initial implementation

Author: dbogunowicz

src/deepsparse/transformers/utils/__init__.py

@@ -12,5 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .decoder_kv_cache import *
+
 # flake8: noqa
 from .kv_cache_ort import *

src/deepsparse/transformers/utils/decoder_kv_cache.py

@@ -0,0 +1,111 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# 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.
+
+from typing import Any, Dict, Optional
+
+from deepsparse.pipeline import SUPPORTED_PIPELINE_ENGINES
+from deepsparse.transformers.utils.kv_cache_ort import KVCacheORT
+
+
+__all__ = ["DecoderKVCache"]
+
+
+class DecoderKVCache:
+    def __init__(self, engine_type: str):
+        """
+        The goal of DecoderKVCache is to provide a common
+        interface for the KVCache objects used
+        by the NLDecoderEngine
+
+        :param engine_type: The engine type to use for the decoder
+        """
+        if engine_type not in SUPPORTED_PIPELINE_ENGINES:
+            raise ValueError(f"Unsupported engine type: {engine_type}")
+        elif engine_type != "onnxruntime":
+            raise NotImplementedError(f"Unsupported engine type: {engine_type}")
+        self._kv_cache_type = KVCacheORT
+
+        self._kv_cache = None
+        self._session_id = None
+        self._frozen_position = None
+        self._num_tokens = None
+
+    def setup_session(
+        self,
+        session_id: str,
+        state: Dict[str, Any],
+        num_tokens: int,
+        frozen_position=Optional[int],
+    ):
+        """
+        Setup the session that will be used to transform
+        the input and output cache values
+
+        :param session_id: The session id to use for the current
+            session
+        :param state: The state of the cache. This is a dictionary
+            that maps the name of the cache array to the cache array.
+            The cache tensor is a numpy array of shape
+            [batch_size, num_heads, sequence_length, hidden_size]
+        :param num_tokens: The number of tokens processed so far,
+            corresponding to the number of "non-blank" entries in the
+            kv cache array.
+        :param frozen_position: The position along the sequence length axis
+            that is frozen and thus, once it is occupied by a "non-blank"
+            cache entry, it cannot be removed from the cache.
+        """
+        self.session_id = session_id
+        self._num_tokens = num_tokens
+        self._frozen_position = frozen_position
+        self._initialize_kv_cache(state)
+
+    def update_session(self, state: Dict[str, Any]):
+        """
+        Update the session with the new state of the cache
+
+        :param state: The state of the cache. This is a dictionary
+            that maps the name of the cache array to the cache array.
+            The cache tensor is a numpy array of shape
+            [batch_size, num_heads, sequence_length, hidden_size]
+        """
+        self._num_tokens += 1
+        self._initialize_kv_cache(state)
+
+    @property
+    def session_id(self):
+        if self._session_id is None:
+            raise ValueError("Attempted to access session_id before setting up session")
+        return self._session_id
+
+    @property
+    def cached_inputs(self):
+        if self._kv_cache is None:
+            raise ValueError(
+                "Attempted to access cached inputs before setting up session"
+            )
+        # TODO: Not sure whether this is the appropriate place
+        # to invoke the shift_last method, to reconsider
+        self._kv_cache.shift_last()
+        return self._kv_cache.state
+
+    @session_id.setter
+    def session_id(self, session_id: str):
+        self._session_id = session_id
+
+    def _initialize_kv_cache(self, state: Dict[str, Any]):
+        self._kv_cache = KVCacheORT(
+            state=state,
+            num_tokens=self._num_tokens,
+            frozen_position=self._frozen_position,
+        )

tests/deepsparse/transformers/utils/test_decoder_kv_cache.py

@@ -0,0 +1,47 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# 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.
+
+import numpy as np
+
+import pytest
+from deepsparse.transformers.utils import DecoderKVCache
+
+
+@pytest.mark.parametrize(
+    "state, num_tokens, state_shifted, new_state, new_state_shifted",
+    [
+        (
+            {"dummy_cache_name": np.array([[[[0], [0], [1], [2], [3]]]])},
+            3,
+            {"dummy_cache_name": np.array([[[[0], [1], [2], [3]]]])},
+            {"dummy_cache_name": np.array([[[[0], [1], [2], [3], [4]]]])},
+            {"dummy_cache_name": np.array([[[[1], [2], [3], [4]]]])},
+        ),
+    ],
+)
+def test_kv_cache_ort_shift(
+    state, num_tokens, state_shifted, new_state, new_state_shifted
+):
+    decoder_kv_cache = DecoderKVCache(engine_type="onnxruntime")
+    decoder_kv_cache.setup_session(
+        session_id="some_id", state=state, num_tokens=num_tokens
+    )
+    cache = decoder_kv_cache.cached_inputs
+    for k, v in cache.items():
+        assert np.array_equal(v, state_shifted[k])
+
+    decoder_kv_cache.update_session(state=new_state)
+    cache = decoder_kv_cache.cached_inputs
+    for k, v in cache.items():
+        assert np.array_equal(v, new_state_shifted[k])