Skip to content

armgit / dspy

Public Imported from GitHub
0

DSPy: The framework for programming—not prompting—language models

Python

fix(rlm): propagate context to batched subqueries (#10007)

Isaac Miller587cf8e0c745
diff --git a/dspy/predict/rlm.py b/dspy/predict/rlm.py
index 8906d03a0..46cb54d06 100644
--- a/dspy/predict/rlm.py
+++ b/dspy/predict/rlm.py
@@ -11,6 +11,7 @@ Reference: "Recursive Language Models" (Zhang, Kraska, Khattab, 2025)
from __future__ import annotations
import base64
+import contextvars
import logging
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
@@ -266,7 +267,10 @@ class RLM(Module):
results: dict[int, str] = {}
with ThreadPoolExecutor(max_workers=max_workers) as executor:
- future_to_idx = {executor.submit(_query_lm, p): i for i, p in enumerate(prompts)}
+ future_to_idx = {
+ executor.submit(contextvars.copy_context().run, _query_lm, prompt): index
+ for index, prompt in enumerate(prompts)
+ }
for future in as_completed(future_to_idx):
idx = future_to_idx[future]
try:
diff --git a/tests/predict/test_rlm.py b/tests/predict/test_rlm.py
index e5a25497d..18c0990ee 100644
--- a/tests/predict/test_rlm.py
+++ b/tests/predict/test_rlm.py
@@ -233,6 +233,31 @@ class TestRLMInitialization:
assert results[0].startswith("[ERROR]")
assert "LM failed" in results[0]
+ def test_batched_query_inherits_request_context(self):
+ import contextvars
+
+ import dspy
+
+ request_marker = contextvars.ContextVar("request_marker", default="global")
+
+ class TaggedLM:
+ def __init__(self, tag):
+ self.tag = tag
+
+ def __call__(self, prompt):
+ return [f"{self.tag}:{request_marker.get()}"]
+
+ dspy.configure(lm=TaggedLM("global"))
+ tools = RLM("context -> answer")._make_llm_tools()
+
+ with dspy.context(lm=TaggedLM("request-local")):
+ request_marker.set("request-local")
+ assert tools["llm_query"]("one") == "request-local:request-local"
+ assert tools["llm_query_batched"](["one", "two"]) == [
+ "request-local:request-local",
+ "request-local:request-local",
+ ]
+
def test_tools_call_counter_is_thread_safe(self):
"""Test that the LLM call counter is thread-safe for concurrent llm_query_batched calls.