# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: Apache-2.0 # Timing notes (measured locally): # - GPU-1 subset (`-m "gpu_1 and not gpu_2"`): 130.43s total for 3 tests. # These tests load a real model and can be slow/flaky when GPU resources are contended, # so we set explicit pytest timeouts to fail fast on hangs (see per-test markers below). import logging import os import time from typing import Any, Dict, Optional import pytest from tests.router.common import ( # utilities _test_router_basic, _test_router_decisions, _test_router_indexers_sync, generate_random_suffix, get_runtime, ) from tests.utils.constants import DefaultPort from tests.utils.managed_process import ManagedProcess from tests.utils.port_utils import allocate_ports, deallocate_ports logger = logging.getLogger(__name__) MODEL_NAME = "TinyLlama/TinyLlama-1.1B-Chat-v1.0" pytestmark = [ pytest.mark.e2e, pytest.mark.vllm, pytest.mark.model(MODEL_NAME), ] SPEEDUP_RATIO = 10.0 NUM_REQUESTS = 10 BLOCK_SIZE = 16 def allocate_frontend_ports(request, count: int) -> list[int]: """Allocate random free frontend ports for xdist-safe execution.""" ports = allocate_ports(count, DefaultPort.FRONTEND.value) request.addfinalizer(lambda: deallocate_ports(ports)) return ports # Shared test payload for all tests TEST_PAYLOAD: Dict[str, Any] = { "model": MODEL_NAME, "messages": [ { "role": "user", "content": "In a quiet meadow tucked between rolling hills, a plump gray rabbit nibbled on clover beneath the shade of a gnarled oak tree. Its ears twitched at the faint rustle of leaves, but it remained calm, confident in the safety of its burrow just a few hops away. The late afternoon sun warmed its fur, and tiny dust motes danced in the golden light as bees hummed lazily nearby. Though the rabbit lived a simple life, every day was an adventure of scents, shadows, and snacks—an endless search for the tastiest patch of greens and the softest spot to nap.", } ], "stream": True, "max_tokens": 10, } # Shared vLLM configuration for all tests # gpu_memory_utilization limits actual VRAM allocation (required for multi-worker on same GPU) VLLM_ARGS: Dict[str, Any] = { "block_size": BLOCK_SIZE, "model": MODEL_NAME, "gpu_memory_utilization": 0.4, # Limit VRAM allocation per worker "max_model_len": 1024, # Limit context length to reduce KV cache size "enforce_eager": True, # Disable CUDA graphs for faster startup & lower memory } class VLLMProcess: """Manages vLLM workers using dynamo.vllm (HTTP API + KV events). This is a drop-in replacement for MockerProcess that uses real vLLM workers. The key difference: dynamo.vllm automatically handles: - HTTP API serving - KV cache event publishing (ZMQ → NATS bridge) - Integration with dynamo.frontend router """ def __init__( self, request, vllm_args: Optional[Dict[str, Any]] = None, num_workers: int = 2, single_gpu: bool = False, data_parallel_size: Optional[int] = None, request_plane: str = "tcp", store_backend: str = "etcd", durable_kv_events: bool = False, ): """Initialize vLLM workers with dynamo integration. Args: request: pytest request fixture for log directory vllm_args: Configuration dict with keys: - block_size: KV cache block size (default: 16) - model: Model name/path (default: TinyLlama-1.1B) - gpu_memory_utilization: Fraction of GPU memory to allocate (optional) - num_gpu_blocks_override: Cap on number of KV cache blocks (optional) - max_model_len: Maximum sequence length (optional) - enforce_eager: Disable CUDA graphs (default: False) num_workers: Number of vLLM worker processes single_gpu: If True, all workers share GPU 0 data_parallel_size: If set, enables data parallelism with this many ranks (num_workers must equal data_parallel_size) request_plane: Request plane to use ("nats", "tcp", or "http"). Defaults to "tcp". store_backend: Storage backend to use ("etcd" or "file"). Defaults to "etcd". durable_kv_events: If True, use JetStream for durable KV events. Defaults to False (NATS Core mode). """ # Generate unique namespace for isolation namespace_suffix = generate_random_suffix() self.namespace = f"test-namespace-{namespace_suffix}" self.component_name = "backend" self.endpoint = f"dyn://{self.namespace}.{self.component_name}.generate" self.num_workers = num_workers self.data_parallel_size = data_parallel_size self.worker_processes = [] self.store_backend = store_backend if vllm_args is None: vllm_args = {} block_size = vllm_args.get("block_size", BLOCK_SIZE) model = vllm_args.get("model", MODEL_NAME) gpu_memory_utilization = vllm_args.get("gpu_memory_utilization") num_gpu_blocks_override = vllm_args.get("num_gpu_blocks_override") max_model_len = vllm_args.get("max_model_len") enforce_eager = vllm_args.get("enforce_eager", False) self.model_name = model # Create vLLM worker processes # Matches test.sh behavior: # - When data_parallel_size is set, launch one process per DP rank # - Each process gets --data-parallel-rank and --data-parallel-size # - Each process runs on its own GPU via CUDA_VISIBLE_DEVICES # - --connector nixl enables KV cache transfer between ranks for worker_idx in range(num_workers): # Calculate GPU device for this process if single_gpu: # Force all processes to GPU 0 (for single-GPU testing) gpu_device = "0" elif data_parallel_size is not None: # Worker sees dp_rank GPUs (each DP rank gets its own GPU) worker_start_gpu = worker_idx * data_parallel_size gpu_device = ",".join( str(i) for i in range( worker_start_gpu, worker_start_gpu + data_parallel_size ) ) else: # No DP; worker sees one GPU gpu_device = str(worker_idx) command = [ "python3", "-m", "dynamo.vllm", "--model", model, "--block-size", str(block_size), ] # Disable CUDA graphs for faster startup & lower memory if enforce_eager: command.append("--enforce-eager") # Limit VRAM allocation (required for multi-worker on same GPU) if gpu_memory_utilization is not None: command.extend( ["--gpu-memory-utilization", str(gpu_memory_utilization)] ) # Add optional max_model_len if specified if max_model_len is not None: command.extend(["--max-model-len", str(max_model_len)]) # Cap block count for predictable KV cache behavior if num_gpu_blocks_override is not None: command.extend( ["--num-gpu-blocks-override", str(num_gpu_blocks_override)] ) if data_parallel_size is not None: # Add DP configuration for external load balancing # See: https://docs.vllm.ai/en/v0.10.0/serving/data_parallel_deployment.html#external-load-balancing command.extend( [ "--data-parallel-size", str(data_parallel_size), # "--data-parallel-address", "127.0.0.1", # Required for DP coordination # "--data-parallel-rpc-port", "13345", # RPC port for DP coordination # "--connector", "nixl", # Required for KV transfer between DP ranks ] ) # Use --durable-kv-events to enable JetStream mode (local indexer disabled) if durable_kv_events: command.append("--durable-kv-events") env = os.environ.copy() # Copy parent environment env_vars = { "CUDA_VISIBLE_DEVICES": gpu_device, "DYN_NAMESPACE": self.namespace, "DYN_REQUEST_PLANE": request_plane, "DYN_VLLM_KV_EVENT_PORT": str(20080 + worker_idx), "VLLM_NIXL_SIDE_CHANNEL_PORT": str(20090 + worker_idx), "PYTHONHASHSEED": "0", # for deterministic event id's } # Add DYN_FILE_KV if using file storage backend if self.store_backend == "file" and "DYN_FILE_KV" in os.environ: env_vars["DYN_FILE_KV"] = os.environ["DYN_FILE_KV"] env.update(env_vars) # Create managed process for the worker process = ManagedProcess( command=command, env=env, timeout=120, # Allow time for model loading display_output=True, health_check_ports=[], health_check_urls=[], log_dir=request.node.name, terminate_all_matching_process_names=False, ) self.worker_processes.append(process) if data_parallel_size is not None: logger.info( f"Created {data_parallel_size} DP ranks per worker on GPU(s) {gpu_device} " f"(gpu_mem={gpu_memory_utilization}) " f"with endpoint: {self.endpoint}" ) else: logger.info( f"Created vLLM worker {worker_idx} on GPU {gpu_device} " f"(gpu_mem={gpu_memory_utilization}) " f"with endpoint: {self.endpoint}" ) def __enter__(self): """Start all vLLM worker processes with sequential initialization. Workers are started sequentially with a delay between each to avoid NIXL/UCX resource contention during initialization. This prevents UCX shared memory handle allocation failures when multiple workers try to initialize simultaneously on the same GPU. """ logger.info( f"[VLLMProcess] Starting {len(self.worker_processes)} worker processes sequentially..." ) # Start each process sequentially, waiting for NIXL initialization before next for i, process in enumerate(self.worker_processes): logger.info(f"[VLLMProcess] Starting vLLM worker {i}...") try: # Manually initialize the process without blocking on health checks process._logger = logging.getLogger(process.__class__.__name__) process._command_name = process.command[0] os.makedirs(process.log_dir, exist_ok=True) log_name = f"{process._command_name}.log.txt" process._log_path = os.path.join(process.log_dir, log_name) if process.data_dir: process._remove_directory(process.data_dir) process._terminate_all_matching_process_names() logger.info( f"[VLLMProcess] Launching process {i} (pid will be assigned)..." ) process._start_process() # Start the process but don't wait logger.info( f"[VLLMProcess] Worker {i} launched with PID: {process.proc.pid if process.proc else 'unknown'}" ) time.sleep(process.delayed_start) # Wait for NIXL initialization before starting next worker # This prevents UCX shared memory contention if i < len(self.worker_processes) - 1: nixl_init_delay = 5 # seconds logger.info( f"[VLLMProcess] Waiting {nixl_init_delay}s for worker {i} to initialize NIXL before starting next worker..." ) time.sleep(nixl_init_delay) except Exception: logger.exception(f"[VLLMProcess] Failed to start worker {i}") # Clean up on failure try: process.__exit__(None, None, None) except Exception as cleanup_err: logger.warning(f"[VLLMProcess] Error during cleanup: {cleanup_err}") raise logger.info( f"[VLLMProcess] All {len(self.worker_processes)} workers launched with sequential initialization." ) logger.info("[VLLMProcess] Waiting for health checks to complete...") # Now wait for health checks for all processes for i, process in enumerate(self.worker_processes): logger.info(f"[VLLMProcess] Checking health for worker {i}...") try: elapsed = process._check_ports(process.timeout) process._check_urls(process.timeout - elapsed) process._check_funcs(process.timeout - elapsed) logger.info(f"[VLLMProcess] Worker {i} health checks passed") except Exception: logger.error(f"[VLLMProcess] Worker {i} health check failed") # Clean up all processes on failure self.__exit__(None, None, None) raise logger.info( "[VLLMProcess] All workers started successfully and passed health checks!" ) return self def __exit__(self, exc_type, exc_val, exc_tb): """Stop all vLLM worker processes gracefully.""" for i, process in enumerate(self.worker_processes): logger.info(f"Stopping vLLM worker {i}") process.__exit__(exc_type, exc_val, exc_tb) # Add delay to ensure full cleanup of NATS/ETCD/ZMQ resources # This prevents test isolation issues when running multiple tests logger.info("Waiting for vLLM worker resources to fully clean up...") time.sleep(2) @pytest.mark.pre_merge @pytest.mark.gpu_1 @pytest.mark.timeout(150) # ~3x average (~43s/test), rounded up @pytest.mark.parametrize("request_plane", ["tcp"], indirect=True) def test_vllm_kv_router_basic( request, runtime_services_dynamic_ports, predownload_models, set_ucx_tls_no_mm, request_plane, ): """ Quick e2e sanity test for KV router with vLLM engine instances. Tests both NATS and TCP request planes. """ # runtime_services starts etcd and nats N_VLLM_WORKERS = 2 logger.info( f"Starting vLLM KV router test with {N_VLLM_WORKERS} workers using request_plane={request_plane}" ) with VLLMProcess( request, vllm_args=VLLM_ARGS, num_workers=N_VLLM_WORKERS, single_gpu=True, # fit workers into one GPU request_plane=request_plane, ) as vllm_workers: # Start vLLM workers logger.info(f"Starting {N_VLLM_WORKERS} vLLM workers") logger.info(f"All vLLM workers using namespace: {vllm_workers.namespace}") # Run basic router test (starts router internally and waits for workers to be ready) frontend_port = allocate_frontend_ports(request, 1)[0] _test_router_basic( engine_workers=vllm_workers, block_size=BLOCK_SIZE, request=request, frontend_port=frontend_port, test_payload=TEST_PAYLOAD, num_requests=NUM_REQUESTS, frontend_timeout=180, # 3 minutes should be plenty for TinyLlama store_backend="etcd", # Explicit for clarity request_plane=request_plane, ) @pytest.mark.pre_merge @pytest.mark.gpu_1 @pytest.mark.timeout(150) # ~3x average (~43s/test), rounded up @pytest.mark.parametrize("request_plane", ["tcp"], indirect=True) def test_router_decisions_vllm_multiple_workers( request, runtime_services_dynamic_ports, predownload_models, set_ucx_tls_no_mm, request_plane, ): # runtime_services starts etcd and nats logger.info("Starting vLLM router prefix reuse test with two workers") N_WORKERS = 2 with VLLMProcess( request, vllm_args=VLLM_ARGS, num_workers=N_WORKERS, single_gpu=True, # Worker uses GPU 0 request_plane=request_plane, ) as vllm_workers: # Start 2 worker processes on the same GPU logger.info("Starting 2 vLLM worker processes on single GPU (gpu_mem=0.4)") logger.info(f"All vLLM workers using namespace: {vllm_workers.namespace}") # Get runtime and create endpoint runtime = get_runtime(request_plane=request_plane) namespace = runtime.namespace(vllm_workers.namespace) component = namespace.component("backend") endpoint = component.endpoint("generate") _test_router_decisions( vllm_workers, endpoint, MODEL_NAME, request, test_dp_rank=False ) @pytest.mark.gpu_2 @pytest.mark.nightly @pytest.mark.parametrize("request_plane", ["tcp"], indirect=True) @pytest.mark.timeout(600) # 10 min max (multi-GPU + DP startup variance) def test_router_decisions_vllm_dp( request, runtime_services_dynamic_ports, predownload_models, set_ucx_tls_no_mm, request_plane, ): """Validate KV cache prefix reuse with vLLM by sending progressive requests with overlapping prefixes. Same flow as test_router_decisions_vllm_multiple_workers; force first request to (worker_id, dp_rank=1). Dump events from router and verify: * All but one (worker_id, dp_rank) should have no events (due to prefix reuse) * The (worker_id, dp_rank) with events should have exactly 4 events (one per request) * All events should be on the forced (worker_id, dp_rank=1) (verifying forced routing and prefix reuse) """ N_WORKERS = 1 DP_SIZE = 2 with VLLMProcess( request, vllm_args=VLLM_ARGS, num_workers=N_WORKERS, # Ignored when data_parallel_size is set single_gpu=False, data_parallel_size=DP_SIZE, # Creates DP_SIZE processes (one per rank) request_plane=request_plane, ) as vllm_workers: logger.info("Starting 2 vLLM DP ranks (dp_size=2) (gpu_mem=0.4)") logger.info(f"All vLLM workers using namespace: {vllm_workers.namespace}") # Get runtime and create endpoint runtime = get_runtime(request_plane=request_plane) # Use the namespace from the vLLM workers namespace = runtime.namespace(vllm_workers.namespace) component = namespace.component("backend") # endpoint is backend.generate endpoint = component.endpoint("generate") _test_router_decisions( vllm_workers, endpoint, MODEL_NAME, request, test_dp_rank=True ) @pytest.mark.pre_merge @pytest.mark.gpu_1 @pytest.mark.timeout(150) # ~3x average (~43s/test), rounded up @pytest.mark.parametrize( "store_backend,durable_kv_events,request_plane", [ ("etcd", False, "tcp"), ], ids=["nats_core"], indirect=["durable_kv_events", "request_plane"], ) def test_vllm_indexers_sync( request, runtime_services_dynamic_ports, predownload_models, file_storage_backend, set_ucx_tls_no_mm, store_backend, durable_kv_events, request_plane, ): """ Test that two KV routers have synchronized indexer states after processing requests with vLLM workers. This test verifies that both routers converge to the same internal state. Tests with configuration: - nats_core: etcd backend, local indexer with NATS Core, TCP request plane (includes NATS interruption/recovery testing) """ # runtime_services_dynamic_ports handles NATS and etcd startup nats_process, _etcd_process = runtime_services_dynamic_ports logger.info( f"Starting vLLM indexers sync test: store_backend={store_backend}, " f"durable_kv_events={durable_kv_events}, request_plane={request_plane}" ) N_VLLM_WORKERS = 2 with VLLMProcess( request, vllm_args=VLLM_ARGS, num_workers=N_VLLM_WORKERS, single_gpu=True, # fit workers into one GPU request_plane=request_plane, store_backend=store_backend, durable_kv_events=durable_kv_events, ) as vllm_workers: # Start vLLM workers logger.info(f"Starting {N_VLLM_WORKERS} vLLM workers") logger.info(f"All vLLM workers using namespace: {vllm_workers.namespace}") # Use the common test implementation (creates its own runtimes for each router) # Note: Consumer verification is done inside _test_router_indexers_sync while routers are alive # When using durable_kv_events=True, use JetStream mode for the router _test_router_indexers_sync( engine_workers=vllm_workers, block_size=BLOCK_SIZE, model_name=MODEL_NAME, num_workers=N_VLLM_WORKERS, store_backend=store_backend, request_plane=request_plane, test_nats_interruption=not durable_kv_events, nats_server=nats_process if not durable_kv_events else None, durable_kv_events=durable_kv_events, ) logger.info("vLLM indexers sync test completed successfully")