[Text Generation] Optimize the slow update method in the KVCacheDec…

…oder (#1190)

* initial commit

* Nit: docstring typo

* fix style

src/deepsparse/transformers/utils/decoder_kv_cache.py

@@ -91,7 +91,7 @@ def update(
     ):
         """
         Updating the session is identical with taking the kv cache
-        output of from the forward pass and restructuring it, so it
+        output from the forward pass and restructuring it, so it
         can be directly used as input for the next forward pass.
 
         :param state: The state of the cache. This is a dictionary
@@ -103,79 +103,78 @@ def update(
             Corresponds to `input_ids.shape[1]`
         """
         self.total_num_processed_tokens += input_ids_len
-        total_cache_capacity = state[list(state.keys())[0]].shape[
+
+        input_state_capacity = state[list(state.keys())[0]].shape[
             self._sequence_len_axis
         ]
-        # total_capacity = num_tokens (num of non-blank tokens) +
-        # + num_padded_entries (num of blank tokens)
+        num_entries_to_delete = input_ids_len
+
+        # compute the number of blank (padded) entries in the cache
         num_padded_entries = max(
-            0, total_cache_capacity - self.total_num_processed_tokens
+            0, input_state_capacity - self.total_num_processed_tokens
         )
-        num_entries_to_delete = input_ids_len
+        # compute how many of those entries need to be deleted
+        num_padded_entries_to_delete = min(num_padded_entries, num_entries_to_delete)
 
-        if num_padded_entries:
-            """
-            Transforms input KV cache that contains blank entries.
-            It removes the rightmost blank entries from the cache.
-            Example 1:
-            (entries in the cache denote the order in which they were
-            added to the cache, zero is to denote a blank entry)
-            ```
-            state["state_name"]: (1, 1, 5, 1) = array([[[[0], [0], [1], [2], [3]]]])
-            -> num_padded_entries = 2
-            -> num_entries_to_delete = 1
-            -> num_padded_entries > num_entries_to_delete
-            # there are more blank entries than entries to delete
-            results in:
-            state["state_name"]: (1, 1, 4, 1) = array([[[[0], [1], [2], [3]]]])
-            ```
-            Example 2:
-            ```
-            state["state_name"]: (1, 1, 6, 1) = array([[[[0], [0], [0], [1], [2], [3]]]]) # noqa: E501
-            -> num_padded_entries = 3
-            -> num_entries_to_delete = 5
-            -> num_padded_entries < num_entries_to_delete
-            # there are less blank entries than entries to delete
-            results in:
-            state["state_name"]: (1, 1, 3, 1) = array([[[[1], [2], [3]]]])
-            ```
-            """
-            num_padded_entries_to_delete = min(
-                num_padded_entries, num_entries_to_delete
-            )
-            idxs_to_remove = [
-                num_padded_entries - i - 1 for i in range(num_padded_entries_to_delete)
-            ]
-            # if we had fewer blank entries than entries to delete,
-            # we updated the number of entries to delete to a non-zero value
-            num_entries_to_delete = max(0, num_entries_to_delete - num_padded_entries)
-            # update the state of the cache
-            state = self._delete_entries(state, idxs_to_remove)
-
-        if num_entries_to_delete:
-            """
-            Transforms the input KV cache that has been totally
-            filled with non-blank entries.
-            Example:
-            ```
-            state["state_name"]: (1, 1, 5, 1) = array([[[[1], [2], [3], [4], [5]]]])
-            num_entries_to_delete = 2
-            if self.freeze_first_position == False:
-                state["state_name"]: (1, 1, 3, 1) = array([[[[3], [4], [5]]]])
-            else:
-
-                state["state_name"]: (1, 1, 3, 1) = array([[[[1], [4], [5]]]])
-            ```
-            """
-            idxs_to_remove = [
-                i + int(self._freeze_first_position)
-                for i in range(num_entries_to_delete)
-            ]
-
-            state = self._delete_entries(state, idxs_to_remove)
+        # if we had fewer padded entries than num_entries_to_delete,
+        # we additionally are forced to delete some non-padded entries (the oldest ones)
+        num_non_padded_entries_to_delete = max(
+            0, num_entries_to_delete - num_padded_entries
+        )
+
+        for name, cache_array in state.items():
+            if num_padded_entries_to_delete:
+                cache_array = self.remove_padded_entries(
+                    cache_array, num_padded_entries_to_delete
+                )
+            if num_non_padded_entries_to_delete:
+                cache_array = self.remove_non_padded_entries(
+                    cache_array, num_entries_to_delete
+                )
+            state[name] = numpy.ascontiguousarray(cache_array)
 
         self._state = state
 
+    def remove_padded_entries(
+        self, cache_array: numpy.ndarray, num_padded_entries_to_delete: int
+    ):
+        """
+        Remove the num_padded_entries_to_delete entries from the cache array.
+        This function assumes that the cache_array has the number
+        of padded (blank) entries that is equal/larger than
+        num_padded_entries_to_delete.
+
+        :param cache_array: The cache array to be modified.
+        :param num_padded_entries_to_delete: The number of padded entries to delete.
+        """
+        return cache_array[:, :, num_padded_entries_to_delete:, :]
+
+    def remove_non_padded_entries(
+        self, cache_array: numpy.ndarray, num_non_padded_entries_to_delete: int
+    ):
+        """
+        Remove the num_non_padded_entries_to_delete entries from the cache array.
+        This function assumes that the cache_array has no padded (blank) entries and
+        thus we are forced to delete the oldest entries from the cache.
+
+        If self._freeze_first_position is set to True, that means that the oldest
+        entry in the cache_array is the one that corresponds to the BOS token. Because
+        we want to keep that entry in the cache, we will delete the oldest entry
+        starting from the second oldest entry.
+        """
+        new_cache_array = cache_array[
+            :,
+            :,
+            bool(self._freeze_first_position) + num_non_padded_entries_to_delete :,
+            :,
+        ]
+        if self._freeze_first_position:
+            bos_entries = cache_array[:, :, :1, :]
+            new_cache_array = numpy.concatenate(
+                [bos_entries, new_cache_array], axis=self._sequence_len_axis
+            )
+        return new_cache_array
+
     def set_capacity(self, capacity: int):
         """
         Enforce a new total capacity for the state
@@ -212,14 +211,6 @@ def set_capacity(self, capacity: int):
 
         self._state = state
 
-    def _delete_entries(
-        self, state: Dict[str, Any], indices: List[int]
-    ) -> Dict[str, Any]:
-        for key, value in state.items():
-            state[key] = numpy.delete(value, indices, axis=self._sequence_len_axis)
-            state[key] = numpy.ascontiguousarray(state[key])
-        return state
-
     def _add_entries(
         self, state: Dict[str, Any], indices: List[int], padding_value: int = 0
     ) -> Dict[str, Any]: