Support Quantization of Big Saved Model for TF Backend (#1396)

Signed-off-by: zehao-intel <zehao.huang@intel.com>

neural_compressor/adaptor/tensorflow.py

@@ -648,6 +648,7 @@ def quantize(self, tune_cfg, model, data_loader, q_func=None):
                     fp32_ops=self.fp32_ops,
                     bf16_ops=self.bf16_ops,
                     data_loader=data_loader,
+                    calib_func=q_func,
                     qdq_enabled=self.qdq_enabled,
                     new_api=self.new_api,
                     performance_only=self.performance_only,
@@ -670,6 +671,7 @@ def quantize(self, tune_cfg, model, data_loader, q_func=None):
                     fp32_ops=self.fp32_ops,
                     bf16_ops=self.bf16_ops,
                     data_loader=data_loader,
+                    calib_func=q_func,
                     qdq_enabled=self.qdq_enabled,
                     new_api=self.new_api,
                     performance_only=self.performance_only,
@@ -693,6 +695,7 @@ def quantize(self, tune_cfg, model, data_loader, q_func=None):
                 fp32_ops=self.fp32_ops,
                 bf16_ops=self.bf16_ops,
                 data_loader=data_loader,
+                calib_func=q_func,
                 qdq_enabled=self.qdq_enabled,
                 new_api=self.new_api,
                 performance_only=self.performance_only,
@@ -761,15 +764,15 @@ def _dump_model_op_stats(self, model_graphdef):
             if i.op in fp32_op_list:
                 if "T" not in i.attr and i.op != "Cast":
                     continue
-                if i.attr["T"].type == dtypes.bfloat16:
-                    res[i.op]["BF16"] += 1
-                elif i.attr["T"].type in (dtypes.quint8, dtypes.qint8):
-                    res[i.op]["INT8"] += 1
-                elif i.op == "Cast":
+                if i.op == "Cast":
                     if i.attr["DstT"].type == dtypes.bfloat16:
                         res[i.op]["BF16"] += 1
                     elif i.attr["DstT"].type == dtypes.float32:
                         res[i.op]["FP32"] += 1
+                elif i.attr["T"].type == dtypes.bfloat16:
+                    res[i.op]["BF16"] += 1
+                elif i.attr["T"].type in (dtypes.quint8, dtypes.qint8):
+                    res[i.op]["INT8"] += 1
                 else:
                     res[i.op]["FP32"] += 1
 
@@ -1815,7 +1818,6 @@ def smooth_quant(
         model,
         dataloader,
         calib_iter=1,
-        tune_cfg=None,
         alpha=0.5,
         folding=False,
         percentile=99.999,
@@ -1832,7 +1834,6 @@ def smooth_quant(
             model: original model
             dataloader: the calibration dataloader
             calib_iter: how many steps of iterations on the dataloader to move forward
-            tune_cfg: quantization config
             alpha: smooth alpha in SmoothQuant, 1.0 will fallback to SPIQ
             folding: whether insert mul(False) or just allow foldable layers(True) for SmoothQuant
             percentile: percentile of calibration to remove outliers
@@ -1852,6 +1853,11 @@ def smooth_quant(
         if self.smooth_quant_model is not None:
             return self.smooth_quant_model
 
+        if model.model_type == "llm_saved_model":
+            return self.smooth_quant_LLM(
+                model, dataloader, calib_iter, alpha, folding, percentile, op_types, scales_per_op
+            )
+
         # Do a pre-optimization before smooth quant
         from .tf_utils.graph_rewriter.generic.pre_optimize import PreOptimization
 
@@ -1860,6 +1866,7 @@ def smooth_quant(
         model.graph_def = self.pre_optimized_model.graph_def
 
         # Get the nodes list which can't be quantized from tune_cfg
+        tune_cfg = None
         black_nodes = []
         if tune_cfg is not None:
             self._tuning_cfg_to_fw(tune_cfg)
@@ -1887,6 +1894,81 @@ def smooth_quant(
         self.smooth_quant_model = model
         return self.smooth_quant_model
 
+    def smooth_quant_LLM(
+        self,
+        model,
+        dataloader,
+        calib_iter=1,
+        alpha=0.5,
+        folding=False,
+        percentile=99.999,
+        op_types=["MatMul", "Conv2D"],
+        scales_per_op=True,
+    ):
+        """Convert the model by smooth quant.
+
+        Args:
+            model: original model of TensorflowLLMModel object.
+            calib_iter: how many steps of iterations on the dataloader to move forward.
+            tune_cfg: quantization config.
+            alpha: smooth alpha in SmoothQuant, 1.0 will fallback to SPIQ.
+            folding: whether insert mul(False) or just allow foldable layers(True) for SmoothQuant.
+            percentile: percentile of calibration to remove outliers.
+            op_types: The op types whose input tensor will be dumped.
+            scales_per_op: True, each op will have an individual scale, mainly for accuracy.
+                           False, ops with the same input will share a scale, mainly for performance.
+
+        Returns:
+            model: A smoothed Tensorflow model.
+        """
+        # Do a pre-optimization before smooth quant
+        from .tf_utils.graph_rewriter.generic.pre_optimize import PreOptimization
+
+        self.pre_optimizer_handle = PreOptimization(model, self.new_api, self.device)
+        self.pre_optimized_model = self.pre_optimizer_handle.get_optimized_model(self.itex_mode)
+        model.graph_def = self.pre_optimized_model.graph_def
+
+        # Get the nodes list which can't be quantized from tune_cfg
+        tune_cfg = None
+        black_nodes = []
+        if tune_cfg is not None:
+            self._tuning_cfg_to_fw(tune_cfg)
+            black_nodes = [node for node in self.quantize_config if self.quantize_config[node] == "fp32"]
+
+        # only support per-tensor MatMul now
+        op_types = ["MatMul"]
+        llm_temp_dir = self.work_dir + "/temp_saved_model"
+        # Run calibration to get max values per channel
+        from .tf_utils.smooth_quant_calibration import SmoothQuantCalibrationLLM
+
+        calibration = SmoothQuantCalibrationLLM(
+            model._model,
+            dataloader,
+            calib_iter,
+            op_types,
+            percentile,
+            black_nodes,
+            llm_temp_dir,
+            model.weight_name_mapping,
+        )
+        max_vals_per_channel, sq_target_node_names, sq_weight_tensor_dict, sq_graph_def = calibration(
+            model.input_node_names, model.output_node_names
+        )
+
+        # Calculate the smooth quant scaler and insert Mul op into the graph
+        from .tf_utils.smooth_quant_scaler import SmoothQuantScalerLLM
+
+        scaler = SmoothQuantScalerLLM(sq_graph_def, alpha, scales_per_op, op_types)
+        sq_graph_def, sq_weight_scale_dict, mul_list = scaler.transform(
+            max_vals_per_channel, sq_weight_tensor_dict, sq_target_node_names
+        )
+        model.graph_def = sq_graph_def
+        model.model_path = llm_temp_dir
+        model.sq_weight_scale_dict = sq_weight_scale_dict
+        self.smooth_quant_mul_ops.extend(mul_list)
+        self.smooth_quant_model = model
+        return self.smooth_quant_model
+
 
 @adaptor_registry
 class Tensorflow_ITEXAdaptor(TensorFlowAdaptor):
@@ -1945,6 +2027,7 @@ def quantize(self, tune_cfg, model, data_loader, q_func=None):
                     fp32_ops=self.fp32_ops,
                     bf16_ops=self.bf16_ops,
                     data_loader=data_loader,
+                    calib_func=q_func,
                     itex_mode=self.itex_mode,
                     qdq_enabled=self.qdq_enabled,
                     new_api=self.new_api,
@@ -1992,6 +2075,7 @@ def quantize(self, tune_cfg, model, data_loader, q_func=None):
                 fp32_ops=self.fp32_ops,
                 bf16_ops=self.bf16_ops,
                 data_loader=data_loader,
+                calib_func=q_func,
                 itex_mode=self.itex_mode,
                 qdq_enabled=self.qdq_enabled,
                 new_api=self.new_api,

neural_compressor/adaptor/tf_utils/graph_converter.py

@@ -102,6 +102,7 @@ def __init__(
         fp32_ops=[],
         bf16_ops=[],
         data_loader=None,
+        calib_func=None,
         fake_quant=False,
         itex_mode=False,
         qdq_enabled=False,
@@ -116,6 +117,7 @@ def __init__(
         :param fp32_ops: fall back to fp32 dtype op list
         :param bf16_ops: fall back to bf16 dtype op list
         :param data_loader: for calibration phase used dataloader
+        :param calib_func: for calibration phase used function
         :param fake_quant: for quantization-aware training model conversion to default model
         """
         self.model = model
@@ -139,6 +141,7 @@ def __init__(
         self._calibration_data = []
         self._fp32_print_data = []
         self.data_loader = data_loader
+        self.calib_func = calib_func
         self._check_tf_version()
         self._check_args()
 
@@ -157,6 +160,7 @@ def __init__(
         self._gen_tmp_filenames()
         self._kl_op_dict = {}
         self._kl_keys = []
+        self._llm_weight_minmax = {}
         self._print_node_mapping = {}
         self._enable_kl_op_names = [k for k in self.op_wise_config if self.op_wise_config[k][1] == "kl"]
         self.scale_info = {}
@@ -193,6 +197,14 @@ def _inference(self, model):
         Args:
             model(TensorflowBaseModel): input TensorflowBaseModel
         """
+        if self.calib_func:
+            self.calib_func(model.model)
+            return
+
+        if model.model_type == "llm_saved_model":
+            self._inference_llm(model)
+            return
+
         # ITEX optimization has broken INC calibration process.
         # INC needs turn off ITEX optimization pass in calibration stage.
         # TODO ITEX will provide API to replace setting environment variable.
@@ -281,6 +293,24 @@ def check_shape(tensor, data):
                 break
         os.environ["ITEX_REMAPPER"] = "1"
 
+    def _inference_llm(self, model):
+        input_tensor_names = model.input_tensor_names
+        auto_trackable = model.model
+        infer = auto_trackable.signatures["serving_default"]
+        for idx, (inputs, _) in enumerate(self.data_loader):
+            feed_dict = {}
+            if len(input_tensor_names) == 1:
+                feed_dict[input_tensor_names[0]] = inputs
+            else:
+                assert len(input_tensor_names) == len(inputs), "inputs len must equal with input_tensor"
+                for i, input_tensor_name in enumerate(input_tensor_names):
+                    feed_dict[input_tensor_name] = inputs[i]
+
+            _ = infer(**feed_dict)
+
+            if idx >= self.calib_iteration:
+                break
+
     def _check_tf_version(self):
         """Check if the installed tensorflow version is supported."""
         is_supported_version = False
@@ -849,6 +879,9 @@ def _insert_qdq_pairs(self):
                 self._inference(self._sampling_model)
             self._calibration_data = Helper.gen_valid_sampling_log(tmp_dump_file)
 
+        if hasattr(self._sampling_model, "_weight_tensor_minmax_dict"):
+            self._llm_weight_minmax = self._sampling_model.weight_tensor_minmax_dict
+
         del sampling_graph_def
         del output_tensor_names
         del self._sampling_model
@@ -868,6 +901,7 @@ def _insert_qdq_pairs(self):
             self.device,
             self.performance_only,
             self.itex_mode,
+            self._llm_weight_minmax,
         ).do_transformation()
 
     def _convert_qdq(self):

neural_compressor/adaptor/tf_utils/graph_rewriter/qdq/insert_qdq_pattern.py

@@ -46,6 +46,7 @@ def __init__(
         device,
         performance_only,
         itex_mode,
+        llm_weight_minmax,
     ):
         """Initialization."""
         super().__init__(model)
@@ -58,6 +59,7 @@ def __init__(
         self.device = device
         self.performance_only = performance_only
         self.itex_mode = itex_mode
+        self.llm_weight_minmax = llm_weight_minmax
         self.node_details = namedtuple("node_details", ["node", "output"])
         self.node_name_mapping = {}
         self.check_op_list = {
@@ -548,6 +550,24 @@ def _insert_qdq_pattern_for_weight_node(
                 # qint8_tensor = np.clip(qint8_tensor, -127, 127).astype(np.int8)
                 min_value = -range_value
                 max_value = range_value
+        elif weight_node.op == "ReadVariableOp":
+            min_value = self.llm_weight_minmax[weight_node.name][0]
+            max_value = self.llm_weight_minmax[weight_node.name][1]
+            min_value *= range_coefficent
+            max_value *= range_coefficent
+            min_value = min(min_value, 0.0)
+            if min_value == max_value:
+                if abs(min_value) < 0.000001:
+                    max_value = min_value + 1.0
+                elif min_value > 0:
+                    max_value = 2 * min_value
+                else:
+                    max_value = min_value / 2.0
+            range_value = np.max(np.abs([min_value, max_value]))
+            # qint8_tensor = (np.around(float_tensor * 127.0 / range_value)).astype(np.int8)
+            # qint8_tensor = np.clip(qint8_tensor, -127, 127).astype(np.int8)
+            min_value = -range_value
+            max_value = range_value
         elif host_op_type == "DepthwiseConv2dNative":
             float_tensor = tensor_util.MakeNdarray(weight_node.attr["value"].tensor)
             # get the max values based on dim 0 and 1 for depthwise conv

neural_compressor/adaptor/tf_utils/graph_util.py

@@ -1044,8 +1044,33 @@ def gen_per_iter(data):
                 res.append(mixed_str)
             return res
 
+        def separate(line):
+            """This function is to separate the strings.
+
+            Example:
+                ';slice__print__;__max:[1];slice__print__;__min:[-1]' -->
+                [';slice__print__;__max:[1]', ';slice__print__;__min:[-1]']
+            """
+            separated_lines = []
+            for subline in line.split("];"):
+                if not subline.startswith(";"):
+                    subline = ";" + subline
+                if not subline.endswith("]"):
+                    subline += "]"
+                separated_lines.append(subline)
+            return separated_lines
+
         with open(log_path) as f:
-            valid_data = [i.strip() for i in f.readlines() if i.startswith(";")]
+            valid_data = []
+            for i in f.readlines():
+                if not i.startswith(";"):
+                    continue
+                line = i.strip()
+                if line.find("];") != 0:
+                    separated_lines = separate(line)
+                    valid_data += separated_lines
+                else:
+                    valid_data.append(line)
 
         first_line = valid_data[0].rsplit(":")[0]