Update instruct_pipeline.py

instruct_pipeline.py

@@ -1,8 +1,16 @@
+import logging
 import re
+from typing import List
 
 import numpy as np
 from transformers import Pipeline, PreTrainedTokenizer
 
+from transformers.utils import is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+
+logger = logging.getLogger(__name__)
 
 INSTRUCTION_KEY = "### Instruction:"
 RESPONSE_KEY = "### Response:"
@@ -53,9 +61,22 @@ class InstructionTextGenerationPipeline(Pipeline):
     def __init__(
         self, *args, do_sample: bool = True, max_new_tokens: int = 256, top_p: float = 0.92, top_k: int = 0, **kwargs
     ):
-        super().__init__(*args, do_sample=do_sample, max_new_tokens=max_new_tokens, top_p=top_p, top_k=top_k, **kwargs)
-
-    def _sanitize_parameters(self, return_instruction_text=False, **generate_kwargs):
+        """Initialize the pipeline
+
+        Args:
+            do_sample (bool, optional): Whether or not to use sampling. Defaults to True.
+            max_new_tokens (int, optional): Max new tokens after the prompt to generate. Defaults to 128.
+            top_p (float, optional): If set to float < 1, only the smallest set of most probable tokens with
+                probabilities that add up to top_p or higher are kept for generation. Defaults to 0.92.
+            top_k (int, optional): The number of highest probability vocabulary tokens to keep for top-k-filtering.
+                Defaults to 0.
+        """
+        super().__init__(*args, do_sample=do_sample, max_new_tokens=max_new_tokens, top_p=top_p, top_k=top_k,
+                         **kwargs)
+
+    def _sanitize_parameters(self,
+                             return_full_text: bool = None,
+                             **generate_kwargs):
         preprocess_params = {}
 
         # newer versions of the tokenizer configure the response key as a special token.  newer versions still may
@@ -79,10 +100,12 @@ class InstructionTextGenerationPipeline(Pipeline):
         forward_params = generate_kwargs
         postprocess_params = {
             "response_key_token_id": response_key_token_id,
-            "end_key_token_id": end_key_token_id,
-            "return_instruction_text": return_instruction_text,
+            "end_key_token_id": end_key_token_id
         }
 
+        if return_full_text is not None:
+            postprocess_params["return_full_text"] = return_full_text
+
         return preprocess_params, forward_params, postprocess_params
 
     def preprocess(self, instruction_text, **generate_kwargs):
@@ -98,63 +121,92 @@ class InstructionTextGenerationPipeline(Pipeline):
     def _forward(self, model_inputs, **generate_kwargs):
         input_ids = model_inputs["input_ids"]
         attention_mask = model_inputs.get("attention_mask", None)
+
+        if input_ids.shape[1] == 0:
+            input_ids = None
+            attention_mask = None
+            in_b = 1
+        else:
+            in_b = input_ids.shape[0]
+
         generated_sequence = self.model.generate(
             input_ids=input_ids.to(self.model.device),
-            attention_mask=attention_mask,
+            attention_mask=attention_mask.to(self.model.device) if attention_mask is not None else None,
             pad_token_id=self.tokenizer.pad_token_id,
             **generate_kwargs,
-        )[0].cpu()
+        )
+
+        out_b = generated_sequence.shape[0]
+        if self.framework == "pt":
+            generated_sequence = generated_sequence.reshape(in_b, out_b // in_b, *generated_sequence.shape[1:])
+        elif self.framework == "tf":
+            generated_sequence = tf.reshape(generated_sequence, (in_b, out_b // in_b, *generated_sequence.shape[1:]))
+
         instruction_text = model_inputs.pop("instruction_text")
         return {"generated_sequence": generated_sequence, "input_ids": input_ids, "instruction_text": instruction_text}
 
-    def postprocess(self, model_outputs, response_key_token_id, end_key_token_id, return_instruction_text):
-        sequence = model_outputs["generated_sequence"]
+    def postprocess(self, model_outputs, response_key_token_id, end_key_token_id, return_full_text: bool = False):
+
+        generated_sequence = model_outputs["generated_sequence"][0]
         instruction_text = model_outputs["instruction_text"]
 
-        # The response will be set to this variable if we can identify it.
-        decoded = None
+        generated_sequence: List[List[int]] = generated_sequence.numpy().tolist()
+        records = []
+        for sequence in generated_sequence:
 
-        # If we have token IDs for the response and end, then we can find the tokens and only decode between them.
-        if response_key_token_id and end_key_token_id:
-            # Find where "### Response:" is first found in the generated tokens.  Considering this is part of the
-            # prompt, we should definitely find it.  We will return the tokens found after this token.
-            response_pos = None
-            response_positions = np.where(sequence == response_key_token_id)[0]
-            if len(response_positions) == 0:
-                pass
-            else:
-                response_pos = response_positions[0]
-
-            if response_pos:
-                # Next find where "### End" is located.  The model has been trained to end its responses with this
-                # sequence (or actually, the token ID it maps to, since it is a special token).  We may not find
-                # this token, as the response could be truncated.  If we don't find it then just return everything
-                # to the end.  Note that even though we set eos_token_id, we still see the this token at the end.
-                end_pos = None
-                end_positions = np.where(sequence == end_key_token_id)[0]
-                if len(end_positions) > 0:
-                    end_pos = end_positions[0]
-
-                decoded = self.tokenizer.decode(sequence[response_pos + 1 : end_pos]).strip()
-        else:
-            # Otherwise we'll decode everything and use a regex to find the response and end.
+            # The response will be set to this variable if we can identify it.
+            decoded = None
 
-            fully_decoded = self.tokenizer.decode(sequence)
+            # If we have token IDs for the response and end, then we can find the tokens and only decode between them.
+            if response_key_token_id and end_key_token_id:
+                # Find where "### Response:" is first found in the generated tokens.  Considering this is part of the
+                # prompt, we should definitely find it.  We will return the tokens found after this token.
+                try:
+                    response_pos = sequence.index(response_key_token_id)
+                except ValueError:
+                    logger.warn(f"Could not find response key {response_key_token_id} in: {sequence}")
+                    response_pos = None
 
-            # The response appears after "### Response:".  The model has been trained to append "### End" at the
-            # end.
-            m = re.search(r"#+\s*Response:\s*(.+?)#+\s*End", fully_decoded, flags=re.DOTALL)
+                if response_pos:
+                    # Next find where "### End" is located.  The model has been trained to end its responses with this
+                    # sequence (or actually, the token ID it maps to, since it is a special token).  We may not find
+                    # this token, as the response could be truncated.  If we don't find it then just return everything
+                    # to the end.  Note that even though we set eos_token_id, we still see the this token at the end.
+                    try:
+                        end_pos = sequence.index(end_key_token_id)
+                    except ValueError:
+                        end_pos = None
 
-            if m:
-                decoded = m.group(1).strip()
-            else:
-                # The model might not generate the "### End" sequence before reaching the max tokens.  In this case,
-                # return everything after "### Response:".
-                m = re.search(r"#+\s*Response:\s*(.+)", fully_decoded, flags=re.DOTALL)
-                if m:
-                    decoded = m.group(1).strip()
+                    decoded = self.tokenizer.decode(sequence[response_pos + 1 : end_pos]).strip()
 
-        if return_instruction_text:
-            return {"instruction_text": instruction_text, "generated_text": decoded}
+            if not decoded:
+                # Otherwise we'll decode everything and use a regex to find the response and end.
 
-        return {'generated_text': decoded}
+                fully_decoded = self.tokenizer.decode(sequence)
+
+                # The response appears after "### Response:".  The model has been trained to append "### End" at the
+                # end.
+                m = re.search(r"#+\s*Response:\s*(.+?)#+\s*End", fully_decoded, flags=re.DOTALL)
+
+                if m:
+                    decoded = m.group(1).strip()
+                else:
+                    # The model might not generate the "### End" sequence before reaching the max tokens.  In this case,
+                    # return everything after "### Response:".
+                    m = re.search(r"#+\s*Response:\s*(.+)", fully_decoded, flags=re.DOTALL)
+                    if m:
+                        decoded = m.group(1).strip()
+                    else:
+                        logger.warn(f"Failed to find response in:\n{fully_decoded}")
+
+            # If the full text is requested, then append the decoded text to the original instruction.
+            # This technically isn't the full text, as we format the instruction in the prompt the model has been
+            # trained on, but to the client it will appear to be the full text.
+            if return_full_text:
+                decoded = f"{instruction_text}\n{decoded}"
+
+            rec = {"generated_text": decoded}
+
+            records.append(rec)
+
+        return records