Kimirock/PaddleNLP
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix typos (#10483)

Browse Source
This commit is contained in:
co63oc
2025-04-24 12:02:41 +08:00
committed by GitHub
parent 3a23130fa0
commit e8a19d361a
43 changed files with 89 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
CONTRIBUTING.md
View File

@@ -72,7 +72,7 @@ PaddleNLP 使用 [Git 分支模型](http://nvie.com/posts/a-successful-git-branc
类似于以下内容：
```text
➜ (my-virtual-env) git commit -m "commiting my cool feature"
➜ (my-virtual-env) git commit -m "committing my cool feature"
black....................................................................Passed
isort....................................................................Passed
flake8...................................................................Passed
@@ -91,7 +91,7 @@ PaddleNLP 使用 [Git 分支模型](http://nvie.com/posts/a-successful-git-branc
但大多数时候事情并没有那么顺利。当您的代码或文档不符合标准时，`pre-commit` 检查将失败。
```text
➜ (my-virtual-env) git commit -m "commiting my cool feature"
➜ (my-virtual-env) git commit -m "committing my cool feature"
black....................................................................Passed
isort....................................................................Failed
- hook id: isort

2
csrc/gpu/moe/fused_moe/cutlass_kernels/moe_gemm/fused_moe_gemm_kernels_template.h
View File

@@ -725,7 +725,7 @@ void MoeGemmRunner<T, WeightType>::run_gemm<EpilogueTag>(
gemmConfigManager.addBestConfig(gemmId, profile_total_rows, best_config);
chosen_config = best_config;
} else {
PADDLE_FATAL("[MoE Configure Search] find no one avaliable config.");
PADDLE_FATAL("[MoE Configure Search] find no one available config.");
}
}
dispatch_to_arch<EpilogueTag>(A,

2
paddlenlp/__init__.py
View File

@@ -35,7 +35,7 @@ if "datasets" in sys.modules.keys():
logger.warning(
"Detected that datasets module was imported before paddlenlp. "
"This may cause PaddleNLP datasets to be unavalible in intranet. "
"This may cause PaddleNLP datasets to be unavailable in intranet. "
"Please import paddlenlp before datasets module to avoid download issues"
)
import paddle

2
paddlenlp/dataaug/char.py
View File

@@ -436,7 +436,7 @@ class CharInsert(BaseAugment):
return sentences
def _generate_sequence(self, output_seq_tokens, aug_tokens, p):
"""Genearte the sequences according to the mapping list"""
"""Generate the sequences according to the mapping list"""
for aug_token in aug_tokens:
idx, token = aug_token
if p == 0:

14
paddlenlp/dataaug/sentence.py
View File

@@ -33,9 +33,9 @@ __all__ = [
class SentenceGenerate:
"""
SentenceGenerate is a sentence-level data augmentation strategy
that generates simialr sentences according to the input sequence.
The strattegy first generates several sentences, and then chooses
the top n simialr sentences by the model.
that generates similar sentences according to the input sequence.
The strategy first generates several sentences, and then chooses
the top n similar sentences by the model.
Args:
model_name (str):
@@ -82,7 +82,7 @@ class SentenceGenerate:
@paddle.no_grad()
def _generate_similar_sentence(self, sequence, model, tokenizer):
"""Generates generate_n similar sentences from the provided sequence, and chooose the best create_n similar sentences."""
"""Generates generate_n similar sentences from the provided sequence, and choose the best create_n similar sentences."""
# Generate generate_n similar sentences
generated_sequences = [sequence]
@@ -134,7 +134,7 @@ class SentenceSummarize:
token probabilities in the "sampling" strategy. Default to 1.0,
which means no effect.
use_fp16_decoding: (bool): Whether to use fp16 for decoding.
Only works when faster entry is avalible. Default to False.
Only works when faster entry is available. Default to False.
kwargs (dict): Additional keyword arguments refer to ..taskflow.text_summarization.TextSummarization
"""
@@ -337,8 +337,8 @@ class SentenceBackTranslate:
class SentenceBackTranslateAPI:
"""
SentenceBackTranslateAPI is a sentence-level data augmentation strategy
that translates the input sequence into one langugage, and backtranslate
back into the sourche language by baidu translate api.
that translates the input sequence into one language, and back-translate
back into the source language by baidu translate api.
Args:
src_lang (str):

4
paddlenlp/dataaug/word.py
View File

@@ -170,7 +170,7 @@ class WordSubstitute(BaseAugment):
return substitute_dict
def _generate_sequence(self, output_seq_tokens, aug_tokens):
"""Genearte the sequences according to the mapping list"""
"""Generate the sequences according to the mapping list"""
for aug_token in aug_tokens:
idx, token = aug_token
output_seq_tokens[int(idx)] = token
@@ -503,7 +503,7 @@ class WordInsert(BaseAugment):
return sentences
def _generate_sequence(self, output_seq_tokens, aug_tokens, p):
"""Genearte the sequences according to the mapping list"""
"""Generate the sequences according to the mapping list"""
for aug_token in aug_tokens:
idx, token = aug_token
if p == 0:

4
paddlenlp/experimental/autonlp/README_en.md
View File

@@ -50,8 +50,8 @@ Args:
- text_column (string, required): Name of the column that contains the input text.
- label_column (string, required): Name of the column that contains the target variable to predict.
- language (string, required): language of the text
- metric_for_best_model (string, optional): the name of the metrc for selecting the best model.
- greater_is_better (bool, optional): Whether better models should have a greater metric or not. Use in conjuction with `metric_for_best_model`.
- metric_for_best_model (string, optional): the name of the metric for selecting the best model.
- greater_is_better (bool, optional): Whether better models should have a greater metric or not. Use in conjunction with `metric_for_best_model`.
- problem_type (str, optional): Select among ["multi_class", "multi_label"] based on the nature of your problem
- output_dir (str, optional): Output directory for the experiments, defaults to "autpnlp_results"
- verbosity: (int, optional): controls the verbosity of the run. Defaults to 1, which let the workers log to the driver.To reduce the amount of logs, use verbosity > 0 to set stop the workers from logging to the driver.

2
paddlenlp/experimental/autonlp/auto_trainer_base.py
View File

@@ -45,7 +45,7 @@ class AutoTrainerBase(metaclass=ABCMeta):
eval_dataset (Dataset, required): Evaluation dataset, must contains the 'text_column' and 'label_column' specified below
language (string, required): language of the text
metric_for_best_model (string, optional): the name of the metric for selecting the best model.
greater_is_better (bool, required): Whether better models should have a greater metric or not. Use in conjuction with `metric_for_best_model`.
greater_is_better (bool, required): Whether better models should have a greater metric or not. Use in conjunction with `metric_for_best_model`.
output_dir (str, optional): Output directory for the experiments, defaults to "autpnlp_results"
verbosity: (int, optional): controls the verbosity of the run. Defaults to 1, which let the workers log to the driver.To reduce the amount of logs,
use verbosity > 0 to set stop the workers from logging to the driver.

6
paddlenlp/experimental/autonlp/text_classification.py
View File

@@ -59,8 +59,8 @@ class AutoTrainerForTextClassification(AutoTrainerBase):
eval_dataset (Dataset, required): Evaluation dataset, must contains the 'text_column' and 'label_column' specified below
text_column (string, required): Name of the column that contains the input text.
label_column (string, required): Name of the column that contains the target variable to predict.
metric_for_best_model (string, optional): the name of the metrc for selecting the best model. Default to 'eval_accuracy'.
greater_is_better (bool, optional): Whether better models should have a greater metric or not. Use in conjuction with `metric_for_best_model`.
metric_for_best_model (string, optional): the name of the metric for selecting the best model. Default to 'eval_accuracy'.
greater_is_better (bool, optional): Whether better models should have a greater metric or not. Use in conjunction with `metric_for_best_model`.
problem_type (str, optional): Select among ["multi_class", "multi_label"] based on the nature of your problem
kwargs (dict, optional): Additional keyword arguments passed along to the specific task.
language (string, required): language of the text.
@@ -639,7 +639,7 @@ class AutoTrainerForTextClassification(AutoTrainerBase):
f"Taskflow config saved to {export_path}. You can use the Taskflow config to create a Taskflow instance for inference"
)
logger.info(f"Exported trial_id: {trial_id} to export_path: {export_path} sucessfully!")
logger.info(f"Exported trial_id: {trial_id} to export_path: {export_path} successfully!")
if os.path.exists(self.training_path):
logger.info("Removing training checkpoints to conserve disk space")

6
paddlenlp/experimental/faster_tokenizer.py
View File

@@ -29,10 +29,10 @@ __all__ = ["to_tensor", "to_vocab_buffer", "FasterTokenizer"]
def to_tensor(string_values, name="text"):
"""
Create the tensor that the value holds the list of string.
NOTICE: The value will be holded in the cpu place.
NOTICE: The value will be held in the cpu place.
Args:
string_values(list[string]): The value will be setted to the tensor.
string_values(list[string]): The value will be set to the tensor.
name(string): The name of the tensor.
"""
tensor = paddle.Tensor(core.VarDesc.VarType.STRING, [], name, core.VarDesc.VarType.STRINGS, False)
@@ -43,7 +43,7 @@ def to_tensor(string_values, name="text"):
def to_vocab_buffer(vocab_dict, name):
"""
Create the tensor that the value holds the map, the type of key is the string.
NOTICE: The value will be holded in the cpu place.
NOTICE: The value will be held in the cpu place.
Args:
vocab_dict(dict): The value will be setted to the tensor.

2
paddlenlp/experimental/transformers/bloom/modeling.py
View File

@@ -392,7 +392,7 @@ class BloomModelInferenceModel(BloomPreTrainedModel):
elif k.endswith("mlp.dense_4h_to_h.bias"):
self.transformer_block.ffn2_biases[idx].set_value(paddle.to_tensor(v))
else:
raise ValueError("Unknow weight {}".format(k))
raise ValueError("Unknown weight {}".format(k))
class BloomLMHead(nn.Layer):

2
paddlenlp/experimental/transformers/chatglm/modeling.py
View File

@@ -481,7 +481,7 @@ class ChatGLMStackDyBatch(nn.Layer):
elif k.endswith("mlp.dense_4h_to_h.bias"):
self.transformer_block.ffn2_biases[idx].set_value(v.astype(dtype))
else:
print("Unknow weight {}".format(k))
print("Unknown weight {}".format(k))
@register_base_model

2
paddlenlp/experimental/transformers/deepseek_v2/modeling.py
View File

@@ -145,7 +145,7 @@ class DeepseekScalingRotaryEmbedding(nn.Layer):
class DeepseekScalingRotaryEmbeddingXPU(nn.Layer):
"""
RotaryEmbedding XPU Implemention. In XPU, cos and sin must be computed in cpu.
RotaryEmbedding XPU Implementation. In XPU, cos and sin must be computed in cpu.
"""
def __init__(

2
paddlenlp/experimental/transformers/generation_utils.py
View File

@@ -102,7 +102,7 @@ class GenerationInferenceModel(GenerationMixin):
config.get("logits_processors", None),
precache_input_spec,
]
# use "==" to distingusih between chatglm and chatglm_v2.
# use "==" to distinguish between chatglm and chatglm_v2.
if self.config["model_type"] and "chatglm" == self.config.model_type.lower():
input_spec[2] = paddle.static.InputSpec(
shape=[None, None, None], dtype="int64", name="position_ids"

2
paddlenlp/experimental/transformers/gpt/modeling.py
View File

@@ -432,7 +432,7 @@ class GPTInferenceModel(GPTPretrainedModel):
elif k.endswith("linear2.bias"):
self.transformer_block.ffn2_biases[idx].set_value(v.astype(dtype))
else:
raise ValueError("Unknow weight {}".format(k))
raise ValueError("Unknown weight {}".format(k))
class GPTForCausalLMInferenceModel(GenerationInferenceModel, GPTPretrainedModel):

2
paddlenlp/generation/configuration_utils.py
View File

@@ -121,7 +121,7 @@ class GenerationConfig:
use_fast: (bool, optional): Whether to use fast entry of model
for FastGeneration. Default to False.
use_fp16_decoding: (bool, optional): Whether to use fp16 for decoding.
Only works when fast entry is avalible. Default to False.
Only works when fast entry is available. Default to False.
trunc_input: (bool, optional): Whether to truncate the inputs from
output sequences . Default to True.
model_kwargs (dict): It can be used to specify additional kwargs

2
paddlenlp/generation/streamers.py
View File

@@ -149,7 +149,7 @@ class TextStreamer(BaseStreamer):
class TextIteratorStreamer(TextStreamer):
"""
Streamer that stores print-ready text in a queue, to be used by a downstream application as an iterator. This is
useful for applications that benefit from acessing the generated text in a non-blocking way (e.g. in an interactive
useful for applications that benefit from accessing the generated text in a non-blocking way (e.g. in an interactive
Gradio demo).
Parameters:

2
paddlenlp/generation/utils.py
View File

@@ -1223,7 +1223,7 @@ class GenerationMixin(object):
next_tokens = paddle.multinomial(probs)
if self.config.tensor_parallel_degree > 1:
# Maybe no need to broadcast if seed is set correclty.
# Maybe no need to broadcast if seed is set correctly.
from paddle.distributed import fleet
try:

2
paddlenlp/metrics/dureader.py
View File

@@ -127,7 +127,7 @@ def compute_predictions(
nbest.append(_NbestPrediction(text=final_text, start_logit=pred.start_logit, end_logit=pred.end_logit))
# if we didn't inlude the empty option in the n-best, inlcude it
# if we didn't include the empty option in the n-best, include it
# In very rare edge cases we could have no valid predictions. So we
# just create a nonce prediction in this case to avoid failure.

2
paddlenlp/ops/optimizer/adamwdl.py
View File

@@ -188,7 +188,7 @@ class AdamWDL(AdamW):
ratio = self.set_param_lr_fun(param_and_grad[0])
param_and_grad[0].optimize_attr["learning_rate"] *= ratio
# excute Adam op
# execute Adam op
res = super(AdamWDL, self)._append_optimize_op(block, param_and_grad)
param_and_grad[0].optimize_attr["learning_rate"] = prev_lr
return res

2
paddlenlp/ops/triton_ops/fused_moe.py
View File

@@ -934,7 +934,7 @@ def get_default_config(
"num_stages": 4,
}
else:
# Block-wise quant: BLOCK_SIZE_K must be divisable by block_shape[1]
# Block-wise quant: BLOCK_SIZE_K must be divisible by block_shape[1]
config = {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": block_shape[0],

2
paddlenlp/ops/triton_ops/triton_utils.py
View File

@@ -98,7 +98,7 @@ def extract_triton_kernel(kernel, file_name):
elif type(kernel) == triton.runtime.autotuner.Autotuner:
fn = kernel.fn.fn
else:
AssertionError("error occures")
AssertionError("error occurs")
py_script = textwrap.dedent(inspect.getsource(fn))
# @triton.jit must only appear once

2
paddlenlp/peft/lokr/lokr_layers.py
View File

@@ -196,7 +196,7 @@ def factorization(dimension: int, factor: int = -1) -> Tuple[int, int]:
In LoRA with Kroneckor Product, first value is a value for weight scale.
second value is a value for weight.
Becuase of non-commutative property, A⊗B ≠ B⊗A. Meaning of two matrices is slightly different.
Because of non-commutative property, A⊗B ≠ B⊗A. Meaning of two matrices is slightly different.
examples)
factor

6
paddlenlp/peft/lora/auto_lora_model.py
View File

@@ -318,7 +318,7 @@ class LoRAAutoModel(nn.Layer):
if self.is_pipelinemodel and merge_tensor_parallel and self.lora_config.tensor_parallel_degree > 1:
merge_tensor_parallel = False
logger.warning(
"Pipeline parallism does not support merge_tensor_parallel. Set merge_tensor_parallel to False."
"Pipeline parallelism does not support merge_tensor_parallel. Set merge_tensor_parallel to False."
)
variant = kwargs.get("variant", None)
@@ -580,7 +580,7 @@ class LoRAAutoModel(nn.Layer):
for k, v in config["mp_config"]["parallelize_plan"].items():
assert (
k not in final_config["mp_config"]["parallelize_plan"].keys()
), f"sublayer mp_config shuld be a subset of model but got sublayer config {config['mp_config']} and model config {final_config['mp_config']}."
), f"sublayer mp_config should be a subset of model but got sublayer config {config['mp_config']} and model config {final_config['mp_config']}."
final_config["mp_config"]["parallelize_plan"][k] = v
if "sp_config" in config and config["sp_config"] is not None:
if final_config["sp_config"] is None:
@@ -589,7 +589,7 @@ class LoRAAutoModel(nn.Layer):
for k, v in config["sp_config"]["parallelize_plan"].items():
assert (
k not in final_config["sp_config"]["parallelize_plan"].keys()
), f"sublayer sp_config shuld be a subset of model but got sublayer config {config['sp_config']} and model config {final_config['sp_config']}."
), f"sublayer sp_config should be a subset of model but got sublayer config {config['sp_config']} and model config {final_config['sp_config']}."
final_config["sp_config"]["parallelize_plan"][k] = v
if "pp_config" in config and config["pp_config"] is not None:

2
paddlenlp/peft/lora/lora_model.py
View File

@@ -420,7 +420,7 @@ class LoRAModel(nn.Layer):
if self.is_pipelinemodel and merge_tensor_parallel and self.lora_config.tensor_parallel_degree > 1:
merge_tensor_parallel = False
logger.warning(
"Pipeline parallism does not support merge_tensor_parallel. Set merge_tensor_parallel to False."
"Pipeline parallelism does not support merge_tensor_parallel. Set merge_tensor_parallel to False."
)
variant = kwargs.get("variant", None)

2
paddlenlp/peft/vera/vera_model.py
View File

@@ -111,7 +111,7 @@ class VeRAModel(nn.Layer):
if self.is_pipelinemodel and merge_tensor_parallel and self.vera_config.tensor_parallel_degree > 1:
merge_tensor_parallel = False
logger.warning(
"Pipeline parallism does not support merge_tensor_parallel. Set merge_tensor_parallel to False."
"Pipeline parallelism does not support merge_tensor_parallel. Set merge_tensor_parallel to False."
)
variant = kwargs.get("variant", None)

2
paddlenlp/prompt/template.py
View File

@@ -503,7 +503,7 @@ class SoftTemplate(Template):
continuous token id sequence for each part in template.
Returns:
`List[Dict[str, str]]`: Template with continuous prompt formated as {"soft": "xxx"}.
`List[Dict[str, str]]`: Template with continuous prompt formatted as {"soft": "xxx"}.
`Tuple[Dict[int, int], List[List[int]], int]`:
- Mapping from continuous ids to word ids for initialization.
- Continuous ids for each part. Id 0 denotes none-continuous part.

2
paddlenlp/rl/trainer/ppo_trainer.py
View File

@@ -1208,7 +1208,7 @@ class PPOTrainer(Trainer):
if dp_degree * sharding_degree == 1:
return micro_batches
# otherwise, need to balance batch accross DP and Sharding groups
# otherwise, need to balance batch across DP and Sharding groups
try:
hcg = fleet.get_hybrid_communicate_group()
sharding_parallel_group = hcg.get_sharding_parallel_group()

14
paddlenlp/rl/trainer/rl_trainer.py
View File

@@ -625,12 +625,12 @@ class RLTrainer(Trainer):
eval_tp_rank = max(tensor_parallel_rank, 0)
group_nums = self.args.logical_process_index // old_dp_workers * eval_tp_size + eval_tp_rank
self._data_trans_group = create_data_trans_group(global_rank, group_nums)
# just for compatiable with old code
# just for compatible with old code
self._policy_model_eval_group = self._data_trans_group
def get_model(self, train=False):
"""
model visitor wrapps PipelineParalle and Inference model to do evaulation
model visitor wraps PipelineParallel and Inference model to do evaluation
and generation.
"""
if train:
@@ -785,10 +785,10 @@ class RLTrainer(Trainer):
NOTE: This is transparent to users.
When using a mixed loss we often want to get the separated loss metrics,
thus we mark loss type of each training step to separate them. This is
not necessary since the loss would be returnd after each training step.
However when using PipelienParallel, the loss returned is 0 when not reach
not necessary since the loss would be returned after each training step.
However when using PipelineParallel, the loss returned is 0 when not reach
accumulated step and the loss returned at accumulated step is a mixed loss.
To separate loss metrics in PipelienParallel:
To separate loss metrics in PipelineParallel:
1. We hack PipelineParallel._forward_step to record actual loss for each
step in a list (only in training and not in evaluation currently).
2. We mark the loss type only once for each step using `loss_step_indice`
@@ -805,7 +805,7 @@ class RLTrainer(Trainer):
"""
Return a dict mapping loss name to value of current training step. This
is mainly to get loss for metric logging, and it would not affect the
training. This is mostly helpful to PipelienParallel with a mixed loss
training. This is mostly helpful to PipelineParallel with a mixed loss
in which the loss returned is 0 when not reach accumulated step and the
loss returned at accumulated step is a mixed loss.
NOTE: 1. Only when reaching accumulated step the losses returned are
@@ -830,7 +830,7 @@ class RLTrainer(Trainer):
mix_loss = paddle.stack(model._step_losses)
model._step_losses = None
else:
# The tessor shape is not actor_model.accumulate_steps
# The tensor shape is not actor_model.accumulate_steps
# (args.accu_steps) but actor_trainer.args.accu_steps,
# since actor_model is created with global pp_config
# using global args.accu_steps which is only half of

2
paddlenlp/rl/utils/comm_utils.py
View File

@@ -139,7 +139,7 @@ def karmarkar_karp(seqlen_list: List[int], k_partitions: int, equal_size: bool):
def get_seqlen_balanced_partitions(seqlen_list: List[int], k_partitions: int, equal_size: bool):
"""get order of seq lengths to make partitions balanced, this is
used in balacing sum of seqlength across dp ranks and microbatches
used in balancing sum of seqlength across dp ranks and microbatches
Parameters:
seqlen_list (List[int]):
seq lengths of each items

2
paddlenlp/transformers/albert/tokenizer.py
View File

@@ -353,7 +353,7 @@ class AlbertTokenizer(PretrainedTokenizer):
Args:
ids (list): A list of ids to be converted.
skip_special_tokens (bool, optional):
Whether or not to skip specical tokens. Defaults to `False`.
Whether or not to skip special tokens. Defaults to `False`.
Returns:
list: A list of converted tokens.

2
paddlenlp/transformers/configuration_utils.py
View File

@@ -1236,7 +1236,7 @@ def get_configuration_file(configuration_files: List[str]) -> str:
configuration_file = CONFIG_NAME
# FIXME: (wj-Mcat) remove the hard dependency of `packaging` which can compare
# the version of package, also be uesed in `transfromer`.
# the version of package, also be used in `transformer`.
# **But**, we don't support version compare function now. so remove the hard dependency.
from packaging import version

18
paddlenlp/transformers/conversion_utils.py
View File

@@ -439,7 +439,7 @@ def normal_fuse_split_tp(weight, tensor_parallel_degree, tensor_parallel_rank=No
"""
There're three types of MultiHeadAttention QKV Layout in Transfomers
There're three types of MultiHeadAttention QKV Layout in Transformers
tensor_parallel_qkv = [q1, k1, v1, q2, k2, v2]
naive_merged_qkv = [q1, q1, k1, k2, v1, v2]
@@ -692,7 +692,7 @@ class StateDictNameMapping:
return self.action == "transpose"
def should_merge_last_two_dim(self) -> bool:
"""check that wether merge last two dim"""
"""check that whether merge last two dim"""
return self.action == "merge_last_two_dim"
def run(self, state_dict: dict[str, ndarray], name: str) -> ndarray:
@@ -764,7 +764,7 @@ class TensorInfoSaver:
"""output the summary info into different terminal
Args:
output_path (Optional[str], optional): the dir/file of sumamry file. Defaults to None.
output_path (Optional[str], optional): the dir/file of summary file. Defaults to None.
"""
if output_path and os.path.isdir(output_path):
output_path = os.path.join(output_path, "tensor_summary.xlsx")
@@ -800,10 +800,10 @@ class TensorInfoSaver:
class LogitHooker:
"""hooks for pytorch model and paddle model, used to generate the logits of elment layers"""
"""hooks for pytorch model and paddle model, used to generate the logits of element layers"""
def __init__(self, mappings: List[StateDictNameMapping], tensor_info_saver: Optional[TensorInfoSaver] = None):
"""registe the logit hooks to compare the inputs * outputs model
"""register the logit hooks to compare the inputs * outputs model
Args:
mappings (List[StateDictNameMapping]): the mappings between paddle & pytorch model
@@ -1006,7 +1006,7 @@ class LogitComparer:
pytorch_model: Union[Module, Dict[str, ndarray]],
name_mappings: List[StateDictNameMapping],
):
"""compare the pytorch and paddle mdoel state with name mappings
"""compare the pytorch and paddle model state with name mappings
Args:
paddle_model (Union[Layer, Dict[str, ndarray]]): paddle model instance
@@ -1045,7 +1045,7 @@ class LogitComparer:
"""compare the logit of pytorch & paddle model
Returns:
bool: if the logits is absolutly same
bool: if the logits is absolutely same
"""
PaddleModel, PytorchModel = self.get_paddle_pytorch_model_classes()
paddle_model = PaddleModel.from_pretrained(self.input_dir)
@@ -1120,7 +1120,7 @@ class LogitComparer:
if is_torch_available() and is_transformers_available():
result = self.compare_logits()
if result is True:
logger.info("the logits between pytorch model and paddle model is absolutly same")
logger.info("the logits between pytorch model and paddle model is absolutely same")
else:
logger.error(
"the logits between pytorch model and paddle model is not same, please check it out more carefully."
@@ -1237,7 +1237,7 @@ class ConversionMixin:
if state_dict is None:
with device_guard("cpu"):
state_dict = paddle.load(weight_file, return_numpy=False)
logger.info("Starting to convert orignal state_dict to tensor parallel state_dict.")
logger.info("Starting to convert original state_dict to tensor parallel state_dict.")
state_keys_map = cls._resolve_prefix_keys(name_action_mappings.keys(), state_dict.keys(), ignore_error)

2
paddlenlp/transformers/convert_slow_tokenizer.py
View File

@@ -274,7 +274,7 @@ def bytes_to_unicode():
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""

2
paddlenlp/transformers/dallebart/modeling.py
View File

@@ -922,7 +922,7 @@ class DalleBartForConditionalGeneration(DalleBartPretrainedModel):
use_fast: (bool, optional): Whether to use fast entry of model
for FastGeneration. Default to False.
use_fp16_decoding: (bool, optional): Whether to use fp16 for decoding.
Only works when fast entry is avalible. Default to False.
Only works when fast entry is available. Default to False.
condition_scale (float, optional): The scale of super conditioning. See
`this twitter <https://twitter.com/RiversHaveWings/status/1478093658716966912>`__
Default to 1.0.

4
paddlenlp/transformers/image_transforms.py
View File

@@ -452,7 +452,7 @@ 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
@@ -494,7 +494,7 @@ def corners_to_center_format(bboxes_corners: TensorType) -> TensorType:
"""
Converts bounding boxes from corners format to center format.
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)
center format: contains the coordinate for the center of the box and its the width, height dimensions
(center_x, center_y, width, height)

2
paddlenlp/transformers/mamba/tokenizer.py
View File

@@ -33,7 +33,7 @@ def bytes_to_unicode():
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""

4
paddlenlp/transformers/model_outputs.py
View File

@@ -278,11 +278,11 @@ def _transformer_encoder_fwd(
# MultiHeadAttention not so efficiently, and maybe optimize it later.
if cache is None and getattr(self, "_use_cache", False):
cache = [tuple(self.layers[0].gen_cache(src))] * len(self.layers)
# To be compatible with `TransformerEncoder.forward`, `_use_cache` defualts
# To be compatible with `TransformerEncoder.forward`, `_use_cache` defaults
# to True when cache is not None.
new_caches = [] if cache is not None and getattr(self, "_use_cache", True) else None
all_attentions = [] if output_attentions else None
# NOTE: Also includes embeding output which is same as HF.
# NOTE: Also includes embedding output which is same as HF.
all_hidden_states = [output] if output_hidden_states else None
for i, mod in enumerate(self.layers):
# if output has no gradient, recompute is unnecessary

18
paddlenlp/transformers/model_utils.py
View File

@@ -591,19 +591,19 @@ def _partion_for_pipeline_mode(keys):
keys = list(keys)
start_idx = -1
prefix_str = None
parttion_map = {}
partition_map = {}
for k in keys:
prefix = layer_prefix(k)
if prefix != prefix_str:
prefix_str = prefix
start_idx += 1
parttion_map[k] = start_idx
partition_map[k] = start_idx
# if only one parttion, we don't parttion it
# if only one partition, we don't partition it
if start_idx < 1:
return {keys[i]: i for i in range(len(keys))}
return parttion_map
return partition_map
def shard_checkpoint(
@@ -808,7 +808,7 @@ def load_sharded_checkpoint(model, folder, variant=None, strict=True, prefer_saf
def faster_set_state_dict(model, state_dict, strict_dtype=True):
# the state_dict will be destroied.
# the state_dict will be destroyed.
unused_keys = set(state_dict.keys())
unset_keys = set(model.state_dict().keys())
with paddle.no_grad():
@@ -891,7 +891,7 @@ def _convert_state_dict_dtype_and_shape(state_dict, model_to_load):
if key in list(state_dict.keys()):
if isinstance(state_dict[key], np.ndarray):
raise ValueError(
"convert_state_dict_dtype expected paddle.Tensor not numpy.ndarray, plase convert numpy.ndarray to paddle.Tensor"
"convert_state_dict_dtype expected paddle.Tensor not numpy.ndarray, please convert numpy.ndarray to paddle.Tensor"
)
# confirm parameter cast is executed on the same device as model
# TODO: cast(FP32 -> FP16) has diff on different devices, need to fix it
@@ -1879,7 +1879,7 @@ class PretrainedModel(Layer, GenerationMixin, ConversionMixin):
elif "pytorch_model.bin" in str(resolved_archive_file):
if not from_hf_hub and not convert_from_torch:
raise ValueError(
f"Download pytorch wight in "
f"Download pytorch weight in "
f" {resolved_archive_file}. Please set convert_from_torch=True in from_pretrained. eg, Model.from_pretrained(model_name, convert_from_torch=True) "
)
@@ -2832,7 +2832,7 @@ class PretrainedModel(Layer, GenerationMixin, ConversionMixin):
for k, v in config["mp_config"]["parallelize_plan"].items():
assert (
k not in final_config["mp_config"]["parallelize_plan"].keys()
), f"sublayer mp_config shuld be a subset of model but got sublayer config {config['mp_config']} and model config {final_config['mp_config']}."
), f"sublayer mp_config should be a subset of model but got sublayer config {config['mp_config']} and model config {final_config['mp_config']}."
final_config["mp_config"]["parallelize_plan"][k] = v
if "sp_config" in config and config["sp_config"] is not None:
if final_config["sp_config"] is None:
@@ -2841,7 +2841,7 @@ class PretrainedModel(Layer, GenerationMixin, ConversionMixin):
for k, v in config["sp_config"]["parallelize_plan"].items():
assert (
k not in final_config["sp_config"]["parallelize_plan"].keys()
), f"sublayer sp_config shuld be a subset of model but got sublayer config {config['sp_config']} and model config {final_config['sp_config']}."
), f"sublayer sp_config should be a subset of model but got sublayer config {config['sp_config']} and model config {final_config['sp_config']}."
final_config["sp_config"]["parallelize_plan"][k] = v
if "pp_config" in config and config["pp_config"] is not None:
if isinstance(config["pp_config"]["split_spec"], str):

2
paddlenlp/transformers/optimization.py
View File

@@ -173,7 +173,7 @@ class CosineDecayWithWarmup(LambdaDecay):
Creates a learning rate scheduler, which increases learning rate linearly
from 0 to given `learning_rate`, after this warmup period learning rate
would be decreased following the values of the cosine function. If
`with_hard_restarts` is True, the cosine function could have serveral hard
`with_hard_restarts` is True, the cosine function could have several hard
restarts.
Args:

6
paddlenlp/transformers/tokenizer_utils.py
View File

@@ -200,7 +200,7 @@ def is_chinese_char(cp):
def _is_nonnormalized_char(char):
"""Check whther `chars` is a non-normalized character."""
"""Check whether `chars` is a non-normalized character."""
cp = ord(char)
if (
(0xFF00 <= cp <= 0xFFEF)
@@ -807,7 +807,7 @@ class ChatTemplateMixin:
def _extract_non_learnable_parts(self, origin_msg: List[Dict[str, str]], split_s: List[str]):
"""Split the entire chat by specified words. Extract the non-learnable parts."""
# distingish and replace the special words in original string to an uncompiled form: Like | -> \|
# distinguish and replace the special words in original string to an uncompiled form: Like | -> \|
regex_pattern = "|".join(map(re.escape, split_s))
# splited by replaced specified words
non_learnable_parts = re.split(
@@ -1585,7 +1585,7 @@ class PretrainedTokenizer(ChatTemplateMixin, PretrainedTokenizerBase):
sequence = ids + pair_ids if pair else ids
token_type_ids = [0] * len(ids) + ([0] * len(pair_ids) if pair else [])
encoded_inputs["offset_mapping"] = offset_mapping
# Build output dictionnary
# Build output dictionary
encoded_inputs["input_ids"] = sequence
if return_token_type_ids:
encoded_inputs["token_type_ids"] = token_type_ids

4
paddlenlp/transformers/tokenizer_utils_base.py
View File

@@ -1561,7 +1561,7 @@ class PretrainedTokenizerBase(SpecialTokensMixin):
# From HF Hub or AI Studio
if from_hf_hub or from_aistudio:
# Only include the necessary resource files specified by the tokenizer cls
# Deep copy to avoid modifiying the class attributes
# Deep copy to avoid modifying the class attributes
vocab_files = copy.deepcopy(cls.resource_files_names)
vocab_files["tokenizer_config_file"] = cls.tokenizer_config_file
@@ -3119,7 +3119,7 @@ class PretrainedTokenizerBase(SpecialTokensMixin):
sequence = ids + pair_ids if pair else ids
token_type_ids = [0] * len(ids) + ([0] * len(pair_ids) if pair else [])
# Build output dictionnary
# Build output dictionary
encoded_inputs["input_ids"] = sequence
if return_token_type_ids:
encoded_inputs["token_type_ids"] = token_type_ids

10
paddlenlp/transformers/utils.py
View File

@@ -205,10 +205,10 @@ def adapt_stale_fwd_patch(self, name, value):
if self.__module__.startswith("paddlenlp"):
warnings.warn(
f"The `forward` method of {self.__class__ if isinstance(self, Layer) else self} is patched and the patch "
"might be based on an old oversion which missing some "
"might be based on an old version which missing some "
f"arguments compared with the latest, such as {new_args}. "
"We automatically add compatibility on the patch for "
"these arguemnts, and maybe the patch should be updated."
"these arguments, and maybe the patch should be updated."
)
else:
warnings.warn(
@@ -216,7 +216,7 @@ def adapt_stale_fwd_patch(self, name, value):
"is patched and the patch might be conflict with patches made "
f"by paddlenlp which seems have more arguments such as {new_args}. "
"We automatically add compatibility on the patch for "
"these arguemnts, and maybe the patch should be updated."
"these arguments, and maybe the patch should be updated."
)
if isinstance(self, Layer) and inspect.isfunction(value):
@@ -450,7 +450,7 @@ def paddlenlp_hub_download(
# Download from custom model url
if is_url(repo_id):
# check wether the target file exist in the comunity bos server
# check wether the target file exist in the community bos server
if url_file_exists(repo_id):
logger.info(f"Downloading {repo_id}")
weight_file_path = get_path_from_url_with_filelock(repo_id, cache_dir)
@@ -483,7 +483,7 @@ def paddlenlp_hub_download(
community_model_file_path = "/".join(url_list)
assert is_url(community_model_file_path)
# check wether the target file exist in the comunity bos server
# check wether the target file exist in the community bos server
if url_file_exists(community_model_file_path):
logger.info(f"Downloading {community_model_file_path}")
weight_file_path = get_path_from_url_with_filelock(community_model_file_path, cache_dir)