Make HF conversion automatically add missing imports (#1241)

llmfoundry/models/mpt/modeling_mpt.py

@@ -40,6 +40,8 @@
     except Exception as e:
         raise e
 
+import logging
+
 from transformers import PreTrainedModel, PreTrainedTokenizerBase
 from transformers.modeling_outputs import (
     BaseModelOutputWithPast,
@@ -62,31 +64,23 @@
 from llmfoundry.models.layers.custom_embedding import SharedEmbedding
 from llmfoundry.models.layers.layer_builders import build_norm
 from llmfoundry.models.mpt.configuration_mpt import MPTConfig
+from llmfoundry.models.utils.act_ckpt import (
+    build_act_ckpt_mod_to_blocks,
+    check_mapping_blocks_overlap,
+    pass_on_block_idx,
+)
 from llmfoundry.models.utils.config_moe_args import config_moe_args
 from llmfoundry.models.utils.mpt_param_count import (
     mpt_get_active_params,
     mpt_get_total_params,
 )
 
-# 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.
+# Import the fcs and param_init_fns here so that the recursive code creating the files for hf checkpoints can find them
+# These are the exceptions because fc.py and param_init_fns.py are not imported in any other place in the import tree
 # isort: off
-from llmfoundry.models.utils.meta_init_context import \
-    init_empty_weights  # type: ignore (see note)
-from llmfoundry.models.utils.param_init_fns import (
-    generic_param_init_fn_,  # type: ignore (see note)
-)
-from llmfoundry.models.layers.ffn import resolve_ffn_act_fn  # type: ignore (see note)
-from llmfoundry.models.layers.fc import fcs  # type: ignore (see note)
-
-from llmfoundry.models.utils.act_ckpt import (
-    pass_on_block_idx,
-    build_act_ckpt_mod_to_blocks,
-    check_mapping_blocks_overlap,
-)
-
-import logging
+from llmfoundry.models.layers.fc import fcs  #  type: ignore
+from llmfoundry.models.utils.param_init_fns import generic_param_init_fn_  # type: ignore
+# isort: on
 
 log = logging.getLogger(__name__)
 

llmfoundry/utils/huggingface_hub_utils.py

@@ -3,6 +3,7 @@
 
 import ast
 import importlib
+import json
 import os
 from typing import Optional, Sequence
 
@@ -139,6 +140,80 @@ def process_file(
     return new_files_to_process
 
 
+def get_all_relative_imports(file_path: str) -> set[str]:
+    """Get all relative imports from a file.
+
+    Args:
+        file_path (str): The file to process.
+
+    Returns:
+        set[str]: The relative imports.
+    """
+    with open(file_path, 'r', encoding='utf-8') as f:
+        source = f.read()
+
+    tree = ast.parse(source)
+    relative_imports = set()
+    for node in ast.walk(tree):
+        if isinstance(
+            node,
+            ast.ImportFrom,
+        ) and node.module is not None and node.level == 1:
+            relative_imports.add(node.module)
+
+    return relative_imports
+
+
+def add_relative_imports(
+    file_path: str,
+    relative_imports_to_add: set[str],
+) -> None:
+    """Add relative imports to a file.
+
+    Args:
+        file_path (str): The file to add to.
+        relative_imports_to_add (set[str]): The set of relative imports to add
+    """
+    # Get the directory name where all the files are located
+    dir_name = os.path.dirname(file_path)
+
+    with open(file_path, 'r', encoding='utf-8') as f:
+        source = f.read()
+
+    tree = ast.parse(source)
+
+    for relative_import in relative_imports_to_add:
+        import_path = os.path.join(dir_name, relative_import + '.py')
+        # Open up the file we are adding an import to find something to import from it
+        with open(import_path, 'r', encoding='utf-8') as f:
+            import_source = f.read()
+        import_tree = ast.parse(import_source)
+        item_to_import = None
+        for node in ast.walk(import_tree):
+            # Find the first function or class
+            if isinstance(node,
+                          ast.FunctionDef) or isinstance(node, ast.ClassDef):
+                # Get its name to import it
+                item_to_import = node.name
+                break
+
+        if item_to_import is None:
+            item_to_import = '*'
+
+        # This will look like `from .relative_import import item_to_import`
+        import_node = ast.ImportFrom(
+            module=relative_import,
+            names=[ast.alias(name=item_to_import, asname=None)],
+            level=1,
+        )
+
+        # Insert near the top of the file, but past the from __future__ import
+        tree.body.insert(2, import_node)
+
+    with open(file_path, 'w', encoding='utf-8') as f:
+        f.write(ast.unparse(tree))
+
+
 def edit_files_for_hf_compatibility(
     folder: str,
     flatten_imports_prefix: Sequence[str] = ('llmfoundry',),
@@ -158,9 +233,27 @@ def edit_files_for_hf_compatibility(
         remove_imports_prefix (Sequence[str], optional): Sequence of prefixes to remove. Takes precedence over flattening.
             Defaults to ('composer', 'omegaconf', 'llmfoundry.metrics', 'llmfoundry.utils.builders').
     """
+    listed_dir = os.listdir(folder)
+
+    # Try to acquire the config file to determine which python file is the entrypoint file
+    config_file_exists = 'config.json' in listed_dir
+    with open(os.path.join(folder, 'config.json'), 'r') as _f:
+        config = json.load(_f)
+
+    # 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
+            split_path = value.split('.')
+            if len(split_path) > 1:
+                entrypoint_files.add(split_path[0] + '.py')
+
     files_to_process = [
         os.path.join(folder, filename)
-        for filename in os.listdir(folder)
+        for filename in listed_dir
         if filename.endswith('.py')
     ]
     files_processed_and_queued = set(files_to_process)
@@ -178,3 +271,22 @@ def edit_files_for_hf_compatibility(
                 if file not in files_processed_and_queued:
                     files_to_process.append(file)
                     files_processed_and_queued.add(file)
+
+    # For each entrypoint, determine which imports are missing, and add them
+    # This is because HF does not recursively search imports when determining
+    # which files to copy into its modules cache
+    all_relative_imports = {
+        os.path.splitext(os.path.basename(f))[0]
+        for f in files_processed_and_queued
+    }
+    for entrypoint in entrypoint_files:
+        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])
+        missing_relative_imports = all_relative_imports - existing_relative_imports
+        add_relative_imports(
+            os.path.join(folder, entrypoint),
+            missing_relative_imports,
+        )