Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

add low_cpu_mem_usage in cli (#1146)

Heng Guo0b74fe1529d0
diff --git a/auto_round/__main__.py b/auto_round/__main__.py
index c468f716d..554b62704 100644
--- a/auto_round/__main__.py
+++ b/auto_round/__main__.py
@@ -135,6 +135,7 @@ class BasicArgumentParser(argparse.ArgumentParser):
help="Enable memory-efficient mode by offloading intermediate features to CPU. "
"Useful when working with large models that don't fit in GPU memory.",
)
+ basic.add_argument("--low_cpu_mem_usage", action="store_true", help="Lower CPU memory mode. Defaults to False.")
basic.add_argument(
"--format",
default="auto_round",
@@ -615,6 +616,7 @@ def tune(args):
batch_size=args.batch_size,
gradient_accumulate_steps=args.gradient_accumulate_steps,
low_gpu_mem_usage=args.low_gpu_mem_usage,
+ low_cpu_mem_usage=args.low_cpu_mem_usage,
device_map=args.device_map,
enable_torch_compile=enable_torch_compile,
seed=args.seed,
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index 327c82b07..4a860ae10 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -652,8 +652,9 @@ class BaseCompressor(object):
and self.iters > 0
):
logger.info(
+ "%s",
"'enable_torch_compile' is set to `False` by default. "
- "Enabling it can reduce tuning cost by 20%%, but it might throw an exception."
+ "Enabling it can reduce tuning cost by 20%, but it might throw an exception.",
)
if (self.data_type.startswith("fp") or self.act_data_type.startswith("fp")) and self.enable_torch_compile: