Integrate AutoRound v0.3 (#1925)

Signed-off-by: Kaihui-intel <[email protected]>

Author: Kaihui-intel

neural_compressor/torch/algorithms/weight_only/autoround.py

@@ -31,69 +31,95 @@ class AutoRoundQuantizer(Quantizer):
     def __init__(
         self,
         quant_config: dict = {},
-        enable_full_range: bool = False,
+        enable_full_range: bool = False,  ##for symmetric, TODO support later
         batch_size: int = 8,
         amp: bool = True,
-        device=None,
+        device: str = None,
         lr_scheduler=None,
+        dataset: Union[str, list, tuple, torch.utils.data.DataLoader] = "NeelNanda/pile-10k",
         enable_quanted_input: bool = True,
         enable_minmax_tuning: bool = True,
         lr: float = None,
         minmax_lr: float = None,
-        low_gpu_mem_usage: bool = True,
+        low_gpu_mem_usage: bool = False,
         iters: int = 200,
         seqlen: int = 2048,
-        n_samples: int = 512,
+        nsamples: int = 128,
         sampler: str = "rand",
         seed: int = 42,
-        n_blocks: int = 1,
+        nblocks: int = 1,
         gradient_accumulate_steps: int = 1,
         not_use_best_mse: bool = False,
         dynamic_max_gap: int = -1,
         data_type: str = "int",
         scale_dtype: str = "fp16",
+        multimodal: bool = False,
+        act_bits: int = 32,
+        act_group_size: int = None,
+        act_sym: bool = None,
+        act_dynamic: bool = True,
+        low_cpu_mem_usage: bool = False,
         **kwargs,
     ):
         """Init a AutQRoundQuantizer object.
 
         Args:
-        quant_config (dict): Configuration for weight quantization (default is None).
-        quant_config={
-                    'layer1':##layer_name
-                    {
-                        'data_type': 'int',
-                        'bits': 4,
-                        'group_size': 32,
-                        'sym': False,
+            quant_config (dict): Configuration for weight quantization (default is None).
+            quant_config={
+                        'layer1':##layer_name
+                        {
+                            'data_type': 'int',
+                            'bits': 4,
+                            'group_size': 32,
+                            'sym': False,
+                            'act_data_type': None,
+                            'act_bits': 32,
+                            'act_sym': None,
+                            'act_dynamic': True,
+                        }
+                        ...,
                     }
-                    ...
-                }
-            keys:
-                data_type (str): The data type to be used (default is "int").
-                bits (int): Number of bits for quantization (default is 4).
-                group_size (int): Size of the quantization group (default is 128).
-                sym (bool): Whether to use symmetric quantization. (default is None).
-        enable_full_range (bool): Whether to enable full range quantization (default is False).
-        batch_size (int): Batch size for training (default is 8).
-        amp (bool): Whether to use automatic mixed precision (default is True). Automatically detect and set.
-        device: The device to be used for tuning (default is None). Automatically detect and set.
-        lr_scheduler: The learning rate scheduler to be used.
-        use_quant_input (bool): Whether to use quantized input data (default is True).
-        enable_minmax_tuning (bool): Whether to enable min-max tuning (default is True).
-        lr (float): The learning rate (default is 0.005).
-        minmax_lr (float): The learning rate for min-max tuning (default is None).
-        low_gpu_mem_usage (bool): Whether to use low GPU memory (default is True).
-        iters (int): Number of iterations (default is 200).
-        seqlen (int): Length of the sequence.
-        n_samples (int): Number of samples (default is 512).
-        sampler (str): The sampling method (default is "rand").
-        seed (int): The random seed (default is 42).
-        n_blocks (int): Number of blocks (default is 1).
-        gradient_accumulate_steps (int): Number of gradient accumulation steps (default is 1).
-        not_use_best_mse (bool): Whether to use mean squared error (default is False).
-        dynamic_max_gap (int): The dynamic maximum gap (default is -1).
-        scale_dtype (str): The data type of quantization scale to be used (default is "float16"), different kernels
-                            have different choices.
+                keys:
+                    data_type (str): The data type to be used (default is "int").
+                    bits (int): Number of bits for quantization (default is 4).
+                    group_size (int): Size of the quantization group (default is 128).
+                    sym (bool): Whether to use symmetric quantization. (default is None).
+            bits (int): Number of bits for quantization (default is 4).
+            group_size (int): Size of the quantization group (default is 128).
+            sym (bool): Whether symmetric quantization is to be used (default is False).
+            enable_full_range (bool): Whether to enable full range quantization (default is False).
+            batch_size (int): Batch size for training (default is 8).
+            amp (bool): Whether to use automatic mixed precision (default is True).
+            device: The device to be used for tuning (default is "auto").
+            lr_scheduler: The learning rate scheduler to be used.
+            dataset (str): The default dataset name (default is "NeelNanda/pile-10k").
+            enable_quanted_input (bool): Whether to use the output of the previous quantized block as
+                the input for the current block (default is True).
+            enable_minmax_tuning (bool): Whether to enable weight min-max tuning (default is True).
+            lr (float): The learning rate (default is None, will be set to 1.0/iters).
+            minmax_lr (float): The learning rate for min-max tuning
+                (default is None, it will be set to lr automatically).
+            low_gpu_mem_usage (bool): Whether to use low GPU memory (default is True).
+            iters (int): Number of iterations (default is 200).
+            seqlen (int): Data length of the sequence for tuning (default is 2048).
+            nsamples (int): Number of samples (default is 128).
+            sampler (str): The sampling method (default is "rand").
+            seed (int): The random seed (default is 42).
+            nblocks (int): Number of blocks (default is 1).
+            gradient_accumulate_steps (int): Number of gradient accumulation steps (default is 1).
+            not_use_best_mse (bool): Whether to use mean squared error (default is False).
+            dynamic_max_gap (int): The dynamic maximum gap (default is -1).
+            data_type (str): The data type to be used (default is "int").
+            scale_dtype (str): The data type of quantization scale to be used (default is "float16"), different kernels
+                have different choices.
+            multimodal(bool): Enable multimodal model quantization, (default is "False").
+            act_bits (int): Number of bits for activation quantization. Default is 32.
+            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.
+
+        Returns:
+            The quantized model.
         """
         super().__init__(quant_config)
         self.tokenizer = None
@@ -109,15 +135,21 @@ def __init__(
         self.low_gpu_mem_usage = low_gpu_mem_usage
         self.iters = iters
         self.seqlen = seqlen
-        self.n_samples = n_samples
+        self.nsamples = nsamples
         self.sampler = sampler
         self.seed = seed
-        self.n_blocks = n_blocks
+        self.nblocks = nblocks
         self.gradient_accumulate_steps = gradient_accumulate_steps
         self.not_use_best_mse = not_use_best_mse
         self.dynamic_max_gap = dynamic_max_gap
         self.data_type = data_type
         self.scale_dtype = scale_dtype
+        self.multimodal = multimodal
+        self.act_bits = act_bits
+        self.act_group_size = act_group_size
+        self.act_sym = act_sym
+        self.act_dynamic = act_dynamic
+        self.low_cpu_mem_usage = low_cpu_mem_usage
 
     def prepare(self, model: torch.nn.Module, *args, **kwargs):
         """Prepares a given model for quantization.
@@ -137,7 +169,7 @@ def convert(self, model: torch.nn.Module, *args, **kwargs):
             model=model,
             tokenizer=None,
             dataset=dataloader,
-            weight_config=self.quant_config or {},
+            layer_config=self.quant_config or {},
             enable_full_range=self.enable_full_range,
             batch_size=self.batch_size,
             amp=self.amp,
@@ -150,23 +182,29 @@ def convert(self, model: torch.nn.Module, *args, **kwargs):
             low_gpu_mem_usage=self.low_gpu_mem_usage,
             iters=self.iters,
             seqlen=self.seqlen,
-            n_samples=self.n_samples,
+            nsamples=self.nsamples,
             sampler=self.sampler,
             seed=self.seed,
-            n_blocks=self.n_blocks,
+            nblocks=self.nblocks,
             gradient_accumulate_steps=self.gradient_accumulate_steps,
             not_use_best_mse=self.not_use_best_mse,
             dynamic_max_gap=self.dynamic_max_gap,
             data_type=self.data_type,
             scale_dtype=self.scale_dtype,
+            multimodal=self.multimodal,
+            act_bits=self.act_bits,
+            act_group_size=self.act_group_size,
+            act_sym=self.act_sym,
+            act_dynamic=self.act_dynamic,
+            low_cpu_mem_usage=self.low_cpu_mem_usage,
         )
         model, weight_config = rounder.quantize()
         model.autoround_config = weight_config
         model = pack_model(model, weight_config, device=self.device, inplace=True)
         return model
 
 
-def get_dataloader(tokenizer, seqlen, dataset_name="NeelNanda/pile-10k", seed=42, bs=8, n_samples=512):
+def get_dataloader(tokenizer, seqlen, dataset_name="NeelNanda/pile-10k", seed=42, bs=8, nsamples=128):
     """Generate a DataLoader for calibration using specified parameters.
 
     Args:
@@ -186,6 +224,6 @@ def get_dataloader(tokenizer, seqlen, dataset_name="NeelNanda/pile-10k", seed=42
     from auto_round.calib_dataset import get_dataloader  # pylint: disable=E0401
 
     dataloader = get_dataloader(
-        tokenizer, seqlen, dataset_name="NeelNanda/pile-10k", seed=seed, bs=bs, n_samples=n_samples
+        tokenizer, seqlen, dataset_name="NeelNanda/pile-10k", seed=seed, bs=bs, nsamples=nsamples
     )
     return dataloader

neural_compressor/torch/quantization/algorithm_entry.py

@@ -572,6 +572,10 @@ def autoround_quantize_entry(
                 "bits": quant_config.bits,
                 "sym": quant_config.use_sym,
                 "group_size": quant_config.group_size,
+                "act_bits": quant_config.act_bits,
+                "act_group_size": quant_config.act_group_size,
+                "act_sym": quant_config.act_sym,
+                "act_dynamic": quant_config.act_dynamic,
             }
             enable_full_range = quant_config.enable_full_range
             batch_size = quant_config.batch_size
@@ -583,14 +587,16 @@ def autoround_quantize_entry(
             low_gpu_mem_usage = quant_config.low_gpu_mem_usage
             iters = quant_config.iters
             seqlen = quant_config.seqlen
-            n_samples = quant_config.n_samples
+            nsamples = quant_config.nsamples
             sampler = quant_config.sampler
             seed = quant_config.seed
-            n_blocks = quant_config.n_blocks
+            nblocks = quant_config.nblocks
             gradient_accumulate_steps = quant_config.gradient_accumulate_steps
             not_use_best_mse = quant_config.not_use_best_mse
             dynamic_max_gap = quant_config.dynamic_max_gap
             scale_dtype = quant_config.scale_dtype
+            multimodal = quant_config.multimodal
+            low_cpu_mem_usage = quant_config.use_layer_wise
 
     kwargs.pop("example_inputs")
 
@@ -608,14 +614,16 @@ def autoround_quantize_entry(
         low_gpu_mem_usage=low_gpu_mem_usage,
         iters=iters,
         seqlen=seqlen,
-        n_samples=n_samples,
+        nsamples=nsamples,
         sampler=sampler,
         seed=seed,
-        n_blocks=n_blocks,
+        nblocks=nblocks,
         gradient_accumulate_steps=gradient_accumulate_steps,
         not_use_best_mse=not_use_best_mse,
         dynamic_max_gap=dynamic_max_gap,
         scale_dtype=scale_dtype,
+        multimodal=multimodal,
+        low_cpu_mem_usage=low_cpu_mem_usage,
     )
     model = quantizer.execute(model=model, mode=mode, *args, **kwargs)
     model.qconfig = configs_mapping

neural_compressor/torch/quantization/config.py

@@ -735,8 +735,8 @@ class AutoRoundConfig(TorchBaseConfig):
         "minmax_lr",
         "iters",
         "seqlen",
-        "n_samples",
-        "n_blocks",
+        "nsamples",
+        "nblocks",
         "gradient_accumulate_steps",
         "not_use_best_mse",
         "dynamic_max_gap",
@@ -750,6 +750,10 @@ def __init__(
         use_sym: bool = False,
         group_size: int = 128,
         # AUTOROUND
+        act_bits: int = 32,
+        act_group_size: int = None,
+        act_sym: bool = None,
+        act_dynamic: bool = True,
         enable_full_range: bool = False,
         batch_size: int = 8,
         lr_scheduler=None,
@@ -759,16 +763,17 @@ def __init__(
         minmax_lr: float = None,
         low_gpu_mem_usage: bool = True,
         iters: int = 200,
-        seqlen: int = 512,
-        n_samples: int = 512,
+        seqlen: int = 2048,
+        nsamples: int = 128,
         sampler: str = "rand",
         seed: int = 42,
-        n_blocks: int = 1,
+        nblocks: int = 1,
         gradient_accumulate_steps: int = 1,
         not_use_best_mse: bool = False,
         dynamic_max_gap: int = -1,
         scale_dtype: str = "fp16",
         use_layer_wise: bool = False,
+        multimodal: bool = False,
         white_list: Optional[List[OP_NAME_OR_MODULE_TYPE]] = DEFAULT_WHITE_LIST,
     ):
         """Init AUTOROUND weight-only quantization config.
@@ -778,6 +783,10 @@ def __init__(
             bits (int): Number of bits used to represent weights, default is 4.
             use_sym (bool): Indicates whether weights are symmetric, default is False.
             group_size (int): Size of weight groups, default is 128.
+            act_bits (int): Number of bits for activation quantization. Default is 32.
+            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.
             enable_full_range (bool): Whether to enable full range quantization (default is False).
             batch_size (int): Batch size for training (default is 8).
             lr_scheduler: The learning rate scheduler to be used.
@@ -788,21 +797,27 @@ def __init__(
             low_gpu_mem_usage (bool): Whether to use low GPU memory (default is True).
             iters (int): Number of iterations (default is 200).
             seqlen (int): Length of the sequence.
-            n_samples (int): Number of samples (default is 512).
+            nsamples (int): Number of samples (default is 512).
             sampler (str): The sampling method (default is "rand").
             seed (int): The random seed (default is 42).
-            n_blocks (int): Number of blocks (default is 1).
+            nblocks (int): Number of blocks (default is 1).
             gradient_accumulate_steps (int): Number of gradient accumulation steps (default is 1).
             not_use_best_mse (bool): Whether to use mean squared error (default is False).
             dynamic_max_gap (int): The dynamic maximum gap (default is -1).
             scale_dtype (str): The data type of quantization scale to be used (default is "float16"), different kernels
                         have different choices.
+            use_layer_wise (bool): Enables quantize model per layer. Defaults to False.
+            multimodal(bool): Enable multimodal model quantization, (default is "False").
         """
         super().__init__(white_list=white_list)
         self.dtype = dtype
         self.bits = bits
         self.use_sym = use_sym
         self.group_size = group_size
+        self.act_bits = act_bits
+        self.act_group_size = act_group_size
+        self.act_sym = act_sym
+        self.act_dynamic = act_dynamic
         self.enable_full_range = enable_full_range
         self.batch_size = batch_size
         self.lr_scheduler = lr_scheduler
@@ -813,15 +828,16 @@ def __init__(
         self.low_gpu_mem_usage = low_gpu_mem_usage
         self.iters = iters
         self.seqlen = seqlen
-        self.n_samples = n_samples
+        self.nsamples = nsamples
         self.sampler = sampler
         self.seed = seed
-        self.n_blocks = n_blocks
+        self.nblocks = nblocks
         self.gradient_accumulate_steps = gradient_accumulate_steps
         self.not_use_best_mse = not_use_best_mse
         self.dynamic_max_gap = dynamic_max_gap
         self.scale_dtype = scale_dtype
         self.use_layer_wise = use_layer_wise
+        self.multimodal = multimodal
         self._post_init()
 
     @classmethod
@@ -1526,7 +1542,7 @@ def get_woq_tuning_config() -> list:
         the list of WOQ quant config.
     """
     RTN_G32ASYM = RTNConfig(use_sym=False, group_size=32)
-    AUTO_ROUND_CONFIG = AutoRoundConfig(use_sym=False, group_size=32)
+    AUTO_ROUND_CONFIG = AutoRoundConfig(use_sym=False, group_size=32, seqlen=512)
     GPTQ_G32ASYM = GPTQConfig(use_sym=False, group_size=32)
     AWQ_G32ASYM = AWQConfig(use_sym=False, group_size=32)
     return [RTN_G32ASYM, AUTO_ROUND_CONFIG, GPTQ_G32ASYM, AWQ_G32ASYM]