ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'main' into main

Browse Source 
Former-commit-id: 7be442f37d53a0c6324728fa1fa8e2c84d7f0fa5
This commit is contained in:
hoshi-hiyouga
2024-07-01 21:01:09 +08:00
committed by GitHub
parent 6c185a2c57 973cf8e980
commit a715490c2a
176 changed files with 4760 additions and 1322 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/llamafactory/hparams/__init__.py
View File

@@ -1,3 +1,17 @@
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .data_args import DataArguments
from .evaluation_args import EvaluationArguments
from .finetuning_args import FinetuningArguments

28
src/llamafactory/hparams/data_args.py
View File

@@ -1,3 +1,20 @@
# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
#
# This code is inspired by the HuggingFace's transformers library.
# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass, field
from typing import Literal, Optional
@@ -28,10 +45,6 @@ class DataArguments:
default=1024,
metadata={"help": "The cutoff length of the tokenized inputs in the dataset."},
)
reserved_label_len: int = field(
default=1,
metadata={"help": "The minimum cutoff length reserved for the tokenized labels in the dataset."},
)
train_on_prompt: bool = field(
default=False,
metadata={"help": "Whether to disable the mask on the prompt or not."},
@@ -90,15 +103,16 @@ class DataArguments:
"help": "Whether or not to pack the sequences without cross-contamination attention for efficient training."
},
)
tool_format: Optional[str] = field(
default=None,
metadata={"help": "Tool format to use for constructing function calling examples."},
)
tokenized_path: Optional[str] = field(
default=None,
metadata={"help": "Path to save or load the tokenized datasets."},
)
def __post_init__(self):
if self.reserved_label_len >= self.cutoff_len:
raise ValueError("`reserved_label_len` must be smaller than `cutoff_len`.")
if self.streaming and self.val_size > 1e-6 and self.val_size < 1:
raise ValueError("Streaming mode should have an integer val size.")

14
src/llamafactory/hparams/evaluation_args.py
View File

@@ -1,3 +1,17 @@
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from dataclasses import dataclass, field
from typing import Literal, Optional

49
src/llamafactory/hparams/finetuning_args.py
View File

@@ -1,5 +1,19 @@
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass, field
from typing import Literal, Optional
from typing import List, Literal, Optional
@dataclass
@@ -94,6 +108,18 @@ class LoraArguments:
default=False,
metadata={"help": "Whether or not to use the weight-decomposed lora method (DoRA)."},
)
pissa_init: bool = field(
default=False,
metadata={"help": "Whether or not to initialize a PiSSA adapter."},
)
pissa_iter: int = field(
default=16,
metadata={"help": "The number of iteration steps performed by FSVD in PiSSA. Use -1 to disable it."},
)
pissa_convert: bool = field(
default=False,
metadata={"help": "Whether or not to convert the PiSSA adapter to a normal LoRA adapter."},
)
create_new_adapter: bool = field(
default=False,
metadata={"help": "Whether or not to create a new adapter with randomly initialized weight."},
@@ -319,20 +345,19 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
return [item.strip() for item in arg.split(",")]
return arg
self.freeze_trainable_modules = split_arg(self.freeze_trainable_modules)
self.freeze_extra_modules = split_arg(self.freeze_extra_modules)
self.lora_alpha = self.lora_alpha or self.lora_rank * 2
self.lora_target = split_arg(self.lora_target)
self.additional_target = split_arg(self.additional_target)
self.galore_target = split_arg(self.galore_target)
self.freeze_trainable_modules: List[str] = split_arg(self.freeze_trainable_modules)
self.freeze_extra_modules: Optional[List[str]] = split_arg(self.freeze_extra_modules)
self.lora_alpha: int = self.lora_alpha or self.lora_rank * 2
self.lora_target: List[str] = split_arg(self.lora_target)
self.additional_target: Optional[List[str]] = split_arg(self.additional_target)
self.galore_target: List[str] = split_arg(self.galore_target)
self.freeze_vision_tower = self.freeze_vision_tower or self.train_mm_proj_only
self.use_ref_model = self.stage == "dpo" and self.pref_loss not in ["orpo", "simpo"]
assert self.finetuning_type in ["lora", "freeze", "full"], "Invalid fine-tuning method."
assert self.ref_model_quantization_bit in [None, 8, 4], "We only accept 4-bit or 8-bit quantization."
assert self.reward_model_quantization_bit in [None, 8, 4], "We only accept 4-bit or 8-bit quantization."
self.use_ref_model = self.pref_loss not in ["orpo", "simpo"]
if self.stage == "ppo" and self.reward_model is None:
raise ValueError("`reward_model` is necessary for PPO training.")
@@ -354,5 +379,11 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
if self.loraplus_lr_ratio is not None and self.finetuning_type != "lora":
raise ValueError("`loraplus_lr_ratio` is only valid for LoRA training.")
if self.pissa_init and self.finetuning_type != "lora":
raise ValueError("`pissa_init` is only valid for LoRA training.")
if self.pissa_init and (self.stage in ["ppo", "kto"] or self.use_ref_model):
raise ValueError("Cannot use PiSSA for current training stage.")
if self.train_mm_proj_only and self.finetuning_type != "full":
raise ValueError("`train_mm_proj_only` is only valid for full training.")

14
src/llamafactory/hparams/generating_args.py
View File

@@ -1,3 +1,17 @@
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import asdict, dataclass, field
from typing import Any, Dict, Optional

64
src/llamafactory/hparams/model_args.py
View File

@@ -1,5 +1,28 @@
# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
#
# This code is inspired by the HuggingFace's transformers library.
# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import asdict, dataclass, field
from typing import Any, Dict, Literal, Optional
from typing import TYPE_CHECKING, Any, Dict, Literal, Optional, Union
from typing_extensions import Self
if TYPE_CHECKING:
import torch
@dataclass
@@ -22,6 +45,10 @@ class ModelArguments:
)
},
)
adapter_folder: Optional[str] = field(
default=None,
metadata={"help": "The folder containing the adapter weights to load."},
)
cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Where to store the pre-trained models downloaded from huggingface.co or modelscope.cn."},
@@ -50,6 +77,10 @@ class ModelArguments:
default=True,
metadata={"help": "Whether or not to use memory-efficient model loading."},
)
quantization_method: Literal["bitsandbytes", "hqq", "eetq"] = field(
default="bitsandbytes",
metadata={"help": "Quantization method to use for on-the-fly quantization."},
)
quantization_bit: Optional[int] = field(
default=None,
metadata={"help": "The number of bits to quantize the model using bitsandbytes."},
@@ -70,7 +101,7 @@ class ModelArguments:
default=None,
metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."},
)
flash_attn: Literal["off", "sdpa", "fa2", "auto"] = field(
flash_attn: Literal["auto", "disabled", "sdpa", "fa2"] = field(
default="auto",
metadata={"help": "Enable FlashAttention for faster training and inference."},
)
@@ -127,13 +158,9 @@ class ModelArguments:
metadata={"help": "Whether or not to disable CUDA graph in the vLLM engine."},
)
vllm_max_lora_rank: int = field(
default=8,
default=32,
metadata={"help": "Maximum rank of all LoRAs in the vLLM engine."},
)
vllm_dtype: Literal["auto", "float16", "bfloat16", "float32"] = field(
default="auto",
metadata={"help": "Data type for model weights and activations in the vLLM engine."},
)
offload_folder: str = field(
default="offload",
metadata={"help": "Path to offload model weights."},
@@ -142,6 +169,10 @@ class ModelArguments:
default=True,
metadata={"help": "Whether or not to use KV cache in generation."},
)
infer_dtype: Literal["auto", "float16", "bfloat16", "float32"] = field(
default="auto",
metadata={"help": "Data type for model weights and activations at inference."},
)
hf_hub_token: Optional[str] = field(
default=None,
metadata={"help": "Auth token to log in with Hugging Face Hub."},
@@ -192,9 +223,9 @@ class ModelArguments:
)
def __post_init__(self):
self.compute_dtype = None
self.device_map = None
self.model_max_length = None
self.compute_dtype: Optional["torch.dtype"] = None
self.device_map: Optional[Union[str, Dict[str, Any]]] = None
self.model_max_length: Optional[int] = None
if self.split_special_tokens and self.use_fast_tokenizer:
raise ValueError("`split_special_tokens` is only supported for slow tokenizers.")
@@ -208,11 +239,18 @@ class ModelArguments:
if self.new_special_tokens is not None: # support multiple special tokens
self.new_special_tokens = [token.strip() for token in self.new_special_tokens.split(",")]
assert self.quantization_bit in [None, 8, 4], "We only accept 4-bit or 8-bit quantization."
assert self.export_quantization_bit in [None, 8, 4, 3, 2], "We only accept 2/3/4/8-bit quantization."
if self.export_quantization_bit is not None and self.export_quantization_dataset is None:
raise ValueError("Quantization dataset is necessary for exporting.")
def to_dict(self) -> Dict[str, Any]:
return asdict(self)
@classmethod
def copyfrom(cls, old_arg: Self, **kwargs) -> Self:
arg_dict = old_arg.to_dict()
arg_dict.update(**kwargs)
new_arg = cls(**arg_dict)
new_arg.compute_dtype = old_arg.compute_dtype
new_arg.device_map = old_arg.device_map
new_arg.model_max_length = old_arg.model_max_length
return new_arg

65
src/llamafactory/hparams/parser.py
View File

@@ -1,3 +1,20 @@
# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
#
# This code is inspired by the HuggingFace's transformers library.
# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import os
import sys
@@ -8,6 +25,7 @@ import transformers
from transformers import HfArgumentParser, Seq2SeqTrainingArguments
from transformers.integrations import is_deepspeed_zero3_enabled
from transformers.trainer_utils import get_last_checkpoint
from transformers.training_args import ParallelMode
from transformers.utils import is_torch_bf16_gpu_available
from transformers.utils.versions import require_version
@@ -65,13 +83,13 @@ def _verify_model_args(model_args: "ModelArguments", finetuning_args: "Finetunin
if model_args.adapter_name_or_path is not None and finetuning_args.finetuning_type != "lora":
raise ValueError("Adapter is only valid for the LoRA method.")
if model_args.use_unsloth and is_deepspeed_zero3_enabled():
raise ValueError("Unsloth is incompatible with DeepSpeed ZeRO-3.")
if model_args.quantization_bit is not None:
if finetuning_args.finetuning_type != "lora":
raise ValueError("Quantization is only compatible with the LoRA method.")
if finetuning_args.pissa_init:
raise ValueError("Please use scripts/pissa_init.py to initialize PiSSA for a quantized model.")
if model_args.resize_vocab:
raise ValueError("Cannot resize embedding layers of a quantized model.")
@@ -100,7 +118,7 @@ def _check_extra_dependencies(
require_version("galore_torch", "To fix: pip install galore_torch")
if finetuning_args.use_badam:
require_version("badam", "To fix: pip install badam")
require_version("badam>=1.2.1", "To fix: pip install badam>=1.2.1")
if finetuning_args.plot_loss:
require_version("matplotlib", "To fix: pip install matplotlib")
@@ -162,6 +180,12 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
):
raise ValueError("PPO only accepts wandb or tensorboard logger.")
if training_args.parallel_mode == ParallelMode.NOT_DISTRIBUTED:
raise ValueError("Please launch distributed training with `llamafactory-cli` or `torchrun`.")
if training_args.deepspeed and training_args.parallel_mode != ParallelMode.DISTRIBUTED:
raise ValueError("Please use `FORCE_TORCHRUN=1` to launch DeepSpeed training.")
if training_args.max_steps == -1 and data_args.streaming:
raise ValueError("Please specify `max_steps` in streaming mode.")
@@ -171,32 +195,31 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
if training_args.do_train and model_args.quantization_device_map == "auto":
raise ValueError("Cannot use device map for quantized models in training.")
if finetuning_args.use_dora and model_args.use_unsloth:
raise ValueError("Unsloth does not support DoRA.")
if finetuning_args.pissa_init and is_deepspeed_zero3_enabled():
raise ValueError("PiSSA is incompatible with DeepSpeed ZeRO-3.")
if finetuning_args.pure_bf16:
if not is_torch_bf16_gpu_available():
raise ValueError("This device does not support `pure_bf16`.")
if training_args.fp16 or training_args.bf16:
raise ValueError("Turn off mixed precision training when using `pure_bf16`.")
if is_deepspeed_zero3_enabled():
raise ValueError("`pure_bf16` is incompatible with DeepSpeed ZeRO-3.")
if (
finetuning_args.use_galore
and finetuning_args.galore_layerwise
and training_args.parallel_mode.value == "distributed"
and training_args.parallel_mode == ParallelMode.DISTRIBUTED
):
raise ValueError("Distributed training does not support layer-wise GaLore.")
if (
finetuning_args.use_badam
and finetuning_args.badam_mode == "layer"
and training_args.parallel_mode.value == "distributed"
):
raise ValueError("Layer-wise BAdam does not yet support distributed training, use ratio-wise BAdam.")
if finetuning_args.use_badam and training_args.parallel_mode == ParallelMode.DISTRIBUTED:
if finetuning_args.badam_mode == "ratio":
raise ValueError("Radio-based BAdam does not yet support distributed training, use layer-wise BAdam.")
elif not is_deepspeed_zero3_enabled():
raise ValueError("Layer-wise BAdam only supports DeepSpeed ZeRO-3 training.")
if (finetuning_args.use_galore or finetuning_args.use_badam) and training_args.deepspeed is not None:
raise ValueError("GaLore and BAdam are incompatible with DeepSpeed yet.")
if finetuning_args.use_galore and training_args.deepspeed is not None:
raise ValueError("GaLore is incompatible with DeepSpeed yet.")
if model_args.infer_backend == "vllm":
raise ValueError("vLLM backend is only available for API, CLI and Web.")
@@ -204,6 +227,9 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
if model_args.visual_inputs and data_args.packing:
raise ValueError("Cannot use packing in MLLM fine-tuning.")
if model_args.use_unsloth and is_deepspeed_zero3_enabled():
raise ValueError("Unsloth is incompatible with DeepSpeed ZeRO-3.")
_verify_model_args(model_args, finetuning_args)
_check_extra_dependencies(model_args, finetuning_args, training_args)
@@ -233,7 +259,7 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
# Post-process training arguments
if (
training_args.parallel_mode.value == "distributed"
training_args.parallel_mode == ParallelMode.DISTRIBUTED
and training_args.ddp_find_unused_parameters is None
and finetuning_args.finetuning_type == "lora"
):
@@ -293,7 +319,7 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
training_args.local_rank,
training_args.device,
training_args.n_gpu,
training_args.parallel_mode.value == "distributed",
training_args.parallel_mode == ParallelMode.DISTRIBUTED,
str(model_args.compute_dtype),
)
)
@@ -332,6 +358,7 @@ def get_infer_args(args: Optional[Dict[str, Any]] = None) -> _INFER_CLS:
if model_args.export_dir is not None and model_args.export_device == "cpu":
model_args.device_map = {"": torch.device("cpu")}
model_args.model_max_length = data_args.cutoff_len
else:
model_args.device_map = "auto"