[TextGeneration] Samling arguments for generation

src/deepsparse/pipeline.py

@@ -232,7 +232,7 @@ def __call__(self, *args, **kwargs) -> BaseModel:
                     f"Inputs parsed to {type(pipeline_inputs)}"
                 )
             # batch size of the inputs may be `> self._batch_size` at this point
-            engine_inputs: List[numpy.ndarray] = self.process_inputs(pipeline_inputs)
+            engine_inputs = self.process_inputs(pipeline_inputs)
             if isinstance(engine_inputs, tuple):
                 engine_inputs, context = engine_inputs
             else:
@@ -494,7 +494,9 @@ def split_engine_inputs(
         return split_engine_inputs(items, batch_size)
 
     def engine_forward(
-        self, engine_inputs: List[numpy.ndarray], context: Dict = {}
+        self,
+        engine_inputs: List[numpy.ndarray],
+        context: Dict = {},
     ) -> List[numpy.ndarray]:
         """
         :param engine_inputs: list of numpy inputs to Pipeline engine forward

src/deepsparse/transformers/engines/nl_decoder_engine.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional
 
 import numpy
 from transformers import AutoTokenizer
@@ -23,7 +23,6 @@
 from deepsparse.transformers.utils.helpers import generate_session_id
 from deepsparse.transformers.utils.timings import TextGenerationTimings
 from deepsparse.utils import TimerManager
-from deepsparse.utils.data import numpy_softmax
 from deepsparse.utils.onnx import (
     CACHE_INPUT_PREFIX,
     CACHE_OUTPUT_PREFIX,
@@ -179,7 +178,7 @@ def __call__(
         self,
         inp: List[numpy.ndarray],
         val_inp: bool = True,
-    ) -> Tuple[numpy.ndarray, numpy.ndarray]:
+    ) -> numpy.ndarray:
         """
         The main entry point for running the engine.
 
@@ -207,10 +206,7 @@ def __call__(
         else:
             logits = out[0]
 
-        # select batch idx 0, batch is always 1
-        token = self.generate_token(logits=logits[0, -1, :])
-
-        return token, logits
+        return logits
 
     def __str__(self):
         return f"{self.__class__.__name__}: {self.engine}"
@@ -233,22 +229,6 @@ def transfer_cache_state(self, cache: DecoderKVCache):
         cache.set_capacity(self.cache_length)
         self.kv_cache = cache
 
-    def generate_token(self, logits: numpy.ndarray) -> numpy.ndarray:
-        """
-        Samples a token from the logits using the sampling temperature.
-
-        :param logits: the logits from the model with shape (vocab_size,)
-        :return: the sampled token
-        """
-        if self.deterministic:
-            return numpy.argmax(logits)
-
-        logits /= self.sampling_temperature
-
-        probs = numpy_softmax(logits)
-
-        return numpy.random.choice(len(probs), p=probs)
-
     def reset_kv_cache(self):
         """
         Resets the kv cache state.

src/deepsparse/transformers/pipelines/text_generation.py

@@ -42,6 +42,7 @@
     pad_to_fixed_length,
 )
 from deepsparse.transformers.utils.timings import TextGenerationTimings
+from deepsparse.transformers.utils.token_generator import TokenGenerator
 from deepsparse.utils.onnx import default_cached_outputs
 
 
@@ -107,6 +108,29 @@ class Config:
         " tokens is generated). Set to `None` to ignore this parameter."
         " Default is `None`.",
     )
+    top_p: Optional[float] = Field(
+        default=0.0,
+        description="Used for filtering generated tokens. Keep the"
+        " tokens where its cumulative probability is >= top_p"
+        " Default set to 0.0",
+    )
+    top_k: Optional[int] = Field(
+        default=0,
+        description="Used for filtering generated tokens. Keep"
+        " top_k generated tokens. Default set to 0",
+    )
+    presence_penalty: Optional[float] = Field(
+        default=0.0,
+        description="Penalty applied for generating new token. Any existing"
+        " token results in the subtraction of its corresponding logit value."
+        " Default set to 0.0",
+    )
+    frequency_penalty: Optional[float] = Field(
+        default=0.0,
+        description="Penalty applied for generating new token. Existing"
+        " token frequencies summed to subtraction the logit of its"
+        " corresponding logit value. Default set to 0.0.",
+    )
 
 
 class TextGenerationOutput(BaseModel):
@@ -420,7 +444,12 @@ def process_inputs(self, inputs: TextGenerationInput) -> List[numpy.ndarray]:
             include_prompt_logits=inputs.include_prompt_logits,
             callback=inputs.callback,
             stop=inputs.stop,
+            top_p=inputs.top_p,
+            top_k=inputs.top_k,
+            presence_penalty=inputs.presence_penalty,
+            frequency_penalty=inputs.frequency_penalty,
         )
+
         return engine_input, postprocessing_kwargs
 
     def process_engine_outputs(
@@ -441,7 +470,9 @@ def process_engine_outputs(
         return TextGenerationOutput(sequences=sequences, logits=logits)
 
     def engine_forward(
-        self, engine_inputs: List[numpy.ndarray], context: Dict
+        self,
+        engine_inputs: List[numpy.ndarray],
+        context: Dict,
     ) -> Tuple[numpy.ndarray, numpy.ndarray]:
         """
         Run the forward pass on the engine.
@@ -456,20 +487,37 @@ def engine_forward(
         # as such, a new context needs to be created since we are no longer in the
         # main thread. That is why `engine_` is prepended to each of the timer phase
         # names in this context
+
         with self.timer_manager.new_timer_context(total_inference=False) as timer:
             streamer = context.get("streamer")
 
             if not self.cache_support_enabled:
-                tokens, prompt_logits = self.multitoken_engine(engine_inputs)
-                return numpy.array([tokens]), prompt_logits
+                prompt_logits = self.multitoken_engine(engine_inputs)
+                token_generator = TokenGenerator(
+                    logits_shape=prompt_logits[-1].shape[-1],
+                    deterministic=self.deterministic,
+                    **context,
+                )
+                for prompt_logit in prompt_logits:
+                    token_generator.generate(prompt_logit)
+                return numpy.array([self.tokens]), prompt_logits
 
             else:
                 # run the prompt through
                 with timer.time(TextGenerationTimings.PROMPT_PREFILL):
-                    tokens, prompt_logits = self.prompt_inference(engine_inputs)
+                    prompt_logits = self.prompt_inference(engine_inputs)
+
+            tokens = engine_inputs[0][engine_inputs[1].nonzero()].tolist()
+            token_generator = TokenGenerator(
+                logits_shape=prompt_logits[-1].shape[-1],
+                tokens=tokens,
+                deterministic=self.deterministic,
+                **context,
+            )
+            token_generator.generate(prompt_logits[-1][0, -1, :])
 
             if streamer is not None:
-                streamer.put(numpy.array(tokens))
+                streamer.put(numpy.array(token_generator.tokens))
 
             # create the generated output
             max_tokens = (
@@ -480,7 +528,7 @@ def engine_forward(
 
             # last prompt token is the first generated token
             # add it to generated tokens, and the logits
-            generated_tokens = [tokens[-1]]
+            generated_tokens = [token_generator.tokens[-1]]
             generated_logits = (
                 prompt_logits
                 if context.get("include_prompt_logits")
@@ -492,8 +540,10 @@ def engine_forward(
             with timer.time(TextGenerationTimings.TOKEN_GENERATION):
                 while len(generated_tokens) < max_tokens:
                     with timer.time(TextGenerationTimings.TOKEN_GENERATION_SINGLE):
-                        token, logits = self.autoregressive_inference(tokens)
-                    tokens.append(token)
+                        logits = self.autoregressive_inference(
+                            tokens=token_generator.tokens
+                        )
+                        token = token_generator.generate(logits=logits[0, -1, :])
                     generated_tokens.append(token)
                     generated_logits.append(logits)
 
@@ -528,7 +578,8 @@ def engine_forward(
         )
 
     def prompt_inference(
-        self, engine_inputs: List[numpy.ndarray]
+        self,
+        engine_inputs: List[numpy.ndarray],
     ) -> Tuple[List[int], List[numpy.ndarray]]:
         """
         An inference run that processes the prompt through the
@@ -545,14 +596,13 @@ def prompt_inference(
         tokens = engine_inputs[0][engine_inputs[1].nonzero()].tolist()
 
         prompt_logits = []
-        new_token = None
         num_tokens_processed = 0
 
         if len(tokens) > self.prompt_sequence_length and self.enable_multitoken_prefill:
             self.multitoken_engine.reset_kv_cache()
             for engine_inputs in self.engine_inputs_for_prefill(tokens):
-                new_token, new_logits = self.multitoken_engine(engine_inputs)
-                num_tokens_processed += self.prompt_sequence_length
+                new_logits = self.multitoken_engine(engine_inputs)
+                num_tokens_processed += self.prompt_processing_sequence_length
                 prompt_logits.append(new_logits)
 
         if num_tokens_processed:
@@ -569,13 +619,11 @@ def prompt_inference(
             with self.timer_manager.current.time(
                 TextGenerationTimings.PROMPT_PREFILL_SINGLE
             ):
-                new_token, new_logits = self.autoregressive_inference(run_tokens)
+                new_logits = self.autoregressive_inference(run_tokens)
 
             prompt_logits.append(new_logits)
 
-        tokens.append(new_token)
-
-        return tokens, prompt_logits
+        return prompt_logits
 
     def autoregressive_inference(
         self,
@@ -612,9 +660,9 @@ def autoregressive_inference(
             engine_inputs_map[name] for name in self.engine.onnx_input_names_no_cache
         ]
 
-        generated_token, generated_logits = self.engine(engine_inputs)
+        generated_logits = self.engine(engine_inputs)
 
-        return generated_token, generated_logits
+        return generated_logits
 
     def engine_inputs_for_prefill(
         self, tokens: List[int]

src/deepsparse/transformers/pipelines/token_classification.py

@@ -522,7 +522,6 @@ def _get_tag(self, entity_name: str) -> Tuple[str, str]:
         return bi, tag
 
     def _group_entities(self, entities: List[dict]) -> List[dict]:
-
         entity_groups = []
         entity_group_disagg = []