Feature/seq gpt (#507)

modelscope/pipelines/nlp/__init__.py

@@ -27,7 +27,7 @@ if TYPE_CHECKING:
     from .translation_quality_estimation_pipeline import TranslationQualityEstimationPipeline
     from .text_error_correction_pipeline import TextErrorCorrectionPipeline
     from .word_alignment_pipeline import WordAlignmentPipeline
-    from .text_generation_pipeline import TextGenerationPipeline, TextGenerationT5Pipeline
+    from .text_generation_pipeline import TextGenerationPipeline, TextGenerationT5Pipeline, SeqGPTPipeline
     from .fid_dialogue_pipeline import FidDialoguePipeline
     from .token_classification_pipeline import TokenClassificationPipeline
     from .translation_pipeline import TranslationPipeline
@@ -79,7 +79,7 @@ else:
         'text_error_correction_pipeline': ['TextErrorCorrectionPipeline'],
         'word_alignment_pipeline': ['WordAlignmentPipeline'],
         'text_generation_pipeline':
-        ['TextGenerationPipeline', 'TextGenerationT5Pipeline'],
+        ['TextGenerationPipeline', 'TextGenerationT5Pipeline', 'SeqGPTPipeline'],
         'fid_dialogue_pipeline': ['FidDialoguePipeline'],
         'token_classification_pipeline': ['TokenClassificationPipeline'],
         'translation_pipeline': ['TranslationPipeline'],

modelscope/pipelines/nlp/text_generation_pipeline.py

@@ -4,6 +4,7 @@
 import os
 from typing import Any, Dict, Optional, Union
 
+import re
 import torch
 from transformers import GenerationConfig
 
@@ -28,6 +29,7 @@ __all__ = [
     'ChatGLM6bV2TextGenerationPipeline',
     'QWenChatPipeline',
     'QWenTextGenerationPipeline',
+    'SeqGPTPipeline'
 ]
 
 
@@ -418,3 +420,46 @@ class QWenTextGenerationPipeline(Pipeline):
     # format the outputs from pipeline
     def postprocess(self, input, **kwargs) -> Dict[str, Any]:
         return input
+
+
+@PIPELINES.register_module(
+    group_key=Tasks.text_generation, module_name='seqgpt')
+class SeqGPTPipeline(Pipeline):
+
+    def __init__(self, model: Union[Model, str], **kwargs):
+        from modelscope.models.nlp import BloomForTextGeneration
+        from modelscope.utils.hf_util import AutoTokenizer
+
+        if isinstance(model, str):
+            model_dir = snapshot_download(
+                model) if not os.path.exists(model) else model
+            model = Model.from_pretrained(model_dir)
+        self.model = model
+        self.model.eval()
+        self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
+
+        super().__init__(model=model, **kwargs)
+
+    def _sanitize_parameters(self, **pipeline_parameters):
+        return {}, pipeline_parameters, {}
+
+    def preprocess(self, inputs, **preprocess_params) -> Dict[str, Any]:
+        return inputs
+
+    # define the forward pass
+    def forward(self, prompt: str, **forward_params) -> Dict[str, Any]:
+        # gen & decode
+        prompt += '[GEN]'
+        input_ids = self.tokenizer(prompt, return_tensors="pt", padding=True, truncation=True, max_length=1024)
+        input_ids = input_ids.input_ids.cuda()
+        outputs = self.model.generate(input_ids, num_beams=4, do_sample=False, max_new_tokens=256)
+        decoded_sentences = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)
+        decoded_sentence = decoded_sentences[0]
+        decoded_sentence = decoded_sentence[len(prompt):]
+        return {
+            OutputKeys.TEXT: decoded_sentence
+        }
+
+    # format the outputs from pipeline
+    def postprocess(self, input, **kwargs) -> Dict[str, Any]:
+        return input

requirements/nlp.txt

@@ -1,6 +1,5 @@
 boto3
 embeddings
-en_core_web_sm>=2.3.5
 filelock
 ftfy
 jieba>=0.42.1

tests/pipelines/test_text_generation.py

@@ -45,6 +45,8 @@ class TextGenerationTest(unittest.TestCase):
 
         self.llama_model_id = 'skyline2006/llama-7b'
         self.llama_input = 'My name is Merve and my favorite'
+        self.seqgpt_model_id = 'damo/nlp_seqgpt-560m'
+        self.ecomgpt_model_id = 'damo/nlp_seqgpt-560m'
 
     def run_pipeline_with_model_instance(self, model_id, input):
         model = Model.from_pretrained(model_id)
@@ -320,6 +322,22 @@ class TextGenerationTest(unittest.TestCase):
             self.llama_input,
             run_kwargs={'max_length': 64})
 
+    @unittest.skipUnless(test_level() >= 0, 'skip test in current test level')
+    def test_seqgpt_with_model_name(self):
+        inputs = {'task': '抽取', 'text': '杭州欢迎你。', 'labels': '地名'}
+        PROMPT_TEMPLATE = '输入: {text}\n{task}: {labels}\n输出: '
+        prompt = PROMPT_TEMPLATE.format(**inputs)
+        self.run_pipeline_with_model_id(self.seqgpt_model_id, prompt)
+
+    @unittest.skipUnless(test_level() >= 0, 'skip test in current test level')
+    def test_ecomgpt_with_model_name(self):
+        PROMPT_TEMPLATE = "Below is an instruction that describes a task. " + \
+                          "Write a response that appropriately completes the request.\n\n" + \
+                          "### Instruction:\n{text}\n{instruction}\n\n### Response:"
+        inputs = {'instruction': 'Classify the sentence, candidate labels: product, brand', 'text': '照相机'}
+        prompt = PROMPT_TEMPLATE.format(**inputs)
+        self.run_pipeline_with_model_id(self.ecomgpt_model_id, prompt)
+
 
 if __name__ == '__main__':
     unittest.main()