[modeling utils] revamp `from_pretrained(..., low_cpu_mem_usage=True)` + tests (#16657)

* add low_cpu_mem_usage tests

* wip: revamping

* wip

* install /usr/bin/time

* wip

* cleanup

* cleanup

* cleanup

* cleanup

* cleanup

* fix assert

* put the wrapper back

* cleanup; switch to bert-base-cased

* Trigger CI

* Trigger CI

.circleci/config.yml

@@ -217,7 +217,7 @@ jobs:
                   keys:
                       - v0.4-torch-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
-            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
+            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng time
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
             - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html

src/transformers/modeling_utils.py

@@ -400,6 +400,95 @@ def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
     return error_msgs
 
 
+def find_submodule_and_param_name(model, long_key, start_prefix):
+    """
+    A helper util to find the last sub-module and the param/buffer name. If `start_prefix` is supplied it'll be removed
+    from the start of the key
+    """
+
+    if len(start_prefix) > 0 and long_key.startswith(start_prefix):
+        long_key = ".".join(long_key.split(".")[1:])
+
+    split_key = long_key.split(".")
+    submodule = model
+    while len(split_key) > 1:
+        if hasattr(submodule, split_key[0]):
+            submodule = getattr(submodule, split_key[0])
+            del split_key[0]
+        else:
+            submodule = None
+            break
+    if submodule == model:
+        submodule = None
+    return submodule, split_key[0]
+
+
+def _move_model_to_meta(model, loaded_state_dict_keys, start_prefix):
+    """
+    Moves `loaded_state_dict_keys` in model to meta device which frees up the memory taken by those params.
+
+    `start_prefix` is used for models which insert their name into model keys, e.g. `bert` in
+    `bert.pooler.dense.weight`
+
+    """
+
+    # meta device was added in pt=1.9
+    require_version_core("torch>=1.9")
+
+    # dematerialize param storage for keys that are going to be replaced by state_dict, by
+    # putting those on the meta device
+    for k in loaded_state_dict_keys:
+        submodule, param_name = find_submodule_and_param_name(model, k, start_prefix)
+        if submodule is not None:
+            # selectively switch to the meta device only those params/buffers that will
+            # be next replaced from state_dict. This a complex way to do p.to_("meta")
+            # since we have no in-place to_ for tensors.
+            new_val = getattr(submodule, param_name)
+            if isinstance(new_val, torch.nn.Parameter):
+                # isinstance returns False for Params on meta device, so switch after the check
+                new_val = torch.nn.Parameter(new_val.to("meta"))
+            else:
+                new_val = new_val.to("meta")
+            setattr(submodule, param_name, new_val)
+
+
+def _load_state_dict_into_meta_model(model, state_dict, loaded_state_dict_keys, start_prefix):
+    """
+    This is somewhat similar to `_load_state_dict_into_model`, but deals with a model that has some or all of its
+    params on a `meta` device. It replaces the model params with the data from the `state_dict`, while moving the
+    params back to the normal device, but only for `loaded_state_dict_keys`.
+
+    `start_prefix` is used for models which insert their name into model keys, e.g. `bert` in
+    `bert.pooler.dense.weight`
+
+    """
+
+    # XXX: remaining features to implement to be fully compatible with _load_state_dict_into_model
+    # - deepspeed zero 3 support
+    # - need to copy metadata if any - see _load_state_dict_into_model
+    # - handling error_msgs - mimicking the error handling in module._load_from_state_dict()
+    # - Is there a situation where some keys aren't in `loaded_state_dict_keys` and in which case
+    #   they won't get loaded.
+
+    if is_deepspeed_zero3_enabled():
+        raise ValueError("low_cpu_mem_usage arg cannot currently be used with DeepSpeed ZeRO-3")
+
+    error_msgs = []
+
+    # materialize state_dict entries one by one on CPU
+    for k in loaded_state_dict_keys:
+        if k in state_dict:
+            submodule, param_name = find_submodule_and_param_name(model, k, start_prefix)
+            if submodule is not None:
+                param_dtype = getattr(submodule, param_name).dtype
+                new_val = state_dict[k].to(param_dtype)
+                if isinstance(getattr(submodule, param_name), torch.nn.Parameter):
+                    new_val = torch.nn.Parameter(new_val)
+                setattr(submodule, param_name, new_val)
+
+    return error_msgs
+
+
 class ModuleUtilsMixin:
     """
     A few utilities for `torch.nn.Modules`, to be used as a mixin.
@@ -1529,7 +1618,24 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         >>> model = BertModel.from_pretrained("./tf_model/my_tf_checkpoint.ckpt.index", from_tf=True, config=config)
         >>> # Loading from a Flax checkpoint file instead of a PyTorch model (slower)
         >>> model = BertModel.from_pretrained("bert-base-uncased", from_flax=True)
-        ```"""
+        ```
+
+        * `low_cpu_mem_usage` algorithm:
+
+        This is an experimental function that loads the model using ~1x model size CPU memory
+
+        Here is how it works:
+
+        1. save which state_dict keys we have
+        2. drop state_dict before the model is created, since the latter takes 1x model size CPU memory
+        3. after the model has been instantiated switch to the meta device all params/buffers that
+        are going to be replaced from the loaded state_dict
+        4. load state_dict 2nd time
+        5. replace the params/buffers from the state_dict
+
+        Currently, it can't handle deepspeed ZeRO stage 3 and ignores loading errors
+
+        """
         config = kwargs.pop("config", None)
         state_dict = kwargs.pop("state_dict", None)
         cache_dir = kwargs.pop("cache_dir", None)
@@ -1778,6 +1884,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             if not is_sharded and state_dict is None:
                 # Time to load the checkpoint
                 state_dict = load_state_dict(resolved_archive_file)
+
             # set dtype to instantiate the model under:
             # 1. If torch_dtype is not None, we use that dtype
             # 2. If torch_dtype is "auto", we auto-detect dtype from the loaded state_dict, by checking its first
@@ -1801,13 +1908,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                         )
                 dtype_orig = cls._set_default_torch_dtype(torch_dtype)
 
-            if low_cpu_mem_usage:
-                # save the keys
             if is_sharded:
                 loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
             else:
                 loaded_state_dict_keys = [k for k in state_dict.keys()]
-                    del state_dict  # free CPU memory - will reload again later
+            if low_cpu_mem_usage:
+                state_dict = None
 
         config.name_or_path = pretrained_model_name_or_path
 
@@ -1825,11 +1931,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             with no_init_weights(_enable=_fast_init):
                 model = cls(config, *model_args, **model_kwargs)
 
-        if from_pt:
-            # restore default dtype
-            if dtype_orig is not None:
-                torch.set_default_dtype(dtype_orig)
-
         if from_tf:
             if resolved_archive_file.endswith(".index"):
                 # Load from a TensorFlow 1.X checkpoint - provided by original authors
@@ -1859,17 +1960,20 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 raise
         elif from_pt:
 
-            if low_cpu_mem_usage:
-                cls._load_pretrained_model_low_mem(model, loaded_state_dict_keys, resolved_archive_file)
-            else:
+            # restore default dtype
+            if dtype_orig is not None:
+                torch.set_default_dtype(dtype_orig)
+
             model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
                 model,
                 state_dict,
+                loaded_state_dict_keys,  # XXX: rename?
                 resolved_archive_file,
                 pretrained_model_name_or_path,
                 ignore_mismatched_sizes=ignore_mismatched_sizes,
                 sharded_metadata=sharded_metadata,
                 _fast_init=_fast_init,
+                low_cpu_mem_usage=low_cpu_mem_usage,
             )
 
         # make sure token embedding weights are still tied if needed
@@ -1894,16 +1998,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         cls,
         model,
         state_dict,
+        loaded_keys,
         resolved_archive_file,
         pretrained_model_name_or_path,
         ignore_mismatched_sizes=False,
         sharded_metadata=None,
         _fast_init=True,
+        low_cpu_mem_usage=False,
     ):
         # Retrieve missing & unexpected_keys
         model_state_dict = model.state_dict()
         expected_keys = list(model_state_dict.keys())
-        loaded_keys = list(state_dict.keys()) if state_dict is not None else sharded_metadata["all_checkpoint_keys"]
         prefix = model.base_model_prefix
 
         def _fix_key(key):
@@ -1994,9 +2099,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                             (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
                         )
                         del state_dict[checkpoint_key]
-
             return mismatched_keys
 
+        if low_cpu_mem_usage:
+            model_state_dict = None  # free references to model's params to allow memory freeing
+            _move_model_to_meta(model, loaded_keys, start_prefix)
+
         if state_dict is not None:
             # Whole checkpoint
             mismatched_keys = _find_mismatched_keys(
@@ -2009,7 +2117,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             )
             error_msgs = _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
         else:
-            # Sharded checkpoint
+            # Sharded checkpoint or whole but low_cpu_mem_usage==True
+
             # This should always be a list but, just to be sure.
             if not isinstance(resolved_archive_file, list):
                 resolved_archive_file = [resolved_archive_file]
@@ -2018,6 +2127,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             mismatched_keys = []
             for shard_file in resolved_archive_file:
                 state_dict = load_state_dict(shard_file)
+
+                if low_cpu_mem_usage:
+                    model_state_dict = model.state_dict()
+
                 # Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
                 # matching the weights in the model.
                 mismatched_keys += _find_mismatched_keys(
@@ -2028,6 +2141,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     remove_prefix_from_model,
                     ignore_mismatched_sizes,
                 )
+
+                if low_cpu_mem_usage:
+                    error_msgs += _load_state_dict_into_meta_model(
+                        model_to_load, state_dict, loaded_keys, start_prefix
+                    )
+                else:
                     error_msgs += _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
 
         if len(error_msgs) > 0:
@@ -2093,13 +2212,13 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         return retrieved_modules
 
     @staticmethod
-    def _load_pretrained_model_low_mem(model, loaded_state_dict_keys, resolved_archive_file):
+    def _load_pretrained_model_low_mem(model, loaded_state_dict_keys, resolved_archive_file, start_prefix=""):
         """
         This is an experimental function that loads the model using ~1.x model size CPU memory
 
-        Before it gets called we do:
+        Before you call it do:
 
-        1. save which state_dict keys we have
+        1. save which state_dict keys are available
         2. drop state_dict before model is created, since the latter takes 1x model size memory
 
         Here then we continue:
@@ -2110,58 +2229,11 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
         Currently, it doesn't handle missing_keys, unexpected_keys, mismatched_keys. It can't handle deepspeed.
         """
-        require_version_core("torch>=1.9")
-        if is_deepspeed_zero3_enabled():
-            raise ValueError("low_cpu_mem_usage arg cannot be used with DeepSpeed ZeRO-3")
 
-        # a helper util to find the last sub-module and the param/buffer name
-        def find_submodule_and_param_name(model, long_key):
-            split_key = long_key.split(".")
-            submodule = model
-            while len(split_key) > 1:
-                if hasattr(submodule, split_key[0]):
-                    submodule = getattr(submodule, split_key[0])
-                    del split_key[0]
-                else:
-                    submodule = None
-                    break
-            return submodule, split_key[0]
-
-        # dematerialize param storage for keys that are going to be replaced by state_dict, by
-        # putting those on the meta device
-        for k in loaded_state_dict_keys:
-            submodule, param_name = find_submodule_and_param_name(model, k)
-            if submodule is not None:
-                # selectively switch to the meta device only those params/buffers that will
-                # be next replaced from state_dict. This a complex way to do p.to_("meta")
-                # since we have no in-place to_ for tensors.
-                new_val = getattr(submodule, param_name)
-                if isinstance(new_val, torch.nn.Parameter):
-                    # isinstance returns False for Params on meta device, so switch after the check
-                    new_val = torch.nn.Parameter(new_val.to("meta"))
-                else:
-                    new_val = new_val.to("meta")
-                setattr(submodule, param_name, new_val)
-
-        # only now can load state_dict(s)
-        if not isinstance(resolved_archive_file, list):
-            resolved_archive_file = [resolved_archive_file]
-
-        for archive_file in resolved_archive_file:
-            state_dict = torch.load(archive_file, map_location="cpu")
-
-            # materialize state_dict entries one by one on CPU
-            for k in loaded_state_dict_keys:
-                if k in state_dict:
-                    submodule, param_name = find_submodule_and_param_name(model, k)
-                    if submodule is not None:
-                        param_dtype = getattr(submodule, param_name).dtype
-                        new_val = state_dict[k].to(param_dtype)
-                        if isinstance(getattr(submodule, param_name), torch.nn.Parameter):
-                            new_val = torch.nn.Parameter(new_val)
-                        setattr(submodule, param_name, new_val)
-
-            del state_dict
+        _move_model_to_meta(model, loaded_state_dict_keys, start_prefix)
+        state_dict = load_state_dict(resolved_archive_file)
+        error_msgs = _load_state_dict_into_meta_model(model, state_dict, loaded_state_dict_keys, start_prefix)
+        return error_msgs
 
     @classmethod
     def register_for_auto_class(cls, auto_class="AutoModel"):

src/transformers/testing_utils.py

@@ -17,6 +17,7 @@ import inspect
 import logging
 import os
 import re
+import shlex
 import shutil
 import sys
 import tempfile
@@ -667,6 +668,20 @@ def require_librosa(test_case):
         return test_case
 
 
+def cmd_exists(cmd):
+    return shutil.which(cmd) is not None
+
+
+def require_usr_bin_time(test_case):
+    """
+    Decorator marking a test that requires `/usr/bin/time`
+    """
+    if not cmd_exists("/usr/bin/time"):
+        return unittest.skip("test requires /usr/bin/time")(test_case)
+    else:
+        return test_case
+
+
 def get_gpu_count():
     """
     Return the number of available gpus (regardless of whether torch, tf or jax is used)
@@ -1178,6 +1193,39 @@ class TestCasePlus(unittest.TestCase):
 
         return tmp_dir
 
+    def python_one_liner_max_rss(self, one_liner_str):
+        """
+        Runs the passed python one liner (just the code) and returns how much max cpu memory was used to run the
+        program.
+
+        Args:
+            one_liner_str (`string`):
+                a python one liner code that gets passed to `python -c`
+
+        Returns:
+            max cpu memory bytes used to run the program. This value is likely to vary slightly from run to run.
+
+        Requirements:
+            this helper needs `/usr/bin/time` to be installed (`apt install time`)
+
+        Example:
+
+        ```
+        one_liner_str = 'from transformers import AutoModel; AutoModel.from_pretrained("t5-large")'
+        max_rss = self.python_one_liner_max_rss(one_liner_str)
+        ```
+        """
+
+        if not cmd_exists("/usr/bin/time"):
+            raise ValueError("/usr/bin/time is required, install with `apt install time`")
+
+        cmd = shlex.split(f"/usr/bin/time -f %M python -c '{one_liner_str}'")
+        with CaptureStd() as cs:
+            execute_subprocess_async(cmd, env=self.get_env())
+        # returned data is in KB so convert to bytes
+        max_rss = int(cs.err.split("\n")[-2].replace("stderr: ", "")) * 1024
+        return max_rss
+
     def tearDown(self):
 
         # get_auto_remove_tmp_dir feature: remove registered temp dirs

tests/test_modeling_common.py

@@ -52,6 +52,7 @@ from transformers.testing_utils import (
     is_staging_test,
     require_torch,
     require_torch_multi_gpu,
+    require_usr_bin_time,
     slow,
     torch_device,
 )
@@ -2489,6 +2490,56 @@ class ModelUtilsTest(TestCasePlus):
         for p1, p2 in zip(model.parameters(), ref_model.parameters()):
             self.assertTrue(torch.allclose(p1, p2))
 
+    def test_from_pretrained_low_cpu_mem_usage_functional(self):
+        # test that we can use `from_pretrained(..., low_cpu_mem_usage=True)` with normal and
+        # sharded models
+
+        mnames = [
+            "hf-internal-testing/tiny-random-bert-sharded",
+            "hf-internal-testing/tiny-random-bert",
+        ]
+        for mname in mnames:
+            _ = BertModel.from_pretrained(mname, low_cpu_mem_usage=True)
+
+    @require_usr_bin_time
+    def test_from_pretrained_low_cpu_mem_usage_measured(self):
+        # test that `from_pretrained(..., low_cpu_mem_usage=True)` uses less cpu memory than default
+
+        mname = "bert-base-cased"
+
+        preamble = "from transformers import AutoModel"
+        one_liner_str = f'{preamble}; AutoModel.from_pretrained("{mname}", low_cpu_mem_usage=False)'
+        max_rss_normal = self.python_one_liner_max_rss(one_liner_str)
+        # print(f"{max_rss_normal=}")
+
+        one_liner_str = f'{preamble};  AutoModel.from_pretrained("{mname}", low_cpu_mem_usage=True)'
+        max_rss_low_mem = self.python_one_liner_max_rss(one_liner_str)
+        # print(f"{max_rss_low_mem=}")
+
+        diff_bytes = max_rss_normal - max_rss_low_mem
+        diff_percent = diff_bytes / max_rss_low_mem
+        # print(f"{diff_bytes=}, {diff_percent=}")
+        # ideally we would compare that the diff is close to ~1x checkpoint size in bytes, but
+        # measuring cpu memory on linux is very tricky and inconsistent, so instead let's check that
+        # it's at least 15% less cpu memory consumed
+
+        self.assertGreater(
+            diff_percent,
+            0.15,
+            "should use less CPU memory for low_cpu_mem_usage=True, "
+            f"but got max_rss_normal={max_rss_normal} and max_rss_low_mem={max_rss_low_mem}",
+        )
+
+        # if you want to compare things manually, let's first look at the size of the model in bytes
+        # model = BertModel.from_pretrained(mname, low_cpu_mem_usage=False)
+        # total_numel = sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())
+        # total_bytes = total_numel * 4  # 420MB
+        # Now the diff_bytes should be very close to total_bytes, but the reports are inconsistent.
+        # The easiest way to test this is to switch the model and torch.load to do all the work on
+        # gpu - that way one can measure exactly the total and peak memory used. Perhaps once we add
+        # functionality to load models directly on gpu, this test can be rewritten to use torch's
+        # cuda memory tracking and then we should be able to do a much more precise test.
+
     def test_cached_files_are_used_when_internet_is_down(self):
         # A mock response for an HTTP head request to emulate server down
         response_mock = mock.Mock()