riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-03 21:00:08 +06:00
Code Issues Actions 30 Packages Projects Releases Wiki Activity

chore: enhance message descriptions in parameters,comments,logs and docstrings (#36554)

Browse Source 
* chore: enhance message descriptons in parameters,comments,logs and docstrings

* chore: enhance message descriptons in parameters,comments,logs and docstrings

* Update src/transformers/hf_argparser.py

* Update src/transformers/keras_callbacks.py

---------

Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
This commit is contained in:
Afanti 2025-03-06 19:02:35 +08:00 committed by GitHub
parent 6966fa1901
commit 9e84b38135
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
17 changed files with 55 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/transformers/configuration_utils.py
View File

@ -191,7 +191,7 @@ class PretrainedConfig(PushToHubMixin):
v5.
loss_type (`str`, *optional*):
The type of loss that the model should use. It should be in `LOSS_MAPPING`'s keys, otherwise the loss will
be automatically infered from the model architecture.
be automatically inferred from the model architecture.
"""
model_type: str = ""
@ -254,7 +254,7 @@ class PretrainedConfig(PushToHubMixin):
if num_labels is not None and len(self.id2label) != num_labels:
logger.warning(
f"You passed along `num_labels={num_labels}` with an incompatible id to label map: "
f"{self.id2label}. The number of labels wil be overwritten to {self.num_labels}."
f"{self.id2label}. The number of labels will be overwritten to {self.num_labels}."
)
self.id2label = {int(key): value for key, value in self.id2label.items()}
# Keys are always strings in JSON so convert ids to int here.
@ -1094,7 +1094,7 @@ class PretrainedConfig(PushToHubMixin):
is_default_in_config = is_default_generation_value = None
parameter_value = getattr(self_decoder_config, parameter_name)
# Three cases in which is okay for the model config to hold generation config parameters:
# 1. The parameter is set to `None`, effectivelly delegating its value to the generation config
# 1. The parameter is set to `None`, effectively delegating its value to the generation config
if parameter_value is None:
continue
# 2. If we have a default config, then the instance should hold the same generation defaults

2
src/transformers/convert_slow_tokenizer.py
View File

@ -1727,5 +1727,5 @@ def convert_slow_tokenizer(transformer_tokenizer, from_tiktoken=False) -> Tokeni
raise ValueError(
f"Converting from Tiktoken failed, if a converter for SentencePiece is available, provide a model path "
f"with a SentencePiece tokenizer.model file."
f"Currently available slow->fast convertors: {list(SLOW_TO_FAST_CONVERTERS.keys())}"
f"Currently available slow->fast converters: {list(SLOW_TO_FAST_CONVERTERS.keys())}"
)

2
src/transformers/hf_argparser.py
View File

@ -201,7 +201,7 @@ class HfArgumentParser(ArgumentParser):
else:
kwargs["required"] = True
elif field.type is bool or field.type == Optional[bool]:
# Copy the currect kwargs to use to instantiate a `no_*` complement argument below.
# Copy the correct kwargs to use to instantiate a `no_*` complement argument below.
# We do not initialize it here because the `no_*` alternative must be instantiated after the real argument
bool_kwargs = copy(kwargs)

2
src/transformers/image_transforms.py
View File

@ -585,7 +585,7 @@ def center_to_corners_format(bboxes_center: TensorType) -> TensorType:
center format: contains the coordinate for the center of the box and its width, height dimensions
(center_x, center_y, width, height)
corners format: contains the coodinates for the top-left and bottom-right corners of the box
corners format: contains the coordinates for the top-left and bottom-right corners of the box
(top_left_x, top_left_y, bottom_right_x, bottom_right_y)
"""
# Function is used during model forward pass, so we use the input framework if possible, without

2
src/transformers/image_utils.py
View File

@ -545,7 +545,7 @@ def default_sample_indices_fn(metadata: VideoMetadata, num_frames=None, fps=None
Args:
metadata (`VideoMetadata`):
`VideoMetadata` object containing metadat about the video, such as "total_num_frames" or "fps".
`VideoMetadata` object containing metadata about the video, such as "total_num_frames" or "fps".
num_frames (`int`, *optional*):
Number of frames to sample uniformly.
fps (`int`, *optional*):

10
src/transformers/modeling_flash_attention_utils.py
View File

@ -137,9 +137,9 @@ def prepare_fa2_from_position_ids(query, key, value, position_ids):
"""
This function returns necessary arguments to call `flash_attn_varlen_func`.
All three query, key, value states will be flattened.
Cummulative lengths of each examples in the batch will be extracted from position_ids.
Cumulative lengths of each examples in the batch will be extracted from position_ids.
NOTE: ideally cummulative lengths should be prepared at the data collator stage
NOTE: ideally cumulative lengths should be prepared at the data collator stage
Arguments:
query (`torch.Tensor`):
@ -268,7 +268,7 @@ def _flash_attention_forward(
softmax_scale (`float`, *optional*):
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
use_top_left_mask (`bool`, defaults to `False`):
flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference.
flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference.
softcap (`float`, *optional*):
Softcap for the attention logits, used e.g. in gemma2.
deterministic (`bool`, *optional*):
@ -374,9 +374,9 @@ class FlashAttentionKwargs(TypedDict, total=False):
Attributes:
cu_seq_lens_q (`torch.LongTensor`, *optional*)
Gets cumlative sequence length for query state.
Gets cumulative sequence length for query state.
cu_seq_lens_k (`torch.LongTensor`, *optional*)
Gets cumlative sequence length for key state.
Gets cumulative sequence length for key state.
max_length_q (`int`, *optional*):
Maximum sequence length for query state.
max_length_k (`int`, *optional*):

8
src/transformers/modeling_flax_utils.py
View File

@ -367,7 +367,7 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
def to_fp32(self, params: Union[Dict, FrozenDict], mask: Any = None):
r"""
Cast the floating-point `parmas` to `jax.numpy.float32`. This method can be used to explicitly convert the
Cast the floating-point `params` to `jax.numpy.float32`. This method can be used to explicitly convert the
model parameters to fp32 precision. This returns a new `params` tree and does not cast the `params` in place.
Arguments:
@ -394,7 +394,7 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
def to_fp16(self, params: Union[Dict, FrozenDict], mask: Any = None):
r"""
Cast the floating-point `parmas` to `jax.numpy.float16`. This returns a new `params` tree and does not cast the
Cast the floating-point `params` to `jax.numpy.float16`. This returns a new `params` tree and does not cast the
`params` in place.
This method can be used on GPU to explicitly convert the model parameters to float16 precision to do full
@ -510,7 +510,7 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
`bool`: Whether this model can generate sequences with `.generate()`.
"""
# Detects whether `prepare_inputs_for_generation` has been overwritten, which is a requirement for generation.
# Alternativelly, the model can also have a custom `generate` function.
# Alternatively, the model can also have a custom `generate` function.
if "GenerationMixin" in str(cls.prepare_inputs_for_generation) and "GenerationMixin" in str(cls.generate):
return False
return True
@ -968,7 +968,7 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
)
cls._missing_keys = missing_keys
# Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
# Mismatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
# matching the weights in the model.
mismatched_keys = []
for key in state.keys():

4
src/transformers/modeling_gguf_pytorch_utils.py
View File

@ -373,7 +373,7 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False, model_to_lo
# to add this patch to ensure things work correctly on our side.
if "llama" in architecture and "mistral" in model_name:
updated_architecture = "mistral"
# FIXME: Currnetly this implementation is only for flan-t5 architecture.
# FIXME: Currently this implementation is only for flan-t5 architecture.
# It needs to be developed for supporting legacy t5.
elif "t5" in architecture or "t5encoder" in architecture:
parsed_parameters["config"]["is_gated_act"] = True
@ -437,7 +437,7 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False, model_to_lo
logger.info(f"Some keys were not parsed and added into account {gguf_key} | {value}")
# retrieve config vocab_size from tokenizer
# Pleas refer to https://github.com/huggingface/transformers/issues/32526 for more details
# Please refer to https://github.com/huggingface/transformers/issues/32526 for more details
if "vocab_size" not in parsed_parameters["config"]:
tokenizer_parameters = parsed_parameters["tokenizer"]
if "tokens" in tokenizer_parameters:

6
src/transformers/modeling_tf_utils.py
View File

@ -795,7 +795,7 @@ def load_tf_shard(model, model_layer_map, resolved_archive_file, ignore_mismatch
ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`): Whether to ignore the mismatched keys
Returns:
`keras.models.Model`: Three lists, one for the layers that were found and succesfully restored (from the
`keras.models.Model`: Three lists, one for the layers that were found and successfully restored (from the
shard file), one for the mismatched layers, and another one for the unexpected layers.
"""
saved_weight_names_set = set()
@ -868,7 +868,7 @@ def load_tf_shard(model, model_layer_map, resolved_archive_file, ignore_mismatch
f"Unable to load weights from TF checkpoint file for '{resolved_archive_file}' "
f"at '{resolved_archive_file}'. "
"If you tried to load a TF model from a sharded checkpoint, you should try converting the model "
"by loading it in pytorch and saving it localy. A convertion script should be realeased soon."
"by loading it in pytorch and saving it locally. A convertion script should be released soon."
)
@ -1391,7 +1391,7 @@ class TFPreTrainedModel(keras.Model, TFModelUtilsMixin, TFGenerationMixin, PushT
`bool`: Whether this model can generate sequences with `.generate()`.
"""
# Detects whether `prepare_inputs_for_generation` has been overwritten, which is a requirement for generation.
# Alternativelly, the model can also have a custom `generate` function.
# Alternatively, the model can also have a custom `generate` function.
if "GenerationMixin" in str(cls.prepare_inputs_for_generation) and "GenerationMixin" in str(cls.generate):
return False
return True

24
src/transformers/modeling_utils.py
View File

@ -1324,7 +1324,7 @@ def _find_mismatched_keys(
and state_dict[checkpoint_key].numel() * 2 == model_state_dict[model_key].numel()
):
# This skips size mismatches for 4-bit weights. Two 4-bit values share an 8-bit container, causing size differences.
# Without matching with module type or paramter type it seems like a practical way to detect valid 4bit weights.
# Without matching with module type or parameter type it seems like a practical way to detect valid 4bit weights.
pass
else:
mismatched_keys.append(
@ -1616,7 +1616,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
3. SDPA implementation, if available and supported by the model type. (`LlamaSdpaAttention` for example)
4. The default model's implementation otherwise (`LlamaAttention` for example) .
"""
# Here we use config._attn_implementation_internal to check whether the attention implementation was explicitely set by the user.
# Here we use config._attn_implementation_internal to check whether the attention implementation was explicitly set by the user.
# The property `PretrainedConfig._attn_implementation` is never `None`, for backward compatibility (always fall back on "eager").
# The `hasattr` here is used as some Transformers tests for some reason do not call PretrainedConfig __init__ (e.g. test_no_super_init_config_and_model)
requested_attn_implementation = None
@ -2207,7 +2207,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
if new_num_tokens is None and pad_to_multiple_of is None:
return model_embeds
# Since we are basically resuing the same old embeddings with new weight values, gathering is required
# Since we are basically reusing the same old embeddings with new weight values, gathering is required
is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
if is_deepspeed_zero3_enabled() and not is_quantized:
import deepspeed
@ -2574,7 +2574,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
sample_shape=(added_num_tokens,)
).to(old_embeddings.weight.dtype)
else:
# Otherwise, just initialize with the mean. because distribtion will not be created.
# Otherwise, just initialize with the mean. because distribution will not be created.
new_embeddings.weight.data[-1 * added_num_tokens :, :] = (
mean_embeddings[None, :].repeat(added_num_tokens, 1).to(old_embeddings.weight.dtype)
)
@ -2593,7 +2593,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
new_lm_head.weight.data = new_lm_head.weight.data.T
old_lm_head.weight.data = old_lm_head.weight.data.T
# The same initilization logic as Embeddings.
# The same initialization logic as Embeddings.
self._init_added_embeddings_weights_with_mean(
old_lm_head, new_lm_head, old_lm_head_dim, old_num_tokens, added_num_tokens
)
@ -2740,7 +2740,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
"""
if self.supports_gradient_checkpointing:
# For old GC format (transformers < 4.35.0) for models that live on the Hub
# we will fall back to the overwritten `_set_gradient_checkpointing` methid
# we will fall back to the overwritten `_set_gradient_checkpointing` method
_is_using_old_format = "value" in inspect.signature(self._set_gradient_checkpointing).parameters
if not _is_using_old_format:
self._set_gradient_checkpointing(enable=False)
@ -2979,7 +2979,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
if ignore_key in state_dict.keys():
del state_dict[ignore_key]
# Rename state_dict keys before saving to file. Do nothing unless overriden in a particular model.
# Rename state_dict keys before saving to file. Do nothing unless overridden in a particular model.
# (initially introduced with TimmWrapperModel to remove prefix and make checkpoints compatible with timm)
state_dict = self._fix_state_dict_keys_on_save(state_dict)
@ -4998,7 +4998,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
shard_file, is_quantized=is_quantized, map_location="meta", weights_only=weights_only
)
# Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
# Mismatched 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(
state_dict,
@ -5321,13 +5321,13 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
"""
Tensor parallelize the model across the given device mesh. This function is a helper to be called after the model
was already loaded in memory, note however that this means that each process will first initialize the whole model,
then parallelize it accross devices. Thus there is a huge waste of GPU memory, and this can lead to OOM at loading time.
then parallelize it across devices. Thus there is a huge waste of GPU memory, and this can lead to OOM at loading time.
Calling `from_pretrained(..., tp_plan="auto")` is prefered, and will parallelize module-by-module during initialization,
Calling `from_pretrained(..., tp_plan="auto")` is preferred, and will parallelize module-by-module during initialization,
so that the expected per-device memory spike at loading time is not larger than the final model size on each device.
Tensor parallelize the model across the given device mesh. This function is a helper to be called after the model
was already loaded in memory, note however that this means that each process will first initialize the whole model,
then parallelize it accross devices. Thus there is a huge waste of GPU memory, and this can lead to OOM at loading time.
then parallelize it across devices. Thus there is a huge waste of GPU memory, and this can lead to OOM at loading time.
Args:
device_mesh (`torch.distributed.DeviceMesh`):
@ -5869,7 +5869,7 @@ def unwrap_model(model: nn.Module, recursive: bool = False) -> nn.Module:
def expand_device_map(device_map, param_names, start_prefix):
"""
Expand a device map to return the correspondance parameter name to device.
Expand a device map to return the correspondence parameter name to device.
"""
new_device_map = {}
param_names = [p[len(start_prefix) :] for p in param_names if p.startswith(start_prefix)]

6
src/transformers/processing_utils.py
View File

@ -901,7 +901,7 @@ class ProcessorMixin(PushToHubMixin):
```python
tokenizer = tokenizer_class(..., {"padding": "max_length"})
image_processor = image_processor_class(...)
processor(tokenizer, image_processor) # will pass max_length unless overriden by kwargs at call
processor(tokenizer, image_processor) # will pass max_length unless overridden by kwargs at call
```
4) defaults kwargs specified at processor level have lowest priority.
```python
@ -1205,7 +1205,7 @@ class ProcessorMixin(PushToHubMixin):
video models might want to specify in the prompt the duration of video or which frame indices at which timestamps
were sampled. This information cannot be accessed before the video is loaded.
For most models it is a no-op, and must be overriden by model processors which require special processing.
For most models it is a no-op, and must be overridden by model processors which require special processing.
Args:
conversation (`List[Dict, str, str]`):
@ -1372,7 +1372,7 @@ class ProcessorMixin(PushToHubMixin):
if tokenize:
# Tokenizer's `apply_chat_template` never adds special tokens when tokenizing
# But processor's `apply_chat_template` didn't have an option to tokenize, so users had to format the prompt
# and pass it to the processor. Users thus never worried about special tokens relying on processor hadnling
# and pass it to the processor. Users thus never worried about special tokens relying on processor handling
# everything internally. The below line is to keep BC for that and be able to work with model that have
# special tokens in the template (consistent with tokenizers). We dont want to raise warning, it will flood command line
# without actionable solution for users

6
src/transformers/testing_utils.py
View File

@ -2407,7 +2407,7 @@ class SubprocessCallException(Exception):
def run_command(command: List[str], return_stdout=False):
"""
Runs `command` with `subprocess.check_output` and will potentially return the `stdout`. Will also properly capture
if an error occured while running `command`
if an error occurred while running `command`
"""
try:
output = subprocess.check_output(command, stderr=subprocess.STDOUT)
@ -2541,7 +2541,7 @@ def hub_retry(max_attempts: int = 5, wait_before_retry: Optional[float] = 2):
requests.exceptions.RequestException,
) as err:
logger.error(
f"Test failed with {err} at try {retry_count}/{max_attempts} as it couldn't connect to the specied Hub repository."
f"Test failed with {err} at try {retry_count}/{max_attempts} as it couldn't connect to the specified Hub repository."
)
if wait_before_retry is not None:
time.sleep(wait_before_retry)
@ -2661,7 +2661,7 @@ def run_test_using_subprocess(func):
The following contains utils to run the documentation tests without having to overwrite any files.
The `preprocess_string` function adds `# doctest: +IGNORE_RESULT` markers on the fly anywhere a `load_dataset` call is
made as a print would otherwise fail the corresonding line.
made as a print would otherwise fail the corresponding line.
To skip cuda tests, make sure to call `SKIP_CUDA_DOCTEST=1 pytest --doctest-modules <path_to_files_to_test>
"""

2
src/transformers/tokenization_utils_fast.py
View File

@ -708,7 +708,7 @@ class PreTrainedTokenizerFast(PreTrainedTokenizerBase):
added_tokens_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + ADDED_TOKENS_FILE
)
# make sure to be foward compatible
# make sure to be forward compatible
added_vocab = {tok: index for tok, index in self.added_tokens_encoder.items() if index >= self.vocab_size}
if added_vocab:
with open(added_tokens_file, "w", encoding="utf-8") as f:

6
src/transformers/trainer.py
View File

@ -2266,7 +2266,7 @@ class Trainer:
(self.model_wrapped,) = release_memory(self.model_wrapped)
self.model_wrapped = self.model
# Check for DeepSpeed *after* the intial pass and modify the config
# Check for DeepSpeed *after* the initial pass and modify the config
if self.is_deepspeed_enabled:
# Temporarily unset `self.args.train_batch_size`
original_bs = self.args.per_device_train_batch_size
@ -2826,7 +2826,7 @@ class Trainer:
# Checkpoint must have been saved with the old smp api.
if hasattr(self.args, "fp16") and self.args.fp16 is True:
logger.warning(
"Enabling FP16 and loading from smp < 1.10 checkpoint together is not suppported."
"Enabling FP16 and loading from smp < 1.10 checkpoint together is not supported."
)
state_dict = torch.load(
weights_file,
@ -4091,7 +4091,7 @@ class Trainer:
A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The
dictionary also contains the epoch number which comes from the training state.
"""
# handle multipe eval datasets
# handle multiple eval datasets
override = eval_dataset is not None
eval_dataset = eval_dataset if override else self.eval_dataset
if isinstance(eval_dataset, dict):

2
src/transformers/trainer_callback.py
View File

@ -88,7 +88,7 @@ class TrainerState:
impact the way data will be logged in TensorBoard.
stateful_callbacks (`List[StatefulTrainerCallback]`, *optional*):
Callbacks attached to the `Trainer` that should have their states be saved or restored.
Relevent callbacks should implement a `state` and `from_state` function.
Relevant callbacks should implement a `state` and `from_state` function.
"""
epoch: Optional[float] = None

8
src/transformers/trainer_pt_utils.py
View File

@ -1231,8 +1231,8 @@ class AcceleratorConfig:
all workers.
use_seedable_sampler (`bool`, *optional*, defaults to `True`):
Whether or not use a fully seedable random sampler ([`accelerate.data_loader.SeedableRandomSampler`]). Ensures
training results are fully reproducable using a different sampling technique. While seed-to-seed results
may differ, on average the differences are neglible when using multiple different seeds to compare. Should
training results are fully reproducible using a different sampling technique. While seed-to-seed results
may differ, on average the differences are negligible when using multiple different seeds to compare. Should
also be ran with [`~utils.set_seed`] for the best results.
gradient_accumulation_kwargs (`dict`, *optional*):
Additional kwargs to configure gradient accumulation, see [`accelerate.utils.GradientAccumulationPlugin`].
@ -1284,8 +1284,8 @@ class AcceleratorConfig:
default=True,
metadata={
"help": "Whether or not use a fully seedable random sampler ([`accelerate.data_loader.SeedableRandomSampler`])."
"Ensures training results are fully reproducable using a different sampling technique. "
"While seed-to-seed results may differ, on average the differences are neglible when using"
"Ensures training results are fully reproducible using a different sampling technique. "
"While seed-to-seed results may differ, on average the differences are negligible when using"
"multiple different seeds to compare. Should also be ran with [`~utils.set_seed`] for the best results."
},
)

14
src/transformers/training_args.py
View File

@ -542,7 +542,7 @@ class TrainingArguments:
all-gathers.
- use_orig_params (`bool`, *optional*, defaults to `True`)
If `"True"`, allows non-uniform `requires_grad` during init, which means support for interspersed
frozen and trainable paramteres. Useful in cases such as parameter-efficient fine-tuning. Please
frozen and trainable parameters. Useful in cases such as parameter-efficient fine-tuning. Please
refer this
[blog](https://dev-discuss.pytorch.org/t/rethinking-pytorch-fully-sharded-data-parallel-fsdp-from-first-principles/1019
- sync_module_states (`bool`, *optional*, defaults to `True`)
@ -604,8 +604,8 @@ class TrainingArguments:
all workers.
- use_seedable_sampler (`bool`, *optional*, defaults to `True`):
Whether or not use a fully seedable random sampler ([`accelerate.data_loader.SeedableRandomSampler`]). Ensures
training results are fully reproducable using a different sampling technique. While seed-to-seed results
may differ, on average the differences are neglible when using multiple different seeds to compare. Should
training results are fully reproducible using a different sampling technique. While seed-to-seed results
may differ, on average the differences are negligible when using multiple different seeds to compare. Should
also be ran with [`~utils.set_seed`] for the best results.
- use_configured_state (`bool`, *optional*, defaults to `False`):
Whether or not to use a pre-configured `AcceleratorState` or `PartialState` defined before calling `TrainingArguments`.
@ -1278,7 +1278,7 @@ class TrainingArguments:
default=None,
metadata={
"help": (
"Config to be used with the internal Accelerator object initializtion. The value is either a "
"Config to be used with the internal Accelerator object initialization. The value is either a "
"accelerator json config file (e.g., `accelerator_config.json`) or an already loaded json file as `dict`."
)
},
@ -1528,7 +1528,7 @@ class TrainingArguments:
neftune_noise_alpha: Optional[float] = field(
default=None,
metadata={
"help": "Activates neftune noise embeddings into the model. NEFTune has been proven to drastically improve model performances for instrcution fine-tuning. Check out the original paper here: https://arxiv.org/abs/2310.05914 and the original code here: https://github.com/neelsjain/NEFTune. Only supported for `PreTrainedModel` and `PeftModel` classes."
"help": "Activates neftune noise embeddings into the model. NEFTune has been proven to drastically improve model performances for instruction fine-tuning. Check out the original paper here: https://arxiv.org/abs/2310.05914 and the original code here: https://github.com/neelsjain/NEFTune. Only supported for `PreTrainedModel` and `PeftModel` classes."
},
)
@ -1584,7 +1584,7 @@ class TrainingArguments:
# Parse in args that could be `dict` sent in from the CLI as a string
for field in _VALID_DICT_FIELDS:
passed_value = getattr(self, field)
# We only want to do this if the str starts with a bracket to indiciate a `dict`
# We only want to do this if the str starts with a bracket to indicate a `dict`
# else its likely a filename if supported
if isinstance(passed_value, str) and passed_value.startswith("{"):
loaded_dict = json.loads(passed_value)
@ -1849,7 +1849,7 @@ class TrainingArguments:
torch.backends.cudnn.allow_tf32 = True
else:
logger.warning(
"The speedups for torchdynamo mostly come wih GPU Ampere or higher and which is not detected here."
"The speedups for torchdynamo mostly come with GPU Ampere or higher and which is not detected here."
)
if self.framework == "pt" and is_torch_available() and self.tf32 is not None:
if self.tf32: