use huggingface standard generation for tnx (deepjavalibrary#778)

engines/python/setup/djl_python/streaming_utils.py

@@ -23,10 +23,10 @@ def get_stream_generator(execution_engine: str):
         ## execution_engine passed to this function is not the same engine specified in serving.properties
         ## in djl-serving. For e.g Accelerate and neuronx use Python as the engine serving.properties
         ## The engine here refers to backend model parallel framework.
-        if execution_engine in {"DeepSpeed", "Accelerate"}:
+        if execution_engine in {
+                "DeepSpeed", "Accelerate", "transformers-neuronx"
+        }:
             return StreamingUtils._hf_model_stream_generator
-        elif execution_engine == "transformers-neuronx":
-            return StreamingUtils._transformers_neuronx_stream_generator
         else:
             raise ValueError(
                 f"{execution_engine} engine is not supported for streaming")
@@ -55,6 +55,12 @@ def _hf_model_stream_generator(model, tokenizer, inputs, **kwargs):
                                           dtype=torch.long,
                                           device=input_ids.device)
         stop_generation = False
+        engine = None
+        if "engine" in kwargs.keys():
+            engine = kwargs["engine"]
+
+        if "transformers-neuronx" == engine:
+            model.reset_generation()
 
         if generic_model_class == "CausalLM":
             input_length = input_ids.shape[1]
@@ -83,10 +89,17 @@ def _hf_model_stream_generator(model, tokenizer, inputs, **kwargs):
 
             if generic_model_class == "CausalLM":
                 attention_mask_curr = attention_mask[:, :curr_length]
-                outputs = model.forward(input_ids=input_ids,
-                                        attention_mask=attention_mask_curr,
-                                        past_key_values=past_key_values,
-                                        use_cache=True)
+                model_inputs = {
+                    "input_ids": input_ids,
+                    "attention_mask": attention_mask_curr,
+                    "past_key_values": past_key_values,
+                    "use_cache": True
+                }
+                if "transformers-neuronx" == engine:
+                    model_inputs = model.prepare_inputs_for_generation(
+                        **model_inputs)
+                    model_inputs["return_dict"] = True
+                outputs = model.forward(**model_inputs)
 
             if generic_model_class == "Seq2SeqLM":
                 outputs = model.forward(
@@ -108,7 +121,8 @@ def _hf_model_stream_generator(model, tokenizer, inputs, **kwargs):
 
             all_decoder_input_ids = torch.cat(
                 [all_decoder_input_ids, token_ids], dim=1)
-            past_key_values = outputs.past_key_values
+            if engine is None:
+                past_key_values = outputs.past_key_values
             new_tokens_count += 1
 
             not_eos_token_ids = (token_ids != tokenizer.eos_token_id).view(
@@ -136,40 +150,6 @@ def _hf_model_stream_generator(model, tokenizer, inputs, **kwargs):
 
             yield token_text
 
-    @staticmethod
-    @torch.inference_mode()
-    def _transformers_neuronx_stream_generator(model, tokenizer, inputs,
-                                               **kwargs):
-        sequence_length = kwargs.get("seq_length",
-                                     StreamingUtils.DEFAULT_MAX_NEW_TOKENS)
-        top_k = kwargs.get("top_k", 50)
-        tokenized_inputs = tokenizer(inputs, return_tensors="pt", padding=True)
-        input_ids = tokenized_inputs["input_ids"]
-        model.reset()
-        eos_token_id = model.config.eos_token_id
-        # populate key/value caches according to the prompt text
-        _, start = input_ids.shape
-        position_ids = torch.arange(start, dtype=torch.int32)
-        next_token_scores = model(input_ids, position_ids)
-
-        tokens = [input_ids]
-        for cur_len in range(start, sequence_length):
-            # don't sample EOS
-            next_token_scores[:, eos_token_id] = -float('inf')
-
-            # Remove all tokens with a probability less than the last token of the top-k
-            topk_values, topk_indices = torch.topk(next_token_scores, top_k)
-            probs = torch.nn.functional.softmax(topk_values, dim=-1)
-            inputs_in_topk = torch.multinomial(probs,
-                                               num_samples=1,
-                                               replacement=True)
-            inputs = torch.gather(topk_indices, 1, inputs_in_topk)
-            tokens.append(inputs)
-            token_text = tokenizer.batch_decode(inputs)
-            position_ids = torch.as_tensor([cur_len], dtype=torch.int32)
-            next_token_scores = model(inputs, position_ids)
-            yield token_text
-
     @staticmethod
     def _has_met_stopping_criteria(not_eos_token_ids, current_token_count,
                                    max_new_tokens):

engines/python/setup/djl_python/transformers-neuronx.py

@@ -10,7 +10,6 @@
 # or in the "LICENSE.txt" file accompanying this file. This file is distributed on an "AS IS"
 # BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, express or implied. See the License for
 # the specific language governing permissions and limitations under the License.
-import torch
 import tempfile
 import os
 import logging
@@ -188,10 +187,9 @@ def infer(self, inputs):
             if self.enable_streaming:
                 stream_generator = StreamingUtils.get_stream_generator(
                     "transformers-neuronx")
-                model_kwargs["seq_length"] = parameters.get("max_length", 128)
-                # TODO: switch to new HF model interface
+                model_kwargs["engine"] = "transformers-neuronx"
                 outputs.add_stream_content(
-                    stream_generator(self.model.model, self.tokenizer,
+                    stream_generator(self.model, self.tokenizer,
                                      input_text, **model_kwargs))
                 return outputs