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

fix extra config (#847)

Heng Guo70233bcb07ac
diff --git a/auto_round/__main__.py b/auto_round/__main__.py
index 78a8fc9d6..07bc3f273 100644
--- a/auto_round/__main__.py
+++ b/auto_round/__main__.py
@@ -436,16 +436,17 @@ def tune(args):
extra_config = ExtraConfig()
tuning_config = TuningExtraConfig(
amp=not args.disable_amp,
- lr=args.lr,
- minmax_lr=args.minmax_lr,
- enable_quanted_input=not args.disable_quanted_input,
- nblocks=args.nblocks,
+ disable_opt_rtn=args.disable_opt_rtn,
+ enable_alg_ext=args.enable_alg_ext,
enable_minmax_tuning=not args.disable_minmax_tuning,
enable_norm_bias_tuning=args.enable_norm_bias_tuning,
+ enable_quanted_input=not args.disable_quanted_input,
enable_deterministic_algorithms=args.enable_deterministic_algorithms,
+ lr=args.lr,
+ minmax_lr=args.minmax_lr,
+ mem_per_param_scale=args.mem_per_param_scale,
+ nblocks=args.nblocks,
to_quant_block_names=args.to_quant_block_names,
- disable_opt_rtn=args.disable_opt_rtn,
- enable_alg_ext=args.enable_alg_ext,
scale_dtype=args.scale_dtype,
)
scheme_config = SchemeExtraConfig(
@@ -459,6 +460,8 @@ def tune(args):
act_dynamic=act_dynamic,
super_bits=args.super_bits,
super_group_size=args.super_group_size,
+ quant_lm_head=args.quant_lm_head,
+ fp_layers=args.fp_layers,
)
mllm_config = MLLMExtraConfig(
quant_nontext_module=args.quant_nontext_module, extra_data_dir=args.extra_data_dir, template=args.template
@@ -480,7 +483,6 @@ def tune(args):
device_map=args.device_map,
enable_torch_compile=enable_torch_compile,
seed=args.seed,
- fp_layers=args.fp_layers,
not_use_best_mse=args.not_use_best_mse,
enable_adam=args.adam,
extra_config=extra_config,
diff --git a/auto_round/compressors/config.py b/auto_round/compressors/config.py
index cc9081ce9..5b40735ff 100644
--- a/auto_round/compressors/config.py
+++ b/auto_round/compressors/config.py
@@ -42,7 +42,6 @@ class ExtraConfig:
minmax_lr: float = None,
mem_per_param_scale: int = None,
nblocks: int = 1,
- quant_lm_head: bool = False,
to_quant_block_names: Union[str, list, None] = None,
scale_dtype: str = "fp16",
# scheme
@@ -58,6 +57,8 @@ class ExtraConfig:
super_bits: int = None,
super_group_size: int = None,
static_kv_dtype: Union[str, torch.dtype] = None,
+ quant_lm_head: bool = False,
+ fp_layers: str = None,
# mllm
processor: Callable = None,
image_processor: Callable = None,
@@ -116,7 +117,6 @@ class ExtraConfig:
minmax_lr=minmax_lr,
mem_per_param_scale=mem_per_param_scale,
nblocks=nblocks,
- quant_lm_head=quant_lm_head,
to_quant_block_names=to_quant_block_names,
scale_dtype=scale_dtype,
)
@@ -133,6 +133,8 @@ class ExtraConfig:
super_bits=super_bits,
super_group_size=super_group_size,
static_kv_dtype=static_kv_dtype,
+ quant_lm_head=quant_lm_head,
+ fp_layers=fp_layers,
)
self.mllm_config = MLLMExtraConfig(
processor=processor,
@@ -232,7 +234,6 @@ class TuningExtraConfig(BaseExtraConfig):
minmax_lr: float = None
mem_per_param_scale: int = None
nblocks: int = 1
- quant_lm_head: bool = False
to_quant_block_names: Union[str, list, None] = None
scale_dtype: str = "fp16"
@@ -251,6 +252,8 @@ class SchemeExtraConfig(BaseExtraConfig):
super_bits: int = None
super_group_size: int = None
static_kv_dtype: Union[str, torch.dtype] = None
+ quant_lm_head: bool = False
+ fp_layers: str = None
@dataclass