Merge pull request #161 from Modalities/feat/initialization

Feat: Various Configurable Initializations

README.md

@@ -83,6 +83,7 @@ In the following, we list the already implemented, planned and in-progress featu
 |--------------------------------|------------------|-------------------------------------------------------------------------------------------------------------------|
 | SwiGLU                         | supported         | A nonlinear activation function combining Gated Linear Units (GLU) with Swish for enhancing model capacity and learning efficiency. |
 | Weight Decay                   | supported        | Regularization technique that adds a penalty on the size of weights, encouraging smaller weights to reduce overfitting and improve generalization. |
+| Weight Initialization          | supported        | Choose between different, configurable weight initialization techniques to stabilize training. |
 | RMSNorm (pre-normalization)    | supported        | Normalizes the pre-activation weights in a layer to stabilize training, often used as an alternative to LayerNorm for improved training dynamics. |
 | Rotary Positional Embeddings (RoPE) | supported  | Encodes sequence position information into attention mechanisms, preserving relative positional information and improving model's understanding of sequence order. |
 | Grouped-query Attention (GQA)  | supported    | Enhances attention mechanisms by grouping queries to reduce computation and memory footprint while maintaining or improving performance. |

config_files/text_generation/text_generation_config_torch.yaml

@@ -62,6 +62,7 @@ raw_model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

config_files/text_generation/text_generation_overfitted_de.yaml

@@ -65,6 +65,7 @@ raw_model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

config_files/training/config_example_coca.yaml

@@ -42,7 +42,7 @@ train_dataset:
   component_key: dataset
   variant_key: dummy_dataset
   config:
-    num_samples: 4
+    num_samples: 64
     sample_definition:
       - sample_key: images
         sample_shape: [3, 224, 224]
@@ -55,7 +55,7 @@ val_dataset:
   component_key: dataset
   variant_key: dummy_dataset
   config:
-    num_samples: 4
+    num_samples: 32
     sample_definition:
       - sample_key: images
         sample_shape: [3, 224, 224]
@@ -170,7 +170,23 @@ wrapped_model:
     sharding_strategy: FULL_SHARD
     block_names: [TransformerBlock, VisionTransformerBlock]
 
-model:
+model: 
+  component_key: model
+  variant_key: model_initialized
+  config:
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+    model_initializer:
+      component_key: model_initialization
+      variant_key: composed
+      config:
+        model_type: coca
+        weight_init_type: plain
+        mean: 0.0
+        std: 0.02
+
+model_raw:
   component_key: model
   variant_key: coca
   config:
@@ -215,9 +231,6 @@ model:
     n_vision_queries: 256
     bias_attn_pool: False
     epsilon_attn_pool: 1e-5
-    weight_init:
-      mean: 0.0
-      std: 0.02
 
 scheduler:
   component_key: scheduler
@@ -229,7 +242,7 @@ scheduler:
     max_lr: 6e-4
     div_factor: 10
     final_div_factor: 1
-    total_steps: 4
+    total_steps: 64
     pct_start: 0.01
     anneal_strategy: cos
 

config_files/training/config_example_mem_map_dataset.yaml

@@ -167,6 +167,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

config_files/training/config_example_openGPTx_dataset.yaml

@@ -155,6 +155,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
 
 scheduler:
   type_hint: StepLR

config_files/training/config_gpt2_small_overfitting_de.yaml

@@ -172,6 +172,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

config_files/training/config_gpt2_small_overfitting_de_abs_pos_emb.yaml

@@ -172,6 +172,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

config_files/training/config_gpt2_small_redpajama_DE_1048576.yaml

@@ -172,6 +172,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

config_files/training/config_lorem_ipsum.yaml

@@ -178,7 +178,24 @@ wrapped_model:
     sharding_strategy: FULL_SHARD
     block_names: [GPT2Block]
 
-model:
+model: 
+  component_key: model
+  variant_key: model_initialized
+  config:
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+    model_initializer:
+      component_key: model_initialization
+      variant_key: composed
+      config:
+        model_type: gpt2
+        weight_init_type: scaled_embed
+        mean: 0.0
+        std: 0.02
+        num_layers: ${model_raw.config.n_layer}
+
+model_raw:
   component_key: model
   variant_key: gpt2
   config:
@@ -198,33 +215,30 @@ model:
       qkv_transforms:
         - type_hint: RotaryTransform
           config:
-            n_embd: ${model.config.n_embd}
-            n_head: ${model.config.n_head_q} #it has to be head_q here
+            n_embd: ${model_raw.config.n_embd}
+            n_head: ${model_raw.config.n_head_q} #it has to be head_q here
             seq_length_dim: -2
     attention_implementation: manual
     activation_type: gelu
-    weight_init:
-      mean: 0.0
-      std: 0.02
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm
       config:
-        ndim: ${model.config.n_embd}
+        ndim: ${model_raw.config.n_embd}
         bias: true
         epsilon: 1e-5
     ffn_norm:
       component_key: layer_norm
       variant_key: rms_norm
       config:
-        ndim: ${model.config.n_embd}
+        ndim: ${model_raw.config.n_embd}
         bias: true
         epsilon: 1e-5
     lm_head_norm:
       component_key: layer_norm
       variant_key: rms_norm
       config:
-        ndim: ${model.config.n_embd}
+        ndim: ${model_raw.config.n_embd}
         bias: true
         epsilon: 1e-5
 

examples/getting_started/example_config.yaml

@@ -181,6 +181,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

examples/getting_started/example_text_generation_config.yaml

@@ -66,6 +66,7 @@ raw_model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
     attention_norm:
       component_key: layer_norm
       variant_key: rms_norm

examples/library_usage/config_lorem_ipsum.yaml

@@ -205,6 +205,7 @@ model:
     weight_init:
       mean: 0.0
       std: 0.02
+      type: scaled
 
 optimizer:  
   component_key: optimizer

src/modalities/config/config.py

@@ -17,6 +17,7 @@
     PydanticCollateFnIFType,
     PydanticDatasetIFType,
     PydanticLLMDataLoaderIFType,
+    PydanticModelInitializationIFType,
     PydanticOptimizerIFType,
     PydanticPytorchDeviceType,
     PydanticPytorchModuleType,
@@ -228,6 +229,11 @@ def parse_sharding_strategy_by_name(cls, name):
         return parse_enum_by_name(name=name, enum_type=ShardingStrategy)
 
 
+class WeightInitializedModelConfig(BaseModel):
+    model: PydanticPytorchModuleType
+    model_initializer: PydanticModelInitializationIFType
+
+
 class PreTrainedHFTokenizerConfig(BaseModel):
     pretrained_model_name_or_path: str
     max_length: Annotated[int, Field(strict=True, ge=0)]

src/modalities/config/pydanctic_if_types.py

@@ -17,6 +17,7 @@
 from modalities.logging_broker.subscriber import MessageSubscriberIF
 from modalities.loss_functions import Loss
 from modalities.models.gpt2.collator import CollateFnIF
+from modalities.nn.model_initialization.initialization_if import ModelInitializationIF
 from modalities.tokenization.tokenizer_wrapper import TokenizerWrapper
 from modalities.training.gradient_clipping.gradient_clipper import GradientClipperIF
 
@@ -61,3 +62,4 @@ def __get_pydantic_core_schema__(
 PydanticPytorchDeviceType = Annotated[torch.device, PydanticThirdPartyTypeIF(torch.device)]
 PydanticTextInferenceComponentType = Annotated[TextInferenceComponent, PydanticThirdPartyTypeIF(TextInferenceComponent)]
 PydanticGradientClipperIFType = Annotated[GradientClipperIF, PydanticThirdPartyTypeIF(GradientClipperIF)]
+PydanticModelInitializationIFType = Annotated[ModelInitializationIF, PydanticThirdPartyTypeIF(ModelInitializationIF)]

src/modalities/models/coca/coca_model.py

@@ -1,5 +1,3 @@
-import math
-from functools import partial
 from typing import Annotated, Dict, Tuple
 
 import torch
@@ -10,8 +8,7 @@
 from modalities.models.coca.attention_pooling import AttentionPooling
 from modalities.models.coca.multi_modal_decoder import MultiModalTextDecoder
 from modalities.models.coca.text_decoder import TextDecoder
-from modalities.models.gpt2.gpt2_model import ActivationType, WeightInitializationConfig
-from modalities.models.model import NNModel
+from modalities.models.model import ActivationType, NNModel
 from modalities.models.vision_transformer.vision_transformer_model import VisionTransformer, VisionTransformerConfig
 from modalities.nn.attention import AttentionConfig
 
@@ -45,7 +42,6 @@ class CoCaConfig(BaseModel):
     n_vision_queries: Annotated[int, Field(ge=1)]
     bias_attn_pool: bool
     epsilon_attn_pool: Annotated[float, Field(ge=0.0)]
-    weight_init: WeightInitializationConfig
 
 
 class CoCa(NNModel):
@@ -71,7 +67,6 @@ def __init__(
         epsilon_attn_pool: float,
         vision_encoder_config: VisionTransformerConfig,
         text_decoder_config: TextDecoderConfig,
-        weight_init: WeightInitializationConfig,
     ) -> None:
         super().__init__()
         self.prediction_key = prediction_key
@@ -126,26 +121,6 @@ def __init__(
             attention_config=text_decoder_config.attention_config,
         )
 
-        # init all weights
-        self.apply(partial(self._init_weights, weight_init=weight_init))
-        # apply special scaled init to the residual projections, per GPT-2 paper
-        for pn, p in self.named_parameters():
-            if pn.endswith("c_proj.weight"):
-                torch.nn.init.normal_(
-                    p,
-                    mean=weight_init.mean,
-                    std=weight_init.std
-                    / math.sqrt(2 * (text_decoder_config.n_layer_text + text_decoder_config.n_layer_multimodal_text)),
-                )
-
-    def _init_weights(self, module: nn.Module, weight_init: WeightInitializationConfig):
-        if isinstance(module, nn.Linear):
-            torch.nn.init.normal_(module.weight, mean=weight_init.mean, std=weight_init.std)
-            if module.bias is not None:
-                torch.nn.init.zeros_(module.bias)
-        elif isinstance(module, nn.Embedding):
-            torch.nn.init.normal_(module.weight, mean=weight_init.mean, std=weight_init.std)
-
     def forward(self, inputs: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
         vision_embd, vision_cls_token = self._forward_encode_vision(inputs)
         text_embd, text_cls_token = self._forward_encode_text(inputs)

src/modalities/models/gpt2/gpt2_model.py

@@ -1,7 +1,6 @@
 import math
 from copy import deepcopy
 from enum import Enum
-from functools import partial
 from typing import Annotated, Dict, List, Tuple
 
 import torch
@@ -16,7 +15,7 @@
 
 from modalities.config.pydanctic_if_types import PydanticPytorchModuleType
 from modalities.config.utils import convert_base_model_config_to_dict
-from modalities.models.model import NNModel
+from modalities.models.model import ActivationType, NNModel
 from modalities.util import parse_enum_by_name
 
 # GPT2 implementation taken from nanogpt https://github.com/karpathy/nanoGPT
@@ -108,11 +107,6 @@ class QueryKeyValueTransformType(Enum):
     RotaryTransform = RotaryTransform
 
 
-class ActivationType(str, Enum):
-    GELU = "gelu"
-    FUSED_SWIGLU = "fused_swiglu"
-
-
 class AttentionImplementation(str, Enum):
     MANUAL = "manual"
     PYTORCH_FLASH = "pytorch_flash"
@@ -139,11 +133,6 @@ def parse_sharding_strategy_by_name(cls, name):
     qkv_transforms: List[QueryKeyValueTransformConfig]
 
 
-class WeightInitializationConfig(BaseModel):
-    mean: Annotated[float, Field(strict=True, ge=0.0)]
-    std: Annotated[float, Field(strict=True, ge=0.0)]
-
-
 class GPT2LLMConfig(BaseModel):
     sample_key: str
     prediction_key: str
@@ -165,7 +154,6 @@ class GPT2LLMConfig(BaseModel):
     attention_norm: PydanticPytorchModuleType
     ffn_norm: PydanticPytorchModuleType
     lm_head_norm: PydanticPytorchModuleType
-    weight_init: WeightInitializationConfig
 
     @model_validator(mode="after")
     def check_divisibility(self) -> "GPT2LLMConfig":
@@ -430,7 +418,6 @@ def __init__(
         bias: bool,
         activation_type: ActivationType,
         attention_implementation: AttentionImplementation,
-        weight_init: WeightInitializationConfig,
         attention_config: AttentionConfig,
         attention_norm: nn.Module,
         ffn_norm: nn.Module,
@@ -499,21 +486,6 @@ def __init__(
         # not 100% sure what this is, so far seems to be harmless. TODO investigate
         self.transformer.wte.weight = self.lm_head.weight  # https://paperswithcode.com/method/weight-tying
 
-        # init all weights
-        self.apply(partial(self._init_weights, weight_init=weight_init))
-        # apply special scaled init to the residual projections, per GPT-2 paper
-        for pn, p in self.named_parameters():
-            if pn.endswith("c_proj.weight"):
-                torch.nn.init.normal_(p, mean=weight_init.mean, std=weight_init.std / math.sqrt(2 * n_layer))
-
-    def _init_weights(self, module: nn.Module, weight_init: WeightInitializationConfig):
-        if isinstance(module, nn.Linear):
-            torch.nn.init.normal_(module.weight, mean=weight_init.mean, std=weight_init.std)
-            if module.bias is not None:
-                torch.nn.init.zeros_(module.bias)
-        elif isinstance(module, nn.Embedding):
-            torch.nn.init.normal_(module.weight, mean=weight_init.mean, std=weight_init.std)
-
     def forward_impl(self, inputs: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
         input_ids = inputs[self.sample_key]
         device = input_ids.device