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

adamround bugfix, refine import (#275)

WeiweiZhang1a633aa707090
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index 142090451..ecd0ed0c6 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -1375,7 +1375,7 @@ class AutoOPTRound(AutoRound):
act_group_size (int): Group size for activation quantization. Default is None.
act_sym (bool): Whether to use symmetric activation quantization. Default is None.
act_dynamic (bool): Whether to use dynamic activation quantization. Default is True.
-
+ quant_block_list (list): A list whose elements are list of block's layer names to be quantized.
**kwargs: Additional keyword arguments.
Returns:
@@ -1418,44 +1418,46 @@ class AutoOPTRound(AutoRound):
act_sym: bool = None,
act_dynamic: bool = True,
enable_norm_bias_tuning: bool = False,
+ quant_block_list: list = None,
optimizer="AdamW",
**kwargs,
):
super(AutoOPTRound, self).__init__(
- model,
- tokenizer,
- bits,
- group_size,
- sym,
- layer_config,
- enable_full_range,
- batch_size,
- amp,
- device,
- lr_scheduler,
- dataset,
- enable_quanted_input,
- enable_minmax_tuning,
- lr,
- minmax_lr,
- low_gpu_mem_usage,
- low_cpu_mem_usage,
- iters,
- seqlen,
- nsamples,
- sampler,
- seed,
- nblocks,
- gradient_accumulate_steps,
- not_use_best_mse,
- dynamic_max_gap,
- data_type,
- scale_dtype,
- act_bits,
- act_group_size,
- act_sym,
- act_dynamic,
- enable_norm_bias_tuning,
+ model=model,
+ tokenizer=tokenizer,
+ bits=bits,
+ group_size=group_size,
+ sym=sym,
+ layer_config=layer_config,
+ enable_full_range=enable_full_range,
+ batch_size=batch_size,
+ amp=amp,
+ device=device,
+ lr_scheduler=lr_scheduler,
+ dataset=dataset,
+ enable_quanted_input=enable_quanted_input,
+ enable_minmax_tuning=enable_minmax_tuning,
+ lr=lr,
+ minmax_lr=minmax_lr,
+ low_gpu_mem_usage=low_gpu_mem_usage,
+ low_cpu_mem_usage=low_cpu_mem_usage,
+ iters=iters,
+ seqlen=seqlen,
+ nsamples=nsamples,
+ sampler=sampler,
+ seed=seed,
+ nblocks=nblocks,
+ gradient_accumulate_steps=gradient_accumulate_steps,
+ not_use_best_mse=not_use_best_mse,
+ dynamic_max_gap=dynamic_max_gap,
+ data_type=data_type,
+ scale_dtype=scale_dtype,
+ act_bits=act_bits,
+ act_group_size=act_group_size,
+ act_sym=act_sym,
+ act_dynamic=act_dynamic,
+ enable_norm_bias_tuning=enable_norm_bias_tuning,
+ quant_block_list=quant_block_list,
**kwargs,
)
@@ -1544,7 +1546,7 @@ class AutoAdamRound(AutoOPTRound):
act_group_size (int): Group size for activation quantization. Default is None.
act_sym (bool): Whether to use symmetric activation quantization. Default is None.
act_dynamic (bool): Whether to use dynamic activation quantization. Default is True.
-
+ quant_block_list (list): A list whose elements are list of block's layer names to be quantized.
Returns:
The quantized model.
"""
@@ -1585,45 +1587,48 @@ class AutoAdamRound(AutoOPTRound):
act_sym: bool = None,
act_dynamic: bool = True,
enable_norm_bias_tuning: bool = False,
+ quant_block_list: list = None,
optimizer="AdamW",
**kwargs,
):
super(AutoAdamRound, self).__init__(
- model,
- tokenizer,
- bits,
- group_size,
- sym,
- layer_config,
- enable_full_range,
- batch_size,
- amp,
- device,
- lr_scheduler,
- dataset,
- enable_quanted_input,
- enable_minmax_tuning,
- lr,
- minmax_lr,
- low_gpu_mem_usage,
- low_cpu_mem_usage,
- iters,
- seqlen,
- nsamples,
- sampler,
- seed,
- nblocks,
- gradient_accumulate_steps,
- not_use_best_mse,
- dynamic_max_gap,
- data_type,
- scale_dtype,
- act_bits,
- act_group_size,
- act_sym,
- act_dynamic,
- enable_norm_bias_tuning,
- optimizer,
+ model=model,
+ tokenizer=tokenizer,
+ bits=bits,
+ group_size=group_size,
+ sym=sym,
+ layer_config=layer_config,
+ enable_full_range=enable_full_range,
+ batch_size=batch_size,
+ amp=amp,
+ device=device,
+ lr_scheduler=lr_scheduler,
+ dataset=dataset,
+ enable_quanted_input=enable_quanted_input,
+ enable_minmax_tuning=enable_minmax_tuning,
+ lr=lr,
+ minmax_lr=minmax_lr,
+ low_gpu_mem_usage=low_gpu_mem_usage,
+ low_cpu_mem_usage=low_cpu_mem_usage,
+ iters=iters,
+ seqlen=seqlen,
+ nsamples=nsamples,
+ sampler=sampler,
+ seed=seed,
+ nblocks=nblocks,
+ gradient_accumulate_steps=gradient_accumulate_steps,
+ not_use_best_mse=not_use_best_mse,
+ dynamic_max_gap=dynamic_max_gap,
+ data_type=data_type,
+ scale_dtype=scale_dtype,
+ act_bits=act_bits,
+ act_group_size=act_group_size,
+ act_sym=act_sym,
+ act_dynamic=act_dynamic,
+ enable_norm_bias_tuning=enable_norm_bias_tuning,
+ quant_block_list=quant_block_list,
+ optimizer=optimizer,
**kwargs,
)
+
diff --git a/auto_round/utils.py b/auto_round/utils.py
index 064bed6eb..05593c02a 100644
--- a/auto_round/utils.py
+++ b/auto_round/utils.py
@@ -891,11 +891,11 @@ def get_library_version(library_name):
except pkg_resources.DistributionNotFound:
return f"{library_name} is not installed"
else:
- import importlib_metadata # pylint: disable=E0401
+ import importlib.metadata # pylint: disable=E0401
try:
- version = importlib_metadata.version(library_name)
+ version = importlib.metadata.version(library_name)
return version
- except importlib_metadata.PackageNotFoundError:
+ except importlib.metadata.PackageNotFoundError:
return f"{library_name} is not installed"
@@ -972,3 +972,4 @@ def get_autogptq_packing_qlinear(backend, bits=4, group_size=128, sym=False):
)
return QuantLinear
+