p12475398
/
LLaMA-Factory-Mirror
forked from p04798526/LLaMA-Factory-Mirror
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix zero2 high ram usage

This commit is contained in:
hiyouga 2024-05-19 21:53:54 +08:00
parent 70214b71b1
commit 31a0564d4f
2 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/llamafactory/model/adapter.py
View File

@ -3,7 +3,7 @@ from typing import TYPE_CHECKING
import torch
from peft import LoraConfig, LoraModel, PeftModel, TaskType, get_peft_model
from transformers.integrations import deepspeed_config, is_deepspeed_zero3_enabled
from transformers.integrations import is_deepspeed_zero3_enabled
from transformers.modeling_utils import is_fsdp_enabled
from ..extras.logging import get_logger
@ -43,8 +43,8 @@ def init_adapter(
if finetuning_args.finetuning_type != "lora" and getattr(model, "quantization_method", None):
raise ValueError("You can only use lora for quantized models.")
if deepspeed_config() is not None or is_fsdp_enabled() or finetuning_args.pure_bf16 or finetuning_args.use_badam:
logger.info("DeepSpeed/FSDP/PureBF16/BAdam detected, remaining trainable params as their original precision.")
if is_deepspeed_zero3_enabled() or is_fsdp_enabled() or finetuning_args.pure_bf16 or finetuning_args.use_badam:
logger.info("ZeRO3/FSDP/PureBF16/BAdam detected, remaining trainable params as their original precision.")
cast_trainable_params_to_fp32 = False
else:
logger.info("Upcasting trainable params to float32.")

4
src/llamafactory/model/patcher.py
View File

@ -5,7 +5,7 @@ from typing import TYPE_CHECKING, Any, Dict
import torch
from peft import PeftModel
from transformers import PreTrainedModel, PreTrainedTokenizerBase, is_torch_npu_available
from transformers.integrations import deepspeed_config, is_deepspeed_zero3_enabled
from transformers.integrations import is_deepspeed_zero3_enabled
from transformers.modeling_utils import is_fsdp_enabled
from ..extras.logging import get_logger
@ -72,7 +72,7 @@ def patch_config(
# deepspeed zero3 is not compatible with low_cpu_mem_usage
init_kwargs["low_cpu_mem_usage"] = model_args.low_cpu_mem_usage and (not is_deepspeed_zero3_enabled())
if deepspeed_config() is None and not is_fsdp_enabled(): # set dtype and device map if not use deepspeed or fsdp
if not is_deepspeed_zero3_enabled() and not is_fsdp_enabled(): # cast dtype and device if not use zero3 or fsdp
init_kwargs["torch_dtype"] = model_args.compute_dtype
if init_kwargs["low_cpu_mem_usage"]: # device map requires low_cpu_mem_usage=True