Fix bugs

docs/source_en/Instruction/Supported-models-datasets.md

@@ -579,7 +579,7 @@ The table below provides relevant information about the datasets accessed throug
 - Tags: Tags of the dataset.
 
 
-| Dataset ID | Subset name | Dataset Size | Statistic (token) | Tags | HF Dataset ID |
+| Dataset ID | Subset Name | Dataset Size | Statistic (token) | Tags | HF Dataset ID |
 | ---------- | ----------- | -------------| ------------------| ---- | ------------- |
 |[AI-ModelScope/COIG-CQIA](https://modelscope.cn/datasets/AI-ModelScope/COIG-CQIA)|chinese_traditional<br>coig_pc<br>exam<br>finance<br>douban<br>human_value<br>logi_qa<br>ruozhiba<br>segmentfault<br>wiki<br>wikihow<br>xhs<br>zhihu|44694|331.2±693.8, min=34, max=19288|general, 🔥|-|
 |[AI-ModelScope/CodeAlpaca-20k](https://modelscope.cn/datasets/AI-ModelScope/CodeAlpaca-20k)|default|20022|99.3±57.6, min=30, max=857|code, en|[HuggingFaceH4/CodeAlpaca_20K](https://huggingface.co/datasets/HuggingFaceH4/CodeAlpaca_20K)|

scripts/utils/run_dataset_info.py

@@ -21,7 +21,7 @@ def get_cache_mapping(fpath):
             continue
         items = text.split('|')
         key = items[1] if items[1] != '-' else items[6]
-        key = re.search(r'\[(.+)\]', key).group(1)
+        key = re.search(r'\[(.+?)\]', key).group(1)
         stat = items[3:5]
         if stat[0] == '-':
             stat = ('huge dataset', '-')
@@ -78,7 +78,7 @@ def write_dataset_info() -> None:
     fpaths = ['docs/source/Instruction/支持的模型和数据集.md', 'docs/source_en/Instruction/Supported-models-datasets.md']
     cache_mapping = get_cache_mapping(fpaths[0])
     res_text_list = []
-    res_text_list.append('| Dataset ID | Subset name | Dataset Size | Statistic (token) | Tags | HF Dataset ID |')
+    res_text_list.append('| Dataset ID | Subset Name | Dataset Size | Statistic (token) | Tags | HF Dataset ID |')
     res_text_list.append('| ---------- | ----------- | -------------| ------------------| ---- | ------------- |')
 
     all_keys = list(DATASET_MAPPING.keys())

swift/cli/web_ui.py

@@ -1,5 +1,5 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
-from swift.ui.app import webui_main
+from swift.ui import webui_main
 
 if __name__ == '__main__':
     webui_main()

swift/hub/hub.py

@@ -157,7 +157,6 @@ def upload_folder(
         token: Union[str, bool, None] = None,
         revision: Optional[str] = 'master',
         ignore_patterns: Optional[Union[List[str], str]] = None,
-        run_as_future: bool = False,
         **kwargs,
     ):
         cls.push_to_hub(repo_id, folder_path, path_in_repo, commit_message, commit_description, token, True, revision,
@@ -271,6 +270,7 @@ def push_to_hub(cls,
             repo_id,
             folder_path,
             token or cls.ms_token,
+            private,
             commit_message=commit_message,
             ignore_file_pattern=ignore_patterns,
             revision=revision,

swift/llm/__init__.py

@@ -15,13 +15,11 @@
         deploy_main,
         LoRARequest,
         InferClient,
-        SwiftInfer,
-        SwiftDeploy,
         run_deploy,
         prepare_pt_engine_template,
     )
     from .export import (export_main, merge_lora, quantize_model, export_to_ollama)
-    from .eval import eval_main, SwiftEval
+    from .eval import eval_main
     from .train import sft_main, pt_main, rlhf_main
     from .argument import (EvalArguments, InferArguments, TrainArguments, ExportArguments, DeployArguments,
                            RLHFArguments, WebUIArguments, BaseArguments)
@@ -43,10 +41,10 @@
         'rlhf': ['rlhf_main'],
         'infer': [
             'deploy_main', 'VllmEngine', 'RequestConfig', 'LmdeployEngine', 'PtEngine', 'infer_main', 'LoRARequest',
-            'InferClient', 'SwiftInfer', 'SwiftDeploy', 'run_deploy', 'prepare_pt_engine_template', 'InferEngine'
+            'InferClient', 'run_deploy', 'prepare_pt_engine_template', 'InferEngine'
         ],
         'export': ['export_main', 'merge_lora', 'quantize_model', 'export_to_ollama'],
-        'eval': ['eval_main', 'SwiftEval'],
+        'eval': ['eval_main'],
         'train': ['sft_main', 'pt_main', 'rlhf_main'],
         'argument': [
             'EvalArguments', 'InferArguments', 'TrainArguments', 'ExportArguments', 'WebUIArguments', 'DeployArguments',

swift/llm/argument/base_args/base_args.py

@@ -28,15 +28,15 @@ def get_supported_tuners():
 class BaseArguments(GenerationArguments, QuantizeArguments, DataArguments, TemplateArguments, ModelArguments):
     """
     BaseArguments class is a dataclass that inherits from multiple argument classes:
-    ModelArguments, TemplateArguments, and DataArguments.
+    GenerationArguments, QuantizeArguments, DataArguments, TemplateArguments, ModelArguments.
 
     Args:
         tuner_backend(str): Support peft or unsloth.
         train_type(str): The training type, support all supported tuners and `full`.
         seed (int): Random seed for reproducibility. Default is 42.
         model_kwargs (Optional[str]): Additional keyword arguments for the model. Default is None.
-        use_hf (bool): Flag to determine if Hugging Face should be used. Default is False.
         load_dataset_config (bool): Flag to determine if dataset configuration should be loaded. Default is False.
+        use_hf (bool): Flag to determine if Hugging Face should be used. Default is False.
         hub_token (Optional[str]): SDK token for authentication. Default is None.
         ignore_args_error (bool): Flag to ignore argument errors for notebook compatibility. Default is False.
         use_swift_lora (bool): Use swift lora, a compatible argument
@@ -143,7 +143,6 @@ def _init_weight_type(self, ckpt_dir):
             self.model = ckpt_dir or self.model
 
     def _init_device(self):
-        """Prepare ddp of course"""
         if is_dist():
             if is_torch_npu_available():
                 torch.npu.set_device(self.local_rank)

swift/llm/argument/base_args/data_args.py

@@ -17,17 +17,17 @@ class DataArguments:
     DataArguments class is a dataclass that holds various arguments related to dataset handling and processing.
 
     Args:
-        dataset (List[str]): List of dataset identifiers or paths.
-        val_dataset (List[str]): List of validation dataset identifiers or paths.
+        dataset (List[str]): List of dataset_id, dataset_path or dataset_dir
+        val_dataset (List[str]): List of validation dataset_id, dataset_path or dataset_dir
         split_dataset_ratio (float): Ratio to split the dataset for validation if val_dataset is empty. Default is 0.01.
         data_seed (Optional[int]): Seed for dataset shuffling. Default is None.
         dataset_num_proc (int): Number of processes to use for data loading and preprocessing. Default is 1.
+        streaming (bool): Flag to enable streaming of datasets. Default is False.
         load_from_cache_file (bool): Flag to load dataset from cache file. Default is False.
         download_mode (Literal): Mode for downloading datasets. Default is 'reuse_dataset_if_exists'.
         model_name (List[str]): List containing Chinese and English names of the model. Default is [None, None].
         model_author (List[str]): List containing Chinese and English names of the model author.
             Default is [None, None].
-        streaming (bool): Flag to enable streaming of datasets. Default is False.
         custom_register_path (Optional[str]): Path to custom .py file for dataset registration. Default is None.
         custom_dataset_info (Optional[str]): Path to custom dataset_info.json file. Default is None.
     """

swift/llm/argument/base_args/generation_args.py

@@ -19,6 +19,7 @@ class GenerationArguments:
         top_p (Optional[float]): Top-p (nucleus) sampling parameter. Default is None.
         repetition_penalty (Optional[float]): Penalty for repeated tokens. Default is None.
         num_beams (int): Number of beams for beam search. Default is 1.
+        stream (bool): Flag to indicate if streaming output should be enabled. Default is None.
         stop_words (List[str]): List of stop words to end generation. Default is an empty list.
     """
 

swift/llm/argument/base_args/model_args.py

@@ -2,7 +2,7 @@
 import math
 import os
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Literal, Optional, Union
+from typing import Any, Dict, Literal, Optional, Union
 
 import json
 import torch
@@ -19,14 +19,14 @@ class ModelArguments:
     ModelArguments class is a dataclass that holds various arguments related to model configuration and usage.
 
     Args:
-        model (Optional[str]): Model identifier or path. Default is None.
+        model (Optional[str]): model_id or model_path. Default is None.
         model_type (Optional[str]): Type of the model group. Default is None.
         model_revision (Optional[str]): Revision of the model. Default is None.
-        torch_dtype (Literal): Model data type. Default is None.
+        torch_dtype (Literal): Model parameter dtype. Default is None.
         attn_impl (Literal): Attention implementation to use. Default is None.
         rope_scaling (Literal): Type of rope scaling to use. Default is None.
         device_map (Optional[str]): Configuration for device mapping. Default is None.
-        local_repo_path (Optional[str]): Path to the local repository for model code. Default is None.
+        local_repo_path (Optional[str]): Path to the local github repository for model. Default is None.
     """
     model: Optional[str] = None  # model id or model path
     model_type: Optional[str] = field(

swift/llm/argument/base_args/quant_args.py

@@ -14,8 +14,8 @@ class QuantizeArguments:
     QuantizeArguments is a dataclass that holds the configuration for model quantization.
 
     Args:
-        quant_method (Literal['bnb', 'hqq', 'eetq', 'awq', 'gptq']): The quantization method to be used.
-        quant_bits (Literal[0, 1, 2, 3, 4, 8]): The number of bits to use for quantization.
+        quant_method (Literal['bnb', 'hqq', 'eetq']): The quantization method to be used.
+        quant_bits (Literal[1, 2, 3, 4, 8]): The number of bits to use for quantization.
         hqq_axis (Optional[int]): The axis for hqq quantization.
         bnb_4bit_compute_dtype (Literal['float16', 'bfloat16', 'float32', None]):
             The compute dtype for bnb 4-bit quantization.
@@ -49,6 +49,8 @@ def get_quantization_config(self):
                 load_in_4bit, load_in_8bit = True, False
             elif self.quant_bits == 8:
                 load_in_4bit, load_in_8bit = False, True
+            else:
+                raise ValueError(f'bnb not support quant_bits: {self.quant_bits}')
 
             from transformers import BitsAndBytesConfig
             quantization_config = BitsAndBytesConfig(

swift/llm/argument/base_args/template_args.py

@@ -1,6 +1,6 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 from dataclasses import dataclass, field
-from typing import List, Literal, Optional
+from typing import Literal, Optional
 
 from swift.llm import TEMPLATE_MAPPING
 from swift.utils import get_logger
@@ -14,14 +14,14 @@ class TemplateArguments:
     TemplateArguments class is a dataclass that holds various arguments related to template configuration and usage.
 
     Args:
-        template (Optional[str]): Template identifier. Default is None, meaning use the template of the model_type.
+        template (Optional[str]): Template type. Default is None, meaning use the template of the model_type.
         system (Optional[str]): Override the default system in the template. Default is None.
         max_length (Optional[int]): Maximum length for the template. Default is None.
         truncation_strategy (Literal): Strategy for truncating the template. Default is 'left'.
-        tools_prompt (str): Override the default tools prompt in the template. Default is 'react_en'.
         max_pixels (Optional[int]): Maximum number of pixels for the template. Default is None.
+        tools_prompt (str): Override the default tools prompt in the template. Default is 'react_en'.
         loss_scale (str): Loss scale for training. Default is 'default',
-            meaning only calculate the loss of the response.
+            meaning only calculate the loss of the assistant.
         sequence_parallel_size (int): Size of sequence parallelism. Default is 1.
         use_chat_template (str): Use chat template or default generation template, default True
         template_backend (str): Use swift template or jinja

swift/llm/argument/deploy_args.py

@@ -23,8 +23,8 @@ class DeployArguments(InferArguments):
         owned_by (str): The owner of the deployment. Default is 'swift'.
         served_model_name (Optional[str]): The name of the model being served. Default is None.
         verbose (bool): Whether to log request information. Default is True.
-        log_interval (int): The interval for printing global statistics. Default is 10.
-        max_logprobs(int): Max number of tokens' log probs to return
+        log_interval (int): The interval for printing global statistics. Default is 20.
+        max_logprobs(int): Max number of logprobs to return
     """
     host: str = '0.0.0.0'
     port: int = 8000

swift/llm/argument/eval_args.py

@@ -2,9 +2,7 @@
 import datetime as dt
 import os
 from dataclasses import dataclass, field
-from typing import List, Literal, Optional
-
-import torch
+from typing import List, Optional
 
 from swift.utils import get_logger
 from .base_args import to_abspath
@@ -16,7 +14,7 @@
 @dataclass
 class EvalArguments(DeployArguments):
     """
-    EvalArguments is a dataclass that extends InferArguments and is used to define
+    EvalArguments is a dataclass that extends DeployArguments and is used to define
     the arguments required for evaluating a model.
 
     Args:

swift/llm/argument/export_args.py

@@ -3,9 +3,7 @@
 from dataclasses import dataclass, field
 from typing import Literal, Optional
 
-import torch.distributed as dist
-
-from swift.utils import get_logger, is_dist
+from swift.utils import get_logger
 from .base_args import BaseArguments, to_abspath
 from .merge_args import MergeArguments
 
@@ -19,8 +17,6 @@ class ExportArguments(MergeArguments, BaseArguments):
 
     Args:
         output_dir (Optional[str]): Directory where the output will be saved.
-        to_peft_format (bool): Flag to indicate if the output should be in PEFT format.
-            This argument is useless for now.
         quant_n_samples (int): Number of samples for quantization.
         max_length (int): Sequence length for quantization.
         quant_batch_size (int): Batch size for quantization.
@@ -30,11 +26,12 @@ class ExportArguments(MergeArguments, BaseArguments):
         hub_model_id (Optional[str]): Model ID for the hub.
         hub_private_repo (bool): Flag to indicate if the hub repository is private.
         commit_message (str): Commit message for pushing to the hub.
+        to_peft_format (bool): Flag to indicate if the output should be in PEFT format.
+            This argument is useless for now.
     """
     ckpt_dir: Optional[str] = field(default=None, metadata={'help': '/path/to/your/vx-xxx/checkpoint-xxx'})
     output_dir: Optional[str] = None
 
-    to_peft_format: bool = False
     # awq/gptq
     quant_method: Literal['awq', 'gptq'] = None
     quant_n_samples: int = 256
@@ -52,6 +49,8 @@ class ExportArguments(MergeArguments, BaseArguments):
     hub_model_id: Optional[str] = None
     hub_private_repo: bool = False
     commit_message: str = 'update files'
+    # compat
+    to_peft_format: bool = False
 
     def _init_quant(self):
 

swift/llm/argument/infer_args.py

@@ -21,7 +21,7 @@ class LmdeployArguments:
     LmdeployArguments is a dataclass that holds the configuration for lmdeploy.
 
     Args:
-        tp (int): Tensor parallelism degree. Default is 1.
+        tp (int): Tensor parallelism size. Default is 1.
         session_len(Optional[int]): The session length, default None.
         cache_max_entry_count (float): Maximum entry count for cache. Default is 0.8.
         quant_policy (int): Quantization policy, e.g., 4, 8. Default is 0.
@@ -56,7 +56,7 @@ class VllmArguments:
         pipeline_parallel_size(int): Pipeline parallelism size. Default is 1.
         max_num_seqs (int): Maximum number of sequences. Default is 256.
         max_model_len (Optional[int]): Maximum model length. Default is None.
-        disable_custom_all_reduce (bool): Flag to disable custom all-reduce. Default is True.
+        disable_custom_all_reduce (bool): Flag to disable custom all-reduce. Default is False.
         enforce_eager (bool): Flag to enforce eager execution. Default is False.
         limit_mm_per_prompt (Optional[str]): Limit multimedia per prompt. Default is None.
         vllm_max_lora_rank (int): Maximum LoRA rank. Default is 16.
@@ -101,26 +101,25 @@ class InferArguments(MergeArguments, VllmArguments, LmdeployArguments, BaseArgum
     It is used to define the arguments required for model inference.
 
     Args:
+        ckpt_dir (Optional[str]): Directory to the checkpoint. Default is None.
         infer_backend (Literal): Backend to use for inference. Default is 'pt'.
             Allowed values are 'vllm', 'pt', 'lmdeploy'.
-        ckpt_dir (Optional[str]): Directory to the checkpoint. Default is None.
+        result_path (Optional[str]): Directory to store inference results. Default is None.
         max_batch_size (int): Maximum batch size for the pt engine. Default is 1.
         val_dataset_sample (Optional[int]): Sample size for validation dataset. Default is None.
-        result_dir (Optional[str]): Directory to store inference results. Default is None.
-        stream (Optional[bool]): Flag to indicate if streaming should be enabled. Default is None.
     """
     ckpt_dir: Optional[str] = field(default=None, metadata={'help': '/path/to/your/vx-xxx/checkpoint-xxx'})
     infer_backend: Literal['vllm', 'pt', 'lmdeploy'] = 'pt'
 
     result_path: Optional[str] = None
-    writer_buffer_size: int = 65536  # B
+    writer_buffer_size: int = 65536
     # for pt engine
     max_batch_size: int = 1
 
     # only for inference
     val_dataset_sample: Optional[int] = None
 
-    def get_result_path(self, folder_name, suffix: str = '.jsonl') -> str:
+    def get_result_path(self, folder_name: str, suffix: str = '.jsonl') -> str:
         result_dir = self.ckpt_dir or self.model_dir
         result_dir = to_abspath(os.path.join(result_dir, folder_name))
         os.makedirs(result_dir, exist_ok=True)
@@ -181,5 +180,5 @@ def _parse_lora_modules(self) -> None:
         lora_request_list = []
         for i, lora_module in enumerate(self.lora_modules):
             lora_name, lora_path = lora_module.split('=')
-            lora_request_list.append(LoRARequest(lora_name, i + 1, lora_path))
+            lora_request_list.append(LoRARequest(lora_name, lora_path))
         self.lora_request_list = lora_request_list

swift/llm/argument/rlhf_args.py

@@ -19,7 +19,6 @@ class RLHFArguments(TrainArguments):
         ref_model_revision (Optional[str]): Revision of the reference model. Default is None.
         beta (Optional[float]): Beta parameter for RLHF. Default is None.
         label_smoothing (float): Label smoothing value. Default is 0.
-        loss_type (Optional[str]): Type of loss function. Default is None.
         rpo_alpha (float): Alpha parameter for RPO. Default is 1.
         cpo_alpha (float): Alpha parameter for CPO. Default is 1.
         simpo_gamma (float): Gamma parameter for SimPO. Default is 1.