Bump version to 0.14.0.dev0

llmfoundry/_version.py

@@ -3,4 +3,4 @@
 
 """The LLM Foundry Version."""
 
-__version__ = '0.13.0.dev0'
+__version__ = '0.14.0.dev0'

llmfoundry/command_utils/eval.py

@@ -4,7 +4,6 @@
 import logging
 import os
 import time
-import warnings
 from typing import Any, Optional, Union
 
 import pandas as pd
@@ -37,7 +36,6 @@
     process_init_device,
 )
 from llmfoundry.utils.registry_utils import import_file
-from llmfoundry.utils.warnings import VersionedDeprecationWarning
 
 log = logging.getLogger(__name__)
 
@@ -63,7 +61,6 @@ def evaluate_model(
     callback_configs: Optional[dict[str, Any]],
     metadata: Optional[dict[str, str]],
     logged_config: dict[str, Any],
-    fsdp_config: Optional[dict[str, Any]] = None,
     parallelism_config: Optional[dict[str, Any]] = None,
     should_log_config: bool = True,
     load_path: Optional[str] = None,
@@ -78,18 +75,6 @@ def evaluate_model(
                     'parallelism_config cannot contain deprecated fsdp_config arguments.',
                 )
 
-    if fsdp_config:
-        warnings.warn(
-            VersionedDeprecationWarning(
-                'The argument fsdp_config is deprecated. Please use parallelism_config instead.',
-                remove_version='0.14.0',
-            ),
-        )
-    if fsdp_config and parallelism_config:
-        raise ValueError(
-            'Both fsdp_config and parallelism_config cannot be provided at the same time. Please use parallelism_config.',
-        )
-
     log.info(f'Evaluating model: {model_name}')
     # Build tokenizer and model
     tokenizer_cfg = tokenizer
@@ -127,7 +112,7 @@ def evaluate_model(
     fsdp_config = parallelism_config.get(
         'fsdp_config',
         None,
-    ) if parallelism_config else fsdp_config
+    ) if parallelism_config else None
     if fsdp_config and model.get('load_in_8bit', False):
         raise ValueError(
             'The FSDP config block is not supported when loading ' +
@@ -175,7 +160,7 @@ def evaluate_model(
         callbacks=callbacks,
         loggers=loggers,
         precision=precision,
-        parallelism_config={'fsdp': fsdp_config},
+        parallelism_config=parallelism_config,
         load_path=load_path,
         load_weights_only=True,
         progress_bar=False,
@@ -268,8 +253,6 @@ def evaluate(cfg: DictConfig) -> tuple[list[Trainer], pd.DataFrame]:
     model_configs = eval_config.models
     eval_gauntlet_config = eval_config.eval_gauntlet or eval_config.eval_gauntlet_str
 
-    fsdp_config = eval_config.fsdp_config
-
     # Mandatory Evaluation Parameters
     icl_tasks = eval_config.icl_tasks or eval_config.icl_tasks_str
     if icl_tasks is None:
@@ -345,9 +328,9 @@ def evaluate(cfg: DictConfig) -> tuple[list[Trainer], pd.DataFrame]:
              device_eval_batch_size=eval_config.device_eval_batch_size,
              eval_gauntlet_config=eval_gauntlet_config,
              eval_loader_config=eval_loader_config,
-             fsdp_config=fsdp_config,
              loggers=loggers,
              python_log_level=eval_config.python_log_level,
+             parallelism_config={'fsdp': eval_config.fsdp_config},
              precision=eval_config.precision,
              eval_gauntlet_df=eval_gauntlet_df,
              callback_configs=eval_config.callbacks,

llmfoundry/models/hf/__init__.py

@@ -9,7 +9,6 @@
     prepare_hf_model_for_fsdp,
 )
 from llmfoundry.models.hf.hf_t5 import ComposerHFT5
-from llmfoundry.models.hf.model_wrapper import HuggingFaceModelWithFSDP
 
 __all__ = [
     'BaseHuggingFaceModel',
@@ -18,5 +17,4 @@
     'prepare_hf_causal_lm_model_for_fsdp',
     'prepare_hf_enc_dec_model_for_fsdp',
     'prepare_hf_model_for_fsdp',
-    'HuggingFaceModelWithFSDP',
 ]

tests/eval/test_eval_deprecation.py

@@ -90,36 +90,3 @@ def test_deprecation_warning_with_deprecated_arg(self):
             'parallelism_config cannot contain deprecated fsdp_config arguments.',
             str(context.exception),
         )
-
-    def test_deprecation_warning_with_fsdp_config(self):
-        with warnings.catch_warnings(record=True) as w:
-            warnings.simplefilter('always')
-
-            try:
-                evaluate_model(
-                    **self.common_args,
-                    parallelism_config=None,
-                    fsdp_config={'verbose': True},
-                )
-            except Exception:
-                pass
-
-            self.assertTrue(
-                any(
-                    issubclass(warning.category, VersionedDeprecationWarning)
-                    for warning in w
-                ),
-            )
-
-    def test_error_with_both_fsdp_and_parallelism_config(self):
-        with self.assertRaises(ValueError) as context:
-            evaluate_model(
-                **self.common_args,
-                parallelism_config={'some_arg': True},
-                fsdp_config={'some_arg': True},
-            )
-
-        self.assertIn(
-            'Both fsdp_config and parallelism_config cannot be provided at the same time.',
-            str(context.exception),
-        )

tests/models/test_model.py

@@ -15,7 +15,10 @@
 from accelerate import init_empty_weights
 from composer.core.precision import Precision, get_precision_context
 from composer.distributed.dist_strategy import prepare_fsdp_module
-from composer.models.huggingface import maybe_get_underlying_model
+from composer.models.huggingface import (
+    HuggingFaceModel,
+    maybe_get_underlying_model,
+)
 from composer.optim import DecoupledAdamW
 from composer.utils import (
     FSDPConfig,
@@ -39,7 +42,6 @@
 
 from llmfoundry import ComposerHFCausalLM
 from llmfoundry.layers_registry import norms
-from llmfoundry.models.hf.model_wrapper import HuggingFaceModelWithFSDP
 from llmfoundry.models.layers import build_alibi_bias
 from llmfoundry.models.layers.attention import (
     check_alibi_support,
@@ -2560,7 +2562,7 @@ def test_hf_init(
         False,
     )
 
-    model = HuggingFaceModelWithFSDP(model, tokenizer)
+    model = HuggingFaceModel(model, tokenizer)
 
     batch = gen_random_batch(batch_size, test_cfg)
 
@@ -2609,7 +2611,7 @@ def test_head_dim_8_flash_mqa_attn(batch_size: int = 2):
 
     mpt = MPTForCausalLM(hf_config)
 
-    model = HuggingFaceModelWithFSDP(mpt, tokenizer, shift_labels=True)
+    model = HuggingFaceModel(mpt, tokenizer, shift_labels=True)
 
     model = model.to(test_cfg.device)
     batch = gen_random_batch(batch_size, test_cfg)