LLM Client Configuration

NeMo Curator’s synthetic data generation uses OpenAI-compatible clients to communicate with LLM inference servers. This guide covers client configuration, performance tuning, and integration with various endpoints.

Overview

Two client types are available:

• AsyncOpenAIClient: Recommended for high-throughput batch processing with concurrent requests
• OpenAIClient: Synchronous client for simpler use cases or debugging

For most SDG workloads, use AsyncOpenAIClient to maximize throughput.

Basic Configuration

NVIDIA API Endpoints

from nemo_curator.models.client.openai_client import AsyncOpenAIClient

client = AsyncOpenAIClient(
    api_key="your-nvidia-api-key",  # Or use NVIDIA_API_KEY env var
    base_url="https://integrate.api.nvidia.com/v1",
    max_concurrent_requests=5,
)

Environment Variables

Set your API key as an environment variable to avoid hardcoding credentials:

export NVIDIA_API_KEY="nvapi-..."

The underlying OpenAI client automatically uses the OPENAI_API_KEY environment variable if no api_key is provided. For NVIDIA APIs, explicitly pass the key:

import os

client = AsyncOpenAIClient(
    api_key=os.environ["NVIDIA_API_KEY"],
    base_url="https://integrate.api.nvidia.com/v1",
)

Generation Parameters

Configure LLM generation behavior using GenerationConfig:

from nemo_curator.models.client.llm_client import GenerationConfig

config = GenerationConfig(
    max_tokens=2048,
    temperature=0.7,
    top_p=0.95,
    seed=42,  # For reproducibility
)

Performance Tuning

Concurrency vs. Parallelism

The max_concurrent_requests parameter controls how many API requests the client can have in-flight simultaneously. This interacts with Ray’s distributed workers:

• Client-level concurrency: max_concurrent_requests limits concurrent API calls per worker
• Worker-level parallelism: Ray distributes tasks across multiple workers

# For NVIDIA API endpoints with rate limits
client = AsyncOpenAIClient(
    base_url="https://integrate.api.nvidia.com/v1",
    max_concurrent_requests=3,  # Conservative for cloud APIs
)

Retry Configuration

The client includes automatic retry with exponential backoff for transient errors:

client = AsyncOpenAIClient(
    base_url="https://integrate.api.nvidia.com/v1",
    max_retries=3,        # Number of retry attempts
    base_delay=1.0,       # Base delay in seconds
    timeout=120,          # Request timeout
)

The retry logic handles:

• Rate limit errors (429): Automatic backoff with jitter
• Connection errors: Retry with exponential delay
• Transient failures: Configurable retry attempts

Using Other OpenAI-Compatible Endpoints

The AsyncOpenAIClient works with any OpenAI-compatible API endpoint. Simply configure the base_url and api_key parameters:

# OpenAI API
client = AsyncOpenAIClient(
    base_url="https://api.openai.com/v1",
    api_key="sk-...",  # Or set OPENAI_API_KEY env var
    max_concurrent_requests=5,
)

# Any OpenAI-compatible endpoint
client = AsyncOpenAIClient(
    base_url="http://your-endpoint/v1",
    api_key="your-api-key",
    max_concurrent_requests=5,
)

Local Inference with InferenceServer

To serve models locally and connect them to AsyncOpenAIClient, use NeMo Curator’s built-in Inference Server (Ray Serve + vLLM):

from nemo_curator.core.serve import InferenceModelConfig, InferenceServer
from nemo_curator.models.client.openai_client import AsyncOpenAIClient

config = InferenceModelConfig(
    model_identifier="meta-llama/Llama-3-8B-Instruct",
    engine_kwargs={"tensor_parallel_size": 2},
)

with InferenceServer(models=[config]) as server:
    client = AsyncOpenAIClient(
        base_url=server.endpoint,
        api_key="unused",
        max_concurrent_requests=10,
    )
    # Use client in pipeline stages

Complete Example

import os
from nemo_curator.models.client.openai_client import AsyncOpenAIClient
from nemo_curator.models.client.llm_client import GenerationConfig
from nemo_curator.pipeline import Pipeline
from nemo_curator.stages.synthetic.qa_multilingual_synthetic import QAMultilingualSyntheticStage

# Configure client
client = AsyncOpenAIClient(
    api_key=os.environ.get("NVIDIA_API_KEY"),
    base_url="https://integrate.api.nvidia.com/v1",
    max_concurrent_requests=5,
    max_retries=3,
    base_delay=1.0,
)

# Configure generation
config = GenerationConfig(
    temperature=0.9,
    top_p=0.95,
    max_tokens=2048,
)

# Use in a pipeline stage
pipeline = Pipeline(name="sdg_example")
pipeline.add_stage(
    QAMultilingualSyntheticStage(
        prompt="Generate a Q&A pair about science in {language}.",
        languages=["English", "French", "German"],
        client=client,
        model_name="meta/llama-3.3-70b-instruct",
        num_samples=100,
        generation_config=config,
    )
)

Troubleshooting

Rate Limit Errors

If you encounter frequent 429 errors:

1. Reduce max_concurrent_requests
2. Increase base_delay for longer backoff
3. Consider using a local deployment for high-volume workloads

Connection Timeouts

For slow networks or high-latency endpoints:

client = AsyncOpenAIClient(
    base_url="...",
    timeout=300,  # Increase from default 120 seconds
)

Next Steps

• Multilingual Q&A: Generate multilingual Q&A pairs
• Nemotron-CC: Advanced text transformation pipelines