Add vLLM Lora testing scripts (#1829)

* Add vLLM Lora testing scripts
* Address code review feedback
---------

Signed-off-by: Jiaxin Shan <seedjeffwan@gmail.com>

Author: Jeffwan

test/integration/engine/lora/README.md

@@ -0,0 +1,138 @@
+# vLLM LoRA Feature Test Suite
+
+This test suite validates vLLM's LoRA adapter memory management, including LRU eviction, CPU-GPU swapping, pinning, and API behaviors.
+
+## Prerequisites
+
+- Python 3.10+
+- vLLM installed with LoRA support
+- GPU with sufficient memory (recommended: 24GB+ for Qwen3-8B)
+- Required packages: `pytest`, `pytest-asyncio`, `torch`, `safetensors`, `transformers`, `vllm`
+
+```bash
+pip install pytest pytest-asyncio torch safetensors transformers vllm
+```
+
+## Quick Start
+
+```bash
+# 1. Generate test LoRA adapters
+python create_test_loras.py --base-model Qwen/Qwen3-8B --num-loras 10 --ranks 8,16,32
+
+# 2. Run all tests
+python run_all_tests.py
+```
+
+## Test Coverage
+
+The test suite validates:
+- LRU eviction behavior
+- CPU-GPU swap latency
+- LoRA pinning functionality
+- Memory isolation between LoRA and KV cache
+- Memory budget and pre-allocation
+- Batching constraints with max_loras
+- Concurrent request handling
+- API idempotency
+
+
+## Artifact Preparation
+
+### Option 1: Generate Dummy LoRAs (Recommended for Testing)
+
+Generate dummy LoRA adapters that match your base model's architecture:
+
+```bash
+# Basic usage - creates 10 LoRAs with rank 8
+python create_test_loras.py --base-model Qwen/Qwen3-8B
+
+# Multiple ranks for swap latency testing
+python create_test_loras.py --base-model Qwen/Qwen3-8B --num-loras 10 --ranks 8,16,32
+
+# Include MLP modules for larger LoRAs
+python create_test_loras.py --base-model Qwen/Qwen3-8B \
+    --target-modules q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+
+# Custom output directory
+python create_test_loras.py --base-model Qwen/Qwen3-8B \
+    --output-dir /path/to/loras
+```
+
+#### Estimate LoRA Sizes
+
+```bash
+python create_test_loras.py --base-model Qwen/Qwen3-8B --ranks 8,16,32,64 --estimate
+```
+
+#### Verify Generated Adapters
+
+```bash
+python create_test_loras.py --verify --output-dir /tmp/test_loras
+```
+
+### Option 2: Use Real LoRA from HuggingFace
+
+Edit `test_config.py` to use a real LoRA adapter:
+
+```python
+# In test_config.py
+LORA_MODEL = "your-org/your-lora-adapter"
+```
+
+The test suite automatically falls back to `LORA_MODEL` if generated LoRAs are not found.
+
+## Running Tests
+
+### Run All Tests
+
+```bash
+python run_all_tests.py
+```
+
+Options:
+- `--timeout 600` - Timeout per test file in seconds (default: 600)
+- `--quiet` - Reduce output verbosity
+
+### Run Specific Tests with pytest
+
+```bash
+# All tests in a file
+pytest test_lru_swap_pinning.py -v -s
+
+# Specific test class
+pytest test_memory.py::TestMemoryBudget -v -s
+
+# Specific test method
+pytest test_api.py::TestDynamicLoading::test_reload_same_lora -v -s
+
+# Run tests matching a keyword
+pytest -v -s -k "eviction"
+```
+
+
+## Configuration
+
+Edit `test_config.py` to customize:
+
+```python
+# Base model
+BASE_MODEL = "Qwen/Qwen3-8B"
+
+# Generated LoRA directory
+LORA_BASE_PATH = "/tmp/test_loras"
+
+# Fallback LoRA (if generated not found)
+LORA_MODEL = "mtzig/qwen3-8b-tfdark-lora2"
+
+# Test parameters
+TEST_PROMPT = "Hello, how are you today?"
+MAX_TOKENS = 50
+```
+
+## Troubleshooting
+
+### LoRA Loading Fails
+
+1. Verify LoRA adapters exist: `python create_test_loras.py --verify`
+2. Check base model compatibility
+3. Ensure `max_lora_rank` >= actual LoRA rank