| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import atexit |
| 5 | import contextlib |
| 6 | import copy |
| 7 | import functools |
| 8 | import importlib |
| 9 | import itertools |
| 10 | import json |
| 11 | import os |
| 12 | import random |
| 13 | import signal |
| 14 | import subprocess |
| 15 | import sys |
| 16 | import tempfile |
| 17 | import threading |
| 18 | import time |
| 19 | import warnings |
| 20 | from collections.abc import Callable, Iterable, MutableMapping, Sequence |
| 21 | from contextlib import ExitStack, contextmanager |
| 22 | from multiprocessing import Process, get_context |
| 23 | from pathlib import Path |
| 24 | from typing import TYPE_CHECKING, Any, Literal, cast |
| 25 | |
| 26 | import anthropic |
| 27 | import cloudpickle |
| 28 | import httpx |
| 29 | import openai |
| 30 | import pytest |
| 31 | import requests |
| 32 | import torch |
| 33 | import torch.nn.functional as F |
| 34 | from typing_extensions import ParamSpec |
| 35 | |
| 36 | import vllm.envs as envs |
| 37 | from vllm.distributed import ( |
| 38 | ensure_model_parallel_initialized, |
| 39 | init_distributed_environment, |
| 40 | ) |
| 41 | from vllm.engine.arg_utils import AsyncEngineArgs |
| 42 | from vllm.entrypoints.cli.serve import ServeSubcommand |
| 43 | from vllm.logger import init_logger |
| 44 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 45 | QuantKey, |
| 46 | ) |
| 47 | from vllm.model_executor.model_loader import get_model_loader |
| 48 | from vllm.platforms import current_platform |
| 49 | from vllm.tokenizers import get_tokenizer |
| 50 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 51 | from vllm.utils.mem_constants import GB_bytes |
| 52 | from vllm.utils.network_utils import get_open_port |
| 53 | from vllm.utils.torch_utils import ( |
| 54 | set_random_seed, # noqa: F401 - re-exported for use in test files |
| 55 | ) |
| 56 | from vllm.v1.engine.utils import get_engine_process_shutdown_timeout |
| 57 | |
| 58 | if TYPE_CHECKING: |
| 59 | from vllm.model_executor.kernels.linear import _KernelT |
| 60 | |
| 61 | logger = init_logger(__name__) |
| 62 | |
| 63 | FP8_DTYPE = current_platform.fp8_dtype() |
| 64 | |
| 65 | |
| 66 | if current_platform.is_cuda(): |
| 67 | from vllm.third_party.pynvml import ( |
| 68 | nvmlDeviceGetHandleByIndex, |
| 69 | nvmlDeviceGetHandleByUUID, |
| 70 | nvmlDeviceGetMemoryInfo, |
| 71 | nvmlInit, |
| 72 | nvmlShutdown, |
| 73 | ) |
| 74 | |
| 75 | @contextmanager |
| 76 | def _nvml(): |
| 77 | try: |
| 78 | nvmlInit() |
| 79 | yield |
| 80 | finally: |
| 81 | nvmlShutdown() |
| 82 | else: |
| 83 | |
| 84 | @contextmanager |
| 85 | def _nvml(): |
| 86 | yield |
| 87 | |
| 88 | |
| 89 | VLLM_PATH = Path(__file__).parent.parent |
| 90 | """Path to root of the vLLM repository.""" |
| 91 | |
| 92 | # ROCm: disable prefix caching and eliminate batch variance to reduce |
| 93 | # test flakiness. |
| 94 | ROCM_EXTRA_ARGS = ( |
| 95 | ["--no-enable-prefix-caching", "--max-num-seqs", "1"] |
| 96 | if current_platform.is_rocm() |
| 97 | else [] |
| 98 | ) |
| 99 | # Python-API equivalent of ROCM_EXTRA_ARGS for use with EngineArgs kwargs. |
| 100 | ROCM_ENGINE_KWARGS: dict = ( |
| 101 | {"enable_prefix_caching": False, "max_num_seqs": 1} |
| 102 | if current_platform.is_rocm() |
| 103 | else {} |
| 104 | ) |
| 105 | _TILELANG_TVM_PYTHONPATH_FRAGMENT = os.path.join( |
| 106 | "tilelang", "3rdparty", "tvm", "python" |
| 107 | ) |
| 108 | _SENSITIVE_CLI_ARG_NARGS = {"--api-key": "+", "--hf-token": "?"} |
| 109 | |
| 110 | |
| 111 | def _redact_sensitive_cli_args(args: Sequence[str]) -> list[str]: |
| 112 | redacted_args = list(args) |
| 113 | index = 0 |
| 114 | while index < len(args): |
| 115 | name, separator, _ = args[index].partition("=") |
| 116 | nargs = _SENSITIVE_CLI_ARG_NARGS.get(name.replace("_", "-")) |
| 117 | if nargs is None: |
| 118 | index += 1 |
| 119 | continue |
| 120 | if separator: |
| 121 | redacted_args[index] = f"{name}=***" |
| 122 | index += 1 |
| 123 | if not separator or nargs == "+": |
| 124 | while index < len(args) and not args[index].startswith("-"): |
| 125 | redacted_args[index] = "***" |
| 126 | index += 1 |
| 127 | if nargs == "?": |
| 128 | break |
| 129 | return redacted_args |
| 130 | |
| 131 | |
| 132 | def _sanitize_pythonpath_value(pythonpath: str | None) -> str: |
| 133 | if not pythonpath: |
| 134 | return "" |
| 135 | entries = [] |
| 136 | for entry in pythonpath.split(os.pathsep): |
| 137 | normalized = entry.replace(os.sep, "/") |
| 138 | if _TILELANG_TVM_PYTHONPATH_FRAGMENT.replace(os.sep, "/") in normalized: |
| 139 | continue |
| 140 | entries.append(entry) |
| 141 | return os.pathsep.join(entries) |
| 142 | |
| 143 | |
| 144 | def _sanitize_pythonpath_env(env: MutableMapping[str, str]) -> None: |
| 145 | cleaned = _sanitize_pythonpath_value(env.get("PYTHONPATH")) |
| 146 | if cleaned: |
| 147 | env["PYTHONPATH"] = cleaned |
| 148 | else: |
| 149 | env.pop("PYTHONPATH", None) |
| 150 | |
| 151 | |
| 152 | def _sanitize_current_pythonpath_env() -> None: |
| 153 | _sanitize_pythonpath_env(os.environ) |
| 154 | |
| 155 | |
| 156 | @contextmanager |
| 157 | def _temporarily_sanitized_pythonpath_env(): |
| 158 | original = os.environ.get("PYTHONPATH") |
| 159 | _sanitize_current_pythonpath_env() |
| 160 | try: |
| 161 | yield |
| 162 | finally: |
| 163 | if original is None: |
| 164 | os.environ.pop("PYTHONPATH", None) |
| 165 | else: |
| 166 | os.environ["PYTHONPATH"] = original |
| 167 | |
| 168 | |
| 169 | def requires_spawn_multiprocessing() -> bool: |
| 170 | """Whether this platform requires spawn instead of fork for test processes.""" |
| 171 | return current_platform.is_rocm() or current_platform.is_xpu() |
| 172 | |
| 173 | |
| 174 | def _run_in_new_process_group( |
| 175 | child_process_fxn: Callable[[dict[str, str] | None, str, list[str]], None], |
| 176 | env_dict: dict[str, str] | None, |
| 177 | model: str, |
| 178 | vllm_serve_args: list[str], |
| 179 | ) -> None: |
| 180 | os.setsid() |
| 181 | child_process_fxn(env_dict, model, vllm_serve_args) |
| 182 | |
| 183 | |
| 184 | class RemoteVLLMServer: |
| 185 | """Base class for launching vLLM server subprocesses for testing. |
| 186 | |
| 187 | Subclasses must override ``_create_cli_subcommand`` and |
| 188 | ``_start_server``. |
| 189 | """ |
| 190 | |
| 191 | DUMMY_API_KEY = "token-abc123" # vLLM's OpenAI server does not need API key |
| 192 | _active_servers: set["RemoteVLLMServer"] = set() |
| 193 | _active_servers_lock = threading.RLock() |
| 194 | _cleanup_hooks_registered = False |
| 195 | _signal_hooks_registered = False |
| 196 | _previous_signal_handlers: dict[int, Any] = {} |
| 197 | proc: subprocess.Popen |
| 198 | |
| 199 | def _create_cli_subcommand(self): |
| 200 | """Return a CLISubcommand instance used to parse CLI args.""" |
| 201 | raise NotImplementedError |
| 202 | |
| 203 | def _start_server( |
| 204 | self, model: str, vllm_serve_args: list[str], env_dict: dict[str, str] | None |
| 205 | ) -> None: |
| 206 | """Subclasses override this method to customize server process launch.""" |
| 207 | raise NotImplementedError |
| 208 | |
| 209 | def _pre_download_model(self, model: str, args) -> None: |
| 210 | """Download model weights before starting the server to avoid timeout.""" |
| 211 | is_local = os.path.isdir(model) |
| 212 | if not is_local: |
| 213 | engine_args = AsyncEngineArgs.from_cli_args(args) |
| 214 | model_config = engine_args.create_model_config() |
| 215 | load_config = engine_args.create_load_config() |
| 216 | |
| 217 | model_loader = get_model_loader(load_config) |
| 218 | model_loader.download_model(model_config) |
| 219 | |
| 220 | def _get_process_termination_timeout(self) -> float: |
| 221 | return 15.0 |
| 222 | |
| 223 | def __init__( |
| 224 | self, |
| 225 | model: str, |
| 226 | vllm_serve_args: list[str], |
| 227 | *, |
| 228 | env_dict: dict[str, str] | None = None, |
| 229 | seed: int = 0, |
| 230 | auto_port: bool = True, |
| 231 | max_wait_seconds: float | None = None, |
| 232 | override_hf_configs: dict[str, Any] | None = None, |
| 233 | ) -> None: |
| 234 | if auto_port: |
| 235 | if "-p" in vllm_serve_args or "--port" in vllm_serve_args: |
| 236 | raise ValueError( |
| 237 | "You have manually specified the port when `auto_port=True`." |
| 238 | ) |
| 239 | |
| 240 | # No need for a port if using unix sockets |
| 241 | if "--uds" not in vllm_serve_args: |
| 242 | # Don't mutate the input args |
| 243 | vllm_serve_args = vllm_serve_args + ["--port", str(get_open_port())] |
| 244 | if seed is not None: |
| 245 | if "--seed" in vllm_serve_args: |
| 246 | raise ValueError( |
| 247 | f"You have manually specified the seed when `seed={seed}`." |
| 248 | ) |
| 249 | |
| 250 | vllm_serve_args = vllm_serve_args + ["--seed", str(seed)] |
| 251 | |
| 252 | if override_hf_configs is not None: |
| 253 | vllm_serve_args = vllm_serve_args + [ |
| 254 | "--hf-overrides", |
| 255 | json.dumps(override_hf_configs), |
| 256 | ] |
| 257 | |
| 258 | parser = FlexibleArgumentParser(description="vLLM's remote server.") |
| 259 | subparsers = parser.add_subparsers(required=False, dest="subparser") |
| 260 | parser = self._create_cli_subcommand().subparser_init(subparsers) |
| 261 | args = parser.parse_args(["--model", model, *vllm_serve_args]) |
| 262 | self.uds = args.uds |
| 263 | if args.uds: |
| 264 | self.host = None |
| 265 | self.port = None |
| 266 | else: |
| 267 | self.host = str(args.host or "127.0.0.1") |
| 268 | self.port = int(args.port) |
| 269 | |
| 270 | self.show_hidden_metrics = ( |
| 271 | getattr(args, "show_hidden_metrics_for_version", None) is not None |
| 272 | ) |
| 273 | self._request_shutdown_timeout = float(args.shutdown_timeout) |
| 274 | |
| 275 | with _temporarily_sanitized_pythonpath_env(): |
| 276 | self._pre_download_model(model, args) |
| 277 | self._shutdown_complete = False |
| 278 | |
| 279 | # Record GPU memory before server start so we know what |
| 280 | # "released" looks like. |
| 281 | self._pre_server_gpu_memory = self._get_gpu_memory_used() |
| 282 | if self._pre_server_gpu_memory is not None: |
| 283 | pre_gb = self._pre_server_gpu_memory / 1e9 |
| 284 | print( |
| 285 | f"[{type(self).__name__}] GPU memory before server start: " |
| 286 | f"{pre_gb:.2f} GB" |
| 287 | ) |
| 288 | |
| 289 | self._start_server(model, vllm_serve_args, env_dict) |
| 290 | self._register_active_server() |
| 291 | max_wait_seconds = max_wait_seconds or 480 |
| 292 | try: |
| 293 | self._wait_for_server(url=self.url_for("health"), timeout=max_wait_seconds) |
| 294 | except Exception: |
| 295 | # If the server never became healthy, we must still clean up |
| 296 | # the subprocess tree. Without this, a timeout in __init__ |
| 297 | # leaks the server + EngineCore processes (and their GPU |
| 298 | # memory), because __exit__ is never called when __init__ |
| 299 | # raises inside a ``with`` statement. |
| 300 | self._shutdown() |
| 301 | raise |
| 302 | |
| 303 | def __enter__(self): |
| 304 | return self |
| 305 | |
| 306 | def __exit__(self, exc_type, exc_value, traceback): |
| 307 | self._shutdown() |
| 308 | |
| 309 | def _shutdown(self) -> None: |
| 310 | """Kill the server process tree and wait for GPU memory release. |
| 311 | |
| 312 | Called from both ``__exit__`` (normal path) and ``__init__`` |
| 313 | (when the server fails to start). Must be safe to call even if |
| 314 | the process is already dead. |
| 315 | """ |
| 316 | if self._shutdown_complete: |
| 317 | return |
| 318 | |
| 319 | self._shutdown_complete = True |
| 320 | try: |
| 321 | self._terminate_process_tree() |
| 322 | self._wait_for_gpu_memory_release() |
| 323 | finally: |
| 324 | self._unregister_active_server() |
| 325 | |
| 326 | @classmethod |
| 327 | def _ensure_cleanup_hooks_registered(cls) -> None: |
| 328 | """Register process-exit cleanup for detached server subprocesses.""" |
| 329 | root_cls = RemoteVLLMServer |
| 330 | with root_cls._active_servers_lock: |
| 331 | if not root_cls._cleanup_hooks_registered: |
| 332 | atexit.register(root_cls._shutdown_active_servers) |
| 333 | root_cls._cleanup_hooks_registered = True |
| 334 | |
| 335 | if ( |
| 336 | threading.current_thread() is threading.main_thread() |
| 337 | and not root_cls._signal_hooks_registered |
| 338 | ): |
| 339 | for signum in (signal.SIGTERM, signal.SIGINT): |
| 340 | root_cls._previous_signal_handlers[signum] = signal.getsignal( |
| 341 | signum |
| 342 | ) |
| 343 | signal.signal(signum, root_cls._handle_parent_signal) |
| 344 | root_cls._signal_hooks_registered = True |
| 345 | |
| 346 | def _register_active_server(self) -> None: |
| 347 | """Track this server so parent-process exits still clean it up.""" |
| 348 | RemoteVLLMServer._ensure_cleanup_hooks_registered() |
| 349 | with RemoteVLLMServer._active_servers_lock: |
| 350 | RemoteVLLMServer._active_servers.add(self) |
| 351 | |
| 352 | def _unregister_active_server(self) -> None: |
| 353 | with RemoteVLLMServer._active_servers_lock: |
| 354 | RemoteVLLMServer._active_servers.discard(self) |
| 355 | |
| 356 | @classmethod |
| 357 | def _shutdown_active_servers(cls) -> None: |
| 358 | """Best-effort shutdown for all live RemoteVLLMServer instances.""" |
| 359 | with cls._active_servers_lock: |
| 360 | servers = list(cls._active_servers) |
| 361 | |
| 362 | for server in servers: |
| 363 | with contextlib.suppress(Exception): |
| 364 | server._shutdown() |
| 365 | |
| 366 | @classmethod |
| 367 | def _handle_parent_signal(cls, signum, frame) -> None: |
| 368 | """Clean up detached servers before letting the signal terminate pytest.""" |
| 369 | cls._shutdown_active_servers() |
| 370 | |
| 371 | previous_handler = cls._previous_signal_handlers.get(signum, signal.SIG_DFL) |
| 372 | if callable(previous_handler): |
| 373 | previous_handler(signum, frame) |
| 374 | elif previous_handler == signal.SIG_IGN: |
| 375 | return |
| 376 | elif signum == signal.SIGINT: |
| 377 | raise KeyboardInterrupt |
| 378 | else: |
| 379 | raise SystemExit(128 + signum) |
| 380 | |
| 381 | def _terminate_process_tree(self) -> None: |
| 382 | """Kill the server process tree without waiting for GPU memory release. |
| 383 | |
| 384 | Split out from ``_shutdown`` so that ``shutdown_many`` can run this |
| 385 | phase in parallel for sibling servers and then wait for GPU memory |
| 386 | release once at the end. |
| 387 | """ |
| 388 | pid = self.proc.pid |
| 389 | |
| 390 | # Get the process group ID. Because we used |
| 391 | # start_new_session=True the pgid equals the server's pid. |
| 392 | try: |
| 393 | pgid = os.getpgid(pid) |
| 394 | except (ProcessLookupError, OSError): |
| 395 | pgid = None |
| 396 | |
| 397 | # Phase 1: graceful SIGTERM to the root process |
| 398 | with contextlib.suppress(ProcessLookupError, OSError): |
| 399 | self.proc.terminate() |
| 400 | print(f"[RemoteOpenAIServer] Sent SIGTERM to process {pid}") |
| 401 | |
| 402 | try: |
| 403 | self.proc.wait(timeout=self._get_process_termination_timeout()) |
| 404 | print(f"[RemoteOpenAIServer] Server {pid} terminated gracefully") |
| 405 | except subprocess.TimeoutExpired: |
| 406 | # Phase 2: SIGKILL the entire process group |
| 407 | print( |
| 408 | f"[RemoteOpenAIServer] Server {pid} did not respond " |
| 409 | "to SIGTERM, sending SIGKILL to process group" |
| 410 | ) |
| 411 | if pgid is not None: |
| 412 | with contextlib.suppress(ProcessLookupError, OSError): |
| 413 | os.killpg(pgid, signal.SIGKILL) |
| 414 | else: |
| 415 | self.proc.kill() |
| 416 | |
| 417 | try: |
| 418 | self.proc.wait(timeout=10) |
| 419 | print(f"[RemoteOpenAIServer] Server {pid} killed") |
| 420 | except subprocess.TimeoutExpired: |
| 421 | pass |
| 422 | |
| 423 | # After killing the root process, ensure all children in the |
| 424 | # process group (e.g. EngineCore workers) are also dead. |
| 425 | # On ROCm especially, surviving children hold GPU contexts and |
| 426 | # prevent VRAM from being reclaimed by the driver. |
| 427 | self._kill_process_group_survivors(pgid) |
| 428 | |
| 429 | @classmethod |
| 430 | def shutdown_many(cls, servers: Sequence["RemoteVLLMServer"]) -> None: |
| 431 | """Shut down multiple sibling servers and wait for GPU memory once. |
| 432 | |
| 433 | Test fixtures that hold several ``RemoteVLLMServer`` instances at |
| 434 | once must NOT shut them down by calling each server's ``__exit__`` |
| 435 | sequentially: every server measures total GPU memory across all |
| 436 | visible devices in ``_wait_for_gpu_memory_release``, so the first |
| 437 | server's wait blocks the full timeout because later sibling |
| 438 | servers are still holding GPU memory. |
| 439 | |
| 440 | Instead, this method terminates every server's process tree in |
| 441 | parallel, then runs the GPU-memory-release wait once against the |
| 442 | earliest recorded baseline (memory before any server started). |
| 443 | """ |
| 444 | if not servers: |
| 445 | return |
| 446 | |
| 447 | for server in servers: |
| 448 | server._shutdown_complete = True |
| 449 | |
| 450 | threads = [ |
| 451 | threading.Thread( |
| 452 | target=s._terminate_process_tree, |
| 453 | name=f"shutdown-{s.proc.pid}", |
| 454 | daemon=True, |
| 455 | ) |
| 456 | for s in servers |
| 457 | ] |
| 458 | for t in threads: |
| 459 | t.start() |
| 460 | for t in threads: |
| 461 | t.join() |
| 462 | |
| 463 | # Use the smallest pre-server baseline so the wait targets memory |
| 464 | # usage before *any* of these sibling servers started, not after |
| 465 | # earlier siblings had already allocated. |
| 466 | earliest = min( |
| 467 | servers, |
| 468 | key=lambda s: ( |
| 469 | float("inf") |
| 470 | if s._pre_server_gpu_memory is None |
| 471 | else s._pre_server_gpu_memory |
| 472 | ), |
| 473 | ) |
| 474 | try: |
| 475 | earliest._wait_for_gpu_memory_release() |
| 476 | finally: |
| 477 | for server in servers: |
| 478 | server._unregister_active_server() |
| 479 | |
| 480 | def _kill_process_group_survivors( |
| 481 | self, pgid: int | None, timeout: float = 15.0 |
| 482 | ) -> None: |
| 483 | """SIGKILL any processes still in the server's process group |
| 484 | and wait for them to exit. |
| 485 | |
| 486 | Because the server is launched with ``start_new_session=True``, |
| 487 | all its children (EngineCore, workers, etc.) share the same |
| 488 | pgid. After the root process is killed, stragglers -- especially |
| 489 | on ROCm where GPU contexts linger until the *process* exits -- |
| 490 | must be reaped explicitly. |
| 491 | |
| 492 | Uses ``/proc`` to scan for pgid members so this works even after |
| 493 | the parent has been reaped (unlike ``psutil.Process.children``). |
| 494 | """ |
| 495 | if pgid is None: |
| 496 | return |
| 497 | |
| 498 | # Send SIGKILL to the entire process group one more time. |
| 499 | # This is cheap and harmless if everyone is already dead. |
| 500 | with contextlib.suppress(ProcessLookupError, OSError): |
| 501 | os.killpg(pgid, signal.SIGKILL) |
| 502 | |
| 503 | # Collect surviving PIDs by scanning /proc for matching pgid. |
| 504 | # This works on Linux even after the parent has been waited on |
| 505 | # and is more reliable than psutil.Process(parent).children(). |
| 506 | survivor_pids = self._find_pgid_members(pgid) |
| 507 | |
| 508 | if not survivor_pids: |
| 509 | return |
| 510 | |
| 511 | print( |
| 512 | f"[RemoteOpenAIServer] {len(survivor_pids)} process(es) still " |
| 513 | f"in pgid {pgid} after SIGKILL: {survivor_pids}" |
| 514 | ) |
| 515 | |
| 516 | # Wait for each survivor to actually exit so the GPU driver |
| 517 | # releases its VRAM. |
| 518 | deadline = time.time() + timeout |
| 519 | while survivor_pids and time.time() < deadline: |
| 520 | still_alive = [] |
| 521 | for spid in survivor_pids: |
| 522 | try: |
| 523 | os.kill(spid, 0) # Check if still alive |
| 524 | still_alive.append(spid) |
| 525 | except (ProcessLookupError, OSError): |
| 526 | pass |
| 527 | survivor_pids = still_alive |
| 528 | if survivor_pids: |
| 529 | time.sleep(0.5) |
| 530 | |
| 531 | if survivor_pids: |
| 532 | print( |
| 533 | f"[RemoteOpenAIServer] WARNING: processes {survivor_pids} " |
| 534 | f"in pgid {pgid} could not be killed within {timeout}s" |
| 535 | ) |
| 536 | |
| 537 | @staticmethod |
| 538 | def _find_pgid_members(pgid: int) -> list[int]: |
| 539 | """Return PIDs of all living processes whose pgid matches.""" |
| 540 | members: list[int] = [] |
| 541 | proc_path = Path("/proc") |
| 542 | if not proc_path.is_dir(): |
| 543 | return members |
| 544 | for entry in proc_path.iterdir(): |
| 545 | if not entry.name.isdigit(): |
| 546 | continue |
| 547 | pid = int(entry.name) |
| 548 | try: |
| 549 | if os.getpgid(pid) == pgid: |
| 550 | members.append(pid) |
| 551 | except OSError: |
| 552 | continue |
| 553 | return members |
| 554 | |
| 555 | def _get_gpu_memory_used(self) -> float | None: |
| 556 | """Get device-wide usage across visible devices in bytes. |
| 557 | |
| 558 | On ROCm this initializes HIP in the caller; subsequent multiprocessing |
| 559 | GPU workers must use spawn, as the ROCm server helpers already do. |
| 560 | """ |
| 561 | try: |
| 562 | if current_platform.is_rocm(): |
| 563 | # Use HIP logical devices. On MI355 DPX/NPS2, the primary |
| 564 | # AMD SMI device can expose whole-card memory counters. |
| 565 | total_used = 0 |
| 566 | for i in range(current_platform.device_count()): |
| 567 | free, total = torch.accelerator.get_memory_info(i) |
| 568 | total_used += total - free |
| 569 | return float(total_used) |
| 570 | elif current_platform.is_cuda(): |
| 571 | with _nvml(): |
| 572 | total_used = 0 |
| 573 | device_count = current_platform.device_count() |
| 574 | for i in range(device_count): |
| 575 | handle = nvmlDeviceGetHandleByIndex(i) |
| 576 | mem_info = nvmlDeviceGetMemoryInfo(handle) |
| 577 | total_used += mem_info.used |
| 578 | return total_used |
| 579 | elif current_platform.is_xpu(): |
| 580 | total_used = 0 |
| 581 | device_count = current_platform.device_count() |
| 582 | for i in range(device_count): |
| 583 | free, total = torch.xpu.mem_get_info(i) |
| 584 | total_used += total - free |
| 585 | return total_used |
| 586 | except Exception as e: |
| 587 | print(f"[RemoteOpenAIServer] Could not query GPU memory: {e}") |
| 588 | return None |
| 589 | return None |
| 590 | |
| 591 | def _wait_for_gpu_memory_release( |
| 592 | self, timeout: float = 120.0, log_interval: float = 10.0 |
| 593 | ): |
| 594 | """Wait for GPU memory to drop back toward pre-server levels. |
| 595 | |
| 596 | Waits the full timeout for memory to return close to the |
| 597 | pre-server baseline. Does NOT fall back to a "stabilization" |
| 598 | heuristic -- if memory is still held when the timeout expires, |
| 599 | the test fails so the problem is surfaced immediately rather |
| 600 | than causing cascading OOM failures in every subsequent test. |
| 601 | """ |
| 602 | baseline = self._pre_server_gpu_memory |
| 603 | if baseline is None: |
| 604 | # Can't query GPU memory - nothing to do |
| 605 | return |
| 606 | |
| 607 | # Allow aggregate driver/context growth above the baseline. |
| 608 | headroom_bytes = ( |
| 609 | 4 * 1024 * 1024 * 1024 |
| 610 | if current_platform.is_rocm() |
| 611 | else 2 * 1024 * 1024 * 1024 |
| 612 | ) |
| 613 | target = baseline + headroom_bytes |
| 614 | |
| 615 | start = time.time() |
| 616 | next_log_time = start + log_interval |
| 617 | |
| 618 | while time.time() - start < timeout: |
| 619 | used = self._get_gpu_memory_used() |
| 620 | |
| 621 | if used is None: |
| 622 | return # Can't query, assume ok |
| 623 | |
| 624 | used_gb = used / 1e9 |
| 625 | target_gb = target / 1e9 |
| 626 | elapsed = time.time() - start |
| 627 | |
| 628 | if used <= target: |
| 629 | print( |
| 630 | f"[RemoteOpenAIServer] GPU memory released to " |
| 631 | f"{used_gb:.2f} GB (target: {target_gb:.2f} GB) " |
| 632 | f"in {elapsed:.1f}s" |
| 633 | ) |
| 634 | return |
| 635 | |
| 636 | now = time.time() |
| 637 | if now >= next_log_time: |
| 638 | print( |
| 639 | f"[RemoteOpenAIServer] Waiting for GPU memory release: " |
| 640 | f"{used_gb:.2f} GB (target: {target_gb:.2f} GB) " |
| 641 | f"[{elapsed:.0f}s/{timeout:.0f}s]" |
| 642 | ) |
| 643 | next_log_time = now + log_interval |
| 644 | |
| 645 | time.sleep(1.0) |
| 646 | |
| 647 | # Timeout -- raise so the current test fails with a clear |
| 648 | # message instead of silently poisoning subsequent tests. |
| 649 | final_used = self._get_gpu_memory_used() |
| 650 | final_gb = final_used / 1e9 if final_used else 0.0 |
| 651 | raise RuntimeError( |
| 652 | f"[RemoteOpenAIServer] GPU memory did not release within " |
| 653 | f"{timeout}s. Current: {final_gb:.2f} GB, " |
| 654 | f"target: {target / 1e9:.2f} GB, " |
| 655 | f"baseline: {baseline / 1e9:.2f} GB. " |
| 656 | f"Child processes may still be holding GPU memory." |
| 657 | ) |
| 658 | |
| 659 | def _poll(self) -> int | None: |
| 660 | """Subclasses override this method to customize process polling.""" |
| 661 | return self.proc.poll() |
| 662 | |
| 663 | def _wait_for_server(self, *, url: str, timeout: float): |
| 664 | # run health check |
| 665 | start = time.time() |
| 666 | client = ( |
| 667 | httpx.Client(transport=httpx.HTTPTransport(uds=self.uds)) |
| 668 | if self.uds |
| 669 | else requests |
| 670 | ) |
| 671 | while True: |
| 672 | try: |
| 673 | if client.get(url).status_code == 200: |
| 674 | break |
| 675 | except Exception: |
| 676 | # this exception can only be raised by requests.get, |
| 677 | # which means the server is not ready yet. |
| 678 | # the stack trace is not useful, so we suppress it |
| 679 | # by using `raise from None`. |
| 680 | result = self._poll() |
| 681 | if result is not None and result != 0: |
| 682 | raise RuntimeError("Server exited unexpectedly.") from None |
| 683 | |
| 684 | time.sleep(0.5) |
| 685 | if time.time() - start > timeout: |
| 686 | raise RuntimeError("Server failed to start in time.") from None |
| 687 | |
| 688 | @property |
| 689 | def url_root(self) -> str: |
| 690 | return ( |
| 691 | f"http://{self.uds.split('/')[-1]}" |
| 692 | if self.uds |
| 693 | else f"http://{self.host}:{self.port}" |
| 694 | ) |
| 695 | |
| 696 | def url_for(self, *parts: str) -> str: |
| 697 | path = "/".join(part.strip("/") for part in parts if part) |
| 698 | return f"{self.url_root}/{path}" |
| 699 | |
| 700 | def get_client(self, **kwargs): |
| 701 | if "timeout" not in kwargs: |
| 702 | kwargs["timeout"] = 600 |
| 703 | return openai.OpenAI( |
| 704 | base_url=self.url_for("v1"), |
| 705 | api_key=self.DUMMY_API_KEY, |
| 706 | max_retries=0, |
| 707 | **kwargs, |
| 708 | ) |
| 709 | |
| 710 | def get_async_client(self, **kwargs): |
| 711 | if "timeout" not in kwargs: |
| 712 | kwargs["timeout"] = 600 |
| 713 | return openai.AsyncOpenAI( |
| 714 | base_url=self.url_for("v1"), |
| 715 | api_key=self.DUMMY_API_KEY, |
| 716 | max_retries=0, |
| 717 | **kwargs, |
| 718 | ) |
| 719 | |
| 720 | def get_client_anthropic(self, **kwargs): |
| 721 | if "timeout" not in kwargs: |
| 722 | kwargs["timeout"] = 600 |
| 723 | return anthropic.Anthropic( |
| 724 | base_url=self.url_for(), |
| 725 | api_key=self.DUMMY_API_KEY, |
| 726 | max_retries=0, |
| 727 | **kwargs, |
| 728 | ) |
| 729 | |
| 730 | def get_async_client_anthropic(self, **kwargs): |
| 731 | if "timeout" not in kwargs: |
| 732 | kwargs["timeout"] = 600 |
| 733 | return anthropic.AsyncAnthropic( |
| 734 | base_url=self.url_for(), api_key=self.DUMMY_API_KEY, max_retries=0, **kwargs |
| 735 | ) |
| 736 | |
| 737 | |
| 738 | class RemoteOpenAIServer(RemoteVLLMServer): |
| 739 | """Launches ``vllm serve`` for testing OpenAI-compatible endpoints.""" |
| 740 | |
| 741 | def _get_process_termination_timeout(self) -> float: |
| 742 | engine_timeout = get_engine_process_shutdown_timeout( |
| 743 | self._request_shutdown_timeout, |
| 744 | self._request_shutdown_timeout, |
| 745 | ) |
| 746 | assert engine_timeout is not None |
| 747 | return engine_timeout + super()._get_process_termination_timeout() |
| 748 | |
| 749 | def _create_cli_subcommand(self): |
| 750 | return ServeSubcommand() |
| 751 | |
| 752 | def _start_server( |
| 753 | self, model: str, vllm_serve_args: list[str], env_dict: dict[str, str] | None |
| 754 | ) -> None: |
| 755 | env = os.environ.copy() |
| 756 | # the current process might initialize cuda, |
| 757 | # to be safe, we should use spawn method |
| 758 | env["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" |
| 759 | if env_dict is not None: |
| 760 | env.update(env_dict) |
| 761 | _sanitize_pythonpath_env(env) |
| 762 | serve_cmd = ["vllm", "serve", model, *vllm_serve_args] |
| 763 | redacted_serve_cmd = _redact_sensitive_cli_args(serve_cmd) |
| 764 | print(f"Launching RemoteOpenAIServer with: {' '.join(redacted_serve_cmd)}") |
| 765 | self.proc: subprocess.Popen = subprocess.Popen( |
| 766 | serve_cmd, |
| 767 | env=env, |
| 768 | stdout=sys.stdout, |
| 769 | stderr=sys.stderr, |
| 770 | # Create a dedicated process group so we can kill |
| 771 | # the entire tree (parent + EngineCore + workers) at once. |
| 772 | start_new_session=True, |
| 773 | ) |
| 774 | |
| 775 | |
| 776 | class RemoteLaunchRenderServer(RemoteVLLMServer): |
| 777 | """Launches ``vllm launch render`` for GPU-less serving tests.""" |
| 778 | |
| 779 | def _create_cli_subcommand(self): |
| 780 | return ServeSubcommand() |
| 781 | |
| 782 | def _start_server( |
| 783 | self, model: str, vllm_serve_args: list[str], env_dict: dict[str, str] | None |
| 784 | ) -> None: |
| 785 | env = os.environ.copy() |
| 786 | env["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" |
| 787 | if env_dict is not None: |
| 788 | env.update(env_dict) |
| 789 | _sanitize_pythonpath_env(env) |
| 790 | serve_cmd = ["vllm", "launch", "render", model, *vllm_serve_args] |
| 791 | redacted_serve_cmd = _redact_sensitive_cli_args(serve_cmd) |
| 792 | print( |
| 793 | f"Launching RemoteLaunchRenderServer with: {' '.join(redacted_serve_cmd)}" |
| 794 | ) |
| 795 | self.proc: subprocess.Popen = subprocess.Popen( |
| 796 | serve_cmd, |
| 797 | env=env, |
| 798 | stdout=sys.stdout, |
| 799 | stderr=sys.stderr, |
| 800 | start_new_session=True, |
| 801 | ) |
| 802 | |
| 803 | def _pre_download_model(self, model: str, args) -> None: |
| 804 | """Download only the tokenizer files (no model weights needed).""" |
| 805 | is_local = os.path.isdir(model) |
| 806 | if not is_local: |
| 807 | engine_args = AsyncEngineArgs.from_cli_args(args) |
| 808 | model_config = engine_args.create_model_config() |
| 809 | get_tokenizer( |
| 810 | model_config.tokenizer, |
| 811 | tokenizer_mode=model_config.tokenizer_mode, |
| 812 | trust_remote_code=model_config.trust_remote_code, |
| 813 | revision=model_config.tokenizer_revision, |
| 814 | ) |
| 815 | |
| 816 | def _wait_for_gpu_memory_release( |
| 817 | self, timeout: float = 30.0, log_interval: float = 10.0 |
| 818 | ): |
| 819 | pass # No GPU used |
| 820 | |
| 821 | |
| 822 | class RemoteOpenAIServerCustom(RemoteOpenAIServer): |
| 823 | """Launch test server with custom child process.""" |
| 824 | |
| 825 | def _start_server( |
| 826 | self, model: str, vllm_serve_args: list[str], env_dict: dict[str, str] | None |
| 827 | ) -> None: |
| 828 | method = "spawn" if requires_spawn_multiprocessing() else "fork" |
| 829 | ctx = get_context(method) |
| 830 | self.proc: Process = cast(Any, ctx).Process( |
| 831 | target=_run_in_new_process_group, |
| 832 | args=(self.child_process_fxn, env_dict, model, vllm_serve_args), |
| 833 | ) |
| 834 | with _temporarily_sanitized_pythonpath_env(): |
| 835 | self.proc.start() |
| 836 | |
| 837 | def __init__( |
| 838 | self, |
| 839 | model: str, |
| 840 | vllm_serve_args: list[str], |
| 841 | child_process_fxn: Callable[[dict[str, str] | None, str, list[str]], None], |
| 842 | *, |
| 843 | env_dict: dict[str, str] | None = None, |
| 844 | seed: int = 0, |
| 845 | auto_port: bool = True, |
| 846 | max_wait_seconds: float | None = None, |
| 847 | ) -> None: |
| 848 | """Store custom child process function then invoke superclass |
| 849 | constructor which will indirectly launch it.""" |
| 850 | self.child_process_fxn = child_process_fxn |
| 851 | super().__init__( |
| 852 | model=model, |
| 853 | vllm_serve_args=vllm_serve_args, |
| 854 | env_dict=env_dict, |
| 855 | seed=seed, |
| 856 | auto_port=auto_port, |
| 857 | max_wait_seconds=max_wait_seconds, |
| 858 | ) |
| 859 | |
| 860 | def _poll(self) -> int | None: |
| 861 | return self.proc.exitcode |
| 862 | |
| 863 | def _terminate_process_tree(self) -> None: |
| 864 | pid = self.proc.pid |
| 865 | if pid is None: |
| 866 | return |
| 867 | |
| 868 | pgid: int | None |
| 869 | try: |
| 870 | pgid = os.getpgid(pid) |
| 871 | # _run_in_new_process_group should make the child the group |
| 872 | # leader. Avoid signaling pytest's process group if startup failed |
| 873 | # before os.setsid() ran. |
| 874 | if pgid != pid: |
| 875 | pgid = None |
| 876 | except (ProcessLookupError, OSError): |
| 877 | pgid = None |
| 878 | |
| 879 | with contextlib.suppress(ProcessLookupError, OSError): |
| 880 | self.proc.terminate() |
| 881 | print(f"[RemoteOpenAIServerCustom] Sent SIGTERM to process {pid}") |
| 882 | |
| 883 | self.proc.join(self._get_process_termination_timeout()) |
| 884 | if self.proc.is_alive(): |
| 885 | print( |
| 886 | f"[RemoteOpenAIServerCustom] Server {pid} did not respond " |
| 887 | "to SIGTERM, sending SIGKILL to process group" |
| 888 | ) |
| 889 | if pgid is not None: |
| 890 | with contextlib.suppress(ProcessLookupError, OSError): |
| 891 | os.killpg(pgid, signal.SIGKILL) |
| 892 | else: |
| 893 | self.proc.kill() |
| 894 | self.proc.join(10) |
| 895 | |
| 896 | self._kill_process_group_survivors(pgid) |
| 897 | |
| 898 | |
| 899 | def _test_completion( |
| 900 | client: openai.OpenAI, |
| 901 | model: str, |
| 902 | prompt: str, |
| 903 | token_ids: list[int], |
| 904 | include_seeded_sampling: bool = True, |
| 905 | ): |
| 906 | results = [] |
| 907 | |
| 908 | # test with text prompt |
| 909 | completion = client.completions.create( |
| 910 | model=model, prompt=prompt, max_tokens=5, temperature=0.0 |
| 911 | ) |
| 912 | |
| 913 | results.append( |
| 914 | { |
| 915 | "test": "single_completion", |
| 916 | "text": completion.choices[0].text, |
| 917 | "finish_reason": completion.choices[0].finish_reason, |
| 918 | "usage": completion.usage, |
| 919 | } |
| 920 | ) |
| 921 | |
| 922 | # test using token IDs |
| 923 | completion = client.completions.create( |
| 924 | model=model, |
| 925 | prompt=token_ids, |
| 926 | max_tokens=5, |
| 927 | temperature=0.0, |
| 928 | ) |
| 929 | |
| 930 | results.append( |
| 931 | { |
| 932 | "test": "token_ids", |
| 933 | "text": completion.choices[0].text, |
| 934 | "finish_reason": completion.choices[0].finish_reason, |
| 935 | "usage": completion.usage, |
| 936 | } |
| 937 | ) |
| 938 | |
| 939 | if include_seeded_sampling: |
| 940 | # test seeded random sampling |
| 941 | completion = client.completions.create( |
| 942 | model=model, prompt=prompt, max_tokens=5, seed=33, temperature=1.0 |
| 943 | ) |
| 944 | |
| 945 | results.append( |
| 946 | { |
| 947 | "test": "seeded_sampling", |
| 948 | "text": completion.choices[0].text, |
| 949 | "finish_reason": completion.choices[0].finish_reason, |
| 950 | "usage": completion.usage, |
| 951 | } |
| 952 | ) |
| 953 | |
| 954 | # test seeded random sampling with multiple prompts |
| 955 | completion = client.completions.create( |
| 956 | model=model, |
| 957 | prompt=[prompt, prompt], |
| 958 | max_tokens=5, |
| 959 | seed=33, |
| 960 | temperature=1.0, |
| 961 | ) |
| 962 | |
| 963 | results.append( |
| 964 | { |
| 965 | "test": "seeded_sampling", |
| 966 | "text": [choice.text for choice in completion.choices], |
| 967 | "finish_reason": [ |
| 968 | choice.finish_reason for choice in completion.choices |
| 969 | ], |
| 970 | "usage": completion.usage, |
| 971 | } |
| 972 | ) |
| 973 | |
| 974 | # test simple list |
| 975 | batch = client.completions.create( |
| 976 | model=model, |
| 977 | prompt=[prompt, prompt], |
| 978 | max_tokens=5, |
| 979 | temperature=0.0, |
| 980 | ) |
| 981 | |
| 982 | results.append( |
| 983 | { |
| 984 | "test": "simple_list", |
| 985 | "text0": batch.choices[0].text, |
| 986 | "text1": batch.choices[1].text, |
| 987 | } |
| 988 | ) |
| 989 | |
| 990 | # test streaming |
| 991 | batch = client.completions.create( |
| 992 | model=model, |
| 993 | prompt=[prompt, prompt], |
| 994 | max_tokens=5, |
| 995 | temperature=0.0, |
| 996 | stream=True, |
| 997 | ) |
| 998 | |
| 999 | texts = [""] * 2 |
| 1000 | for chunk in batch: |
| 1001 | assert len(chunk.choices) == 1 |
| 1002 | choice = chunk.choices[0] |
| 1003 | texts[choice.index] += choice.text |
| 1004 | |
| 1005 | results.append( |
| 1006 | { |
| 1007 | "test": "streaming", |
| 1008 | "texts": texts, |
| 1009 | } |
| 1010 | ) |
| 1011 | |
| 1012 | return results |
| 1013 | |
| 1014 | |
| 1015 | def _test_completion_close( |
| 1016 | client: openai.OpenAI, |
| 1017 | model: str, |
| 1018 | prompt: str, |
| 1019 | ): |
| 1020 | results = [] |
| 1021 | |
| 1022 | # test with text prompt |
| 1023 | completion = client.completions.create( |
| 1024 | model=model, prompt=prompt, max_tokens=1, logprobs=5, temperature=0.0 |
| 1025 | ) |
| 1026 | |
| 1027 | logprobs = completion.choices[0].logprobs.top_logprobs[0] |
| 1028 | logprobs = {k: round(v, 2) for k, v in logprobs.items()} |
| 1029 | |
| 1030 | results.append( |
| 1031 | { |
| 1032 | "test": "completion_close", |
| 1033 | "logprobs": logprobs, |
| 1034 | } |
| 1035 | ) |
| 1036 | |
| 1037 | return results |
| 1038 | |
| 1039 | |
| 1040 | def _test_chat( |
| 1041 | client: openai.OpenAI, |
| 1042 | model: str, |
| 1043 | prompt: str, |
| 1044 | ): |
| 1045 | results = [] |
| 1046 | |
| 1047 | messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}] |
| 1048 | |
| 1049 | # test with text prompt |
| 1050 | chat_response = client.chat.completions.create( |
| 1051 | model=model, messages=messages, max_tokens=5, temperature=0.0 |
| 1052 | ) |
| 1053 | |
| 1054 | results.append( |
| 1055 | { |
| 1056 | "test": "completion_close", |
| 1057 | "text": chat_response.choices[0].message.content, |
| 1058 | "finish_reason": chat_response.choices[0].finish_reason, |
| 1059 | "usage": chat_response.usage, |
| 1060 | } |
| 1061 | ) |
| 1062 | |
| 1063 | return results |
| 1064 | |
| 1065 | |
| 1066 | def _test_embeddings( |
| 1067 | client: openai.OpenAI, |
| 1068 | model: str, |
| 1069 | text: str, |
| 1070 | ): |
| 1071 | results = [] |
| 1072 | |
| 1073 | # test with text input |
| 1074 | embeddings = client.embeddings.create( |
| 1075 | model=model, |
| 1076 | input=text, |
| 1077 | encoding_format="float", |
| 1078 | ) |
| 1079 | |
| 1080 | results.append( |
| 1081 | { |
| 1082 | "test": "single_embedding", |
| 1083 | "embedding": embeddings.data[0].embedding, |
| 1084 | "usage": embeddings.usage, |
| 1085 | } |
| 1086 | ) |
| 1087 | |
| 1088 | return results |
| 1089 | |
| 1090 | |
| 1091 | def _test_image_text( |
| 1092 | client: openai.OpenAI, |
| 1093 | model_name: str, |
| 1094 | image_url: str, |
| 1095 | ): |
| 1096 | results = [] |
| 1097 | |
| 1098 | # test pure text input |
| 1099 | messages = [ |
| 1100 | { |
| 1101 | "role": "user", |
| 1102 | "content": [ |
| 1103 | {"type": "text", "text": "How do you feel today?"}, |
| 1104 | ], |
| 1105 | } |
| 1106 | ] |
| 1107 | |
| 1108 | chat_completion = client.chat.completions.create( |
| 1109 | model=model_name, |
| 1110 | messages=messages, |
| 1111 | temperature=0.0, |
| 1112 | max_tokens=1, |
| 1113 | logprobs=True, |
| 1114 | top_logprobs=5, |
| 1115 | ) |
| 1116 | top_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs |
| 1117 | |
| 1118 | for x in top_logprobs: |
| 1119 | x.logprob = round(x.logprob, 2) |
| 1120 | |
| 1121 | results.append( |
| 1122 | { |
| 1123 | "test": "pure_text", |
| 1124 | "logprobs": top_logprobs, |
| 1125 | } |
| 1126 | ) |
| 1127 | |
| 1128 | messages = [ |
| 1129 | { |
| 1130 | "role": "user", |
| 1131 | "content": [ |
| 1132 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1133 | {"type": "text", "text": "What's in this image?"}, |
| 1134 | ], |
| 1135 | } |
| 1136 | ] |
| 1137 | |
| 1138 | chat_completion = client.chat.completions.create( |
| 1139 | model=model_name, |
| 1140 | messages=messages, |
| 1141 | temperature=0.0, |
| 1142 | max_tokens=1, |
| 1143 | logprobs=True, |
| 1144 | top_logprobs=5, |
| 1145 | ) |
| 1146 | top_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs |
| 1147 | |
| 1148 | results.append( |
| 1149 | { |
| 1150 | "test": "text_image", |
| 1151 | "logprobs": top_logprobs, |
| 1152 | } |
| 1153 | ) |
| 1154 | |
| 1155 | return results |
| 1156 | |
| 1157 | |
| 1158 | def compare_two_settings( |
| 1159 | model: str, |
| 1160 | arg1: list[str], |
| 1161 | arg2: list[str], |
| 1162 | env1: dict[str, str] | None = None, |
| 1163 | env2: dict[str, str] | None = None, |
| 1164 | *, |
| 1165 | method: str = "generate", |
| 1166 | max_wait_seconds: float | None = None, |
| 1167 | include_seeded_sampling: bool = True, |
| 1168 | force_v1_runner: bool = False, |
| 1169 | ) -> None: |
| 1170 | """Launch API server with two different sets of arguments/environments |
| 1171 | and compare the results of the API calls. |
| 1172 | |
| 1173 | Args: |
| 1174 | model: The model to test. |
| 1175 | arg1: The first set of arguments to pass to the API server. |
| 1176 | arg2: The second set of arguments to pass to the API server. |
| 1177 | env1: The first set of environment variables to pass to the API server. |
| 1178 | env2: The second set of environment variables to pass to the API server. |
| 1179 | include_seeded_sampling: Whether to include temperature=1.0 seeded |
| 1180 | sampling checks in the default generate comparison. |
| 1181 | force_v1_runner: Whether to pin all compared settings to the v1 model |
| 1182 | runner to avoid mixing model runner differences into correctness |
| 1183 | tests. |
| 1184 | |
| 1185 | """ |
| 1186 | compare_all_settings( |
| 1187 | model, |
| 1188 | [arg1, arg2], |
| 1189 | [env1, env2], |
| 1190 | method=method, |
| 1191 | max_wait_seconds=max_wait_seconds, |
| 1192 | include_seeded_sampling=include_seeded_sampling, |
| 1193 | force_v1_runner=force_v1_runner, |
| 1194 | ) |
| 1195 | |
| 1196 | |
| 1197 | def compare_all_settings( |
| 1198 | model: str, |
| 1199 | all_args: list[list[str]], |
| 1200 | all_envs: list[dict[str, str] | None], |
| 1201 | *, |
| 1202 | method: str = "generate", |
| 1203 | max_wait_seconds: float | None = None, |
| 1204 | include_seeded_sampling: bool = True, |
| 1205 | force_v1_runner: bool = False, |
| 1206 | ) -> None: |
| 1207 | """Launch API server with several different sets of arguments/environments |
| 1208 | and compare the results of the API calls with the first set of arguments. |
| 1209 | |
| 1210 | Args: |
| 1211 | model: The model to test. |
| 1212 | all_args: A list of argument lists to pass to the API server. |
| 1213 | all_envs: A list of environment dictionaries to pass to the API server. |
| 1214 | include_seeded_sampling: Whether to include temperature=1.0 seeded |
| 1215 | sampling checks in the default generate comparison. |
| 1216 | force_v1_runner: Whether to pin all compared settings to the v1 model |
| 1217 | runner to avoid mixing model runner differences into correctness |
| 1218 | tests. |
| 1219 | |
| 1220 | """ |
| 1221 | if force_v1_runner: |
| 1222 | all_envs = [ |
| 1223 | {"VLLM_USE_V2_MODEL_RUNNER": "0", **(env or {})} for env in all_envs |
| 1224 | ] |
| 1225 | |
| 1226 | trust_remote_code = False |
| 1227 | for args in all_args: |
| 1228 | if "--trust-remote-code" in args: |
| 1229 | trust_remote_code = True |
| 1230 | break |
| 1231 | |
| 1232 | tokenizer_mode = "auto" |
| 1233 | for args in all_args: |
| 1234 | if "--tokenizer-mode" in args: |
| 1235 | tokenizer_mode = args[args.index("--tokenizer-mode") + 1] |
| 1236 | break |
| 1237 | |
| 1238 | tokenizer = get_tokenizer( |
| 1239 | model, |
| 1240 | trust_remote_code=trust_remote_code, |
| 1241 | tokenizer_mode=tokenizer_mode, |
| 1242 | ) |
| 1243 | |
| 1244 | can_force_load_format = True |
| 1245 | |
| 1246 | for args in all_args: |
| 1247 | if "--load-format" in args: |
| 1248 | can_force_load_format = False |
| 1249 | break |
| 1250 | |
| 1251 | prompt = "Hello, my name is" |
| 1252 | token_ids = tokenizer(prompt).input_ids |
| 1253 | ref_results: list = [] |
| 1254 | for i, (args, env) in enumerate(zip(all_args, all_envs)): |
| 1255 | if can_force_load_format: |
| 1256 | # we are comparing the results and |
| 1257 | # usually we don't need real weights. |
| 1258 | # we force to use dummy weights by default, |
| 1259 | # and it should work for most of the cases. |
| 1260 | # if not, we can use VLLM_TEST_FORCE_LOAD_FORMAT |
| 1261 | # environment variable to force the load format, |
| 1262 | # e.g. in quantization tests. |
| 1263 | args = args + ["--load-format", envs.VLLM_TEST_FORCE_LOAD_FORMAT] |
| 1264 | compare_results: list = [] |
| 1265 | results = ref_results if i == 0 else compare_results |
| 1266 | with RemoteOpenAIServer( |
| 1267 | model, args, env_dict=env, max_wait_seconds=max_wait_seconds |
| 1268 | ) as server: |
| 1269 | client = server.get_client() |
| 1270 | |
| 1271 | # test models list |
| 1272 | models = client.models.list() |
| 1273 | models = models.data |
| 1274 | served_model = models[0] |
| 1275 | results.append( |
| 1276 | { |
| 1277 | "test": "models_list", |
| 1278 | "id": served_model.id, |
| 1279 | "root": served_model.root, |
| 1280 | } |
| 1281 | ) |
| 1282 | |
| 1283 | if method == "generate": |
| 1284 | results += _test_completion( |
| 1285 | client, |
| 1286 | model, |
| 1287 | prompt, |
| 1288 | token_ids, |
| 1289 | include_seeded_sampling=include_seeded_sampling, |
| 1290 | ) |
| 1291 | elif method == "generate_close": |
| 1292 | results += _test_completion_close(client, model, prompt) |
| 1293 | elif method == "generate_chat": |
| 1294 | results += _test_chat(client, model, prompt) |
| 1295 | elif method == "generate_with_image": |
| 1296 | results += _test_image_text( |
| 1297 | client, |
| 1298 | model, |
| 1299 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/RGBA_comp.png", |
| 1300 | ) |
| 1301 | elif method == "encode": |
| 1302 | results += _test_embeddings(client, model, prompt) |
| 1303 | else: |
| 1304 | raise ValueError(f"Unknown method: {method}") |
| 1305 | |
| 1306 | if i > 0: |
| 1307 | # if any setting fails, raise an error early |
| 1308 | ref_args = all_args[0] |
| 1309 | ref_envs = all_envs[0] |
| 1310 | compare_args = all_args[i] |
| 1311 | compare_envs = all_envs[i] |
| 1312 | for ref_result, compare_result in zip(ref_results, compare_results): |
| 1313 | ref_result = copy.deepcopy(ref_result) |
| 1314 | compare_result = copy.deepcopy(compare_result) |
| 1315 | if "embedding" in ref_result and method == "encode": |
| 1316 | sim = F.cosine_similarity( |
| 1317 | torch.tensor(ref_result["embedding"]), |
| 1318 | torch.tensor(compare_result["embedding"]), |
| 1319 | dim=0, |
| 1320 | ) |
| 1321 | assert sim >= 0.999, ( |
| 1322 | f"Embedding for {model=} are not the same.\n" |
| 1323 | f"cosine_similarity={sim}\n" |
| 1324 | ) |
| 1325 | del ref_result["embedding"] |
| 1326 | del compare_result["embedding"] |
| 1327 | assert ref_result == compare_result, ( |
| 1328 | f"Results for {model=} are not the same.\n" |
| 1329 | f"{ref_args=} {ref_envs=}\n" |
| 1330 | f"{compare_args=} {compare_envs=}\n" |
| 1331 | f"{ref_result=}\n" |
| 1332 | f"{compare_result=}\n" |
| 1333 | ) |
| 1334 | |
| 1335 | |
| 1336 | @contextmanager |
| 1337 | def ensure_current_vllm_config(): |
| 1338 | """Ensures a vllm config is set for the duration of the context. |
| 1339 | |
| 1340 | If a config is already set, this is a no-op. Otherwise, it creates a default |
| 1341 | VllmConfig and sets it for the duration of the context. |
| 1342 | |
| 1343 | Used for tests that call functions which require a vllm config but don't |
| 1344 | need a specific config. |
| 1345 | |
| 1346 | Example: |
| 1347 | with ensure_current_vllm_config(): |
| 1348 | init_distributed_environment(...) |
| 1349 | ensure_model_parallel_initialized(...) |
| 1350 | |
| 1351 | """ |
| 1352 | from vllm.config import ( |
| 1353 | VllmConfig, |
| 1354 | get_current_vllm_config_or_none, |
| 1355 | set_current_vllm_config, |
| 1356 | ) |
| 1357 | |
| 1358 | if get_current_vllm_config_or_none() is not None: |
| 1359 | # Config already set, just yield |
| 1360 | yield |
| 1361 | else: |
| 1362 | # No config set, create a default one for the duration |
| 1363 | with set_current_vllm_config(VllmConfig()): |
| 1364 | yield |
| 1365 | |
| 1366 | |
| 1367 | def init_test_distributed_environment( |
| 1368 | tp_size: int, |
| 1369 | pp_size: int, |
| 1370 | rank: int, |
| 1371 | distributed_init_port: str, |
| 1372 | local_rank: int = -1, |
| 1373 | data_parallel_size: int = 1, |
| 1374 | data_parallel_master_port: int | None = None, |
| 1375 | ) -> None: |
| 1376 | # Note: This function is often called from Ray worker processes, so we |
| 1377 | # can't rely on pytest fixtures to set the config. We check if the config |
| 1378 | # is already set and only create a default one if needed. |
| 1379 | from vllm.config import ( |
| 1380 | VllmConfig, |
| 1381 | get_current_vllm_config_or_none, |
| 1382 | set_current_vllm_config, |
| 1383 | ) |
| 1384 | from vllm.platforms import current_platform |
| 1385 | |
| 1386 | distributed_init_method = f"tcp://localhost:{distributed_init_port}" |
| 1387 | backend = current_platform.dist_backend |
| 1388 | |
| 1389 | if data_parallel_size > 1: |
| 1390 | # For DP we need to set a common DP master port |
| 1391 | from vllm.config.parallel import ParallelConfig |
| 1392 | |
| 1393 | assert data_parallel_master_port is not None, ( |
| 1394 | "data_parallel_master_port is required when data_parallel_size > 1" |
| 1395 | ) |
| 1396 | tp_pp_world = tp_size * pp_size |
| 1397 | parallel_config = ParallelConfig( |
| 1398 | data_parallel_size=data_parallel_size, |
| 1399 | data_parallel_rank=rank // tp_pp_world, |
| 1400 | _data_parallel_master_port_list=[int(data_parallel_master_port)], |
| 1401 | ) |
| 1402 | with set_current_vllm_config(VllmConfig(parallel_config=parallel_config)): |
| 1403 | init_distributed_environment( |
| 1404 | world_size=tp_pp_world, |
| 1405 | rank=rank % tp_pp_world, |
| 1406 | distributed_init_method=distributed_init_method, |
| 1407 | local_rank=local_rank if local_rank >= 0 else rank, |
| 1408 | backend=backend, |
| 1409 | ) |
| 1410 | ensure_model_parallel_initialized(tp_size, pp_size) |
| 1411 | return |
| 1412 | |
| 1413 | if get_current_vllm_config_or_none() is not None: |
| 1414 | # Config already set, use it directly |
| 1415 | init_distributed_environment( |
| 1416 | world_size=pp_size * tp_size, |
| 1417 | rank=rank, |
| 1418 | distributed_init_method=distributed_init_method, |
| 1419 | local_rank=local_rank, |
| 1420 | backend=backend, |
| 1421 | ) |
| 1422 | ensure_model_parallel_initialized(tp_size, pp_size) |
| 1423 | else: |
| 1424 | # No config set, create a default one for the test |
| 1425 | with set_current_vllm_config(VllmConfig()): |
| 1426 | init_distributed_environment( |
| 1427 | world_size=pp_size * tp_size, |
| 1428 | rank=rank, |
| 1429 | distributed_init_method=distributed_init_method, |
| 1430 | local_rank=local_rank, |
| 1431 | backend=backend, |
| 1432 | ) |
| 1433 | ensure_model_parallel_initialized(tp_size, pp_size) |
| 1434 | |
| 1435 | |
| 1436 | def multi_process_parallel( |
| 1437 | monkeypatch: pytest.MonkeyPatch, |
| 1438 | tp_size: int, |
| 1439 | pp_size: int, |
| 1440 | test_target: Any, |
| 1441 | data_parallel_size: int = 1, |
| 1442 | ) -> None: |
| 1443 | import ray |
| 1444 | |
| 1445 | # Using ray helps debugging the error when it failed as compared to |
| 1446 | # multiprocessing. For local Ray workers, putting the repo root on |
| 1447 | # PYTHONPATH is enough and avoids uploading the full source tree, which |
| 1448 | # exceeds Ray's working_dir package size limit on CI. |
| 1449 | env_vars = { |
| 1450 | "PYTHONPATH": os.pathsep.join( |
| 1451 | filter(None, [str(VLLM_PATH), os.environ.get("PYTHONPATH")]) |
| 1452 | ), |
| 1453 | **{env_var: "1" for env_var in current_platform.ray_noset_device_env_vars}, |
| 1454 | } |
| 1455 | ray.init( |
| 1456 | runtime_env={ |
| 1457 | "env_vars": env_vars, |
| 1458 | } |
| 1459 | ) |
| 1460 | |
| 1461 | distributed_init_port = get_open_port() |
| 1462 | # Separate port for the DP master group; only used when data_parallel_size > 1. |
| 1463 | data_parallel_master_port = get_open_port() if data_parallel_size > 1 else None |
| 1464 | world_size = data_parallel_size * tp_size * pp_size |
| 1465 | try: |
| 1466 | refs = [] |
| 1467 | for rank in range(world_size): |
| 1468 | if data_parallel_size > 1: |
| 1469 | refs.append( |
| 1470 | test_target.remote( |
| 1471 | monkeypatch, |
| 1472 | tp_size, |
| 1473 | pp_size, |
| 1474 | rank, |
| 1475 | distributed_init_port, |
| 1476 | data_parallel_size, |
| 1477 | data_parallel_master_port, |
| 1478 | ), |
| 1479 | ) |
| 1480 | else: |
| 1481 | refs.append( |
| 1482 | test_target.remote( |
| 1483 | monkeypatch, |
| 1484 | tp_size, |
| 1485 | pp_size, |
| 1486 | rank, |
| 1487 | distributed_init_port, |
| 1488 | ), |
| 1489 | ) |
| 1490 | ray.get(refs) |
| 1491 | finally: |
| 1492 | ray.shutdown() |
| 1493 | |
| 1494 | |
| 1495 | def assert_rocm_custom_allreduce_backend_state( |
| 1496 | use_aiter_custom_ar: bool, |
| 1497 | quick_reduce_quantization: str, |
| 1498 | ) -> None: |
| 1499 | from vllm.distributed.parallel_state import get_tp_group |
| 1500 | |
| 1501 | device_communicator = get_tp_group().device_communicator |
| 1502 | aiter_ar_comm = device_communicator.aiter_ar_comm |
| 1503 | if use_aiter_custom_ar: |
| 1504 | assert aiter_ar_comm is not None, "AITER CustomAllreduce was not initialized." |
| 1505 | assert not aiter_ar_comm.disabled, "AITER CustomAllreduce is disabled." |
| 1506 | assert device_communicator.ca_comm is None, ( |
| 1507 | "vLLM CustomAllreduce should not be initialized when AITER CA is used." |
| 1508 | ) |
| 1509 | else: |
| 1510 | assert aiter_ar_comm is None, ( |
| 1511 | "AITER CustomAllreduce should not be initialized when disabled." |
| 1512 | ) |
| 1513 | assert device_communicator.ca_comm is not None, ( |
| 1514 | "vLLM CustomAllreduce should be initialized when AITER CA is disabled." |
| 1515 | ) |
| 1516 | |
| 1517 | qr_comm = device_communicator.qr_comm |
| 1518 | assert qr_comm is not None, "QuickReduce communicator was not initialized." |
| 1519 | if quick_reduce_quantization == "NONE": |
| 1520 | assert qr_comm.disabled, "QuickReduce should be disabled." |
| 1521 | else: |
| 1522 | assert not qr_comm.disabled, "QuickReduce should be enabled." |
| 1523 | |
| 1524 | |
| 1525 | def assert_rocm_custom_allreduce_backend_state_on_worker( |
| 1526 | _worker, |
| 1527 | use_aiter_custom_ar: bool, |
| 1528 | quick_reduce_quantization: str, |
| 1529 | ) -> None: |
| 1530 | assert_rocm_custom_allreduce_backend_state( |
| 1531 | use_aiter_custom_ar, quick_reduce_quantization |
| 1532 | ) |
| 1533 | |
| 1534 | |
| 1535 | def get_physical_device_indices(devices: list[int]): |
| 1536 | visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES") |
| 1537 | if visible_devices is None: |
| 1538 | return devices |
| 1539 | |
| 1540 | visible_indices = [int(x) for x in visible_devices.split(",")] |
| 1541 | index_mapping = {i: physical for i, physical in enumerate(visible_indices)} |
| 1542 | return [index_mapping[i] for i in devices if i in index_mapping] |
| 1543 | |
| 1544 | |
| 1545 | def get_nvml_device_handle(device: int): |
| 1546 | visible_devices = os.environ.get("NVIDIA_VISIBLE_DEVICES") |
| 1547 | if visible_devices is not None: |
| 1548 | identifiers = visible_devices.split(",") |
| 1549 | if device < len(identifiers): |
| 1550 | identifier = identifiers[device] |
| 1551 | if identifier.startswith(("GPU-", "MIG-")): |
| 1552 | return nvmlDeviceGetHandleByUUID(identifier) |
| 1553 | |
| 1554 | return nvmlDeviceGetHandleByIndex(device) |
| 1555 | |
| 1556 | |
| 1557 | @_nvml() |
| 1558 | def record_gpu_memory_usage_stats( |
| 1559 | *, |
| 1560 | devices: list[int], |
| 1561 | ) -> dict[int, tuple[float, float]]: |
| 1562 | """Return device-wide used/total GiB; ROCm IDs are HIP logical devices. |
| 1563 | |
| 1564 | ROCm queries initialize HIP in the caller and include other processes' |
| 1565 | allocations on the same memory partition. |
| 1566 | """ |
| 1567 | output: dict[int, tuple[float, float]] = {} |
| 1568 | for device in devices: |
| 1569 | if current_platform.is_rocm(): |
| 1570 | free, total = torch.accelerator.get_memory_info(device) |
| 1571 | gb_used = (total - free) / 2**30 |
| 1572 | gb_total = total / 2**30 |
| 1573 | elif current_platform.is_xpu(): |
| 1574 | # nvml/amdsmi are unavailable on XPU. Query device memory through |
| 1575 | # torch.accelerator.get_memory_info, which the XPU platform patches |
| 1576 | # to return (free, total) bytes via Level Zero. |
| 1577 | free_b, total_b = torch.accelerator.get_memory_info(device) |
| 1578 | gb_used = (total_b - free_b) / 2**30 |
| 1579 | gb_total = total_b / 2**30 |
| 1580 | else: |
| 1581 | dev_handle = get_nvml_device_handle(device) |
| 1582 | mem_info = nvmlDeviceGetMemoryInfo(dev_handle) |
| 1583 | gb_used = mem_info.used / 2**30 |
| 1584 | gb_total = mem_info.total / 2**30 |
| 1585 | output[device] = (gb_used, gb_total) |
| 1586 | return output |
| 1587 | |
| 1588 | |
| 1589 | def wait_for_gpu_memory_to_clear( |
| 1590 | *, |
| 1591 | devices: list[int], |
| 1592 | threshold_bytes: int | dict[int, int] | None = None, |
| 1593 | threshold_ratio: float | dict[int, float] | None = None, |
| 1594 | timeout_s: float = 120, |
| 1595 | stable_duration_s: float = 0, |
| 1596 | stable_tolerance_bytes: int = 512 * 1024**2, |
| 1597 | poll_interval_s: float = 5, |
| 1598 | ) -> None: |
| 1599 | assert threshold_bytes is not None or threshold_ratio is not None |
| 1600 | # HIP already applies device visibility; keep logical IDs and threshold keys. |
| 1601 | if not current_platform.is_rocm(): |
| 1602 | devices = get_physical_device_indices(devices) |
| 1603 | if isinstance(threshold_bytes, int): |
| 1604 | threshold_bytes = {device: threshold_bytes for device in devices} |
| 1605 | elif isinstance(threshold_bytes, dict): |
| 1606 | assert threshold_bytes.keys() == set(devices) |
| 1607 | if isinstance(threshold_ratio, float): |
| 1608 | threshold_ratio = {device: threshold_ratio for device in devices} |
| 1609 | elif isinstance(threshold_ratio, dict): |
| 1610 | assert threshold_ratio.keys() == set(devices) |
| 1611 | if current_platform.is_rocm() and threshold_ratio is not None: |
| 1612 | # ROCm can keep a small runtime/driver footprint resident even after |
| 1613 | # all model allocations are gone. On MI300 this has been observed |
| 1614 | # around 2.5 GiB, which is above a strict 1% idle threshold but nowhere |
| 1615 | # near the amount of free memory needed by the next vLLM runner. |
| 1616 | min_threshold_b = 4 * 1024**3 |
| 1617 | if threshold_bytes is None: |
| 1618 | threshold_bytes = {} |
| 1619 | for device, ratio in threshold_ratio.items(): |
| 1620 | threshold_bytes[device] = max( |
| 1621 | threshold_bytes.get(device, 0), min_threshold_b if ratio < 0.05 else 0 |
| 1622 | ) |
| 1623 | |
| 1624 | start_time = time.time() |
| 1625 | stable_since: float | None = None |
| 1626 | stable_used_bytes: dict[int, int] | None = None |
| 1627 | while True: |
| 1628 | output_raw = record_gpu_memory_usage_stats(devices=devices) |
| 1629 | used_bytes_by_device = { |
| 1630 | device: int(gb_used * 2**30) for device, (gb_used, _) in output_raw.items() |
| 1631 | } |
| 1632 | output = { |
| 1633 | device: f"{gb_used:.02f}/{gb_total:.02f}" |
| 1634 | for device, (gb_used, gb_total) in output_raw.items() |
| 1635 | } |
| 1636 | print("gpu memory used/total (GiB): ", end="") |
| 1637 | for k, v in output.items(): |
| 1638 | print(f"{k}={v}; ", end="") |
| 1639 | print("") |
| 1640 | |
| 1641 | if threshold_bytes is not None and threshold_ratio is not None: |
| 1642 | threshold_gib = { |
| 1643 | device: threshold_b / 2**30 |
| 1644 | for device, threshold_b in threshold_bytes.items() |
| 1645 | } |
| 1646 | threshold = "; ".join( |
| 1647 | f"{device=}: max({threshold_gib[device]:.2f} GiB, " |
| 1648 | f"{threshold_ratio[device]:.3f})" |
| 1649 | for device in devices |
| 1650 | ) |
| 1651 | all_free = all( |
| 1652 | used <= max(threshold_gib[device], total * threshold_ratio[device]) |
| 1653 | for device, (used, total) in output_raw.items() |
| 1654 | ) |
| 1655 | elif threshold_bytes is not None: |
| 1656 | threshold_gib = { |
| 1657 | device: threshold_b / 2**30 |
| 1658 | for device, threshold_b in threshold_bytes.items() |
| 1659 | } |
| 1660 | threshold = "; ".join( |
| 1661 | f"{device=}: {threshold_gib[device]:.2f} GiB" for device in devices |
| 1662 | ) |
| 1663 | all_free = all( |
| 1664 | used <= threshold_gib[device] |
| 1665 | for device, (used, _) in output_raw.items() |
| 1666 | ) |
| 1667 | else: |
| 1668 | assert threshold_ratio is not None |
| 1669 | threshold = "; ".join( |
| 1670 | f"{device=}: {threshold_ratio[device]:.3f}" for device in devices |
| 1671 | ) |
| 1672 | all_free = all( |
| 1673 | used / total <= threshold_ratio[device] |
| 1674 | for device, (used, total) in output_raw.items() |
| 1675 | ) |
| 1676 | |
| 1677 | dur_s = time.time() - start_time |
| 1678 | if all_free: |
| 1679 | if stable_duration_s <= 0: |
| 1680 | print( |
| 1681 | f"Done waiting for free GPU memory on devices {devices=} " |
| 1682 | f"({threshold=}) {dur_s=:.02f}" |
| 1683 | ) |
| 1684 | break |
| 1685 | |
| 1686 | now = time.time() |
| 1687 | if stable_used_bytes is None: |
| 1688 | stable_since = now |
| 1689 | stable_used_bytes = used_bytes_by_device |
| 1690 | else: |
| 1691 | memory_changed = any( |
| 1692 | abs(used_bytes_by_device[device] - stable_used_bytes[device]) |
| 1693 | > stable_tolerance_bytes |
| 1694 | for device in devices |
| 1695 | ) |
| 1696 | if memory_changed: |
| 1697 | stable_since = now |
| 1698 | stable_used_bytes = used_bytes_by_device |
| 1699 | elif ( |
| 1700 | stable_since is not None and now - stable_since >= stable_duration_s |
| 1701 | ): |
| 1702 | print( |
| 1703 | f"Done waiting for stable free GPU memory on devices " |
| 1704 | f"{devices=} ({threshold=}) {dur_s=:.02f}" |
| 1705 | ) |
| 1706 | break |
| 1707 | else: |
| 1708 | stable_since = None |
| 1709 | stable_used_bytes = None |
| 1710 | |
| 1711 | if dur_s >= timeout_s: |
| 1712 | raise ValueError( |
| 1713 | f"Memory of devices {devices=} not free after " |
| 1714 | f"{dur_s=:.02f} ({threshold=})" |
| 1715 | ) |
| 1716 | |
| 1717 | time.sleep(poll_interval_s) |
| 1718 | |
| 1719 | |
| 1720 | def wait_for_memory_to_settle( |
| 1721 | *, |
| 1722 | threshold_ratio: float | dict[int, float] | None = 0.1, |
| 1723 | timeout_s: float = 240, |
| 1724 | ) -> None: |
| 1725 | """Block until ROCm or XPU device VRAM usage drops below ``threshold_ratio``. |
| 1726 | |
| 1727 | ROCm and XPU reclaims GPU memory more lazily than CUDA, so back-to-back model |
| 1728 | loads in a single test process can OOM the *next* engine/model startup |
| 1729 | even after ``cleanup_dist_env_and_memory``. This gives the driver time to |
| 1730 | actually release VRAM before the next allocation. No-op off ROCm. |
| 1731 | """ |
| 1732 | if not current_platform.is_rocm() and not current_platform.is_xpu(): |
| 1733 | return |
| 1734 | |
| 1735 | num_gpus = current_platform.device_count() |
| 1736 | if num_gpus == 0: |
| 1737 | return |
| 1738 | if threshold_ratio is None: |
| 1739 | threshold_ratio = 0.1 |
| 1740 | |
| 1741 | wait_for_gpu_memory_to_clear( |
| 1742 | devices=list(range(num_gpus)), |
| 1743 | threshold_ratio=threshold_ratio, |
| 1744 | timeout_s=timeout_s, |
| 1745 | stable_duration_s=2.0, |
| 1746 | poll_interval_s=1.0, |
| 1747 | ) |
| 1748 | |
| 1749 | |
| 1750 | _P = ParamSpec("_P") |
| 1751 | |
| 1752 | |
| 1753 | def fork_new_process_for_each_test(func: Callable[_P, None]) -> Callable[_P, None]: |
| 1754 | """Decorator to fork a new process for each test function. |
| 1755 | See https://github.com/vllm-project/vllm/issues/7053 for more details. |
| 1756 | """ |
| 1757 | |
| 1758 | @functools.wraps(func) |
| 1759 | def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> None: |
| 1760 | from _pytest.outcomes import Skipped |
| 1761 | |
| 1762 | # Create a unique temporary file to store exception info from child |
| 1763 | # process. Use test function name and process ID to avoid collisions. |
| 1764 | with ( |
| 1765 | tempfile.NamedTemporaryFile( |
| 1766 | delete=False, |
| 1767 | mode="w+b", |
| 1768 | prefix=f"vllm_test_{func.__name__}_{os.getpid()}_", |
| 1769 | suffix=".exc", |
| 1770 | ) as exc_file, |
| 1771 | ExitStack() as delete_after, |
| 1772 | ): |
| 1773 | exc_file_path = exc_file.name |
| 1774 | delete_after.callback(os.remove, exc_file_path) |
| 1775 | |
| 1776 | pid = os.fork() |
| 1777 | print(f"Fork a new process to run a test {pid}") |
| 1778 | if pid == 0: |
| 1779 | # Make the child process the leader of its own process group |
| 1780 | # to avoid sending SIGTERM to the parent process |
| 1781 | os.setpgrp() |
| 1782 | # Parent process responsible for deleting, don't delete |
| 1783 | # in child. |
| 1784 | delete_after.pop_all() |
| 1785 | try: |
| 1786 | func(*args, **kwargs) |
| 1787 | except Skipped as e: |
| 1788 | # convert Skipped to exit code 0 |
| 1789 | print(str(e)) |
| 1790 | os._exit(0) |
| 1791 | except Exception as e: |
| 1792 | import traceback |
| 1793 | |
| 1794 | tb_string = traceback.format_exc() |
| 1795 | |
| 1796 | # Try to serialize the exception object first |
| 1797 | exc_to_serialize: dict[str, Any] |
| 1798 | try: |
| 1799 | # First, try to pickle the actual exception with |
| 1800 | # its traceback. |
| 1801 | exc_to_serialize = {"pickled_exception": e} |
| 1802 | # Test if it can be pickled |
| 1803 | cloudpickle.dumps(exc_to_serialize) |
| 1804 | except (Exception, KeyboardInterrupt): |
| 1805 | # Fall back to string-based approach. |
| 1806 | exc_to_serialize = { |
| 1807 | "exception_type": type(e).__name__, |
| 1808 | "exception_msg": str(e), |
| 1809 | "traceback": tb_string, |
| 1810 | } |
| 1811 | try: |
| 1812 | with open(exc_file_path, "wb") as f: |
| 1813 | cloudpickle.dump(exc_to_serialize, f) |
| 1814 | except Exception: |
| 1815 | # Fallback: just print the traceback. |
| 1816 | print(tb_string) |
| 1817 | os._exit(1) |
| 1818 | else: |
| 1819 | os._exit(0) |
| 1820 | else: |
| 1821 | # After setpgrp(), the child's pgid equals its pid |
| 1822 | pgid = pid |
| 1823 | _pid, _exitcode = os.waitpid(pid, 0) |
| 1824 | # kill all child processes - but they may already have exited cleanly |
| 1825 | with contextlib.suppress(ProcessLookupError): |
| 1826 | os.killpg(pgid, signal.SIGTERM) |
| 1827 | if _exitcode != 0: |
| 1828 | # Try to read the exception from the child process |
| 1829 | exc_info = {} |
| 1830 | if os.path.exists(exc_file_path): |
| 1831 | with ( |
| 1832 | contextlib.suppress(Exception), |
| 1833 | open(exc_file_path, "rb") as f, |
| 1834 | ): |
| 1835 | exc_info = cloudpickle.load(f) |
| 1836 | |
| 1837 | if ( |
| 1838 | original_exception := exc_info.get("pickled_exception") |
| 1839 | ) is not None: |
| 1840 | # Re-raise the actual exception object if it was |
| 1841 | # successfully pickled. |
| 1842 | assert isinstance(original_exception, Exception) |
| 1843 | raise original_exception |
| 1844 | |
| 1845 | if (original_tb := exc_info.get("traceback")) is not None: |
| 1846 | # Use string-based traceback for fallback case |
| 1847 | raise AssertionError( |
| 1848 | f"Test {func.__name__} failed when called with" |
| 1849 | f" args {args} and kwargs {kwargs}" |
| 1850 | f" (exit code: {_exitcode}):\n{original_tb}" |
| 1851 | ) from None |
| 1852 | |
| 1853 | # Fallback to the original generic error |
| 1854 | raise AssertionError( |
| 1855 | f"function {func.__name__} failed when called with" |
| 1856 | f" args {args} and kwargs {kwargs}" |
| 1857 | f" (exit code: {_exitcode})" |
| 1858 | ) from None |
| 1859 | |
| 1860 | return wrapper |
| 1861 | |
| 1862 | |
| 1863 | def _format_subprocess_exit(returncode: int) -> str: |
| 1864 | """Render a subprocess exit code, naming the signal for negative codes.""" |
| 1865 | if returncode >= 0: |
| 1866 | return f"exit code {returncode}" |
| 1867 | try: |
| 1868 | return f"killed by {signal.Signals(-returncode).name} ({returncode})" |
| 1869 | except ValueError: |
| 1870 | return f"exit code {returncode}" |
| 1871 | |
| 1872 | |
| 1873 | # Set on the spawn-child interpreter so the wrapper short-circuits when the |
| 1874 | # child resolves `module.qualname` back to its own decorated form, instead of |
| 1875 | # launching another subprocess. |
| 1876 | _SPAWN_CHILD_ENV = "VLLM_TEST_SPAWN_CHILD" |
| 1877 | |
| 1878 | |
| 1879 | def spawn_new_process_for_each_test(f: Callable[_P, None]) -> Callable[_P, None]: |
| 1880 | """Decorator to spawn a new process for each test function. |
| 1881 | |
| 1882 | Uses subprocess to run each test in a fresh interpreter and propagates |
| 1883 | exceptions back to the parent, so test failures are never silently |
| 1884 | swallowed (fixes https://github.com/vllm-project/vllm/issues/41415). |
| 1885 | |
| 1886 | The child resolves the test function by importing its module and looking |
| 1887 | it up by qualified name, rather than reconstructing it from a cloudpickle |
| 1888 | blob. Pickling the function by value would also pickle its ``__globals__`` |
| 1889 | by value — turning module-level singletons (e.g. |
| 1890 | ``vllm.compilation.counter.compilation_counter``) into stale clones in |
| 1891 | the child, so increments performed by the production code in the child |
| 1892 | would never be observable to the test. |
| 1893 | |
| 1894 | The child inherits the parent's stdout/stderr so its output (engine |
| 1895 | cores, NCCL, CUDA, ...) reaches the test runner live; the Python-level |
| 1896 | traceback or skip reason is serialized to ``tb_file`` for propagation. A |
| 1897 | native crash leaves ``tb_file`` empty — the diagnostic is then only in |
| 1898 | the inherited subprocess output. |
| 1899 | """ |
| 1900 | |
| 1901 | @functools.wraps(f) |
| 1902 | def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> None: |
| 1903 | if os.environ.get(_SPAWN_CHILD_ENV) == "1": |
| 1904 | return f(*args, **kwargs) |
| 1905 | |
| 1906 | with tempfile.NamedTemporaryFile(delete=False, suffix=".tb", mode="wb") as tmp: |
| 1907 | tb_file = tmp.name |
| 1908 | |
| 1909 | try: |
| 1910 | payload = cloudpickle.dumps( |
| 1911 | { |
| 1912 | "module": f.__module__, |
| 1913 | "qualname": f.__qualname__, |
| 1914 | "args": args, |
| 1915 | "kwargs": kwargs, |
| 1916 | "tb_file": tb_file, |
| 1917 | } |
| 1918 | ) |
| 1919 | |
| 1920 | child_script = ( |
| 1921 | "import sys, importlib, cloudpickle, traceback, json\n" |
| 1922 | "try:\n" |
| 1923 | " from _pytest.outcomes import Skipped\n" |
| 1924 | "except ImportError:\n" |
| 1925 | " class Skipped(BaseException): pass\n" |
| 1926 | "data = cloudpickle.loads(sys.stdin.buffer.read())\n" |
| 1927 | "mod = importlib.import_module(data['module'])\n" |
| 1928 | "target = mod\n" |
| 1929 | "for name in data['qualname'].split('.'):\n" |
| 1930 | " target = getattr(target, name)\n" |
| 1931 | "try:\n" |
| 1932 | " target(*data['args'], **data['kwargs'])\n" |
| 1933 | "except Skipped as exc:\n" |
| 1934 | " with open(data['tb_file'], 'w') as fp:\n" |
| 1935 | " json.dump(str(exc), fp)\n" |
| 1936 | " sys.exit(0)\n" |
| 1937 | "except BaseException:\n" |
| 1938 | " with open(data['tb_file'], 'w') as fp:\n" |
| 1939 | " fp.write(traceback.format_exc())\n" |
| 1940 | " sys.exit(1)\n" |
| 1941 | ) |
| 1942 | |
| 1943 | repo_root = str(VLLM_PATH.resolve()) |
| 1944 | env = os.environ.copy() |
| 1945 | env["PYTHONPATH"] = repo_root + os.pathsep + env.get("PYTHONPATH", "") |
| 1946 | env[_SPAWN_CHILD_ENV] = "1" |
| 1947 | |
| 1948 | result = subprocess.run( |
| 1949 | [sys.executable, "-c", child_script], |
| 1950 | input=payload, |
| 1951 | env=env, |
| 1952 | ) |
| 1953 | |
| 1954 | try: |
| 1955 | with open(tb_file) as fp: |
| 1956 | tb = fp.read() |
| 1957 | except OSError: |
| 1958 | tb = "" |
| 1959 | if result.returncode != 0: |
| 1960 | if not tb: |
| 1961 | tb = "<no Python traceback; see subprocess output above>" |
| 1962 | raise RuntimeError( |
| 1963 | f"Test subprocess '{f.__name__}' failed " |
| 1964 | f"({_format_subprocess_exit(result.returncode)}):\n{tb}" |
| 1965 | ) |
| 1966 | if tb: |
| 1967 | pytest.skip(json.loads(tb)) |
| 1968 | finally: |
| 1969 | with contextlib.suppress(OSError): |
| 1970 | os.remove(tb_file) |
| 1971 | |
| 1972 | return wrapper |
| 1973 | |
| 1974 | |
| 1975 | def create_new_process_for_each_test( |
| 1976 | method: Literal["spawn", "fork"] | None = None, |
| 1977 | ) -> Callable[[Callable[_P, None]], Callable[_P, None]]: |
| 1978 | """Creates a decorator that runs each test function in a new process. |
| 1979 | |
| 1980 | Args: |
| 1981 | method: The process creation method. Can be either "spawn" or "fork". |
| 1982 | If not specified, it defaults to "spawn" on ROCm and XPU |
| 1983 | platforms and "fork" otherwise. |
| 1984 | |
| 1985 | Returns: |
| 1986 | A decorator to run test functions in separate processes. |
| 1987 | |
| 1988 | """ |
| 1989 | if method is None: |
| 1990 | method = "spawn" if requires_spawn_multiprocessing() else "fork" |
| 1991 | |
| 1992 | assert method in ["spawn", "fork"], "Method must be either 'spawn' or 'fork'" |
| 1993 | |
| 1994 | if method == "fork": |
| 1995 | return fork_new_process_for_each_test |
| 1996 | |
| 1997 | return spawn_new_process_for_each_test |
| 1998 | |
| 1999 | |
| 2000 | def large_gpu_mark(min_gb: int) -> pytest.MarkDecorator: |
| 2001 | """Get a pytest mark, which skips the test if the GPU doesn't meet |
| 2002 | a minimum memory requirement in GB. |
| 2003 | |
| 2004 | This can be leveraged via `@large_gpu_test` to skip tests in environments |
| 2005 | without enough resources, or called when filtering tests to run directly. |
| 2006 | """ |
| 2007 | try: |
| 2008 | if current_platform.is_cpu(): |
| 2009 | memory_gb = 0 |
| 2010 | else: |
| 2011 | memory_gb = current_platform.get_device_total_memory() / GB_bytes |
| 2012 | except Exception as e: |
| 2013 | warnings.warn( |
| 2014 | f"An error occurred when finding the available memory: {e}", |
| 2015 | stacklevel=2, |
| 2016 | ) |
| 2017 | memory_gb = 0 |
| 2018 | |
| 2019 | return pytest.mark.skipif( |
| 2020 | memory_gb < min_gb, |
| 2021 | reason=( |
| 2022 | f"Need at least {min_gb}GB GPU memory to run the test " |
| 2023 | f"(found {memory_gb}GB)." |
| 2024 | ), |
| 2025 | ) |
| 2026 | |
| 2027 | |
| 2028 | requires_fp8 = pytest.mark.skipif( |
| 2029 | not current_platform.supports_fp8(), |
| 2030 | reason="FP8 is not supported on this GPU (requires Hopper or " |
| 2031 | "Ada architecture, compute capability 8.9+)", |
| 2032 | ) |
| 2033 | |
| 2034 | |
| 2035 | def large_gpu_test(*, min_gb: int): |
| 2036 | """Decorate a test to be skipped if no GPU is available or it does not have |
| 2037 | sufficient memory. |
| 2038 | |
| 2039 | Currently, the CI machine uses L4 GPU which has 24 GB VRAM. |
| 2040 | """ |
| 2041 | mark = large_gpu_mark(min_gb) |
| 2042 | |
| 2043 | def wrapper(f: Callable[_P, None]) -> Callable[_P, None]: |
| 2044 | return mark(f) |
| 2045 | |
| 2046 | return wrapper |
| 2047 | |
| 2048 | |
| 2049 | def multi_gpu_marks(*, num_gpus: int): |
| 2050 | """Get a collection of pytest marks to apply for `@multi_gpu_test`.""" |
| 2051 | test_selector = pytest.mark.distributed(num_gpus=num_gpus) |
| 2052 | test_skipif = pytest.mark.skipif( |
| 2053 | current_platform.device_count() < num_gpus, |
| 2054 | reason=f"Need at least {num_gpus} GPUs to run the test.", |
| 2055 | ) |
| 2056 | |
| 2057 | return [test_selector, test_skipif] |
| 2058 | |
| 2059 | |
| 2060 | def multi_gpu_test(*, num_gpus: int): |
| 2061 | """Decorate a test to be run only when multiple GPUs are available.""" |
| 2062 | marks = multi_gpu_marks(num_gpus=num_gpus) |
| 2063 | |
| 2064 | def wrapper(f: Callable[_P, None]) -> Callable[_P, None]: |
| 2065 | func = create_new_process_for_each_test()(f) |
| 2066 | for mark in reversed(marks): |
| 2067 | func = mark(func) |
| 2068 | |
| 2069 | return func |
| 2070 | |
| 2071 | return wrapper |
| 2072 | |
| 2073 | |
| 2074 | def gpu_tier_mark(*, min_gpus: int = 1, max_gpus: int | None = None): |
| 2075 | """Mark a test to only run when the GPU count falls within [min_gpus, max_gpus]. |
| 2076 | |
| 2077 | Examples: |
| 2078 | @gpu_tier_mark(min_gpus=2) # only on multi-GPU |
| 2079 | @gpu_tier_mark(max_gpus=1) # only on single-GPU |
| 2080 | @gpu_tier_mark(min_gpus=2, max_gpus=4) # 2-4 GPUs only |
| 2081 | |
| 2082 | """ |
| 2083 | gpu_count = current_platform.device_count() |
| 2084 | marks = [] |
| 2085 | |
| 2086 | if min_gpus > 1: |
| 2087 | marks.append(pytest.mark.distributed(num_gpus=min_gpus)) |
| 2088 | |
| 2089 | reasons = [] |
| 2090 | if gpu_count < min_gpus: |
| 2091 | reasons.append(f"Need at least {min_gpus} GPUs (have {gpu_count})") |
| 2092 | if max_gpus is not None and gpu_count > max_gpus: |
| 2093 | reasons.append(f"Need at most {max_gpus} GPUs (have {gpu_count})") |
| 2094 | |
| 2095 | if reasons: |
| 2096 | marks.append(pytest.mark.skipif(True, reason="; ".join(reasons))) |
| 2097 | |
| 2098 | return marks |
| 2099 | |
| 2100 | |
| 2101 | def single_gpu_only(f=None): |
| 2102 | """Skip this test when running in a multi-GPU environment.""" |
| 2103 | marks = gpu_tier_mark(max_gpus=1) |
| 2104 | |
| 2105 | def wrapper(func): |
| 2106 | for mark in reversed(marks): |
| 2107 | func = mark(func) |
| 2108 | return func |
| 2109 | |
| 2110 | return wrapper(f) if f is not None else wrapper |
| 2111 | |
| 2112 | |
| 2113 | def multi_gpu_only(*, num_gpus: int = 2): |
| 2114 | """Skip this test when running on fewer than num_gpus GPUs.""" |
| 2115 | marks = gpu_tier_mark(min_gpus=num_gpus) |
| 2116 | |
| 2117 | def wrapper(f): |
| 2118 | for mark in reversed(marks): |
| 2119 | f = mark(f) |
| 2120 | return f |
| 2121 | |
| 2122 | return wrapper |
| 2123 | |
| 2124 | |
| 2125 | def has_module_attribute(module_name, attribute_name): |
| 2126 | """Helper function to check if a module has a specific attribute.""" |
| 2127 | try: |
| 2128 | module = importlib.import_module(module_name) |
| 2129 | return hasattr(module, attribute_name) |
| 2130 | except ImportError: |
| 2131 | return False |
| 2132 | |
| 2133 | |
| 2134 | def get_attn_backend_list_based_on_platform() -> list[str]: |
| 2135 | if current_platform.is_cuda(): |
| 2136 | return ["FLASH_ATTN", "TRITON_ATTN"] |
| 2137 | elif current_platform.is_rocm(): |
| 2138 | attn_backend_list = ["TRITON_ATTN"] |
| 2139 | try: |
| 2140 | import aiter # noqa: F401 |
| 2141 | |
| 2142 | attn_backend_list.append("ROCM_AITER_FA") |
| 2143 | except Exception: |
| 2144 | print("Skip ROCM_AITER_FA on ROCm as aiter is not installed") |
| 2145 | |
| 2146 | return attn_backend_list |
| 2147 | elif current_platform.is_xpu(): |
| 2148 | return ["FLASH_ATTN", "TRITON_ATTN"] |
| 2149 | else: |
| 2150 | raise ValueError("Unsupported platform") |
| 2151 | |
| 2152 | |
| 2153 | def prep_prompts(batch_size: int, ln_range: tuple[int, int] = (800, 1100)): |
| 2154 | """Generate prompts which a bunch of assignments, |
| 2155 | then asking for the value of one of them. |
| 2156 | The prompt is just under 10k tokens; sliding window is 4k |
| 2157 | so the answer is outside sliding window, but should still be correct. |
| 2158 | |
| 2159 | Args: |
| 2160 | batch_size: number of prompts to generate |
| 2161 | ln_range: an argument to control the length of the prompt |
| 2162 | |
| 2163 | """ |
| 2164 | prompts: list[str] = [] |
| 2165 | answer: list[int] = [] |
| 2166 | indices: list[int] = [] |
| 2167 | random.seed(1) |
| 2168 | for _ in range(batch_size): |
| 2169 | idx = random.randint(30, 90) |
| 2170 | indices.append(idx) |
| 2171 | prompt = ( |
| 2172 | "```python\n# We set a number of variables, " |
| 2173 | f"x{idx} will be important later\n" |
| 2174 | ) |
| 2175 | ln = random.randint(*ln_range) |
| 2176 | for k in range(30, ln): |
| 2177 | v = random.randint(10, 99) |
| 2178 | if k == idx: |
| 2179 | answer.append(v) |
| 2180 | prompt += f"x{k} = {v}\n" |
| 2181 | prompt += f"# Now, we check the value of x{idx}:\n" |
| 2182 | prompt += f"assert x{idx} == " |
| 2183 | prompts.append(prompt) |
| 2184 | return prompts, answer, indices |
| 2185 | |
| 2186 | |
| 2187 | def check_answers( |
| 2188 | indices: list[int], answer: list[int], outputs: list[str], accept_rate: float = 0.7 |
| 2189 | ): |
| 2190 | answer2 = [int(text[0:2].strip()) for text in outputs] |
| 2191 | print(list(zip(indices, zip(answer, answer2)))) |
| 2192 | numok = 0 |
| 2193 | for a1, a2 in zip(answer, answer2): |
| 2194 | if a1 == a2: |
| 2195 | numok += 1 |
| 2196 | frac_ok = numok / len(answer) |
| 2197 | print(f"Num OK: {numok}/{len(answer)} {frac_ok}") |
| 2198 | assert frac_ok >= accept_rate |
| 2199 | |
| 2200 | |
| 2201 | def flat_product(*iterables: Iterable[Any]): |
| 2202 | """Flatten lists of tuples of the cartesian product. |
| 2203 | Useful when we want to avoid nested tuples to allow |
| 2204 | test params to be unpacked directly from the decorator. |
| 2205 | |
| 2206 | Example: |
| 2207 | flat_product([(1, 2), (3, 4)], ["a", "b"]) -> |
| 2208 | [ |
| 2209 | (1, 2, "a"), |
| 2210 | (1, 2, "b"), |
| 2211 | (3, 4, "a"), |
| 2212 | (3, 4, "b"), |
| 2213 | ] |
| 2214 | |
| 2215 | """ |
| 2216 | for element in itertools.product(*iterables): |
| 2217 | normalized = (e if isinstance(e, tuple) else (e,) for e in element) |
| 2218 | yield tuple(itertools.chain(*normalized)) |
| 2219 | |
| 2220 | |
| 2221 | class TestFP8Layer(torch.nn.Module): |
| 2222 | """Test helper for FP8 linear operations. Creates random weights and scales |
| 2223 | based on quantization configuration. |
| 2224 | |
| 2225 | Args: |
| 2226 | weight_shape: Shape of the weight tensor (out_features, in_features). |
| 2227 | activation_quant_key: Activation quantization configuration. |
| 2228 | weight_quant_key: Weight quantization configuration. |
| 2229 | out_dtype: Output dtype. Defaults to current default dtype. |
| 2230 | force_kernel: Optional kernel to force use of specific implementation. |
| 2231 | |
| 2232 | """ |
| 2233 | |
| 2234 | def __init__( |
| 2235 | self, |
| 2236 | weight_shape: tuple[int, int], |
| 2237 | activation_quant_key: QuantKey, |
| 2238 | weight_quant_key: QuantKey, |
| 2239 | input_dtype: torch.dtype, |
| 2240 | out_dtype: torch.dtype | None = None, |
| 2241 | transpose_weights: bool = False, |
| 2242 | device: torch.device | None = None, |
| 2243 | force_kernel: "type[_KernelT] | None" = None, |
| 2244 | ): |
| 2245 | super().__init__() |
| 2246 | from vllm.model_executor.kernels.linear import init_fp8_linear_kernel |
| 2247 | |
| 2248 | self.input_size_per_partition = weight_shape[1] |
| 2249 | self.output_size_per_partition = weight_shape[0] |
| 2250 | self.logical_widths = [self.output_size_per_partition] |
| 2251 | self.orig_dtype = input_dtype |
| 2252 | act_scale_desc = activation_quant_key.scale |
| 2253 | weight_scale_desc = weight_quant_key.scale |
| 2254 | is_block_wise = act_scale_desc.group_shape.is_per_group() |
| 2255 | if is_block_wise: |
| 2256 | block_size = weight_scale_desc.group_shape.col |
| 2257 | weight_scale_shape = weight_shape[0] // block_size |
| 2258 | self.weight_scale_inv = torch.rand( |
| 2259 | (weight_scale_shape, weight_scale_shape), |
| 2260 | dtype=torch.float32, |
| 2261 | device=device, |
| 2262 | ) |
| 2263 | self.weight = torch.rand(weight_shape, device=device).to(dtype=FP8_DTYPE) |
| 2264 | self.input_scale = None |
| 2265 | self.weight_block_size = [block_size, block_size] |
| 2266 | if transpose_weights: |
| 2267 | self.weight = self.weight.t() |
| 2268 | else: |
| 2269 | per_tensor_weights = weight_scale_desc.group_shape.is_per_tensor() |
| 2270 | is_static_activation_scale = act_scale_desc.static |
| 2271 | weight_scale_shape = (1,) if per_tensor_weights else (weight_shape[0], 1) |
| 2272 | self.weight_scale_inv = None |
| 2273 | self.weight_scale = torch.rand( |
| 2274 | weight_scale_shape, dtype=torch.float32, device=device |
| 2275 | ) |
| 2276 | self.input_scale = ( |
| 2277 | torch.rand(1, dtype=torch.float32, device=device) |
| 2278 | if is_static_activation_scale |
| 2279 | else None |
| 2280 | ) |
| 2281 | self.weight = ( |
| 2282 | torch.rand(weight_shape, device=device).to(dtype=FP8_DTYPE).t() |
| 2283 | ) |
| 2284 | self.input_scale_ub = None |
| 2285 | |
| 2286 | out_dtype = torch.get_default_dtype() if out_dtype is None else out_dtype |
| 2287 | |
| 2288 | self.kernel = init_fp8_linear_kernel( |
| 2289 | activation_quant_key=activation_quant_key, |
| 2290 | weight_quant_key=weight_quant_key, |
| 2291 | weight_shape=weight_shape, |
| 2292 | input_dtype=input_dtype, |
| 2293 | out_dtype=out_dtype, |
| 2294 | force_kernel=force_kernel, |
| 2295 | ) |
| 2296 | self.kernel.process_weights_after_loading(self) |
| 2297 | |
| 2298 | def is_quant_fp8_enabled(self) -> bool: |
| 2299 | return self.kernel.quant_fp8.enabled() |
| 2300 | |
| 2301 | def forward( |
| 2302 | self, y: torch.Tensor, bias: torch.Tensor | None = None |
| 2303 | ) -> torch.Tensor: |
| 2304 | return self.kernel.apply_weights(self, y, bias) |