🧼 remove v4.44 deprecations (#34245)

* remove v4.44 deprecations

* PR comments

* deprecations scheduled for v4.50

* hub version update

* make fiuxp

---------

Co-authored-by: Marc Sun <[email protected]>
Co-authored-by: Arthur <[email protected]>

setup.py

@@ -117,7 +117,7 @@
     "fugashi>=1.0",
     "GitPython<3.1.19",
     "hf-doc-builder>=0.3.0",
-    "huggingface-hub>=0.23.2,<1.0",
+    "huggingface-hub>=0.24.0,<1.0",
     "importlib_metadata",
     "ipadic>=1.0.0,<2.0",
     "isort>=5.5.4",

src/transformers/dependency_versions_table.py

@@ -24,7 +24,7 @@
     "fugashi": "fugashi>=1.0",
     "GitPython": "GitPython<3.1.19",
     "hf-doc-builder": "hf-doc-builder>=0.3.0",
-    "huggingface-hub": "huggingface-hub>=0.23.2,<1.0",
+    "huggingface-hub": "huggingface-hub>=0.24.0,<1.0",
     "importlib_metadata": "importlib_metadata",
     "ipadic": "ipadic>=1.0.0,<2.0",
     "isort": "isort>=5.5.4",

src/transformers/modeling_utils.py

@@ -94,7 +94,7 @@
     replace_return_docstrings,
     strtobool,
 )
-from .utils.hub import convert_file_size_to_int, create_and_tag_model_card, get_checkpoint_shard_files
+from .utils.hub import create_and_tag_model_card, get_checkpoint_shard_files
 from .utils.import_utils import (
     ENV_VARS_TRUE_VALUES,
     is_sagemaker_mp_enabled,
@@ -381,92 +381,6 @@ def check_support_param_buffer_assignment(model_to_load, state_dict, start_prefi
     return False
 
 
-def shard_checkpoint(
-    state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB", weights_name: str = WEIGHTS_NAME
-):
-    """
-    Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
-    given size.
-
-    The sub-checkpoints are determined by iterating through the `state_dict` in the order of its keys, so there is no
-    optimization made to make each sub-checkpoint as close as possible to the maximum size passed. For example, if the
-    limit is 10GB and we have weights of sizes [6GB, 6GB, 2GB, 6GB, 2GB, 2GB] they will get sharded as [6GB], [6+2GB],
-    [6+2+2GB] and not [6+2+2GB], [6+2GB], [6GB].
-
-    <Tip warning={true}>
-
-    If one of the model's weight is bigger than `max_shard_size`, it will end up in its own sub-checkpoint which will
-    have a size greater than `max_shard_size`.
-
-    </Tip>
-
-    Args:
-        state_dict (`Dict[str, torch.Tensor]`): The state dictionary of a model to save.
-        max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
-            The maximum size of each sub-checkpoint. If expressed as a string, needs to be digits followed by a unit
-            (like `"5MB"`).
-        weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
-            The name of the model save file.
-    """
-    logger.warning(
-        "Note that `shard_checkpoint` is deprecated and will be removed in v4.44. We recommend you using "
-        "split_torch_state_dict_into_shards from huggingface_hub library"
-    )
-    max_shard_size = convert_file_size_to_int(max_shard_size)
-
-    sharded_state_dicts = [{}]
-    last_block_size = 0
-    total_size = 0
-    storage_id_to_block = {}
-
-    for key, weight in state_dict.items():
-        # when bnb serialization is used the weights in the state dict can be strings
-        # check: https://github.com/huggingface/transformers/pull/24416 for more details
-        if isinstance(weight, str):
-            continue
-        else:
-            storage_id = id_tensor_storage(weight)
-
-        # If a `weight` shares the same underlying storage as another tensor, we put `weight` in the same `block`
-        if storage_id in storage_id_to_block and weight.device != torch.device("meta"):
-            block_id = storage_id_to_block[storage_id]
-            sharded_state_dicts[block_id][key] = weight
-            continue
-
-        weight_size = weight.numel() * dtype_byte_size(weight.dtype)
-        # If this weight is going to tip up over the maximal size, we split, but only if we have put at least one
-        # weight in the current shard.
-        if last_block_size + weight_size > max_shard_size and len(sharded_state_dicts[-1]) > 0:
-            sharded_state_dicts.append({})
-            last_block_size = 0
-
-        sharded_state_dicts[-1][key] = weight
-        last_block_size += weight_size
-        total_size += weight_size
-        storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
-
-    # If we only have one shard, we return it
-    if len(sharded_state_dicts) == 1:
-        return {weights_name: sharded_state_dicts[0]}, None
-
-    # Otherwise, let's build the index
-    weight_map = {}
-    shards = {}
-    for idx, shard in enumerate(sharded_state_dicts):
-        shard_file = weights_name.replace(".bin", f"-{idx+1:05d}-of-{len(sharded_state_dicts):05d}.bin")
-        shard_file = shard_file.replace(
-            ".safetensors", f"-{idx + 1:05d}-of-{len(sharded_state_dicts):05d}.safetensors"
-        )
-        shards[shard_file] = shard
-        for key in shard.keys():
-            weight_map[key] = shard_file
-
-    # Add the metadata
-    metadata = {"total_size": total_size}
-    index = {"metadata": metadata, "weight_map": weight_map}
-    return shards, index
-
-
 def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
     """
     This is the same as

src/transformers/models/blip_2/modeling_blip_2.py

@@ -2203,7 +2203,7 @@ def forward(
             logger.warning_once(
                 "Expanding inputs for image tokens in BLIP-2 should be done in processing. "
                 "Please follow instruction here (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042) to update your BLIP-2 model. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
             attention_mask = torch.cat(
@@ -2326,7 +2326,7 @@ def generate(
             logger.warning_once(
                 "Expanding inputs for image tokens in BLIP-2 should be done in processing. "
                 "Please follow instruction here (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042) to update your BLIP-2 model. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
             attention_mask = torch.cat(

src/transformers/models/blip_2/processing_blip_2.py

@@ -153,7 +153,7 @@ def __call__(
                 logger.warning_once(
                     "Expanding inputs for image tokens in BLIP-2 should be done in processing. "
                     "Please follow instruction here (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042) to update your BLIP-2 model. "
-                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
                 )
 
             # cast to desired return tensors type

src/transformers/models/instructblip/modeling_instructblip.py

@@ -1471,7 +1471,7 @@ def forward(
             logger.warning_once(
                 "Expanding inputs for image tokens in InstructBLIP should be done in processing. "
                 "Please follow instruction here (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042) to update your InstructBLIP model. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
             attention_mask = torch.cat(
@@ -1610,7 +1610,7 @@ def generate(
             logger.warning_once(
                 "Expanding inputs for image tokens in InstructBLIP should be done in processing. "
                 "Please follow instruction here (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042) to update your InstructBLIP model. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             inputs_embeds = torch.cat([language_model_inputs, inputs_embeds.to(language_model_inputs.device)], dim=1)
             attention_mask = torch.cat(

src/transformers/models/instructblip/processing_instructblip.py

@@ -148,7 +148,7 @@ def __call__(
                     logger.warning_once(
                         "Expanding inputs for image tokens in InstructBLIP should be done in processing. "
                         "Please follow instruction here (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042) to update your InstructBLIP model. "
-                        "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                        "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
                     )
 
             # cast to desired return tensors type after concatenating

src/transformers/models/llava/modeling_llava.py

@@ -485,7 +485,7 @@ def forward(
                 "Expanding inputs for image tokens in LLaVa should be done in processing. "
                 "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                 "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             # prefill stage vs decoding stage (legacy behavior copied)
             if input_ids.shape[1] != 1:

src/transformers/models/llava/processing_llava.py

@@ -160,7 +160,7 @@ def __call__(
                     "Expanding inputs for image tokens in LLaVa should be done in processing. "
                     "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                     "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
-                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
                 )
 
         text_inputs = self.tokenizer(prompt_strings, **output_kwargs["text_kwargs"])

src/transformers/models/llava_next/modeling_llava_next.py

@@ -868,7 +868,7 @@ def forward(
                 "Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
                 "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                 "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             if input_ids.shape[1] != 1:
                 inputs_embeds = inputs_embeds.to(image_features.dtype)

src/transformers/models/llava_next/processing_llava_next.py

@@ -143,7 +143,7 @@ def __call__(
                     "Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
                     "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                     "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
-                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
                 )
             else:
                 image_sizes = iter(image_inputs["image_sizes"])

src/transformers/models/megatron_gpt2/checkpoint_reshaping_and_interoperability.py

@@ -21,10 +21,11 @@
 import types
 
 import torch
+from huggingface_hub import split_torch_state_dict_into_shards
 from packaging import version
 
 from transformers import AutoTokenizer, GPT2Config
-from transformers.modeling_utils import WEIGHTS_INDEX_NAME, WEIGHTS_NAME, shard_checkpoint
+from transformers.modeling_utils import WEIGHTS_INDEX_NAME, WEIGHTS_NAME
 
 
 def add_checkpointing_args(parser):
@@ -571,7 +572,15 @@ def convert_checkpoint_from_megatron_to_transformers(args):
 
     # Store the state_dict to file.
     max_shard_size = int(args.max_shard_size) if args.max_shard_size.isdigit() else args.max_shard_size
-    shards, index = shard_checkpoint(output_state_dict, max_shard_size=max_shard_size)
+    state_dict_split = split_torch_state_dict_into_shards(output_state_dict, max_shard_size=max_shard_size)
+    shards = index = None
+    for tensors in state_dict_split.filename_to_tensors.values():
+        shards = {tensor: state_dict[tensor] for tensor in tensors}
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
 
     # Save the model
     for shard_file, shard in shards.items():

src/transformers/models/rwkv/convert_rwkv_checkpoint_to_hf.py

@@ -21,10 +21,10 @@
 import re
 
 import torch
-from huggingface_hub import hf_hub_download
+from huggingface_hub import hf_hub_download, split_torch_state_dict_into_shards
 
 from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerFast, RwkvConfig
-from transformers.modeling_utils import WEIGHTS_INDEX_NAME, shard_checkpoint
+from transformers.modeling_utils import WEIGHTS_INDEX_NAME
 
 
 NUM_HIDDEN_LAYERS_MAPPING = {
@@ -116,7 +116,16 @@ def convert_rmkv_checkpoint_to_hf_format(
     state_dict = convert_state_dict(state_dict)
 
     # 4. Split in shards and save
-    shards, index = shard_checkpoint(state_dict)
+    state_dict_split = split_torch_state_dict_into_shards(state_dict)
+    shards = index = None
+    for tensors in state_dict_split.filename_to_tensors.values():
+        shards = {tensor: state_dict[tensor] for tensor in tensors}
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+
     for shard_file, shard in shards.items():
         torch.save(shard, os.path.join(output_dir, shard_file))
 

src/transformers/models/video_llava/modeling_video_llava.py

@@ -578,7 +578,7 @@ def forward(
                 "Expanding inputs for image tokens in Video-LLaVa should be done in processing. "
                 "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                 "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             if input_ids.shape[1] != 1:
                 for features, frames in ((image_features, 1), (video_features, num_frames)):

src/transformers/models/video_llava/processing_video_llava.py

@@ -149,9 +149,10 @@ def __call__(
         if encoded_images is not None and (self.patch_size is None or self.vision_feature_select_strategy is None):
             logger.warning_once(
                 "Expanding inputs for image tokens in Video-LLaVa should be done in processing. "
-                "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
-                "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.44."
+                "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set "
+                "directly with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = "
+                "{{vision_feature_select_strategy}}`. Using processors without these attributes in the config is "
+                "deprecated and will throw an error in v4.50."
             )
         # Replace the image/video tokens with the expanded token sequence
         elif encoded_images is not None:

src/transformers/models/vipllava/modeling_vipllava.py

@@ -476,7 +476,7 @@ def forward(
             logger.warning_once(
                 "Expanding inputs for image tokens in VipLLaVa should be done in processing. "
                 "Please add `patch_size` and `vision_feature_select_strategy` to the model's image processing config. "
-                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
+                "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
             )
             # prefill stage vs decoding stage (legacy behavior copied)
             if input_ids.shape[1] != 1: