fix vllm gen cfg & set default to llm-ppl (#658)

* fix vllm gen cfg & set default to llm-ppl

* fix code style

* reset llm_framework default value

* fix pre-commit

---------

Co-authored-by: suluyan.sly <suluyan.sly@alibaba-inc.com>

modelscope/pipelines/accelerate/vllm.py

@@ -42,6 +42,24 @@ class Vllm(InferFramework):
                 The string batch or the token list batch to input to the model.
             kwargs: Sampling parameters.
         """
+
+        # convert hf generate config to vllm
+        do_sample = kwargs.pop('do_sample', None)
+        num_beam = kwargs.pop('num_beam', 1)
+        max_length = kwargs.pop('max_length', None)
+        max_new_tokens = kwargs.pop('max_new_tokens', None)
+
+        # for vllm, default to do_sample/greedy(depends on temperature).
+        # for hf, do_sample=false, num_beam=1 -> greedy(default)
+        #         do_sample=ture, num_beam=1 -> sample
+        #         do_sample=false, num_beam>1 -> beam_search
+        if not do_sample and num_beam > 1:
+            kwargs['use_beam_search'] = True
+        if max_length:
+            kwargs['max_tokens'] = max_length - len(prompts[0])
+        if max_new_tokens:
+            kwargs['max_tokens'] = max_new_tokens
+
         from vllm import SamplingParams
         sampling_params = SamplingParams(**kwargs)
         if isinstance(prompts[0], str):

modelscope/pipelines/nlp/llm_pipeline.py

@@ -104,12 +104,20 @@ class LLMPipeline(Pipeline):
 
         if isinstance(model, str) and is_official_hub_path(model):
             logger.info(f'initiate model from location {model}.')
-            if self.llm_framework is not None:
+            if self.llm_framework:
                 model_dir = model if os.path.exists(
                     model) else snapshot_download(model)
-                return self._wrap_infer_framework(model_dir,
-                                                  self.llm_framework)
-            elif is_model(model):
+                try:
+                    model = self._wrap_infer_framework(model_dir,
+                                                       self.llm_framework)
+                    logger.info(f'initiate model with {framework}.')
+                    return model
+                except Exception as e:
+                    self.llm_framework = None
+                    logger.warning(
+                        f'Cannot using llm_framework with {model}, '
+                        f'ignoring llm_framework={self.llm_framework} : {e}')
+            if is_model(model):
                 return Model.from_pretrained(
                     model,
                     invoked_by=Invoke.PIPELINE,