fix(clients): pass num_retries and retry_strategy to streaming LLM calls

[abdelhadi703]

Summary

Fix streaming LLM calls to respect num_retries and retry_strategy parameters.

Problem: dspy.streamify() uses _get_stream_completion_fn() which calls litellm.acompletion(stream=True) without num_retries or retry_strategy. Rate limit errors (429) crash immediately instead of retrying with exponential backoff.

Fix:

1. Add num_retries parameter to _get_stream_completion_fn()
2. Forward num_retries and retry_strategy="exponential_backoff_retry" to litellm.acompletion(stream=True)
3. Update both litellm_completion() and alitellm_completion() to pass num_retries
4. Bonus fix: alitellm_completion() was also missing headers — now added

Changes

• dspy/clients/lm.py:
  • _get_stream_completion_fn: added num_retries=0 parameter
  • stream_completion() closure: pass num_retries and retry_strategy to litellm.acompletion(stream=True)
  • litellm_completion: pass num_retries to _get_stream_completion_fn
  • alitellm_completion: pass num_retries and headers to _get_stream_completion_fn

Reproduction (from issue)

import dspy
lm = dspy.LM("anthropic/claude-opus-4-6", num_retries=5)
dspy.configure(lm=lm)

# Without streamify — retries work ✅
result = module(question="What is 2+2?")

# With streamify — now retries work too ✅
streamed = dspy.streamify(module)
result = streamed(question="What is 2+2?")

Security/Reliability

• No new dependencies
• Graceful degradation: retry_strategy=None when num_retries=0 preserves existing behavior
• Matches pattern used in non-streaming code path

---

Contribution by abdelhadisalmaoui0909@outlook.fr