423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-04-27 18:29:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

[v1] add deepspeed zero3 trigger for low memory usage weight loading (#10300)

Browse Source
This commit is contained in:
jiaqiw09
2026-04-21 14:09:52 +08:00
committed by GitHub
parent f5d739b132
commit 28a6ea1cdc
13 changed files with 160 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llamafactory/v1/config/training_args.py
View File

@@ -54,7 +54,7 @@ class TrainingArguments:
metadata={"help": "Maximum gradient norm for training."},
)
bf16: bool = field(
default=False,
default=True,
metadata={"help": "Use bf16 for training."},
)
batching_strategy: BatchingStrategy = field(
@@ -66,7 +66,7 @@ class TrainingArguments:
metadata={"help": "Number of workers for batching."},
)
enable_activation_checkpointing: bool = field(
default=False,
default=True,
metadata={"help": "Enable activation checkpointing for training."},
)
dist_config: PluginConfig | None = field(