fix NotImplementedError: get_type is not implemented (#2133)

Signed-off-by: Xin He <xinhe3@habana.ai>

Author: xin3he

neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py

@@ -591,6 +591,13 @@ def __init__(self, mod, parent, mod_extra_config, *args, **kwargs):
             setattr(self, "w2_weight", None)
         self.forward = self.forward_orig
 
+    def extra_repr(self) -> str:
+        return extra_representation(
+            self.extra_repr_org(),
+            self.class_name_org,
+            get_current_repr(self),
+        )
+
 
 # This patched module is called by the vllm-mixtral FusedMoE layer
 # we wrap each expert weight with this module since FusedMoE has a single tensor for all experts weights
@@ -853,6 +860,13 @@ def update_measure(self, prev, cur, dim, idx, inp_seq_len):
         measure_output((output,), self._mod_extra_config.outputs)
         return output
 
+    def extra_repr(self) -> str:
+        return extra_representation(
+            self.extra_repr_org(),
+            self.class_name_org,
+            get_current_repr(self),
+        )
+
 
 class PatchedVLLMKVCache(PatchedModuleBase):
     # Module to patch VLLMKVCache module from llama model
@@ -891,6 +905,14 @@ def fetch_from_cache(self, cache, blocks, permutations=None):
         output_cache = self.orig_fetch_from_cache(quant_cache, blocks)
         return self.dequant_output(output_cache)
 
+    def extra_repr(self) -> str:
+        return extra_representation(
+            self.extra_repr_org(),
+            self.class_name_org,
+            get_current_repr(self),
+        )
+
+
 def init_conv(instance, mod_extra_config):
     if instance.quantization_mode in [QuantMode.QUANTIZE, QuantMode.LOAD]:
         instance.quant_input = instance._mod_extra_config.inputs[0]

neural_compressor/torch/algorithms/fp8_quant/patched_module_base.py

@@ -175,32 +175,34 @@ def forward_quant(self, *args, **kwargs):
 
     @classmethod
     def get_module_info(cls) -> ModuleInfo:
-        """Return the module info for the module.
+        """Only necessary for the newly registered patched module that doesn't in _mod_default_dict.
+        Return the module info for the module, which is used to determine the scaling methods for the module.
 
         For example, for linear module, the module info is: ModuleInfo(type="linear", patched_module=cls).
         """
         return ModuleInfo(type=cls.get_type(), patched_module=cls)
 
     @classmethod
-    @abstractmethod
     def get_type(cls) -> str:
-        """Return the type of the patched module.
+        """Only necessary for the newly registered patched module that doesn't in _mod_default_dict.
+        Return the type of the patched module, which is used to determine the scaling methods for the module.
 
         Multiple patched modules can have the same type, and share the same scaling methods.
         """
         raise NotImplementedError("`get_type` is not implemented")
 
     @classmethod
-    @abstractmethod
     def get_module_type(cls) -> ModuleType:
-        """Return the module type for the module.
+        """Only necessary for the newly registered patched module that doesn't in _mod_default_dict.
+        Return the module type for the module, which is used to determine the number of inputs, outputs, and parameters of the module.
 
         The module type is used to determine the number of inputs, outputs, and parameters of the module.
         For example, for linear module, the module type is: ModuleType(1, ["weight"], 1, False).
         """
         raise NotImplementedError("`get_module_type` is not implemented")
 
     def extra_repr(self):
+        """This extra_repr is only for the newly registered patched module that doesn't in _mod_default_dict."""
         return  f"quantization_mode={self.quantization_mode}, " + \
                 f"module_info={self.get_module_info()}, " + \
                 f"module_type={self.get_module_type()}"