Merge remote-tracking branch 'origin/feature/damian/v2/factor_out_transformation_utils' into feature/damian/no_kv_cache

Author: dbogunowicz

src/deepsparse/transformers/helpers.py

@@ -76,6 +76,13 @@ def setup_transformers_pipeline(
         tokenizer.pad_token = tokenizer.eos_token
 
     engine_kwargs = engine_kwargs or {}
+    if engine_kwargs.get("model_path"):
+        raise ValueError(
+            "The engine kwargs already specify "
+            f"a model path: {engine_kwargs['model_path']}, "
+            f"but a model path was also provided: {model_path}. "
+            "Please only provide one."
+        )
     engine_kwargs["model_path"] = model_path
     return model_path, config, tokenizer, engine_kwargs
 
@@ -84,6 +91,7 @@ def setup_onnx_file_path(
     model_path: str,
     sequence_length: int,
     onnx_model_name: Optional[str] = None,
+    task: Optional[str] = None,
 ) -> Tuple[str, transformers.PretrainedConfig, transformers.PreTrainedTokenizer]:
     """
     Parses ONNX model from the `model_path` provided. It additionally
@@ -102,7 +110,9 @@ def setup_onnx_file_path(
     hf_logger_level = hf_logger.level
     hf_logger.setLevel(logging.ERROR)
 
-    config = transformers.PretrainedConfig.from_pretrained(deployment_path)
+    config = transformers.PretrainedConfig.from_pretrained(
+        deployment_path, finetuning_task=task
+    )
     hf_logger.setLevel(hf_logger_level)
 
     trust_remote_code = False

src/deepsparse/transformers/pipelines/pipeline.py

@@ -16,19 +16,18 @@
 Base Pipeline class for transformers inference pipeline
 """
 
-import logging
+
 import warnings
 from pathlib import Path
 from typing import Any, Dict, List, Mapping, Optional, Union
 
 import numpy
 import transformers
-from transformers.models.auto import AutoTokenizer
 
 from deepsparse import Bucketable, Pipeline
+from deepsparse.transformers.helpers import overwrite_transformer_onnx_model_inputs
 from deepsparse.transformers.helpers import (
-    get_deployment_path,
-    overwrite_transformer_onnx_model_inputs,
+    setup_onnx_file_path as setup_onnx_file_path_v2,
 )
 
 
@@ -124,24 +123,15 @@ def setup_onnx_file_path(self) -> str:
 
         :return: file path to the processed ONNX file for the engine to compile
         """
-        deployment_path, onnx_path = get_deployment_path(self.model_path)
-
-        # temporarily set transformers logger to ERROR to avoid
-        # printing misleading warnings
-        hf_logger = logging.getLogger("transformers")
-        hf_logger_level = hf_logger.level
-        hf_logger.setLevel(logging.ERROR)
-        self.config = transformers.PretrainedConfig.from_pretrained(
-            deployment_path,
-            finetuning_task=self.task if hasattr(self, "task") else None,
-        )
-        hf_logger.setLevel(hf_logger_level)
-
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            deployment_path,
-            trust_remote_code=self._trust_remote_code,
-            model_max_length=self.sequence_length,
+        # we will be soon retiring V1 pipelines. This is why I am deciding
+        # to reuse the functions from V2 pipelines in the (soon) legacy pipelines
+        onnx_path, config, tokenizer = setup_onnx_file_path_v2(
+            model_path=self.model_path,
+            sequence_length=self.sequence_length,
+            task=self.task if hasattr(self, "task") else None,
         )
+        self.config = config
+        self.tokenizer = tokenizer
 
         if not self._delay_overwriting_inputs:
             # overwrite onnx graph to given required input shape
@@ -153,12 +143,6 @@ def setup_onnx_file_path(self) -> str:
                 onnx_path, max_length=self.sequence_length
             )
 
-        if not self.config or not self.tokenizer:
-            raise RuntimeError(
-                "Invalid config or tokenizer provided. Please provide "
-                "paths to the files or ensure they exist in the `model_path` provided. "
-                "See `tokenizer` and `config` arguments for details."
-            )
         return onnx_path
 
     def tokens_to_engine_input(

src/deepsparse/transformers/utils/token_generator.py

@@ -77,16 +77,16 @@ def generate(self, logits: numpy.ndarray) -> numpy.ndarray:
         :param logits: the logits from the model with shape (vocab_size,)
         :return: the sampled token
         """
-        if self.deterministic:
-            token = numpy.argmax(logits)
-            self.tokens.append(token)
-            return token
-
         if self.top_k:
             logits = self.apply_top_k(logits)
         if self.top_p:
             logits = self.apply_top_p(logits)
 
+        if self.deterministic:
+            token = numpy.argmax(logits)
+            self.tokens.append(token)
+            return token
+
         if self.sampling_temperature != 1.0:
             logits /= self.sampling_temperature
 

src/deepsparse/v2/text_generation/join_output.py

@@ -33,9 +33,6 @@ def __init__(self, tokenizer):
         self.tokenizer = tokenizer
 
     def run(self, inp: List[CompileGenerationsOutput], **kwargs):
-
-        if not isinstance(inp, list):
-            inp = [[inp]]
         batch_outputs = [x for x in inp[0]]
         generated_tokens = [x.generated_tokens for x in batch_outputs]
         generated_logits = [x.generated_logits for x in batch_outputs]

src/deepsparse/v2/text_generation/prep_for_generation.py

@@ -91,7 +91,6 @@ def run(
             "token_generator": token_generator,
         }
         output = {
-            "logits": prompt_logits,
             "tokens": token_generator.tokens,
             "kv_cache": kv_cache,
             "in_generation": True,

tests/deepsparse/v2/unit/text_generation/conftest.py

@@ -25,7 +25,7 @@
 from deepsparse.v2 import InferenceState, PipelineState
 from deepsparse.v2.text_generation import (
     GenerationDefaults,
-    NlEngineOperator,
+    NLEngineOperator,
     TokenGeneratorOperator,
 )
 
@@ -61,7 +61,7 @@ def single_token_engine_no_internal_cache(text_generation_attributes, model_attr
     seq_length, _ = text_generation_attributes
     _, model_path = model_attributes
 
-    nl_engine_operator = NlEngineOperator(
+    nl_engine_operator = NLEngineOperator(
         sequence_length=seq_length, input_ids_length=1, model_path=model_path
     )
     return nl_engine_operator