[Text Generation] Automatically benchmark in auto-regressive setting (#…

…1142)

* initial commit

* improve logging docstring

* more verbose logging

* add sequence_length as variable

* fixed type annotations and avoided overwriting inputs when no sequence_length is passed

* fix bad merge

* tested

* update defaults

* address Luka comments

---------

Co-authored-by: Luka Govedic <luka.govedic@gmail.com>

src/deepsparse/benchmark/benchmark_model.py

@@ -81,6 +81,14 @@
    zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/base-none \
    --input_shapes "[1,512],[1,512],[1,512]"
 
+##########
+Example on a CodeGen (model with KV cache support)
+from SparseZoo with input_ids_length 10 and sequence length 256:
+deepsparse.benchmark \
+   zoo:nlg/text_generation/codegen_mono-350m/pytorch/
+   huggingface/bigpython_bigquery_thepile/pruned50-none
+   --input_ids_length 10 --sequence_length 256
+
 ##########
 Example on local ONNX model:
 deepsparse.benchmark /PATH/TO/model.onnx
@@ -110,8 +118,10 @@
 from deepsparse.log import set_logging_level
 from deepsparse.utils import (
     generate_random_inputs,
+    has_model_kv_cache,
     model_to_path,
     override_onnx_input_shapes,
+    overwrite_cache_model_inputs,
     parse_input_shapes,
 )
 
@@ -143,6 +153,26 @@ def parse_args():
         default=1,
         help="The batch size to run the analysis for. Must be greater than 0",
     )
+
+    parser.add_argument(
+        "-seq_len",
+        "--sequence_length",
+        type=int,
+        default=2048,
+        help="The sequence length to run the "
+        "KV cache supported model benchmarks for. "
+        "Must be greater than 0, default is 2048",
+    )
+
+    parser.add_argument(
+        "-input_ids_len",
+        "--input_ids_length",
+        type=int,
+        default=1,
+        help="The input ids length to run the "
+        "KV cache supported model benchmarks for. "
+        "Must be greater than 0, default is 1",
+    )
     parser.add_argument(
         "-i",
         "-shapes",
@@ -265,6 +295,8 @@ def load_custom_engine(custom_engine_identifier: str):
 def benchmark_model(
     model_path: str,
     batch_size: int = 1,
+    sequence_length: int = 2048,
+    input_ids_length: int = 1,
     input_shapes: str = "",
     num_cores: int = None,
     scenario: str = "sync",
@@ -290,6 +322,28 @@ def benchmark_model(
 
     orig_model_path = model_path
     model_path = model_to_path(model_path)
+
+    if has_model_kv_cache(model_path):
+        if batch_size != 1:
+            raise ValueError(
+                "Unable to run models with KV cache support "
+                "for batch size different than one."
+                "Please set batch size to 1 and try again"
+            )
+
+        _LOGGER.info(
+            "Found model with KV cache support. "
+            "Benchmarking the autoregressive model with "
+            f"input_ids_length: {input_ids_length} and "
+            f"sequence length: {sequence_length}."
+        )
+
+        model_path, _, _ = overwrite_cache_model_inputs(
+            model_path=model_path,
+            input_ids_length=input_ids_length,
+            sequence_length=sequence_length,
+        )
+
     num_streams = parse_num_streams(num_streams, num_cores, scenario)
 
     # Compile the ONNX into a runnable model
@@ -351,6 +405,8 @@ def benchmark_model(
         "orig_model_path": orig_model_path,
         "model_path": model_path,
         "batch_size": batch_size,
+        "sequence_length": sequence_length,
+        "input_ids_length": input_ids_length,
         "input_shapes": input_shapes,
         "num_cores": num_cores,
         "scenario": scenario,
@@ -376,6 +432,8 @@ def main():
 
     result = benchmark_model(
         model_path=args.model_path,
+        sequence_length=args.sequence_length,
+        input_ids_length=args.input_ids_length,
         batch_size=args.batch_size,
         input_shapes=args.input_shapes,
         num_cores=args.num_cores,
@@ -392,6 +450,10 @@ def main():
     # Results summary
     print("Original Model Path: {}".format(args.model_path))
     print("Batch Size: {}".format(args.batch_size))
+    if args.sequence_length is not None:
+        print("Sequence Length: {}".format(args.sequence_length))
+    if args.input_ids_length is not None:
+        print("Input IDs Length: {}".format(args.input_ids_length))
     print("Scenario: {}".format(args.scenario))
     print(
         "Throughput (items/sec): {:.4f}".format(

src/deepsparse/transformers/engines/nl_decoder_engine.py

@@ -22,17 +22,16 @@
 from deepsparse.transformers.utils.decoder_kv_cache import DecoderKVCache
 from deepsparse.transformers.utils.helpers import (
     generate_session_id,
-    overwrite_onnx_model_inputs,
+    overwrite_onnx_model_inputs_for_kv_cache_models,
 )
 from deepsparse.utils.data import numpy_softmax
+from deepsparse.utils.onnx import CACHE_INPUT_PREFIX, CACHE_OUTPUT_PREFIX
 
 
 _LOGGER = logging.getLogger(__name__)
 
 __all__ = ["NLDecoderEngine"]
 
-_CACHE_INPUT_NAME = "past_key_values"
-
 
 class NLDecoderEngine:
     """
@@ -69,17 +68,17 @@ def __init__(
     ):
         # flag to indicate if the model is quantized or not
         self.kv_cache_data_type = None
-
         (
             onnx_file_path,
             output_indices_to_be_cached,
             kv_cache_data_type,
-        ) = overwrite_onnx_model_inputs(
+        ) = overwrite_onnx_model_inputs_for_kv_cache_models(
             onnx_file_path=onnx_file_path,
             batch_size=engine_args.get("batch_size", 1),
             sequence_length=sequence_length,
             input_ids_length=input_ids_length,
         )
+
         kv_cache_enabled = False
         if sum(output_indices_to_be_cached):
             kv_cache_enabled = True
@@ -129,7 +128,7 @@ def onnx_input_names_no_cache(self) -> List[str]:
         return [
             name
             for name in self.engine.input_names
-            if not name.startswith(_CACHE_INPUT_NAME)
+            if not name.startswith(CACHE_INPUT_PREFIX)
         ]
 
     @property
@@ -284,7 +283,7 @@ def update_kv_cache(
         cache_onnx_names = [
             name
             for name in self.engine.input_names
-            if name.startswith(_CACHE_INPUT_NAME)
+            if name.startswith(CACHE_INPUT_PREFIX)
         ]
         kv_cache_state = {
             name: array for name, array in zip(cache_onnx_names, kv_cache_state)
@@ -302,7 +301,7 @@ def _initialize_kv_cache_state(self, length: int) -> Dict[str, numpy.ndarray]:
         cache_engine_input_index = next(
             i
             for i, name in enumerate(self.engine.input_names)
-            if _CACHE_INPUT_NAME in name
+            if CACHE_INPUT_PREFIX in name
         )
         batch_size, num_attention_heads, _, hidden_dims = self.engine.input_shapes[
             cache_engine_input_index
@@ -314,9 +313,9 @@ def _initialize_kv_cache_state(self, length: int) -> Dict[str, numpy.ndarray]:
         )
 
         cache_keys = [
-            output_name.replace("present", _CACHE_INPUT_NAME)
+            output_name.replace(CACHE_OUTPUT_PREFIX, CACHE_INPUT_PREFIX)
             for output_name in self.engine.output_names
-            if output_name.startswith("present")
+            if output_name.startswith(CACHE_OUTPUT_PREFIX)
         ]
         return {key: empty_kv_cache_tensor for key in cache_keys}
 

src/deepsparse/transformers/utils/helpers.py

@@ -11,10 +11,9 @@
 # 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 logging
 import uuid
-from typing import List, Tuple, Union
+from typing import List, Optional, Tuple, Union
 
 import numpy
 import onnx
@@ -24,21 +23,21 @@
 
 
 __all__ = [
+    "overwrite_onnx_model_inputs_for_kv_cache_models",
     "generate_session_id",
     "pad_to_fixed_length",
     "create_causal_mask",
-    "overwrite_onnx_model_inputs",
 ]
 
 _LOGGER = logging.getLogger(__name__)
 
 
-def overwrite_onnx_model_inputs(
+def overwrite_onnx_model_inputs_for_kv_cache_models(
     onnx_file_path: str,
     sequence_length: int,
     input_ids_length: int,
     batch_size: int = 1,
-) -> Tuple[str, List[int]]:
+) -> Tuple[str, List[int], Optional[int]]:
     """
     Enforces the appropriate input shapes for the onnx model, as well as
     checks whether kv cache is enabled or not.

src/deepsparse/utils/onnx.py

@@ -50,6 +50,8 @@
     "truncate_onnx_model",
     "truncate_onnx_embedding_model",
     "default_cached_outputs",
+    "has_model_kv_cache",
+    "overwrite_cache_model_inputs",
     "CACHE_INPUT_PREFIX",
     "CACHE_OUTPUT_PREFIX",
 ]
@@ -494,3 +496,56 @@ def default_cached_outputs(model_path: str) -> List[bool]:
     assert len(output_names) > 0
 
     return [name.startswith(CACHE_OUTPUT_PREFIX) for name in output_names]
+
+
+def has_model_kv_cache(model: Union[str, ModelProto]) -> bool:
+    """
+    Check whether a model has a KV cache support.
+
+    :param model_path: Path to a model or a model proto.
+    :return True if the model has a KV cache support, False otherwise.
+    """
+    return bool(any(default_cached_outputs(model)))
+
+
+def overwrite_cache_model_inputs(
+    model_path: str,
+    input_ids_length: int,
+    sequence_length: int,
+) -> Tuple[str, List[int], Optional[int]]:
+    """
+    Takes a path to an onnx model and enforces that it has
+    static input dimensions.
+
+    :param model_path: Path to a model.
+    :param input_ids_length: The input_ids length to overwrite the model with.
+    :param sequence_length: The sequence length to overwrite the model with.
+    :return: A tuple that contains:
+        -   the path to the onnx model file that has been overwritten
+            with the new input shapes
+        -   boolean list, where elements are set to True if the
+            corresponding model output should be cached or False
+            if not.
+        -   the data type of the kv cache. If the model does not
+            use kv cache, then the data type is None
+    """
+    from deepsparse.transformers.utils.helpers import (
+        overwrite_onnx_model_inputs_for_kv_cache_models,
+    )
+
+    assert input_ids_length < sequence_length, (
+        f"input_ids_length {input_ids_length} "
+        f"must be less than sequence_length {sequence_length}"
+    )
+
+    (
+        onnx_file_path,
+        output_indices_to_be_cached,
+        kv_cache_data_type,
+    ) = overwrite_onnx_model_inputs_for_kv_cache_models(
+        onnx_file_path=model_path,
+        sequence_length=sequence_length,
+        input_ids_length=input_ids_length,
+    )
+
+    return onnx_file_path, output_indices_to_be_cached, kv_cache_data_type

tests/deepsparse/transformers/pipelines/test_text_generation.py

@@ -22,7 +22,7 @@
 from deepsparse import Pipeline
 from deepsparse.transformers.utils.helpers import (
     create_causal_mask,
-    overwrite_onnx_model_inputs,
+    overwrite_onnx_model_inputs_for_kv_cache_models,
 )
 from deepsparse.utils.onnx import CACHE_INPUT_PREFIX
 from sparsezoo import Model
@@ -216,7 +216,7 @@ def _get_cache_state_ort_kv_cache(model_onnx_path, sequence, model_name):
 
         # setup model and session
         # (run full sequence inference)
-        overwrite_onnx_model_inputs(
+        overwrite_onnx_model_inputs_for_kv_cache_models(
             model_onnx_path, sequence_length=128, input_ids_length=128
         )
         sess = onnxruntime.InferenceSession(model_onnx_path)