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

Update setting of disable opt rtn (#1249)

Weiweicc521cab9776
diff --git a/README.md b/README.md
index 335a8f33f..b2309a975 100644
--- a/README.md
+++ b/README.md
@@ -195,7 +195,8 @@ ar.quantize_and_save(output_dir="./qmodel", format="auto_round")
##### Algorithm Settings
- **`enable_alg_ext` (bool)**: [Experimental Feature] Only for `iters>0`. Enable algorithm variants for specific schemes (e.g., MXFP4/W2A16) that could bring notable improvements. Default is `False`.
-- **`disable_opt_rtn` (bool)**: Use pure RTN mode for specific schemes (e.g., GGUF and WOQ). Default is `None` (improved RTN enabled).
+
+- **`disable_opt_rtn` (bool|None)**: Use pure RTN mode for specific schemes (e.g., GGUF and WOQ). Default is `None`. If None, it defaults to `False` in most cases to improve accuracy, but may be set to `True` due to known issues.
##### Tuning Process Parameters
- **`iters` (int)**: Number of tuning iterations (default is `200`). Common values: 0 (RTN mode), 50 (with lr=5e-3 recommended), 1000. Higher values increase accuracy but slow down tuning.
@@ -355,3 +356,4 @@ Special thanks to open-source low precision libraries such as AutoGPTQ, AutoAWQ,
## 🌟 Support Us
If you find AutoRound helpful, please ⭐ star the repo and share it with your community!
+
diff --git a/auto_round/__main__.py b/auto_round/__main__.py
index 6d95b4864..0a17a05d0 100644
--- a/auto_round/__main__.py
+++ b/auto_round/__main__.py
@@ -253,14 +253,23 @@ class BasicArgumentParser(argparse.ArgumentParser):
action="store_true",
help="Enable PyTorch deterministic algorithms for reproducible results. ",
)
- tuning.add_argument(
+ group = tuning.add_mutually_exclusive_group()
+ group.add_argument(
"--disable_opt_rtn",
- "--disable-opt-rtn",
- action=argparse.BooleanOptionalAction,
+ action="store_const",
+ const=True,
+ dest="disable_opt_rtn",
default=None,
help="Disable optimization for RTN (Round-To-Nearest) mode when iters=0. "
"RTN is fast but less accurate; keeping optimization enabled is recommended.",
)
+ group.add_argument(
+ "--enable_opt_rtn",
+ action="store_const",
+ const=False,
+ dest="disable_opt_rtn",
+ help="Enable optimization for RTN mode when iters=0.",
+ )
scheme = self.add_argument_group("Scheme Arguments")
scheme.add_argument("--bits", default=None, type=int, help="Number of bits for weight quantization. ")
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index 51c850c16..6d69c01c3 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -85,7 +85,7 @@ class AutoRound:
enable_adam: bool = False,
extra_config: ExtraConfig = None,
enable_alg_ext: bool = None,
- disable_opt_rtn: Optional[bool] = None,
+ disable_opt_rtn: bool | None = None,
low_cpu_mem_usage: bool = False,
**kwargs,
) -> BaseCompressor:
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index cd9fc516c..a2af9e011 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -189,7 +189,7 @@ class BaseCompressor(object):
device_map: Union[str, torch.device, int, dict] = 0,
enable_torch_compile: bool = False,
enable_alg_ext: bool = False,
- disable_opt_rtn: Optional[bool] = None,
+ disable_opt_rtn: bool | None = None,
seed: int = 42,
low_cpu_mem_usage: bool = False,
**kwargs,
@@ -397,9 +397,11 @@ class BaseCompressor(object):
and self.data_type == "int"
and disable_opt_rtn is None
):
- logger.warning("for INT8 RTN quantization, set `--disable_opt_rtn` as default.")
+ logger.warning("For INT8 RTN quantization, set `--disable_opt_rtn` as default.")
disable_opt_rtn = True
if disable_opt_rtn is None:
+ if self.iters == 0:
+ logger.info("For the most RTN cases, set `--disable_opt_rtn` to False as default.")
disable_otp_rtn = False
self.minmax_lr = minmax_lr or self.lr
diff --git a/auto_round/compressors/config.py b/auto_round/compressors/config.py
index 4bab246c3..f7e0d92f2 100644
--- a/auto_round/compressors/config.py
+++ b/auto_round/compressors/config.py
@@ -32,7 +32,7 @@ class ExtraConfig:
self,
# tuning
amp: bool = True,
- disable_opt_rtn: Optional[bool] = True,
+ disable_opt_rtn: bool | None = None,
enable_alg_ext: bool = False,
enable_minmax_tuning: bool = True,
enable_norm_bias_tuning: bool = False,
@@ -247,7 +247,7 @@ class BaseExtraConfig:
@dataclass
class TuningExtraConfig(BaseExtraConfig):
amp: bool = True
- disable_opt_rtn: Optional[bool] = True
+ disable_opt_rtn: bool | None = True
enable_alg_ext: bool = False
enable_minmax_tuning: bool = True
enable_norm_bias_tuning: bool = False