DSPy: The framework for programming—not prompting—language models
Python
Fix UsageTracker AttributeError when using ParallelExecutor with dspy.context (#9095)
Copilot90eabeeb9f63
diff --git a/dspy/utils/parallelizer.py b/dspy/utils/parallelizer.pyindex c32f5e3eb..5b5544d3d 100644--- a/dspy/utils/parallelizer.py+++ b/dspy/utils/parallelizer.py@@ -89,10 +89,11 @@ class ParallelExecutor:from dspy.dsp.utils.settings import thread_local_overridesoriginal = thread_local_overrides.get()- token = thread_local_overrides.set({**original, **parent_overrides.copy()})- if parent_overrides.get("usage_tracker"):+ new_overrides = {**original, **parent_overrides.copy()}+ if new_overrides.get("usage_tracker"):# Usage tracker needs to be deep copied across threads so that each thread tracks its own usage- thread_local_overrides.overrides["usage_tracker"] = copy.deepcopy(parent_overrides["usage_tracker"])+ new_overrides["usage_tracker"] = copy.deepcopy(new_overrides["usage_tracker"])+ token = thread_local_overrides.set(new_overrides)try:return index, function(item)diff --git a/tests/utils/test_usage_tracker.py b/tests/utils/test_usage_tracker.pyindex 574d1f740..28e35e40a 100644--- a/tests/utils/test_usage_tracker.py+++ b/tests/utils/test_usage_tracker.py@@ -1,3 +1,5 @@+from unittest import mock+from pydantic import BaseModelimport dspy@@ -325,3 +327,58 @@ def test_merge_usage_entries_with_pydantic_models():assert total_usage["gpt-4o-mini"]["completion_tokens_details"]["audio_tokens"] == 1assert total_usage["gpt-4o-mini"]["completion_tokens_details"]["accepted_prediction_tokens"] == 1assert total_usage["gpt-4o-mini"]["completion_tokens_details"]["rejected_prediction_tokens"] == 1+++def test_parallel_executor_with_usage_tracker():+ """Test that usage tracking works correctly with ParallelExecutor and mocked LM calls."""++ parent_tracker = UsageTracker()++ # Mock LM with different responses+ mock_lm = mock.MagicMock(spec=dspy.LM)+ mock_lm.return_value = ['{"answer": "Mocked answer"}']+ mock_lm.kwargs = {}+ mock_lm.model = "openai/gpt-4o-mini"++ dspy.configure(lm=mock_lm, adapter=dspy.JSONAdapter())++ def task1():+ # Simulate LM usage tracking for task 1+ dspy.settings.usage_tracker.add_usage(+ "openai/gpt-4o-mini",+ {+ "prompt_tokens": 50,+ "completion_tokens": 10,+ "total_tokens": 60,+ },+ )+ return dspy.settings.usage_tracker.get_total_tokens()++ def task2():+ # Simulate LM usage tracking for task 2 with different values+ dspy.settings.usage_tracker.add_usage(+ "openai/gpt-4o-mini",+ {+ "prompt_tokens": 80,+ "completion_tokens": 15,+ "total_tokens": 95,+ },+ )+ return dspy.settings.usage_tracker.get_total_tokens()++ # Execute tasks in parallel+ with dspy.context(track_usage=True, usage_tracker=parent_tracker):+ executor = dspy.Parallel()+ results = executor([(task1, {}), (task2, {})])+ # Verify that the two workers had different usage+ usage1 = results[0]+ usage2 = results[1]++ # Task 1 should have 50 prompt tokens, task 2 should have 80+ assert usage1["openai/gpt-4o-mini"]["prompt_tokens"] == 50+ assert usage1["openai/gpt-4o-mini"]["completion_tokens"] == 10+ assert usage2["openai/gpt-4o-mini"]["prompt_tokens"] == 80+ assert usage2["openai/gpt-4o-mini"]["completion_tokens"] == 15++ # Parent tracker should remain unchanged (workers have independent copies)+ assert len(parent_tracker.usage_data) == 0