Skip to content

armgit / dspy

Public Imported from GitHub
0

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

Python

Fix GPT-5 LM state roundtrip (#9837)

Isaac Miller0aa9eeb95e58
diff --git a/dspy/clients/lm.py b/dspy/clients/lm.py
index ccece1579..974b2568f 100644
--- a/dspy/clients/lm.py
+++ b/dspy/clients/lm.py
@@ -44,6 +44,14 @@ def _get_litellm():
return get_litellm(feature="dspy.LM")
+def _is_openai_reasoning_model(model: str) -> bool:
+ model_family = model.split("/")[-1].lower() if "/" in model else model.lower()
+ return re.match(
+ r"^(?:o[1345](?:-(?:mini|nano|pro))?(?:-\d{4}-\d{2}-\d{2})?|gpt-5(?!-chat)(?:-.*)?)$",
+ model_family,
+ ) is not None
+
+
class LM(BaseLM):
"""
A language model supporting chat or text completion requests for use with DSPy modules.
@@ -111,18 +119,7 @@ class LM(BaseLM):
def _get_initial_kwargs(self, *, temperature, max_tokens, **kwargs) -> dict[str, Any]:
# Override BaseLM's default kwargs shape for LiteLLM/model-family-specific token parameters.
- model_family = self.model.split("/")[-1].lower() if "/" in self.model else self.model.lower()
-
- # Recognize OpenAI reasoning models (o1, o3, o4, gpt-5 family)
- # Exclude non-reasoning variants like gpt-5-chat this is in azure ai foundry
- # Allow date suffixes like -2023-01-01 after model name or mini/nano/pro
- # For gpt-5, use negative lookahead to exclude -chat and allow other suffixes
- model_pattern = re.match(
- r"^(?:o[1345](?:-(?:mini|nano|pro))?(?:-\d{4}-\d{2}-\d{2})?|gpt-5(?!-chat)(?:-.*)?)$",
- model_family,
- )
-
- if model_pattern:
+ if _is_openai_reasoning_model(self.model):
if (temperature and temperature != 1.0) or (max_tokens and max_tokens < 16000):
raise LMConfigurationError(
"OpenAI's reasoning models require passing temperature=1.0 or None and max_tokens >= 16000 or None to "
@@ -418,8 +415,22 @@ class LM(BaseLM):
)
if self.use_developer_role:
state["use_developer_role"] = self.use_developer_role
+ if _is_openai_reasoning_model(self.model) and "max_completion_tokens" in state:
+ state["max_tokens"] = state.pop("max_completion_tokens")
return state
+ @classmethod
+ def load_state(cls, state: dict[str, Any], *, allow_custom_lm_class: bool = False):
+ state = dict(state)
+
+ model = state.get("model")
+ if isinstance(model, str) and _is_openai_reasoning_model(model) and "max_completion_tokens" in state:
+ if "max_tokens" not in state:
+ state["max_tokens"] = state["max_completion_tokens"]
+ state.pop("max_completion_tokens")
+
+ return super().load_state(state, allow_custom_lm_class=allow_custom_lm_class)
+
def _check_truncation(self, results):
if self.model_type != "responses" and any(c.finish_reason == "length" for c in results["choices"]):
logger.warning(
diff --git a/tests/clients/test_lm.py b/tests/clients/test_lm.py
index d5bd2f3e9..71a5bf0cb 100644
--- a/tests/clients/test_lm.py
+++ b/tests/clients/test_lm.py
@@ -519,6 +519,22 @@ def test_dump_state():
}
+def test_reasoning_model_dump_state_uses_constructor_max_tokens():
+ lm = dspy.LM(
+ model="openai/gpt-5-nano",
+ temperature=1.0,
+ max_tokens=16_000,
+ cache=False,
+ num_retries=1,
+ )
+
+ state = lm.dump_state()
+
+ assert lm.kwargs["max_completion_tokens"] == 16_000
+ assert "max_completion_tokens" not in state
+ assert state["max_tokens"] == 16_000
+
+
def test_dump_state_preserves_enabled_developer_role():
lm = dspy.LM("openai/gpt-4o-mini", use_developer_role=True)
@@ -555,6 +571,59 @@ def test_load_state():
assert loaded_lm.dump_state() == lm.dump_state()
+def test_reasoning_model_load_state_round_trips_canonical_state():
+ lm = dspy.LM(
+ model="openai/gpt-5-nano",
+ temperature=1.0,
+ max_tokens=16_000,
+ cache=False,
+ num_retries=1,
+ )
+
+ loaded_lm = dspy.BaseLM.load_state(lm.dump_state())
+
+ assert isinstance(loaded_lm, dspy.LM)
+ assert loaded_lm.kwargs["max_completion_tokens"] == 16_000
+ assert loaded_lm.dump_state() == lm.dump_state()
+
+
+def test_reasoning_model_load_state_accepts_max_completion_tokens_alias():
+ state = {
+ "_dspy_lm_class": "dspy.clients.lm.LM",
+ "model": "openai/gpt-5-nano",
+ "model_type": "chat",
+ "cache": False,
+ "num_retries": 1,
+ "temperature": 1.0,
+ "max_completion_tokens": 16_000,
+ "finetuning_model": None,
+ "launch_kwargs": {},
+ "train_kwargs": {},
+ }
+
+ loaded_lm = dspy.BaseLM.load_state(state)
+
+ assert isinstance(loaded_lm, dspy.LM)
+ assert loaded_lm.kwargs["max_completion_tokens"] == 16_000
+ assert "max_completion_tokens" not in loaded_lm.dump_state()
+ assert loaded_lm.dump_state()["max_tokens"] == 16_000
+
+
+def test_lm_load_state_forwards_allow_custom_lm_class(monkeypatch):
+ calls = []
+ original_load_state = dspy.BaseLM.load_state.__func__
+
+ def spy_load_state(cls, state, *, allow_custom_lm_class=False):
+ calls.append(allow_custom_lm_class)
+ return original_load_state(cls, state, allow_custom_lm_class=allow_custom_lm_class)
+
+ monkeypatch.setattr(dspy.BaseLM, "load_state", classmethod(spy_load_state))
+
+ dspy.LM.load_state(dspy.LM("openai/gpt-4o-mini").dump_state(), allow_custom_lm_class=True)
+
+ assert calls == [True]
+
+
def test_exponential_backoff_retry():
time_counter = []