| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import os |
| 5 | |
| 6 | import multiprocess as mp |
| 7 | import numpy as np |
| 8 | import pytest |
| 9 | import torch |
| 10 | import torch.distributed |
| 11 | |
| 12 | import vllm.envs as envs |
| 13 | from tests.utils import ensure_current_vllm_config |
| 14 | from vllm.distributed.communication_op import tensor_model_parallel_all_reduce # noqa |
| 15 | from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator |
| 16 | from vllm.distributed.device_communicators.pynccl_wrapper import NCCLLibrary |
| 17 | from vllm.distributed.parallel_state import ( |
| 18 | ensure_model_parallel_initialized, |
| 19 | get_tp_group, |
| 20 | get_world_group, |
| 21 | graph_capture, |
| 22 | init_distributed_environment, |
| 23 | ) |
| 24 | from vllm.utils.system_utils import update_environment_variables |
| 25 | |
| 26 | mp.set_start_method("spawn", force=True) |
| 27 | |
| 28 | |
| 29 | def distributed_run(fn, world_size): |
| 30 | number_of_processes = world_size |
| 31 | processes: list[mp.Process] = [] |
| 32 | for i in range(number_of_processes): |
| 33 | env: dict[str, str] = {} |
| 34 | env["RANK"] = str(i) |
| 35 | env["LOCAL_RANK"] = str(i) |
| 36 | env["WORLD_SIZE"] = str(number_of_processes) |
| 37 | env["LOCAL_WORLD_SIZE"] = str(number_of_processes) |
| 38 | env["MASTER_ADDR"] = "localhost" |
| 39 | env["MASTER_PORT"] = "12345" |
| 40 | p = mp.Process(target=fn, args=(env,)) |
| 41 | processes.append(p) |
| 42 | p.start() |
| 43 | |
| 44 | for p in processes: |
| 45 | p.join() |
| 46 | |
| 47 | for p in processes: |
| 48 | assert p.exitcode == 0 |
| 49 | |
| 50 | |
| 51 | def worker_fn_wrapper(fn): |
| 52 | # `multiprocessing.Process` cannot accept environment variables directly |
| 53 | # so we need to pass the environment variables as arguments |
| 54 | # and update the environment variables in the function |
| 55 | def wrapped_fn(env): |
| 56 | update_environment_variables(env) |
| 57 | local_rank = os.environ["LOCAL_RANK"] |
| 58 | device = torch.device(f"cuda:{local_rank}") |
| 59 | torch.accelerator.set_device_index(device) |
| 60 | init_distributed_environment() |
| 61 | fn() |
| 62 | |
| 63 | return wrapped_fn |
| 64 | |
| 65 | |
| 66 | @worker_fn_wrapper |
| 67 | def worker_fn(): |
| 68 | pynccl_comm = PyNcclCommunicator( |
| 69 | get_world_group().cpu_group, device=get_world_group().device |
| 70 | ) |
| 71 | tensor = torch.ones(16, 1024, 1024, dtype=torch.float32).cuda(pynccl_comm.rank) |
| 72 | tensor = pynccl_comm.all_reduce(tensor) |
| 73 | torch.accelerator.synchronize() |
| 74 | assert torch.all(tensor == pynccl_comm.world_size).cpu().item() |
| 75 | |
| 76 | |
| 77 | @pytest.mark.skipif( |
| 78 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 79 | ) |
| 80 | def test_pynccl(): |
| 81 | distributed_run(worker_fn, 2) |
| 82 | |
| 83 | |
| 84 | @worker_fn_wrapper |
| 85 | def multiple_allreduce_worker_fn(): |
| 86 | device = torch.device(f"cuda:{torch.distributed.get_rank()}") |
| 87 | if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP: |
| 88 | # Eager-init path: parent PG has bound_device_id + a CPU backend, |
| 89 | # so split_group is supported. |
| 90 | group = torch.distributed.split_group( |
| 91 | split_ranks=[[0, 1], [2, 3]], backend="cpu:gloo,cuda:nccl" |
| 92 | ) |
| 93 | else: |
| 94 | groups = [ |
| 95 | torch.distributed.new_group(ranks=[0, 1], backend="gloo"), |
| 96 | torch.distributed.new_group(ranks=[2, 3], backend="gloo"), |
| 97 | ] |
| 98 | group = groups[0] if torch.distributed.get_rank() in [0, 1] else groups[1] |
| 99 | pynccl_comm = PyNcclCommunicator(group=group, device=device) |
| 100 | tensor = torch.ones(16, 1024, 1024, dtype=torch.float32, device=device) |
| 101 | # two groups can communicate independently |
| 102 | if torch.distributed.get_rank() in [0, 1]: |
| 103 | tensor = pynccl_comm.all_reduce(tensor) |
| 104 | tensor = pynccl_comm.all_reduce(tensor) |
| 105 | torch.accelerator.synchronize() |
| 106 | assert torch.all(tensor == 4).cpu().item() |
| 107 | else: |
| 108 | tensor = pynccl_comm.all_reduce(tensor) |
| 109 | torch.accelerator.synchronize() |
| 110 | assert torch.all(tensor == 2).cpu().item() |
| 111 | |
| 112 | |
| 113 | @pytest.mark.skipif( |
| 114 | torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test." |
| 115 | ) |
| 116 | def test_pynccl_multiple_allreduce(): |
| 117 | # this tests pynccl for multiple tp groups, in a standalone way |
| 118 | # i.e. call `pynccl_comm.all_reduce` directly |
| 119 | distributed_run(multiple_allreduce_worker_fn, 4) |
| 120 | |
| 121 | |
| 122 | @worker_fn_wrapper |
| 123 | def multiple_allreduce_with_vllm_worker_fn(): |
| 124 | device = torch.device(f"cuda:{torch.distributed.get_rank()}") |
| 125 | with ensure_current_vllm_config(): |
| 126 | ensure_model_parallel_initialized(2, 2) |
| 127 | tensor = torch.ones(16, 1024, 1024, dtype=torch.float32, device=device) |
| 128 | with graph_capture(device=device): |
| 129 | # two tp groups can communicate independently |
| 130 | if torch.distributed.get_rank() in [0, 1]: |
| 131 | tensor = tensor_model_parallel_all_reduce(tensor) |
| 132 | tensor = tensor_model_parallel_all_reduce(tensor) |
| 133 | torch.accelerator.synchronize() |
| 134 | assert torch.all(tensor == 4).cpu().item() |
| 135 | else: |
| 136 | tensor = tensor_model_parallel_all_reduce(tensor) |
| 137 | torch.accelerator.synchronize() |
| 138 | assert torch.all(tensor == 2).cpu().item() |
| 139 | |
| 140 | |
| 141 | @pytest.mark.skipif( |
| 142 | torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test." |
| 143 | ) |
| 144 | def test_pynccl_multiple_allreduce_with_vllm(): |
| 145 | # this tests pynccl for multiple tp groups, together with vllm |
| 146 | # i.e. call `tensor_model_parallel_all_reduce` |
| 147 | distributed_run(multiple_allreduce_with_vllm_worker_fn, 4) |
| 148 | |
| 149 | |
| 150 | @worker_fn_wrapper |
| 151 | def worker_fn_with_cudagraph(): |
| 152 | with torch.no_grad(): |
| 153 | graph = torch.cuda.CUDAGraph() |
| 154 | pynccl_comm = PyNcclCommunicator( |
| 155 | get_world_group().cpu_group, device=get_world_group().device |
| 156 | ) |
| 157 | # run something in the default stream to initialize torch engine |
| 158 | a = torch.ones((4, 4), device=f"cuda:{pynccl_comm.rank}") |
| 159 | torch.accelerator.synchronize() |
| 160 | with torch.cuda.graph(graph): |
| 161 | a_out = pynccl_comm.all_reduce(a) |
| 162 | torch.accelerator.synchronize() |
| 163 | graph.replay() |
| 164 | torch.accelerator.synchronize() |
| 165 | assert torch.all(a_out == pynccl_comm.world_size).cpu().item() |
| 166 | |
| 167 | |
| 168 | @worker_fn_wrapper |
| 169 | def all_gather_worker_fn(): |
| 170 | pynccl_comm = PyNcclCommunicator( |
| 171 | get_world_group().cpu_group, device=get_world_group().device |
| 172 | ) |
| 173 | |
| 174 | rank = pynccl_comm.rank |
| 175 | world_size = pynccl_comm.world_size |
| 176 | device = f"cuda:{pynccl_comm.rank}" |
| 177 | |
| 178 | num_elems = 1000 |
| 179 | tensor = ( |
| 180 | torch.arange(num_elems, dtype=torch.float32, device=device) + rank * num_elems |
| 181 | ) |
| 182 | result = torch.zeros(num_elems * world_size, dtype=torch.float32, device=device) |
| 183 | |
| 184 | expected = torch.cat( |
| 185 | [ |
| 186 | torch.arange(num_elems, dtype=torch.float32) + r * num_elems |
| 187 | for r in range(world_size) |
| 188 | ] |
| 189 | ).to(device) |
| 190 | |
| 191 | pynccl_comm.all_gather(result, tensor) |
| 192 | torch.accelerator.synchronize() |
| 193 | torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8) |
| 194 | |
| 195 | |
| 196 | @pytest.mark.skipif( |
| 197 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 198 | ) |
| 199 | def test_pynccl_all_gather(): |
| 200 | distributed_run(all_gather_worker_fn, 2) |
| 201 | |
| 202 | |
| 203 | @worker_fn_wrapper |
| 204 | def cuda_communicator_all_gather_dim_worker_fn(): |
| 205 | with ensure_current_vllm_config(): |
| 206 | ensure_model_parallel_initialized(2, 1) |
| 207 | |
| 208 | tp_group = get_tp_group() |
| 209 | comm = tp_group.device_communicator |
| 210 | assert comm is not None |
| 211 | |
| 212 | rank = tp_group.rank_in_group |
| 213 | world_size = tp_group.world_size |
| 214 | device = tp_group.device |
| 215 | |
| 216 | shape = (2, 3, 4) |
| 217 | num_elems = 1 |
| 218 | for size in shape: |
| 219 | num_elems *= size |
| 220 | |
| 221 | for dim in (1, -1): |
| 222 | tensor = ( |
| 223 | torch.arange(num_elems, dtype=torch.float32, device=device).reshape(shape) |
| 224 | + rank * num_elems |
| 225 | ) |
| 226 | expected = torch.cat( |
| 227 | [ |
| 228 | torch.arange(num_elems, dtype=torch.float32, device=device).reshape( |
| 229 | shape |
| 230 | ) |
| 231 | + r * num_elems |
| 232 | for r in range(world_size) |
| 233 | ], |
| 234 | dim=dim, |
| 235 | ) |
| 236 | |
| 237 | result = comm.all_gather(tensor, dim=dim) |
| 238 | torch.accelerator.synchronize() |
| 239 | torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8) |
| 240 | |
| 241 | |
| 242 | @pytest.mark.skipif( |
| 243 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 244 | ) |
| 245 | def test_cuda_communicator_all_gather_dim_not_zero(): |
| 246 | distributed_run(cuda_communicator_all_gather_dim_worker_fn, 2) |
| 247 | |
| 248 | |
| 249 | @worker_fn_wrapper |
| 250 | def all_gatherv_worker_fn(): |
| 251 | pynccl_comm = PyNcclCommunicator( |
| 252 | get_world_group().cpu_group, device=get_world_group().device |
| 253 | ) |
| 254 | |
| 255 | rank = pynccl_comm.rank |
| 256 | world_size = pynccl_comm.world_size |
| 257 | device = f"cuda:{pynccl_comm.rank}" |
| 258 | |
| 259 | assert world_size <= 8 |
| 260 | sizes = [81, 20, 57, 52, 81, 5, 49, 49][:world_size] |
| 261 | num_elems = sizes[rank] |
| 262 | tensor = torch.arange(num_elems, dtype=torch.float32, device=device) + rank * 100 |
| 263 | result = torch.zeros(sum(sizes), dtype=torch.float32, device=device) |
| 264 | |
| 265 | expected = torch.cat( |
| 266 | [ |
| 267 | torch.arange(sizes[r], dtype=torch.float32) + r * 100 |
| 268 | for r in range(world_size) |
| 269 | ] |
| 270 | ).to(device) |
| 271 | |
| 272 | pynccl_comm.all_gatherv(result, tensor, sizes=sizes) |
| 273 | torch.accelerator.synchronize() |
| 274 | torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8) |
| 275 | |
| 276 | |
| 277 | @pytest.mark.skipif( |
| 278 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 279 | ) |
| 280 | def test_pynccl_all_gatherv(): |
| 281 | distributed_run(all_gatherv_worker_fn, 2) |
| 282 | |
| 283 | |
| 284 | @worker_fn_wrapper |
| 285 | def reduce_scatter_worker_fn(): |
| 286 | pynccl_comm = PyNcclCommunicator( |
| 287 | get_world_group().cpu_group, device=get_world_group().device |
| 288 | ) |
| 289 | |
| 290 | rank = pynccl_comm.rank |
| 291 | world_size = pynccl_comm.world_size |
| 292 | device = f"cuda:{pynccl_comm.rank}" |
| 293 | |
| 294 | num_elems = 1000 |
| 295 | tensor = ( |
| 296 | torch.arange(num_elems, dtype=torch.float32, device=device) + rank * num_elems |
| 297 | ) |
| 298 | assert num_elems % world_size == 0 |
| 299 | result = torch.zeros(num_elems // world_size, dtype=torch.float32, device=device) |
| 300 | |
| 301 | # Calculate expected result for this rank's chunk |
| 302 | scattered_size = num_elems // world_size |
| 303 | all_tensors = [ |
| 304 | torch.arange(num_elems, dtype=torch.float32) + r * num_elems |
| 305 | for r in range(world_size) |
| 306 | ] |
| 307 | expected = sum( |
| 308 | tensor[rank * scattered_size : (rank + 1) * scattered_size] |
| 309 | for tensor in all_tensors |
| 310 | ).to(device) |
| 311 | |
| 312 | pynccl_comm.reduce_scatter(result, tensor) |
| 313 | torch.accelerator.synchronize() |
| 314 | torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8) |
| 315 | |
| 316 | |
| 317 | @pytest.mark.skipif( |
| 318 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 319 | ) |
| 320 | def test_pynccl_reduce_scatter(): |
| 321 | distributed_run(reduce_scatter_worker_fn, 2) |
| 322 | |
| 323 | |
| 324 | @worker_fn_wrapper |
| 325 | def reduce_scatterv_worker_fn(): |
| 326 | pynccl_comm = PyNcclCommunicator( |
| 327 | get_world_group().cpu_group, device=get_world_group().device |
| 328 | ) |
| 329 | |
| 330 | rank = pynccl_comm.rank |
| 331 | world_size = pynccl_comm.world_size |
| 332 | device = f"cuda:{pynccl_comm.rank}" |
| 333 | |
| 334 | assert world_size <= 8 |
| 335 | sizes = [81, 20, 57, 52, 81, 5, 49, 49][:world_size] |
| 336 | num_elems = sum(sizes) |
| 337 | tensor = torch.arange(num_elems, dtype=torch.float32, device=device) + rank * 100 |
| 338 | result = torch.zeros(sizes[rank], dtype=torch.float32, device=device) |
| 339 | |
| 340 | # Calculate expected result for this rank's chunk |
| 341 | all_tensors = [ |
| 342 | torch.arange(num_elems, dtype=torch.float32) + r * 100 |
| 343 | for r in range(world_size) |
| 344 | ] |
| 345 | sizes_cumsum = np.cumsum(sizes) |
| 346 | start = 0 if rank == 0 else sizes_cumsum[rank - 1] |
| 347 | end = sizes_cumsum[rank] |
| 348 | expected = sum(tensor[start:end] for tensor in all_tensors).to(device) |
| 349 | |
| 350 | pynccl_comm.reduce_scatterv(result, tensor, sizes=sizes) |
| 351 | torch.accelerator.synchronize() |
| 352 | torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8) |
| 353 | |
| 354 | |
| 355 | @pytest.mark.skipif( |
| 356 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 357 | ) |
| 358 | def test_pynccl_reduce_scatterv(): |
| 359 | distributed_run(reduce_scatterv_worker_fn, 2) |
| 360 | |
| 361 | |
| 362 | @pytest.mark.skipif( |
| 363 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 364 | ) |
| 365 | def test_pynccl_with_cudagraph(): |
| 366 | distributed_run(worker_fn_with_cudagraph, 2) |
| 367 | |
| 368 | |
| 369 | @worker_fn_wrapper |
| 370 | def send_recv_worker_fn(): |
| 371 | pynccl_comm = PyNcclCommunicator( |
| 372 | get_world_group().cpu_group, device=get_world_group().device |
| 373 | ) |
| 374 | if pynccl_comm.rank == 0: |
| 375 | tensor = torch.ones(16, 1024, 1024, dtype=torch.float32).cuda(pynccl_comm.rank) |
| 376 | else: |
| 377 | tensor = torch.empty(16, 1024, 1024, dtype=torch.float32).cuda(pynccl_comm.rank) |
| 378 | |
| 379 | if pynccl_comm.rank == 0: |
| 380 | pynccl_comm.send(tensor, dst=(pynccl_comm.rank + 1) % pynccl_comm.world_size) |
| 381 | else: |
| 382 | pynccl_comm.recv(tensor, src=(pynccl_comm.rank - 1) % pynccl_comm.world_size) |
| 383 | torch.accelerator.synchronize() |
| 384 | assert torch.all(tensor == 1).cpu().item() |
| 385 | |
| 386 | |
| 387 | @pytest.mark.skipif( |
| 388 | torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test." |
| 389 | ) |
| 390 | def test_pynccl_send_recv(): |
| 391 | distributed_run(send_recv_worker_fn, 2) |
| 392 | |
| 393 | |
| 394 | @worker_fn_wrapper |
| 395 | def multiple_send_recv_worker_fn(): |
| 396 | device = torch.device(f"cuda:{torch.distributed.get_rank()}") |
| 397 | if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP: |
| 398 | group = torch.distributed.split_group( |
| 399 | split_ranks=[[0, 2], [1, 3]], backend="cpu:gloo,cuda:nccl" |
| 400 | ) |
| 401 | else: |
| 402 | groups = [ |
| 403 | torch.distributed.new_group(ranks=[0, 2], backend="gloo"), |
| 404 | torch.distributed.new_group(ranks=[1, 3], backend="gloo"), |
| 405 | ] |
| 406 | group = groups[0] if torch.distributed.get_rank() in [0, 2] else groups[1] |
| 407 | pynccl_comm = PyNcclCommunicator(group=group, device=device) |
| 408 | if torch.distributed.get_rank() == 0: |
| 409 | tensor = torch.ones(16, 1024, 1024, dtype=torch.float32, device=device) |
| 410 | elif torch.distributed.get_rank() == 1: |
| 411 | tensor = 2 * torch.ones(16, 1024, 1024, dtype=torch.float32, device=device) |
| 412 | else: |
| 413 | tensor = torch.empty(16, 1024, 1024, dtype=torch.float32, device=device) |
| 414 | if torch.distributed.get_rank() in [0, 1]: |
| 415 | pynccl_comm.send(tensor, dst=(pynccl_comm.rank + 1) % pynccl_comm.world_size) |
| 416 | else: |
| 417 | pynccl_comm.recv(tensor, src=(pynccl_comm.rank - 1) % pynccl_comm.world_size) |
| 418 | torch.accelerator.synchronize() |
| 419 | if torch.distributed.get_rank() in [0, 2]: |
| 420 | assert torch.all(tensor == 1).cpu().item() |
| 421 | else: |
| 422 | assert torch.all(tensor == 2).cpu().item() |
| 423 | |
| 424 | |
| 425 | @pytest.mark.skipif( |
| 426 | torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test." |
| 427 | ) |
| 428 | def test_pynccl_multiple_send_recv(): |
| 429 | distributed_run(multiple_send_recv_worker_fn, 4) |
| 430 | |
| 431 | |
| 432 | @pytest.mark.skipif( |
| 433 | torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test." |
| 434 | ) |
| 435 | def test_pynccl_broadcast(): |
| 436 | distributed_run(broadcast_worker_fn, 4) |
| 437 | |
| 438 | |
| 439 | @worker_fn_wrapper |
| 440 | def broadcast_worker_fn(): |
| 441 | # Test broadcast for every root rank. |
| 442 | # Essentially this is an all-gather operation. |
| 443 | pynccl_comm = PyNcclCommunicator( |
| 444 | get_world_group().cpu_group, device=get_world_group().device |
| 445 | ) |
| 446 | recv_tensors = [ |
| 447 | torch.empty(16, 1024, 1024, dtype=torch.float32, device=pynccl_comm.device) |
| 448 | for i in range(pynccl_comm.world_size) |
| 449 | ] |
| 450 | recv_tensors[pynccl_comm.rank] = ( |
| 451 | torch.ones(16, 1024, 1024, dtype=torch.float32, device=pynccl_comm.device) |
| 452 | * pynccl_comm.rank |
| 453 | ) |
| 454 | |
| 455 | for i in range(pynccl_comm.world_size): |
| 456 | pynccl_comm.broadcast(recv_tensors[i], src=i) |
| 457 | # the broadcast op might be launched in a different stream |
| 458 | # need to synchronize to make sure the tensor is ready |
| 459 | torch.accelerator.synchronize() |
| 460 | assert torch.all(recv_tensors[i] == i).cpu().item() |
| 461 | |
| 462 | |
| 463 | def test_ncclGetUniqueId(): |
| 464 | lib = NCCLLibrary() |
| 465 | unique_id = lib.ncclGetUniqueId() |
| 466 | # `list(unique_id.internal)` is something like this: |
| 467 | # [34, -16, 23, 83, 109, -19, 59, 95, 2, 0, -86, 55, 10, -128, 0, 29, 0, |
| 468 | # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 469 | # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 470 | # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 471 | # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 472 | # 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 473 | # as long as the function doesn't raise an exception, we're good |
| 474 | assert unique_id is not None |
| 475 | |
| 476 | |
| 477 | def test_pynccl_suspend_resume_idempotent(): |
| 478 | """Repeated suspend/resume (e.g. staged wake-ups) reach NCCL once each.""" |
| 479 | from unittest.mock import Mock |
| 480 | |
| 481 | comm = object.__new__(PyNcclCommunicator) |
| 482 | comm.disabled = False |
| 483 | comm._suspended = False |
| 484 | comm.comm = object() |
| 485 | comm.nccl = Mock() |
| 486 | |
| 487 | comm.suspend() |
| 488 | comm.suspend() |
| 489 | comm.resume() |
| 490 | comm.resume() |
| 491 | |
| 492 | assert comm.nccl.ncclCommSuspend.call_count == 1 |
| 493 | assert comm.nccl.ncclCommResume.call_count == 1 |
| 494 | |
| 495 | |
| 496 | def test_pynccl_suspend_noop_without_symbol(): |
| 497 | """RCCL / NCCL < 2.29.7 lack ncclCommSuspend; suspend() must no-op, not |
| 498 | crash, and resume() must then no-op too since nothing was suspended.""" |
| 499 | from unittest.mock import Mock |
| 500 | |
| 501 | comm = object.__new__(PyNcclCommunicator) |
| 502 | comm.disabled = False |
| 503 | comm._suspended = False |
| 504 | comm.comm = object() |
| 505 | comm.nccl = Mock() |
| 506 | comm.nccl.has_symbol.return_value = False |
| 507 | |
| 508 | comm.suspend() |
| 509 | comm.resume() |
| 510 | |
| 511 | comm.nccl.ncclCommSuspend.assert_not_called() |
| 512 | comm.nccl.ncclCommResume.assert_not_called() |
| 513 | assert comm._suspended is False |