TensorRT-LLMs/tests/integration/test_lists/qa

Description

This folder contains QA test definitions for TensorRT-LLM, which are executed on a daily/release schedule. These tests focus on end-to-end validation, accuracy verification, disaggregated testing, and performance benchmarking.

Test Categories

QA tests are organized into three main categories:

1. Functional Tests

Functional tests include E2E (end-to-end), accuracy, and disaggregated test cases:

• E2E Tests: Complete workflow validation from model loading to inference output
• Accuracy Tests: Model accuracy verification against reference implementations
• Disaggregated Tests: Distributed deployment and multi-node scenario validation

2. Performance Tests

Performance tests focus on benchmarking and performance validation:

• Baseline performance measurements
• Performance regression detection
• Throughput and latency benchmarking
• Resource utilization analysis

3. Triton Backend Tests

Triton backend tests validate the integration with NVIDIA Triton Inference Server:

• Backend functionality validation
• Model serving capabilities
• API compatibility testing
• Integration performance testing

Dependencies

The following Python packages are required for running QA tests:

pip install mako oyaml rouge_score lm_eval

Dependency Details

• mako: Template engine for test generation and configuration
• oyaml: YAML parser with ordered dictionary support
• rouge_score: ROUGE evaluation metrics for text generation quality assessment
• lm_eval: Language model evaluation framework

Test Files

This directory contains various test configuration files:

Functional Test Lists

• llm_function_full.txt - Primary test list for single node multi-GPU scenarios (all new test cases should be added here)
• llm_function_sanity.txt - Subset of examples for quick torch flow validation
• llm_function_nim.txt - NIM-specific functional test cases
• llm_function_multinode.txt - Multi-node functional test cases
• llm_function_gb20x.txt - GB20X release test cases
• llm_function_rtx6kd.txt - RTX 6000 Ada specific tests

Performance Test Files

• llm_perf_full.yml - Main performance test configuration
• llm_perf_cluster.yml - Cluster-based performance tests
• llm_perf_sanity.yml - Performance sanity checks
• llm_perf_nim.yml - NIM-specific performance tests
• llm_trt_integration_perf.yml - Integration performance tests
• llm_trt_integration_perf_sanity.yml - Integration performance sanity checks

Triton Backend Tests

• llm_triton_integration.txt - Triton backend integration tests

Release-Specific Tests

• llm_digits_func.txt - Functional tests for DIGITS release
• llm_digits_perf.txt - Performance tests for DIGITS release

Test Execution Schedule

QA tests are executed on a regular schedule:

• Weekly: Automated regression testing
• Release: Comprehensive validation before each release
• On-demand: Manual execution for specific validation needs

Running Tests

Manual Execution

To run specific test categories:

# direct to defs folder
cd tests/integration/defs
# Run all fp8 functional test
pytest --no-header -vs --test-list=../test_lists/qa/llm_function_full.txt -k fp8
# Run a single test case
pytest -vs accuracy/test_cli_flow.py::TestLlama3_1_8B::test_auto_dtype

Automated Execution

QA tests are typically executed through CI/CD pipelines with appropriate test selection based on:

• Release requirements
• Hardware availability
• Test priority and scope

Test Guidelines

Adding New Test Cases

• Primary Location: For functional testing, new test cases should be added to llm_function_full.txt first
• Categorization: Test cases should be categorized based on their scope and execution time
• Validation: Ensure test cases are properly validated before adding to any test list