fix: LLM, embedder, and memory framework adapters ignore do_auto_retry=False

Version

develop (and v1.9.0-rc3)

Which installation method(s) does this occur on?

Source

Describe the bug.

RetryMixin provides do_auto_retry: bool = Field(default=True, description="Whether to automatically retry method calls that fail with a retryable error.", exclude=True) and num_retries: int = Field(default=5, gt=0, ...).

Because num_retries validates gt=0, users cannot disable retries by specifying num_retries: 0. Setting do_auto_retry: false in workflow configurations is the documented and tested mechanism to disable retries (as tested in test_retry_mixin.py: test_disabling_retries_uses_do_auto_retry).

While evaluator integrations (langsmith_judge.py, trajectory_evaluator.py, ragas, remote_workflow.py) properly check if config.do_auto_retry:, all runtime framework adapters for LLMs, embedders, and memory editors only check:

if isinstance(llm_config, RetryMixin):
    client = patch_with_retry(...)

(or in nvidia_nat_memmachine, call patch_with_retry unconditionally without any check).

Because all provider configs (OpenAIModelConfig, NIMModelConfig, AzureOpenAIModelConfig, AWSBedrockModelConfig, etc.) inherit from RetryMixin, isinstance(..., RetryMixin) is always True. As a result, setting do_auto_retry: false in YAML or configuration is completely ignored across:

• nvidia_nat_langchain (LLMs & embedders)
• nvidia_nat_llama_index (LLMs & embedders)
• nvidia_nat_agno (LLMs)
• nvidia_nat_autogen (LLMs)
• nvidia_nat_crewai (LLMs)
• nvidia_nat_semantic_kernel (LLMs)
• nvidia_nat_strands (LLMs)
• nvidia_nat_mem0ai (memory)
• nvidia_nat_memmachine (memory)
• nvidia_nat_zep_cloud (memory)

When a retryable error occurs (e.g., HTTP 429, 500, 503), the client retries 5 times with exponential backoff instead of failing immediately as requested by the user.

Expected Behavior

When do_auto_retry: false is configured, client wrappers should not wrap the instance with patch_with_retry, allowing immediate failure and honoring user configuration.

Minimum reproducible example

import asyncio
from unittest.mock import MagicMock, patch
from nat.llm.openai_llm import OpenAIModelConfig
from nat.plugins.langchain.llm import openai_langchain

class DummyChat:
    def __init__(self):
        self.attempts = 0
        self.model_kwargs = {}
    async def ainvoke(self, *args, **kwargs):
        self.attempts += 1
        print(f"Attempt {self.attempts}")
        raise RuntimeError("429 Too Many Requests")

async def main():
    cfg = OpenAIModelConfig(model_name="gpt-4o-mini", do_auto_retry=False)
    mock_builder = MagicMock()
    dummy = DummyChat()

    with patch("langchain_openai.ChatOpenAI", return_value=dummy):
        async with openai_langchain(cfg, mock_builder) as client:
            try:
                await client.ainvoke("hello")
            except Exception:
                print(f"Total attempts made: {dummy.attempts}")

asyncio.run(main())

Relevant log output

Attempt 1
Attempt 2
Attempt 3
Attempt 4
Attempt 5
Total attempts made: 5

(Expected: 1 attempt, but 5 attempts were executed because patch_with_retry was applied despite do_auto_retry=False.)

Other/Misc.

I have verified the fix and will submit a PR linking to this issue.

Code of Conduct

• I agree to follow the NeMo Agent Toolkit Code of Conduct
• I have searched the open bugs and have found no duplicates for this bug report

Start with test_retry_mixin.py and trace patch_with_retry through the runtime adapters named in the issue, including the LLM, embedder, and memory integrations. Compare them with the evaluator integrations that already inspect do_auto_retry. Done means each listed adapter honors do_auto_retry: false and retry behavior is covered by tests.