[Bug] UnsupportedParamsError with COPRO Optimizer and Amazon Bedrock

[Nasreddine]

What happened?

I'm encountering an UnsupportedParamsError when using the COPRO optimizer with Amazon Bedrock models:

• mistral.mistral-small-2402-v1:0
• us.meta.llama3-2-3b-instruct-v1:0

The error occurs due to unsupported parameter {'n': 9}.

At the beginig I get this warning :

2025/04/11 12:02:47 WARNING dspy.adapters.json_adapter: Failed to use structured output format. Falling back to JSON mode. Error: litellm.UnsupportedParamsError: bedrock does not support parameters: {'n': 9}, for model=mistral.mistral-small-2402-v1:0. To drop these, set litellm.drop_params=True or for proxy:
litellm_settings: drop_params: true

Then I get following exception error:

File ~/dev/work/playground/.venv/lib/python3.10/site-packages/dspy/adapters/json_adapter.py:69, in JSONAdapter.call(self, lm, lm_kwargs, signature, demos, inputs)
67 return super().call(lm, lm_kwargs, signature, demos, inputs)
68 except Exception as e:
---> 69 raise RuntimeError(
70 "Both structured output format and JSON mode failed. Please choose a model that supports "
71 f"response_format argument. Original error: {e}"
72 ) from e

RuntimeError: Both structured output format and JSON mode failed. Please choose a model that supports response_format argument. Original error: litellm.UnsupportedParamsError: bedrock does not support parameters: {'n': 9}, for model=mistral.mistral-small-2402-v1:0. To drop these, set litellm.drop_params=True or for proxy:

litellm_settings: drop_params: true

Steps to reproduce

Configure and compile COPRO optimizer with Bedrock.

llm = dspy.LM(model="mistral.mistral-small-2402-v1:0")
dspy.settings.configure(lm=llm)

from dspy.teleprompt import COPRO

teleprompter = COPRO(
    metric=my_metric,
    verbose=True,
)

kwargs = dict(display_progress=True, display_table=1)

compiled_prompt_opt = teleprompter.compile(my_pipline, trainset=my_trainset, eval_kwargs=kwargs)

DSPy version

2.6.17

[arnavsinghvi11]

Hey @Nasreddine , the error trace states you need to drop the n parameter to work with Bedrock models To drop these, set litellm.drop_params=True
Feel free to reference the LiteLLM AWS Bedrock guide and issues in the LiteLLM repo for anything Bedrock related here.

[okhat]

@Nasreddine Why use COPRO btw? Consider more modern optimizers like dspy.SIMBA or MIPRO.

[Nasreddine]

@arnavsinghvi11 I believe LiteLLM is encapsulated within DSPy. Is there a way to pass parameters to it?

@okhat I’m currently testing all the optimizers to determine which one performs best for our task. I’ve already completed testing for MIPRO.