deflake the LM test (#8924)

Author: chenmoneygithub

tests/streaming/test_streaming.py

@@ -157,7 +157,7 @@ def __call__(self, x: str, **kwargs):
         include_final_prediction_in_output_stream=False,
     )
     # Turn off the cache to ensure the stream is produced.
-    with dspy.context(lm=dspy.LM(lm_for_test, cache=False)):
+    with dspy.context(lm=dspy.LM(lm_for_test, cache=False, temperature=0.0)):
         output = program(x="why did a chicken cross the kitchen?")
         all_chunks = []
         async for value in output:
@@ -221,7 +221,7 @@ def __call__(self, x: str, **kwargs):
         include_final_prediction_in_output_stream=False,
     )
     # Turn off the cache to ensure the stream is produced.
-    with dspy.context(lm=dspy.LM(lm_for_test, cache=False), adapter=dspy.JSONAdapter()):
+    with dspy.context(lm=dspy.LM(lm_for_test, cache=False, temperature=0.0), adapter=dspy.JSONAdapter()):
         output = program(x="why did a chicken cross the kitchen?")
         all_chunks = []
         async for value in output:
@@ -289,7 +289,7 @@ def __call__(self, x: str, **kwargs):
         async_streaming=False,
     )
     # Turn off the cache to ensure the stream is produced.
-    with dspy.context(lm=dspy.LM(lm_for_test, cache=False)):
+    with dspy.context(lm=dspy.LM(lm_for_test, cache=False, temperature=0.0)):
         output = program(x="why did a chicken cross the kitchen?")
         all_chunks = []
         for value in output:
@@ -437,6 +437,7 @@ async def completion_side_effect(*args, **kwargs):
         assert all_chunks[22].chunk == "."
         assert all_chunks[22].is_last_chunk is True
 
+
 @pytest.mark.anyio
 async def test_stream_listener_returns_correct_chunk_json_adapter():
     class MyProgram(dspy.Module):