Feature/add support for global defaults

[vladimirrotariu]

Adds global retry defaults for LM calls and docs/tutorial updates.

What’s included

• Settings: default_num_retries, retry_strategy
• LM: falls back to global default when num_retries=None
• LiteLLM: retry_strategy forwarded in chat/text (sync/async) paths
• Tests for fallback/override/strategy passthrough
• Docs: cheatsheet paragraph; tutorial notebook (global_retries)

Example

import dspy

dspy.configure(default_num_retries=4, retry_strategy="exponential_backoff_retry")

lm_default = dspy.LM(model="openai/gpt-4o-mini", cache=False, num_retries=None)
print(lm_default.dump_state()["num_retries"])  # 4

with dspy.context(default_num_retries=6):
    lm_scoped = dspy.LM(model="openai/gpt-4o-mini", cache=False, num_retries=None)
    print(lm_scoped.dump_state()["num_retries"])  # 6

lm_override = dspy.LM(model="openai/gpt-4o-mini", cache=False, num_retries=1)
print(lm_override.dump_state()["num_retries"])  # 1

Notes

• Backwards compatible; per-instance num_retries takes precedence

[vladimirrotariu]

@okhat @chenmoneygithub could you please have a look at this?

[okhat]

Thanks @vladimirrotariu !

My take is that retry_strategy is a very low-level concept. I'd remove that from the global settings.

And I'd rename default_num_retries to lm_num_retries.

That said, my hunch is that these add some complexity to reasoning about the default LM settings and don't serve a huge value (it's easy to just configure them per LM or to create an LM wrapper in your program that sets those defaults).

Do you think we really need these?