Merge pull request #367 from NVIDIA/fix/qa-fixes-6

Fix LangChain warnings and bug affecting Llama-2 example.

nemoguardrails/llm/helpers.py

@@ -66,8 +66,9 @@ def _modify_instance_kwargs(self):
  """
 
  if hasattr(llm_instance, "model_kwargs"):
- llm_instance.model_kwargs["temperature"] = self.temperature
- llm_instance.model_kwargs["streaming"] = self.streaming
+ if isinstance(llm_instance.model_kwargs, dict):
+ llm_instance.model_kwargs["temperature"] = self.temperature
+ llm_instance.model_kwargs["streaming"] = self.streaming
 
  def _call(
  self,

nemoguardrails/llm/providers/providers.py

@@ -69,7 +69,7 @@ def _call(
  )
 
  # Streaming for NeMo Guardrails is not supported in sync calls.
- if self.model_kwargs.get("streaming"):
+ if self.model_kwargs and self.model_kwargs.get("streaming"):
  raise Exception(
  "Streaming mode not supported for HuggingFacePipeline in NeMo Guardrails!"
  )
@@ -100,7 +100,7 @@ async def _acall(
  )
 
  # Handle streaming, if the flag is set
- if self.model_kwargs.get("streaming"):
+ if self.model_kwargs and self.model_kwargs.get("streaming"):
  # Retrieve the streamer object, needs to be set in model_kwargs
  streamer = self.model_kwargs.get("streamer")
  if not streamer:
@@ -153,7 +153,18 @@ async def _acall(self, *args, **kwargs):
 
 def discover_langchain_providers():
  """Automatically discover all LLM providers from LangChain."""
- _providers.update(llms.type_to_cls_dict)
+ # To deal with deprecated stuff and avoid warnings, we compose the type_to_cls_dict here
+ if hasattr(llms, "get_type_to_cls_dict"):
+ type_to_cls_dict = {
+ k: v()
+ for k, v in llms.get_type_to_cls_dict().items()
+ # Exclude deprecated ones
+ if k not in ["mlflow-chat", "databricks-chat"]
+ }
+ else:
+ type_to_cls_dict = llms.type_to_cls_dict
+
+ _providers.update(type_to_cls_dict)
 
  # We make sure we have OpenAI from the right package.
  if "openai" in _providers: