Proposition: Add exponential backoff with jitter for retries

src/smolagents/models.py

@@ -35,8 +35,10 @@
 
 logger = logging.getLogger(__name__)
 
-RETRY_WAIT = 120
+RETRY_WAIT = 60
 RETRY_MAX_ATTEMPTS = 3
+RETRY_EXPONENTIAL_BASE = 2
+RETRY_JITTER = True
 STRUCTURED_GENERATION_PROVIDERS = ["cerebras", "fireworks-ai"]
 CODEAGENT_RESPONSE_FORMAT = {
     "type": "json_schema",
@@ -1102,6 +1104,8 @@ def __init__(
         self.retryer = Retrying(
             max_attempts=RETRY_MAX_ATTEMPTS if retry else 1,
             wait_seconds=RETRY_WAIT,
+            exponential_base=RETRY_EXPONENTIAL_BASE,
+            jitter=RETRY_JITTER,
             retry_predicate=is_rate_limit_error,
             reraise=True,
             before_sleep_logger=(logger, logging.INFO),

src/smolagents/utils.py

@@ -21,6 +21,7 @@
 import json
 import keyword
 import os
+import random
 import re
 import time
 from functools import lru_cache
@@ -516,20 +517,25 @@ def __init__(
         self,
         max_attempts: int = 1,
         wait_seconds: float = 0.0,
+        exponential_base: float = 2.0,
+        jitter: bool = True,
         retry_predicate: Callable[[BaseException], bool] | None = None,
         reraise: bool = False,
         before_sleep_logger: tuple[Logger, int] | None = None,
         after_logger: tuple[Logger, int] | None = None,
     ):
         self.max_attempts = max_attempts
         self.wait_seconds = wait_seconds
+        self.exponential_base = exponential_base
+        self.jitter = jitter
         self.retry_predicate = retry_predicate
         self.reraise = reraise
         self.before_sleep_logger = before_sleep_logger
         self.after_logger = after_logger
 
     def __call__(self, fn, *args: Any, **kwargs: Any) -> Any:
         start_time = time.time()
+        delay = self.wait_seconds
 
         for attempt_number in range(1, self.max_attempts + 1):
             try:
@@ -542,7 +548,7 @@ def __call__(self, fn, *args: Any, **kwargs: Any) -> Any:
                     fn_name = getattr(fn, "__name__", repr(fn))
                     logger.log(
                         log_level,
-                        f"Finished call to '{fn_name}' after {seconds:.3f}(s), this was attempt n°{attempt_number}.",
+                        f"Finished call to '{fn_name}' after {seconds:.3f}(s), this was attempt n°{attempt_number}/{self.max_attempts}.",
                     )
 
                 return result
@@ -564,18 +570,22 @@ def __call__(self, fn, *args: Any, **kwargs: Any) -> Any:
                     fn_name = getattr(fn, "__name__", repr(fn))
                     logger.log(
                         log_level,
-                        f"Finished call to '{fn_name}' after {seconds:.3f}(s), this was attempt n°{attempt_number}.",
+                        f"Finished call to '{fn_name}' after {seconds:.3f}(s), this was attempt n°{attempt_number}/{self.max_attempts}.",
                     )
 
+                # Exponential backoff with jitter
+                # https://cookbook.openai.com/examples/how_to_handle_rate_limits#example-3-manual-backoff-implementation
+                delay *= self.exponential_base * (1 + self.jitter * random.random())
+
                 # Log before sleeping
                 if self.before_sleep_logger:
                     logger, log_level = self.before_sleep_logger
                     fn_name = getattr(fn, "__name__", repr(fn))
                     logger.log(
                         log_level,
-                        f"Retrying {fn_name} in {self.wait_seconds} seconds as it raised {e.__class__.__name__}: {e}.",
+                        f"Retrying {fn_name} in {delay} seconds as it raised {e.__class__.__name__}: {e}.",
                     )
 
                 # Sleep before next attempt
-                if self.wait_seconds > 0:
-                    time.sleep(self.wait_seconds)
+                if delay > 0:
+                    time.sleep(delay)

tests/test_models.py

@@ -419,7 +419,8 @@ def test_retry_on_rate_limit_error(self):
         mock_litellm = MagicMock()
 
         with (
-            patch("smolagents.models.RETRY_WAIT", 1),
+            patch("smolagents.models.RETRY_WAIT", 0.1),
+            patch("smolagents.utils.random.random", side_effect=[0.1, 0.1]),
             patch("smolagents.models.LiteLLMModel.create_client", return_value=mock_litellm),
         ):
             model = LiteLLMModel(model_id="test-model")
@@ -438,22 +439,25 @@ def test_retry_on_rate_limit_error(self):
             # Create a 429 rate limit error
             rate_limit_error = Exception("Error code: 429 - Rate limit exceeded")
 
-            # Mock the litellm client to raise error first, then succeed
-            model.client.completion.side_effect = [rate_limit_error, mock_success_response]
+            # Mock the litellm client to raise an error twice, and then succeed
+            model.client.completion.side_effect = [rate_limit_error, rate_limit_error, mock_success_response]
 
             # Measure time to verify retry wait time
             start_time = time.time()
             result = model.generate(messages)
             elapsed_time = time.time() - start_time
 
-            # Verify that completion was called twice (once failed, once succeeded)
-            assert model.client.completion.call_count == 2
+            # Verify that completion was called thrice (twice failed, once succeeded)
+            assert model.client.completion.call_count == 3
             assert result.content == "Success response"
             assert result.token_usage.input_tokens == 10
             assert result.token_usage.output_tokens == 20
 
-            # Verify that the wait time was around 1s (allow some tolerance)
-            assert 0.9 <= elapsed_time <= 1.2
+            # Verify that the wait time was around
+            # 0.22s (1st retry) [0.1 * 2.0 * (1 + 1 * 0.1)]
+            # + 0.48s (2nd retry) [0.22 * 2.0 * (1 + 1 * 0.1)]
+            # = 0.704s (allow some tolerance)
+            assert 0.67 <= elapsed_time <= 0.73
 
     def test_passing_flatten_messages(self):
         model = LiteLLMModel(model_id="groq/llama-3.3-70b", flatten_messages_as_text=False)