Extendability refactors

llmfoundry/data/finetuning/dataloader.py

@@ -46,7 +46,7 @@
 
 def build_finetuning_dataloader(
     tokenizer: PreTrainedTokenizerBase,
-    device_batch_size: int,
+    device_batch_size: Union[int, float],
     dataset: Dict[str, Any],
     num_workers: int,
     drop_last: bool = False,

llmfoundry/data/text_data.py

@@ -300,7 +300,7 @@ def build_streams(streams: Optional[Dict[str, Any]] = None,):
 
 def build_text_dataloader(
     tokenizer: PreTrainedTokenizerBase,
-    device_batch_size: int,
+    device_batch_size: Union[int, float],
     dataset: Dict[str, Any],
     drop_last: bool,
     num_workers: int,

llmfoundry/models/layers/blocks.py

@@ -212,13 +212,32 @@ def apply_ffn(
         indices = None
         if not self.use_pad_tok_in_ffn and attention_mask is not None:
             assert unpad_input is not None
+            attention_mask = self.slice_attention_mask(attention_mask, seq_len)
             m, indices, _, _ = unpad_input(m, attention_mask)
         n = self.ffn(m)
         if not self.use_pad_tok_in_ffn and attention_mask is not None:
             assert pad_input is not None
             n = pad_input(n, indices, batch_size, seq_len)
         return n
 
+    def slice_attention_mask(
+        self,
+        attention_mask: torch.ByteTensor,
+        seq_len: int,
+    ) -> torch.ByteTensor:
+        """Slice attention mask to the correct size.
+
+        Can be overridden by subclasses to apply different slicing logic.
+
+        Args:
+            attention_mask (torch.ByteTensor): The attention mask.
+            seq_len (int): The sequence length.
+
+        Returns:
+            torch.ByteTensor: The sliced attention mask.
+        """
+        return attention_mask
+
 
 class FusedNormAttentionNorm(nn.Module):
 

llmfoundry/models/mpt/configuration_mpt.py

@@ -14,22 +14,12 @@
     check_alibi_support,
     is_flash_v2_installed,
 )
-
-# NOTE: All utils are imported directly even if unused so that
-# HuggingFace can detect all the needed files to copy into its modules folder.
-# Otherwise, certain modules are missing.
-# isort: off
-from llmfoundry.models.layers.norm import LPLayerNorm  # type: ignore (see note)
-from llmfoundry.models.layers.layer_builders import build_norm, build_fc, build_ffn  # type: ignore (see note)
-from llmfoundry.models.layers.dmoe import dMoE  # 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)
 from llmfoundry.models.utils.config_defaults import (
     attn_config_defaults,
+    fc_type_defaults,
     ffn_config_defaults,
     init_config_defaults,
-    fc_type_defaults,
-)  # type: ignore (see note)
+)
 
 
 class MPTConfig(PretrainedConfig):
@@ -196,6 +186,10 @@ def _set_config_defaults(
                 )
         return config
 
+    def validate_attention_config(self) -> None:
+        if (self.attn_config.get('seq_parallel_world_size', 1) or 1) > 1:
+            raise NotImplementedError('Sequence Parallelism is not supported.')
+
     def _validate_config(self) -> None:
         # set config defaults
         self.attn_config = self._set_config_defaults(
@@ -336,5 +330,5 @@ def _validate_config(self) -> None:
                 raise ImportError(
                     'In order to set `use_pad_tok_in_ffn=False`, please install flash-attn==1.0.9 or flash-attn==2.3.6',
                 )
-        if (self.attn_config.get('seq_parallel_world_size', 1) or 1) > 1:
-            raise NotImplementedError('Sequence Parallelism is not supported.')
+
+        self.validate_attention_config()

llmfoundry/models/mpt/modeling_mpt.py

@@ -80,6 +80,7 @@
 # isort: off
 from llmfoundry.models.layers.fc import fcs  #  type: ignore
 from llmfoundry.models.utils.param_init_fns import generic_param_init_fn_  # type: ignore
+from llmfoundry.models.layers.norm import LPLayerNorm  # type: ignore
 # isort: on
 
 log = logging.getLogger(__name__)
@@ -425,6 +426,10 @@ def __init__(self, config: MPTConfig):
         log.debug(self)
         log.debug(f'Using {self.config.init_config["name"]} initialization.')
 
+    @property
+    def block_class(self) -> Type[MPTBlock]:
+        return MPTBlock
+
     def construct_blocks(self, config: MPTConfig) -> nn.ModuleList:
         """Construct the nn.ModuleList with the Transformer blocks.
 
@@ -437,7 +442,7 @@ def construct_blocks(self, config: MPTConfig) -> nn.ModuleList:
         block_args = self.extract_block_args(config.to_dict())
 
         return nn.ModuleList([
-            MPTBlock(
+            self.block_class(
                 device=config.init_device,
                 **block_args,
             ) for _ in range(config.n_layers)

llmfoundry/registry.py

@@ -222,6 +222,19 @@
     description=_icl_datasets_description,
 )
 
+_config_transforms_description = (
+    'The config_transforms registry is used to register functions that transform the training config before it is passed to the Trainer.'
+    +
+    'Note: By default ALL registered transforms will be applied to the train config and NONE to the eval config. Each transform should return the modified config.'
+)
+config_transforms = create_registry(
+    'llmfoundry',
+    'config_transforms',
+    generic_type=Callable[[Dict[str, Any]], Dict[str, Any]],
+    entry_points=True,
+    description=_config_transforms_description,
+)
+
 __all__ = [
     'loggers',
     'callbacks',
@@ -245,4 +258,5 @@
     'attention_implementations',
     'fcs',
     'icl_datasets',
+    'config_transforms',
 ]

llmfoundry/utils/__init__.py

@@ -1,6 +1,7 @@
 # Copyright 2022 MosaicML LLM Foundry authors
 # SPDX-License-Identifier: Apache-2.0
 
+from llmfoundry.registry import config_transforms
 from llmfoundry.utils.builders import (
     build_algorithm,
     build_callback,
@@ -59,6 +60,11 @@
     experimental_function,
 )
 
+config_transforms.register(
+    'update_batch_size_info',
+    func=update_batch_size_info,
+)
+
 __all__ = [
     'build_algorithm',
     'build_callback',

llmfoundry/utils/builders.py

@@ -63,7 +63,7 @@ def build_evaluators(
     eval_gauntlet_config: Optional[Union[str, Dict[str, Any]]],
     *,
     tokenizer: PreTrainedTokenizerBase,
-    device_eval_batch_size: int,
+    device_eval_batch_size: Union[int, float],
     icl_seq_len: int,
     icl_subset_num_batches: Optional[int],
 ) -> Tuple[List[Evaluator], List[str], Optional[EvalGauntlet]]:
@@ -79,6 +79,10 @@ def build_evaluators(
     logger_keys = []
     eval_gauntlet_callback = None
     if icl_tasks_config is not None:
+        if not isinstance(device_eval_batch_size, int):
+            raise ValueError(
+                'device_eval_batch_size should be an int for icl tasks.',
+            )
         icl_evaluators, logger_keys, eval_gauntlet_callback = build_icl_data_and_gauntlet(
             icl_tasks_config,
             eval_gauntlet_config,
@@ -95,7 +99,7 @@ def build_evaluators(
 def build_eval_loaders(
     eval_loader_config: Union[Dict[str, Any], List[Dict[str, Any]]],
     tokenizer: PreTrainedTokenizerBase,
-    device_eval_batch_size: int,
+    device_eval_batch_size: Union[int, float],
 ) -> List[Evaluator]:
     evaluators: List[Evaluator] = []
     if isinstance(eval_loader_config, list):
@@ -122,7 +126,8 @@ def build_eval_loaders(
             # Load the eval data to fail fast. metrics will get added
             # later in add_metrics_to_eval_loaders, after the model is loaded
             metric_names=[],
-            device_eval_microbatch_size=device_eval_batch_size,
+            # TODO: Fix type in Composer
+            device_eval_microbatch_size=device_eval_batch_size,  # type: ignore
         )
         evaluators.append(eval_loader)
     return evaluators

llmfoundry/utils/config_utils.py

@@ -30,6 +30,7 @@
 
 from llmfoundry.layers_registry import ffns_with_megablocks
 from llmfoundry.models.utils import init_empty_weights
+from llmfoundry.registry import config_transforms
 
 log = logging.getLogger(__name__)
 
@@ -48,7 +49,7 @@ class EvalConfig:
     # Eval Config required parameters:
     models: List[Dict[str, Any]] = MISSING
     max_seq_len: int = MISSING
-    device_eval_batch_size: int = MISSING
+    device_eval_batch_size: Union[int, float] = MISSING
 
     # Eval Config optional parameters:
     code_paths: Optional[List[str]] = None
@@ -101,7 +102,7 @@ class TrainConfig:
     scheduler: Dict[str, Any] = MISSING
     train_loader: Dict[str, Any] = MISSING
     device_train_batch_size: Union[int, float] = MISSING
-    device_eval_batch_size: int = MISSING
+    device_eval_batch_size: Union[int, float] = MISSING
     max_duration: Union[int, str] = MISSING
     eval_interval: Union[int, str] = MISSING
     max_seq_len: int = MISSING
@@ -160,7 +161,7 @@ class TrainConfig:
     save_ignore_keys: Optional[List[str]] = None
 
     # Dataloader
-    device_train_microbatch_size: Union[str, int] = 'auto'
+    device_train_microbatch_size: Union[str, int, float] = 'auto'
     global_train_batch_size: Optional[int] = None
 
     # Eval dataloader
@@ -238,12 +239,49 @@ def to_container(
 T = TypeVar('T')
 
 
+def apply_transforms_to_config(
+    cfg: Dict[str, Any],
+    transforms: Optional[Union[List[Callable[[Dict[str, Any]], Dict[str, Any]]],
+                               List[str], str]],
+) -> Dict[str, Any]:
+    if transforms is None or (
+        isinstance(transforms, list) and len(transforms) == 0
+    ):
+        return cfg
+
+    transform_functions = []
+    if isinstance(transforms, list):
+        for transform in transforms:
+            if isinstance(transform, str):
+                transform_functions.append(config_transforms.get(transform))
+            elif callable(transform):
+                transform_functions.append(transform)
+            else:
+                raise ValueError(
+                    f'Invalid transform: {transform}. Must be a string or callable.',
+                )
+    elif isinstance(transforms, str) and transforms == 'all':
+        transform_functions = [
+            config_transforms.get(transform)
+            for transform in config_transforms.get_all()
+        ]
+    else:
+        raise ValueError(
+            f'Invalid transforms: {transforms}. Must be a list of strings or callables, or ``all``.',
+        )
+
+    for transform in transform_functions:
+        cfg = transform(cfg)
+
+    return cfg
+
+
 def make_dataclass_and_log_config(
     cfg: DictConfig,
     dataclass_constructor: Callable[..., T],
     dataclass_fields: Set[str],
-    transforms: Optional[List[Callable[[Dict[str, Any]], Dict[str,
-                                                              Any]]]] = None,
+    transforms: Optional[Union[List[Callable[[Dict[str, Any]], Dict[str, Any]]],
+                               List[str], str]] = None,
     icl_tasks_required: bool = False,
 ) -> Tuple[Dict[str, Any], T]:
     """Converts a DictConfig to a dataclass and creates a logged config."""
@@ -281,8 +319,10 @@ def make_dataclass_and_log_config(
     logged_cfg: Dict[str, Any] = copy.deepcopy(unstructured_config)
 
     # Apply transforms to the unstructured config before constructing dataclass
-    for transform in transforms or []:
-        unstructured_config = transform(unstructured_config)
+    unstructured_config = apply_transforms_to_config(
+        unstructured_config,
+        transforms,
+    )
 
     logged_cfg.update(unstructured_config, merge=True)
 
@@ -367,20 +407,21 @@ def calculate_batch_size_info(
     data_replication_degree: int = 1,
 ) -> Tuple[Union[int, float], Union[int, float, Literal['auto']], Union[
     int, Literal['auto']]]:
-    if dist.get_world_size() % data_replication_degree != 0:
+
+    world_size = dist.get_world_size()
+    assert isinstance(world_size, int)
+    if world_size % data_replication_degree != 0:
         raise ValueError(
-            f'World size {dist.get_world_size()} is not divisible by data replication degree {data_replication_degree}.',
+            f'World size {world_size} is not divisible by data replication degree {data_replication_degree}.',
         )
-    if global_batch_size % (
-        dist.get_world_size() // data_replication_degree
-    ) != 0:
+    if global_batch_size % (world_size // data_replication_degree) != 0:
         raise ValueError(
-            f'Global batchsize {global_batch_size} is not divisible by {(dist.get_world_size() // data_replication_degree)=} '
+            f'Global batchsize {global_batch_size} is not divisible by {(world_size // data_replication_degree)=} '
             +
             'as a result, the batch size would be truncated, please adjust `global_batch_size` '
-            + f'to be divisible by world size, {dist.get_world_size()}.',
+            + f'to be divisible by world size, {world_size}.',
         )
-    device_batch_size = global_batch_size / dist.get_world_size()
+    device_batch_size = global_batch_size / world_size
     if device_batch_size == round(device_batch_size):
         device_batch_size = round(device_batch_size)
     if device_microbatch_size == 'auto':
@@ -401,7 +442,6 @@ def calculate_batch_size_info(
     return device_batch_size, device_microbatch_size, device_grad_accum
 
 
-# Coming soon: this conversion math will be done inside Composer Trainer
 def update_batch_size_info(cfg: Dict[str, Any]) -> Dict[str, Any]:
     data_replication_degree = 1
     device_train_batch_size, device_train_microbatch_size, device_train_grad_accum = calculate_batch_size_info(

llmfoundry/utils/huggingface_hub_utils.py

@@ -243,10 +243,7 @@ def edit_files_for_hf_compatibility(
     # If the config file exists, the entrypoint files would be specified in the auto map
     entrypoint_files = set()
     if config_file_exists:
-        for key, value in config.get('auto_map', {}).items():
-            # Only keep the modeling entrypoints, e.g. AutoModelForCausalLM
-            if 'model' not in key.lower():
-                continue
+        for value in config.get('auto_map', {}).values():
             split_path = value.split('.')
             if len(split_path) > 1:
                 entrypoint_files.add(split_path[0] + '.py')
@@ -279,15 +276,21 @@ def edit_files_for_hf_compatibility(
         os.path.splitext(os.path.basename(f))[0]
         for f in files_processed_and_queued
     }
+    # Filter out __init__ files
+    all_relative_imports = {
+        relative_import for relative_import in all_relative_imports
+        if relative_import != '__init__'
+    }
     for entrypoint in entrypoint_files:
         file_path = os.path.join(folder, entrypoint)
         if not os.path.exists(file_path):
             continue
         existing_relative_imports = get_all_relative_imports(
             os.path.join(folder, entrypoint),
         )
-        # Add in self so we don't create a circular import
-        existing_relative_imports.add(os.path.splitext(entrypoint)[0])
+        # Add in all entrypoints so we don't create a circular import
+        for sub_entrypoint in entrypoint_files:
+            existing_relative_imports.add(os.path.splitext(sub_entrypoint)[0])
         missing_relative_imports = all_relative_imports - existing_relative_imports
         add_relative_imports(
             os.path.join(folder, entrypoint),

scripts/eval/eval.py

@@ -53,7 +53,7 @@ def evaluate_model(
     seed: int,
     icl_tasks: Union[str, List[Dict[str, Any]]],
     max_seq_len: int,
-    device_eval_batch_size: int,
+    device_eval_batch_size: Union[int, float],
     eval_gauntlet_config: Optional[Union[str, Dict[str, Any]]],
     eval_loader_config: Optional[Union[Dict[str, Any], List[Dict[str, Any]]]],
     fsdp_config: Optional[Dict[str, Any]],