+            )
             if qk_ln:
                 norm_size = self.head_dim * kv_n_heads
-            self.k_ln = norm_class(norm_size, device=device)
+            self.k_ln = build_norm(
+                name=norm_type.lower(),
+                normalized_shape=norm_size,
+                device=device,
+            )
 
         if self.attn_impl == 'flash':
             self.attn_fn = flash_attn_fn

llmfoundry/models/layers/blocks.py

@@ -10,7 +10,7 @@
 
 from llmfoundry.models.layers.attention import ATTN_CLASS_REGISTRY
 from llmfoundry.models.layers.ffn import FFN_CLASS_REGISTRY, build_ffn
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
+from llmfoundry.models.layers.layer_builders import build_norm
 
 try:
     from flash_attn.bert_padding import unpad_input, pad_input  # type: ignore # yapf: disable # isort: skip
@@ -72,7 +72,6 @@ def __init__(
         del kwargs  # unused, just to capture any extra args from the config
         super().__init__()
 
-        norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
         assert isinstance(attn_config['attn_type'], str)
         attn_class = ATTN_CLASS_REGISTRY[attn_config['attn_type']]
 
@@ -88,7 +87,11 @@ def __init__(
             if k not in args_to_exclude_in_attn_class
         }
 
-        self.norm_1 = norm_class(d_model, device=device)
+        self.norm_1 = build_norm(
+            name=norm_type.lower(),
+            normalized_shape=d_model,
+            device=device,
+        )
         self.attn = attn_class(
             d_model=d_model,
             n_heads=n_heads,
@@ -100,7 +103,11 @@ def __init__(
         self.norm_2 = None
         if not getattr(FFN_CLASS_REGISTRY[ffn_config['ffn_type']], '_has_norm',
                        False):
-            self.norm_2 = norm_class(d_model, device=device)
+            self.norm_2 = build_norm(
+                name=norm_type.lower(),
+                normalized_shape=d_model,
+                device=device,
+            )
         self.ffn = build_ffn(
             d_model=d_model,
             expansion_ratio=expansion_ratio,

llmfoundry/models/layers/layer_builders.py

@@ -0,0 +1,25 @@
+# Copyright 2024 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import List, Optional, Union
+
+import torch
+
+from llmfoundry.layers_registry import norms
+from llmfoundry.utils.registry_utils import construct_from_registry
+
+
+def build_norm(
+    name: str,
+    normalized_shape: Union[int, List[int], torch.Size],
+    device: Optional[str] = None,
+):
+    kwargs = {
+        'normalized_shape': normalized_shape,
+        'device': device,
+    }
+
+    return construct_from_registry(name=name,
+                                   registry=norms,
+                                   pre_validation_function=torch.nn.Module,
+                                   kwargs=kwargs)

llmfoundry/models/layers/norm.py

@@ -1,10 +1,14 @@
 # Copyright 2022 MosaicML LLM Foundry authors
 # SPDX-License-Identifier: Apache-2.0
 
-from typing import Dict, List, Optional, Type, Union
+from typing import List, Optional, Union
 
 import torch
 
+from llmfoundry.layers_registry import norms
+
+norms.register(name='layernorm', func=torch.nn.LayerNorm)
+
 
 def _cast_if_autocast_enabled(tensor: torch.Tensor) -> torch.Tensor:
     if torch.is_autocast_enabled():
@@ -18,6 +22,7 @@ def _cast_if_autocast_enabled(tensor: torch.Tensor) -> torch.Tensor:
     return tensor
 
 
+@norms.register_class('low_precision_layernorm')
 class LPLayerNorm(torch.nn.LayerNorm):
 
     def __init__(
@@ -62,6 +67,7 @@ def rms_norm(x: torch.Tensor,
     return output
 
 
+@norms.register_class('rmsnorm')
 class RMSNorm(torch.nn.Module):
 
     def __init__(
@@ -84,6 +90,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
         return rms_norm(x.float(), self.weight, self.eps).to(dtype=x.dtype)
 
 
+@norms.register_class('low_precision_rmsnorm')
 class LPRMSNorm(RMSNorm):
 
     def __init__(
@@ -111,6 +118,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
                             self.eps).to(dtype=x.dtype)
 
 
+@norms.register_class('triton_rmsnorm')
 class TritonRMSNorm(torch.nn.Module):
 
     def __init__(
@@ -150,12 +158,3 @@ def forward(self, x: torch.Tensor):
             prenorm=False,
             residual_in_fp32=False,
         )
-
-
-NORM_CLASS_REGISTRY: Dict[str, Type[torch.nn.Module]] = {
-    'layernorm': torch.nn.LayerNorm,
-    'low_precision_layernorm': LPLayerNorm,
-    'rmsnorm': RMSNorm,
-    'low_precision_rmsnorm': LPRMSNorm,
-    'triton_rmsnorm': TritonRMSNorm,
-}

llmfoundry/models/mpt/configuration_mpt.py

@@ -19,6 +19,9 @@
 from llmfoundry.models.layers.fc import FC_CLASS_REGISTRY  # type: ignore (see note)
 from llmfoundry.models.layers.norm import LPLayerNorm  # type: ignore (see note)
 from llmfoundry.models.layers.ffn import FFN_CLASS_REGISTRY  # type: ignore (see note)
+from llmfoundry.models.layers.layer_builders import build_norm  # type: ignore (see note)
+from llmfoundry.layers_registry import norms  # type: ignore (see note)
+from llmfoundry.utils.registry_utils import construct_from_registry  # type: ignore (see note)
 
 ffn_config_defaults: Dict = {
     'ffn_type': 'mptmlp',

llmfoundry/models/mpt/modeling_mpt.py

@@ -20,7 +20,6 @@
 from composer.utils import dist
 
 from llmfoundry.models.layers.attention import is_flash_v2_installed
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
 
 if is_flash_v2_installed():
     try:  # This try...except is needed because transformers requires it despite the 'if' statement above
@@ -42,11 +41,13 @@
 from transformers.models.llama.modeling_llama import \
     LlamaRotaryEmbedding as HFRotaryEmbedding
 
+from llmfoundry.layers_registry import norms
 from llmfoundry.models.layers.attention import (attn_bias_shape,
                                                 build_attn_bias, gen_slopes)
 from llmfoundry.models.layers.blocks import MPTBlock
 from llmfoundry.models.layers.custom_embedding import SharedEmbedding
 from llmfoundry.models.layers.ffn import build_ffn as build_ffn
+from llmfoundry.models.layers.layer_builders import build_norm
 from llmfoundry.models.mpt.configuration_mpt import MPTConfig
 
 # NOTE: All utils are imported directly even if unused so that
@@ -297,12 +298,11 @@ def __init__(self, config: MPTConfig):
             else:
                 config.init_device = 'meta'
 
-        if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():
-            norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())
+        if config.norm_type.lower() not in norms.get_all():
+            norm_options = ' | '.join(norms.get_all())
             raise NotImplementedError(
                 f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).'
             )
-        norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]
 
         # CogView (https://arxiv.org/abs/2105.13290) and GLM-130B (https://arxiv.org/abs/2210.02414)
         # both report this helping with stabilizing training
@@ -329,7 +329,11 @@ def __init__(self, config: MPTConfig):
             block.max_block_idx = config.n_layers - 1
             pass_on_block_idx(block)
 
-        self.norm_f = norm_class(config.d_model, device=config.init_device)
+        self.norm_f = build_norm(
+            name=config.norm_type.lower(),
+            normalized_shape=config.d_model,
+            device=config.init_device,
+        )
 
         self.rope = config.attn_config['rope']
         self.rope_impl = None

llmfoundry/models/utils/act_ckpt.py

@@ -5,10 +5,10 @@
 
 import torch
 
+from llmfoundry.layers_registry import norms
 from llmfoundry.models.layers.attention import ATTN_CLASS_REGISTRY
 from llmfoundry.models.layers.blocks import MPTBlock
 from llmfoundry.models.layers.ffn import FFN_CLASS_REGISTRY
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
 
 
 def pass_on_block_idx(parent: torch.nn.Module):
@@ -29,12 +29,12 @@ def get_act_ckpt_module(mod_name: str) -> Any:
         mod_type = ATTN_CLASS_REGISTRY[mod_name]
     elif mod_name in FFN_CLASS_REGISTRY:
         mod_type = FFN_CLASS_REGISTRY[mod_name]
-    elif mod_name in NORM_CLASS_REGISTRY:
-        mod_type = NORM_CLASS_REGISTRY[mod_name]
+    elif mod_name in norms:
+        mod_type = norms.get(mod_name)
     else:
         msg = ', '.join(
             list(ATTN_CLASS_REGISTRY.keys()) + list(FFN_CLASS_REGISTRY.keys()) +
-            list(NORM_CLASS_REGISTRY.keys()) + ['MPTBlock'])
+            list(norms.get_all()) + ['MPTBlock'])
         raise ValueError(
             f'{mod_name} (specified in activation_checkpointing_target) is not a recognized option out of available options {msg}.'
         )

llmfoundry/models/utils/param_init_fns.py

@@ -10,8 +10,8 @@
 import torch
 from torch import nn
 
+from llmfoundry.layers_registry import norms
 from llmfoundry.models.layers.fc import FC_CLASS_REGISTRY
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
 
 try:
     import transformer_engine.pytorch as te
@@ -129,7 +129,8 @@ def generic_param_init_fn_(
 
         emb_init_fn_(module.weight)
 
-    elif isinstance(module, tuple(set(NORM_CLASS_REGISTRY.values()))):
+    elif isinstance(module,
+                    tuple(set([norms.get(name) for name in norms.get_all()]))):
         # Norm
         if hasattr(module, 'weight') and isinstance(module.weight,
                                                     torch.Tensor):

llmfoundry/registry.py

@@ -12,6 +12,7 @@
 from transformers import PreTrainedTokenizerBase
 
 from llmfoundry.interfaces import CallbackWithConfig
+from llmfoundry.layers_registry import norms
 from llmfoundry.utils.registry_utils import create_registry
 
 _loggers_description = (
@@ -119,4 +120,5 @@
     'models',
     'metrics',
     'dataloaders',
+    'norms',
 ]

llmfoundry/utils/registry_utils.py

@@ -14,6 +14,7 @@
 __all__ = ['TypedRegistry', 'create_registry', 'construct_from_registry']
 
 T = TypeVar('T')
+TypeBoundT = TypeVar('TypeBoundT', bound=Type)
 
 
 class TypedRegistry(catalogue.Registry, Generic[T]):
@@ -36,6 +37,12 @@ def __call__(self, name: str, func: Optional[T] = None) -> Callable[[T], T]:
     def register(self, name: str, *, func: Optional[T] = None) -> T:
         return super().register(name, func=func)
 
+    def register_class(self,
+                       name: str,
+                       *,
+                       func: Optional[TypeBoundT] = None) -> TypeBoundT:
+        return super().register(name, func=func)
+
     def get(self, name: str) -> T:
         return super().get(name)
 

tests/models/test_model.py

@@ -26,8 +26,9 @@
 from transformers.models.bloom.modeling_bloom import build_alibi_tensor
 
 from llmfoundry import ComposerHFCausalLM
+from llmfoundry.layers_registry import norms
 from llmfoundry.models.hf.model_wrapper import HuggingFaceModelWithFSDP
-from llmfoundry.models.layers import NORM_CLASS_REGISTRY, build_alibi_bias
+from llmfoundry.models.layers import build_alibi_bias
 from llmfoundry.models.layers.attention import (check_alibi_support,
                                                 is_flash_v2_installed)
 from llmfoundry.models.layers.blocks import MPTBlock
@@ -682,7 +683,7 @@ def test_lora_id():
     assert isinstance(model.model, peft.PeftModelForCausalLM)
 
 
-@pytest.mark.parametrize('norm_type', NORM_CLASS_REGISTRY.keys())
+@pytest.mark.parametrize('norm_type', norms.get_all())
 @pytest.mark.parametrize('no_bias', [False, True])
 @pytest.mark.parametrize('tie_word_embeddings', [True, False])
 @pytest.mark.parametrize('expansion_ratio,ffn_hidden_size', [