Rename RTNWeightOnlyConfig to RTNConfig (#1551)

* Rename RTNWeightOnlyConfig to RTNConfig

Signed-off-by: xin3he <xin3.he@intel.com>

Author: xin3he

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/llm/run_clm_no_trainer.py

@@ -230,7 +230,7 @@ def get_user_model():
 
     # 3.x api
     if args.approach == 'weight_only':
-        from neural_compressor.torch import RTNWeightQuantConfig, GPTQConfig, quantize
+        from neural_compressor.torch import RTNConfig, GPTQConfig, quantize
         from neural_compressor.torch.utils.utility import get_double_quant_config
         weight_sym = True if args.woq_scheme == "sym" else False
         double_quant_config_dict = get_double_quant_config(args.double_quant_type, weight_sym=weight_sym)
@@ -243,9 +243,9 @@ def get_user_model():
                         "enable_mse_search": args.woq_enable_mse_search,
                     }
                 )
-                quant_config = RTNWeightQuantConfig.from_dict(double_quant_config_dict)
+                quant_config = RTNConfig.from_dict(double_quant_config_dict)
             else:
-                quant_config = RTNWeightQuantConfig(
+                quant_config = RTNConfig(
                     weight_dtype=args.woq_dtype,
                     weight_bits=args.woq_bits,
                     weight_group_size=args.woq_group_size,
@@ -257,7 +257,7 @@ def get_user_model():
                     double_quant_sym=args.double_quant_sym,
                     double_quant_group_size=args.double_quant_group_size,
                 )
-            quant_config.set_local("lm_head", RTNWeightQuantConfig(weight_dtype="fp32"))
+            quant_config.set_local("lm_head", RTNConfig(weight_dtype="fp32"))
             user_model = quantize(
                 model=user_model, quant_config=quant_config
             )

neural_compressor/common/utility.py

@@ -27,7 +27,7 @@
 # config name
 BASE_CONFIG = "base_config"
 COMPOSABLE_CONFIG = "composable_config"
-RTN_WEIGHT_ONLY_QUANT = "rtn_weight_only_quant"
+RTN = "rtn"
 STATIC_QUANT = "static_quant"
 GPTQ = "gptq"
 FP8_QUANT = "fp8_quant"

neural_compressor/tensorflow/utils.py

@@ -35,7 +35,7 @@ def register_algo(name):
 
     Usage example:
         @register_algo(name=example_algo)
-        def example_algo(model: torch.nn.Module, quant_config: RTNWeightQuantConfig) -> torch.nn.Module:
+        def example_algo(model: torch.nn.Module, quant_config: RTNConfig) -> torch.nn.Module:
             ...
     Args:
         name (str): The name under which the algorithm function will be registered.

neural_compressor/torch/__init__.py

@@ -17,11 +17,11 @@
 
 from neural_compressor.torch.quantization import (
     quantize,
-    RTNWeightQuantConfig,
+    RTNConfig,
     get_default_rtn_config,
     GPTQConfig,
     get_default_gptq_config,
 )
 
 from neural_compressor.common.base_tuning import TuningConfig
-from neural_compressor.torch.autotune import autotune, get_default_tune_config
+from neural_compressor.torch.quantization.autotune import autotune, get_default_tune_config

neural_compressor/torch/algorithms/weight_only/rtn.py

@@ -580,10 +580,10 @@ def quant_weight_w_scale(weight, scale, zp, group_size=-1, dtype="int"):
     return int_weight
 
 
-from neural_compressor.torch.quantization.config import RTNWeightQuantConfig
+from neural_compressor.torch.quantization.config import RTNConfig
 
 
-def apply_rtn_on_single_module(module: torch.nn.Module, quant_config: RTNWeightQuantConfig) -> torch.nn.Module:
+def apply_rtn_on_single_module(module: torch.nn.Module, quant_config: RTNConfig) -> torch.nn.Module:
     # TODO (Yi) remove it
     enable_full_range = quant_config.enable_full_range
     enable_mse_search = quant_config.enable_mse_search

neural_compressor/torch/algorithms/weight_only_algos.py

@@ -18,17 +18,17 @@
 import torch
 
 from neural_compressor.common.logger import Logger
-from neural_compressor.common.utility import GPTQ, RTN_WEIGHT_ONLY_QUANT
-from neural_compressor.torch.quantization.config import GPTQConfig, RTNWeightQuantConfig
+from neural_compressor.common.utility import GPTQ, RTN
+from neural_compressor.torch.quantization.config import GPTQConfig, RTNConfig
 from neural_compressor.torch.utils.utility import fetch_module, register_algo, set_module
 
 logger = Logger().get_logger()
 
 
 ###################### RTN Algo Entry ##################################
-@register_algo(name=RTN_WEIGHT_ONLY_QUANT)
+@register_algo(name=RTN)
 def rtn_quantize_entry(
-    model: torch.nn.Module, configs_mapping: Dict[Tuple[str, callable], RTNWeightQuantConfig], *args, **kwargs
+    model: torch.nn.Module, configs_mapping: Dict[Tuple[str, callable], RTNConfig], *args, **kwargs
 ) -> torch.nn.Module:
     """The main entry to apply rtn quantization."""
     from .weight_only.rtn import apply_rtn_on_single_module

neural_compressor/torch/quantization/__init__.py

@@ -14,7 +14,7 @@
 
 from neural_compressor.torch.quantization.quantize import quantize, quantize_dynamic
 from neural_compressor.torch.quantization.config import (
-    RTNWeightQuantConfig,
+    RTNConfig,
     get_default_rtn_config,
     GPTQConfig,
     get_default_gptq_config,

neural_compressor/torch/autotune.py renamed to neural_compressor/torch/quantization/autotune.py

@@ -20,7 +20,7 @@
 from neural_compressor.common.base_tuning import TuningConfig, evaluator, init_tuning
 from neural_compressor.common.logger import Logger
 from neural_compressor.torch import quantize
-from neural_compressor.torch.quantization.config import GPTQConfig, RTNWeightQuantConfig
+from neural_compressor.torch.quantization.config import GPTQConfig, RTNConfig
 
 logger = Logger().get_logger()
 
@@ -33,7 +33,7 @@
 
 def get_default_tune_config() -> TuningConfig:
     # TODO use the registered default tuning config in the next PR
-    return TuningConfig(quant_configs=[GPTQConfig(weight_bits=[4, 8]), RTNWeightQuantConfig(weight_bits=[4, 8])])
+    return TuningConfig(quant_configs=[GPTQConfig(weight_bits=[4, 8]), RTNConfig(weight_bits=[4, 8])])
 
 
 def autotune(

neural_compressor/torch/quantization/config.py

@@ -24,13 +24,7 @@
 import torch
 
 from neural_compressor.common.base_config import BaseConfig, config_registry, register_config
-from neural_compressor.common.utility import (
-    DEFAULT_WHITE_LIST,
-    FP8_QUANT,
-    GPTQ,
-    OP_NAME_OR_MODULE_TYPE,
-    RTN_WEIGHT_ONLY_QUANT,
-)
+from neural_compressor.common.utility import DEFAULT_WHITE_LIST, FP8_QUANT, GPTQ, OP_NAME_OR_MODULE_TYPE, RTN
 from neural_compressor.torch.utils.constants import PRIORITY_GPTQ, PRIORITY_RTN
 from neural_compressor.torch.utils.utility import is_hpex_avaliable, logger
 
@@ -60,8 +54,8 @@ class OperatorConfig(NamedTuple):
 ######################## RNT Config ###############################
 
 
-@register_config(framework_name=FRAMEWORK_NAME, algo_name=RTN_WEIGHT_ONLY_QUANT, priority=PRIORITY_RTN)
-class RTNWeightQuantConfig(BaseConfig):
+@register_config(framework_name=FRAMEWORK_NAME, algo_name=RTN, priority=PRIORITY_RTN)
+class RTNConfig(BaseConfig):
     """Config class for round-to-nearest weight-only quantization."""
 
     supported_configs: List[OperatorConfig] = []
@@ -80,7 +74,7 @@ class RTNWeightQuantConfig(BaseConfig):
         "double_quant_sym",
         "double_quant_group_size",
     ]
-    name = RTN_WEIGHT_ONLY_QUANT
+    name = RTN
 
     def __init__(
         self,
@@ -137,12 +131,12 @@ def to_dict(self):
 
     @classmethod
     def from_dict(cls, config_dict):
-        return super(RTNWeightQuantConfig, cls).from_dict(config_dict=config_dict, str2operator=str2operator)
+        return super(RTNConfig, cls).from_dict(config_dict=config_dict, str2operator=str2operator)
 
     @classmethod
     def register_supported_configs(cls) -> List[OperatorConfig]:
         supported_configs = []
-        linear_rtn_config = RTNWeightQuantConfig(
+        linear_rtn_config = RTNConfig(
             weight_dtype=["int", "int8", "int4", "nf4", "fp4", "fp4_e2m1_bnb", "fp4_e2m1"],
             weight_bits=[4, 1, 2, 3, 5, 6, 7, 8],
             weight_group_size=[32, -1, 1, 4, 8, 16, 64, 128, 256, 512, 1024],
@@ -173,16 +167,16 @@ def get_model_info(model: torch.nn.Module) -> List[Tuple[str, Callable]]:
 
 
 # TODO(Yi) run `register_supported_configs` for all registered config.
-RTNWeightQuantConfig.register_supported_configs()
+RTNConfig.register_supported_configs()
 
 
-def get_default_rtn_config() -> RTNWeightQuantConfig:
+def get_default_rtn_config() -> RTNConfig:
     """Generate the default rtn config.
 
     Returns:
         the default rtn config.
     """
-    return RTNWeightQuantConfig()
+    return RTNConfig()
 
 
 ######################## GPTQ Config ###############################

neural_compressor/torch/utils/utility.py

@@ -33,7 +33,7 @@ def register_algo(name):
 
     Usage example:
         @register_algo(name=example_algo)
-        def example_algo(model: torch.nn.Module, quant_config: RTNWeightQuantConfig) -> torch.nn.Module:
+        def example_algo(model: torch.nn.Module, quant_config: RTNConfig) -> torch.nn.Module:
             ...
 
     Args: