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

merge data part to the text stream

Browse Source 
Former-commit-id: 7ee20286d9bcc2d5378bfd6bb02cd3648396d873
This commit is contained in:
BUAADreamer
2024-04-25 19:19:59 +08:00
parent 00e2a272ef
commit 3c792174db
13 changed files with 802 additions and 284 deletions
Download Patch File Download Diff File Expand all files Collapse all files

100
src/llmtuner/model/adapter.py
View File

@@ -11,7 +11,7 @@ from .utils.unsloth import get_unsloth_peft_model, load_unsloth_peft_model
if TYPE_CHECKING:
from transformers import PretrainedConfig, PreTrainedModel, AutoModelForVision2Seq
from transformers import PretrainedConfig, PreTrainedModel
from ..hparams import FinetuningArguments, ModelArguments
@@ -21,11 +21,11 @@ logger = get_logger(__name__)
def init_adapter(
config: "PretrainedConfig",
model: Union["PreTrainedModel","AutoModelForVision2Seq"],
model: Union["PreTrainedModel"],
model_args: "ModelArguments",
finetuning_args: "FinetuningArguments",
is_trainable: bool,
) -> Union["PreTrainedModel","AutoModelForVision2Seq"]:
) -> Union["PreTrainedModel"]:
r"""
Initializes the adapters.
@@ -38,7 +38,9 @@ def init_adapter(
logger.info("Adapter is not found at evaluation, load the base model.")
return model
if finetuning_args.finetuning_type != "lora" and getattr(model, "quantization_method", None):
if finetuning_args.finetuning_type != "lora" and getattr(
model, "quantization_method", None
):
raise ValueError("You can only use lora for quantized models.")
if finetuning_args.finetuning_type == "full" and is_trainable:
@@ -49,9 +51,9 @@ def init_adapter(
if finetuning_args.finetuning_type == "freeze" and is_trainable:
logger.info("Fine-tuning method: Freeze")
num_layers = (
getattr(model.config, "num_hidden_layers", None)
or getattr(model.config, "num_layers", None)
or getattr(model.config, "n_layer", None)
getattr(model.config, "num_hidden_layers", None)
or getattr(model.config, "num_layers", None)
or getattr(model.config, "n_layer", None)
)
if not num_layers:
raise ValueError("Current model does not support freeze tuning.")
@@ -66,8 +68,12 @@ def init_adapter(
stride = num_layers // finetuning_args.num_layer_trainable
trainable_layer_ids = range(stride - 1, num_layers + stride - 1, stride)
elif finetuning_args.num_layer_trainable > 0: # fine-tuning the last n layers if num_layer_trainable > 0
trainable_layer_ids = range(num_layers - finetuning_args.num_layer_trainable, num_layers)
elif (
finetuning_args.num_layer_trainable > 0
): # fine-tuning the last n layers if num_layer_trainable > 0
trainable_layer_ids = range(
num_layers - finetuning_args.num_layer_trainable, num_layers
)
else: # fine-tuning the first n layers if num_layer_trainable < 0
trainable_layer_ids = range(-finetuning_args.num_layer_trainable)
@@ -82,11 +88,15 @@ def init_adapter(
for module_name in finetuning_args.name_module_trainable:
if module_name not in freeze_modules:
raise ValueError(
"Module {} is not found, please choose from {}".format(module_name, ", ".join(freeze_modules))
"Module {} is not found, please choose from {}".format(
module_name, ", ".join(freeze_modules)
)
)
for idx in trainable_layer_ids:
trainable_layers.append(".{:d}.{}".format(idx, module_name if module_name != "all" else ""))
trainable_layers.append(
".{:d}.{}".format(idx, module_name if module_name != "all" else "")
)
for name, param in model.named_parameters():
if any(trainable_layer in name for trainable_layer in trainable_layers):
@@ -95,27 +105,43 @@ def init_adapter(
else:
param.requires_grad_(False)
logger.info("Set trainable layers: {}".format(",".join(map(str, trainable_layer_ids))))
logger.info(
"Set trainable layers: {}".format(",".join(map(str, trainable_layer_ids)))
)
if finetuning_args.finetuning_type == "lora":
logger.info("Fine-tuning method: {}".format("DoRA" if finetuning_args.use_dora else "LoRA"))
logger.info(
"Fine-tuning method: {}".format(
"DoRA" if finetuning_args.use_dora else "LoRA"
)
)
adapter_to_resume = None
if model_args.adapter_name_or_path is not None:
is_mergeable = True
if getattr(model, "quantization_method", None): # merge lora in quantized model is unstable
assert len(model_args.adapter_name_or_path) == 1, "Quantized model only accepts a single adapter."
if getattr(
model, "quantization_method", None
): # merge lora in quantized model is unstable
assert (
len(model_args.adapter_name_or_path) == 1
), "Quantized model only accepts a single adapter."
is_mergeable = False
if is_deepspeed_zero3_enabled():
assert len(model_args.adapter_name_or_path) == 1, "Cannot use multiple adapters in DeepSpeed ZeRO-3."
assert (
len(model_args.adapter_name_or_path) == 1
), "Cannot use multiple adapters in DeepSpeed ZeRO-3."
is_mergeable = False
if model_args.use_unsloth:
assert len(model_args.adapter_name_or_path) == 1, "Unsloth model only accepts a single adapter."
assert (
len(model_args.adapter_name_or_path) == 1
), "Unsloth model only accepts a single adapter."
is_mergeable = False
if (is_trainable and not finetuning_args.create_new_adapter) or (not is_mergeable):
if (is_trainable and not finetuning_args.create_new_adapter) or (
not is_mergeable
):
adapter_to_merge = model_args.adapter_name_or_path[:-1]
adapter_to_resume = model_args.adapter_name_or_path[-1]
else:
@@ -132,7 +158,9 @@ def init_adapter(
if adapter_to_resume is not None: # resume lora training
if model_args.use_unsloth:
model = load_unsloth_peft_model(config, model_args, is_trainable=is_trainable)
model = load_unsloth_peft_model(
config, model_args, is_trainable=is_trainable
)
else:
model = PeftModel.from_pretrained(
model,
@@ -141,19 +169,27 @@ def init_adapter(
offload_folder=model_args.offload_folder,
)
if is_trainable and adapter_to_resume is None: # create new lora weights while training
if len(finetuning_args.lora_target) == 1 and finetuning_args.lora_target[0] == "all":
if (
is_trainable and adapter_to_resume is None
): # create new lora weights while training
if (
len(finetuning_args.lora_target) == 1
and finetuning_args.lora_target[0] == "all"
):
target_modules = find_all_linear_modules(model)
else:
target_modules = finetuning_args.lora_target
if finetuning_args.use_llama_pro:
target_modules = find_expanded_modules(model, target_modules, finetuning_args.num_layer_trainable)
target_modules = find_expanded_modules(
model, target_modules, finetuning_args.num_layer_trainable
)
if (
finetuning_args.use_dora
and getattr(model, "quantization_method", None) is not None
and getattr(model, "quantization_method", None) != QuantizationMethod.BITS_AND_BYTES
finetuning_args.use_dora
and getattr(model, "quantization_method", None) is not None
and getattr(model, "quantization_method", None)
!= QuantizationMethod.BITS_AND_BYTES
):
raise ValueError("DoRA is not compatible with PTQ-quantized models.")
@@ -166,7 +202,11 @@ def init_adapter(
module_names.add(name.split(".")[-1])
finetuning_args.additional_target = module_names
logger.warning("Vocab has been resized, add {} to trainable params.".format(",".join(module_names)))
logger.warning(
"Vocab has been resized, add {} to trainable params.".format(
",".join(module_names)
)
)
peft_kwargs = {
"r": finetuning_args.lora_rank,
@@ -193,6 +233,10 @@ def init_adapter(
param.data = param.data.to(torch.float32)
if model_args.adapter_name_or_path is not None:
logger.info("Loaded adapter(s): {}".format(",".join(model_args.adapter_name_or_path)))
logger.info(
"Loaded adapter(s): {}".format(
",".join(model_args.adapter_name_or_path)
)
)
return model
return model

26
src/llmtuner/model/loader.py
View File

@@ -1,6 +1,12 @@
from typing import TYPE_CHECKING, Any, Dict, Union
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, AutoProcessor, AutoModelForVision2Seq
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
AutoProcessor,
AutoModelForVision2Seq,
)
from trl import AutoModelForCausalLMWithValueHead
from ..extras.logging import get_logger
@@ -62,10 +68,14 @@ def load_tokenizer(model_args: "ModelArguments") -> "PreTrainedTokenizer":
dict(additional_special_tokens=model_args.new_special_tokens),
replace_additional_special_tokens=False,
)
logger.info("Add {} to special tokens.".format(",".join(model_args.new_special_tokens)))
logger.info(
"Add {} to special tokens.".format(",".join(model_args.new_special_tokens))
)
if num_added_tokens > 0 and not model_args.resize_vocab:
model_args.resize_vocab = True
logger.warning("New tokens have been added, changed `resize_vocab` to True.")
logger.warning(
"New tokens have been added, changed `resize_vocab` to True."
)
patch_tokenizer(tokenizer)
return tokenizer
@@ -111,7 +121,7 @@ def load_model(
finetuning_args: "FinetuningArguments",
is_trainable: bool = False,
add_valuehead: bool = False,
) -> Union["PreTrainedModel", "AutoModelForVision2Seq"]:
) -> Union["PreTrainedModel"]:
r"""
Loads pretrained model.
"""
@@ -170,8 +180,10 @@ def load_model(
trainable_params, all_param = count_parameters(model)
if is_trainable:
param_stats = "trainable params: {:d} || all params: {:d} || trainable%: {:.4f}".format(
trainable_params, all_param, 100 * trainable_params / all_param
param_stats = (
"trainable params: {:d} || all params: {:d} || trainable%: {:.4f}".format(
trainable_params, all_param, 100 * trainable_params / all_param
)
)
else:
param_stats = "all params: {:d}".format(all_param)
@@ -185,4 +197,4 @@ def load_model(
)
)
return model
return model