Merge pull request #169 from BrandonGroth/quant_refactor_perCh

new functions are isolated from current functions, safe to merge. But we will need a new example to demo how to use the new `_rc` functions.

Author: chichun-charlie-liu

.gitignore

@@ -17,6 +17,9 @@ qcfg.json
 configs
 pytest.out
 
+# Log file
+fms_mo.log
+
 # IDEs
 .vscode/
 .idea/

fms_mo/quant/quantizers.py

@@ -513,9 +513,12 @@ def forward(
 
         if istraining:
             # only recalc clipvals under training mode
+            num_bits_int = (
+                num_bits.item() if isinstance(num_bits, torch.Tensor) else num_bits
+            )
             SAWBcode_mapping = {8: 803, 4: 403, 2: 103}
             if num_bits in [2, 4, 8]:
-                sawb_code = SAWBcode_mapping[num_bits]
+                sawb_code = SAWBcode_mapping[num_bits_int]
                 clip_val, _ = sawb_params_code(
                     num_bits, sawb_code, input_tensor, perCh=True
                 )
@@ -552,9 +555,13 @@ def forward(
                 clip_val.dtype
             )  # NOTE return will be a fp32 tensor; function only support float()
         else:
-            output = torch.quantize_per_channel(
-                input_tensor, scale, zero_point, 0, torch.qint8
-            ).int_repr()
+            output = (
+                torch.quantize_per_channel(
+                    input_tensor, scale, zero_point, 0, torch.qint8
+                )
+                .int_repr()
+                .clamp(int_l, int_u)
+            )
             # NOTE return will be a torch.int8 tensor
 
         return output
@@ -2540,7 +2547,7 @@ def asymmetric_linear_quantization_params(
         return scale, zero_point
 
 
-def clamp(input_tensor: torch.FloatTensor, clamp_min, clamp_max, inplace=False):
+def clamp(input_tensor: torch.Tensor, clamp_min, clamp_max, inplace=False):
     """
     Returns:
         Clamped Torch Tensor.

fms_mo/quant_refactor/base_model.py

@@ -0,0 +1,114 @@
+# Copyright The FMS Model Optimizer Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Base QModel Class
+"""
+
+# Standard
+# pylint: disable=keyword-arg-before-vararg
+import logging
+
+# Third Party
+import torch
+
+logger = logging.getLogger(__name__)
+
+
+class Qmodel:  # do not inherit nn.Module, or self.model will not show up in __dict__
+    """
+    A wrapper for fms_mo model, mainly for user API simplification purpose.
+    Everything is the same as the original model, but we can add new member functions.
+    Make sure the naming will be unique enough so that we won't override any existing functions
+    in Huggingface models.
+    """
+
+    def __init__(self, original_model, qcfg=None, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.org_attr = dir(original_model)
+        self.model = original_model
+        if qcfg:
+            self.qcfg = qcfg
+
+    def __getattr__(self, name: str):
+        if name in self.org_attr:
+            logger.info(f"Trying to access self.{name}, forward to self.model.{name}")
+            return getattr(self.model, name)
+            # NOTE: this self.model is in __dict__, so it will not trigger __getattr__
+            # recursively.!!
+
+    def __call__(self, *args, **kwargs):
+        logger.info(
+            "Make this object callable, but actually just calling self.model.__call__()"
+        )
+        return self.model(*args, **kwargs)
+
+    def __repr__(self):
+        OKCYAN = "\033[96m"
+        ENDC = "\033[0m"
+        rep_txt = f"{OKCYAN}FMSMO_Qmodel_wrapper({ENDC}\n{self.model.__repr__()}{OKCYAN}){ENDC}"
+        return rep_txt
+
+    def to(self, tar_dev: torch.device):
+        """
+        Demonstrate that we can override a function in original model
+        it should not call __getattr__(), i.e. will not see the printout from that func
+
+        Args:
+            tar_dev (torch.device): A new device
+
+        Returns:
+            Qmodel: Moved model to tar_dev
+        """
+        logger.info(
+            f"Override a function in original model. moving the model to a new device {tar_dev}"
+        )
+        return self.model.to(tar_dev)
+
+    def save_model_in_pt_fmt(
+        self, filename: str = "model.pt", exam_inp: torch.Tensor = None
+    ):
+        """
+        Save entire model to file
+
+        Args:
+            filename (str, optional): File path to save model. Defaults to "model.pt".
+            exam_inp (torch.Tensor, optional): Example input for model. Defaults to None.
+        """
+        # NOTE self.qcfg has a lot of info already, like transformers_version
+        # NOTE cannot save wrapped self, can only save self.model...
+        save_dict = {"model": self.model}
+        if exam_inp:
+            save_dict["exam_inp"] = exam_inp
+        torch.save(save_dict, filename)
+        logger.info(f"{filename} saved successfully.")
+
+    def save_statedict_in_pt_fmt(
+        self,
+        filename: str = "model.pt",
+    ):
+        """
+        Save the model state dict to file
+
+        Args:
+            filename (str, optional): File path to save model state dict. Defaults to "model.pt".
+        """
+        torch.save(self.model.state_dict(), filename)
+        logger.info(f"model.state_dict() is saved to {filename} successfully.")
+
+    def run_gptq(self):
+        """
+        Check model is supported by AutoGPTQ first
+        """
+        return