已合并
[2/2] Add tp cases #18965
dilililiwhy创建于 2025年3月14日
[2/2] Add tp cases #18965
已合并
从refs/pull/18965/head合入到master
共 1 个文件变更+565-0
| @@ -0,0 +1,565 @@ | |||
| 1 | +# Copyright (c) Meta Platforms, Inc. and affiliates | ||
| 2 | +# Owner(s): ["oncall: distributed"] | ||
| 3 | + | ||
| 4 | +import itertools | ||
| 5 | +from copy import deepcopy | ||
| 6 | +from typing import NamedTuple, Optional | ||
| 7 | + | ||
| 8 | +import torch | ||
| 9 | +import torch.distributed as dist | ||
| 10 | +import torch.nn.functional as F | ||
| 11 | +from torch.distributed._tensor import ( | ||
| 12 | + DeviceMesh, | ||
| 13 | + distribute_tensor, | ||
| 14 | + DTensor, | ||
| 15 | + Replicate, | ||
| 16 | + Shard, | ||
| 17 | +) | ||
| 18 | +from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import ( | ||
| 19 | + checkpoint_wrapper, | ||
| 20 | + CheckpointImpl, | ||
| 21 | +) | ||
| 22 | +from torch.distributed.tensor.debug import CommDebugMode | ||
| 23 | +from torch.distributed.tensor.parallel import ( | ||
| 24 | + ColwiseParallel, | ||
| 25 | + loss_parallel, | ||
| 26 | + parallelize_module, | ||
| 27 | + RowwiseParallel, | ||
| 28 | +) | ||
| 29 | +from torch.distributed.tensor.parallel.input_reshard import input_reshard | ||
| 30 | +from torch.testing._internal.common_utils import ( | ||
| 31 | + instantiate_parametrized_tests, | ||
| 32 | + parametrize, | ||
| 33 | + run_tests, | ||
| 34 | +) | ||
| 35 | +from torch.testing._internal.distributed._tensor.common_dtensor import ( | ||
| 36 | + DTensorTestBase, | ||
| 37 | + MLPModule, | ||
| 38 | + ModelArgs, | ||
| 39 | + skip_unless_torch_gpu, | ||
| 40 | + Transformer, | ||
| 41 | +) | ||
| 42 | + | ||
| 43 | +import torch_npu | ||
| 44 | +from torch_npu.testing.common_distributed import with_comms, skipIfUnsupportMultiNPU | ||
| 45 | + | ||
| 46 | + | ||
| 47 | +c10d_functional = torch.ops.c10d_functional | ||
| 48 | +reduce_scatter, all_gather, all_reduce = ( | ||
| 49 | + c10d_functional.reduce_scatter_tensor, | ||
| 50 | + c10d_functional.all_gather_into_tensor, | ||
| 51 | + c10d_functional.all_reduce, | ||
| 52 | +) | ||
| 53 | + | ||
| 54 | + | ||
| 55 | +class ExpCommCounts(NamedTuple): | ||
| 56 | + fwd: Optional[dict] = None | ||
| 57 | + bwd: Optional[dict] = None | ||
| 58 | + optim: Optional[dict] = None | ||
| 59 | + | ||
| 60 | + | ||
| 61 | +class DistTensorParallelExampleTest(DTensorTestBase): | ||
| 62 | + | ||
| 63 | + def world_size(self): | ||
| 64 | + return 2 | ||
| 65 | + | ||
| 66 | + def _check_module(self, m1, m2, check_grad=False): | ||
| 67 | + named_parameters = dict(m1.named_parameters()) | ||
| 68 | + for name, param_m2 in m2.named_parameters(): | ||
| 69 | + self.assertTrue(name in named_parameters) | ||
| 70 | + param_m1 = named_parameters[name] | ||
| 71 | + if check_grad: | ||
| 72 | + param_m2 = param_m2.grad | ||
| 73 | + param_m1 = param_m1.grad | ||
| 74 | + if isinstance(param_m2, DTensor): | ||
| 75 | + replicate = [Replicate()] | ||
| 76 | + param_m2 = param_m2.redistribute( | ||
| 77 | + device_mesh=param_m2.device_mesh, placements=replicate | ||
| 78 | + ).to_local() | ||
| 79 | + self.assertEqual(param_m2, param_m1) | ||
| 80 | + | ||
| 81 | + def _test_mlp_training_e2e(self, is_seq_parallel=False, recompute_activation=False): | ||
| 82 | + inp_size = [8, 10] | ||
| 83 | + # Ensure all tp ranks have same input. | ||
| 84 | + rng_seed = self.rank if is_seq_parallel else 0 | ||
| 85 | + torch.manual_seed(rng_seed) | ||
| 86 | + inp = torch.rand(*inp_size, device=self.device_type) | ||
| 87 | + model = MLPModule(self.device_type) | ||
| 88 | + model_tp = deepcopy(model) | ||
| 89 | + | ||
| 90 | + # Ensure model are initialized the same way. | ||
| 91 | + self._check_module(model, model_tp) | ||
| 92 | + | ||
| 93 | + # Shard module and initialize optimizer. | ||
| 94 | + LR = 0.25 | ||
| 95 | + device_mesh = DeviceMesh( | ||
| 96 | + self.device_type, | ||
| 97 | + torch.arange(0, self.world_size), | ||
| 98 | + ) | ||
| 99 | + parallelize_plan = { | ||
| 100 | + "net1": ( | ||
| 101 | + ColwiseParallel(input_layouts=Shard(0)) | ||
| 102 | + if is_seq_parallel | ||
| 103 | + else ColwiseParallel() | ||
| 104 | + ), | ||
| 105 | + "net2": ( | ||
| 106 | + RowwiseParallel(output_layouts=Shard(0)) | ||
| 107 | + if is_seq_parallel | ||
| 108 | + else RowwiseParallel() | ||
| 109 | + ), | ||
| 110 | + } | ||
| 111 | + model_tp = parallelize_module(model_tp, device_mesh, parallelize_plan) | ||
| 112 | + if recompute_activation: | ||
| 113 | + model_tp = input_reshard( | ||
| 114 | + checkpoint_wrapper( | ||
| 115 | + model_tp, checkpoint_impl=CheckpointImpl.NO_REENTRANT | ||
| 116 | + ), | ||
| 117 | + device_mesh, | ||
| 118 | + None if is_seq_parallel else 0, | ||
| 119 | + ) | ||
| 120 | + optim = torch.optim.SGD(model.parameters(), lr=LR) | ||
| 121 | + optim_tp = torch.optim.SGD(model_tp.parameters(), lr=LR) | ||
| 122 | + | ||
| 123 | + output = model(inp) | ||
| 124 | + output.sum().backward() | ||
| 125 | + | ||
| 126 | + comm_mode = CommDebugMode() | ||
| 127 | + with comm_mode: | ||
| 128 | + output_tp = model_tp(inp) | ||
| 129 | + output_tp.sum().backward() | ||
| 130 | + | ||
| 131 | + self.assertEqual(output, output_tp) | ||
| 132 | + if is_seq_parallel: | ||
| 133 | + self.assertEqual( | ||
| 134 | + comm_mode.get_comm_counts()[c10d_functional.all_gather_into_tensor], 2 | ||
| 135 | + ) | ||
| 136 | + self.assertEqual( | ||
| 137 | + comm_mode.get_comm_counts()[c10d_functional.reduce_scatter_tensor], 1 | ||
| 138 | + ) | ||
| 139 | + else: | ||
| 140 | + self.assertEqual(comm_mode.get_comm_counts()[c10d_functional.all_reduce], 1) | ||
| 141 | + | ||
| 142 | + if is_seq_parallel: | ||
| 143 | + # Sum gradients from different ranks, since input | ||
| 144 | + # are different across ranks for sequence parallel. | ||
| 145 | + dist.all_reduce(model.net1.weight.grad) | ||
| 146 | + dist.all_reduce(model.net1.bias.grad) | ||
| 147 | + dist.all_reduce(model.net2.weight.grad) | ||
| 148 | + dist.all_reduce(model.net2.bias.grad) | ||
| 149 | + | ||
| 150 | + # Ensure gradients are same. | ||
| 151 | + self._check_module(model, model_tp, check_grad=True) | ||
| 152 | + | ||
| 153 | + optim.step() | ||
| 154 | + optim_tp.step() | ||
| 155 | + | ||
| 156 | + # Ensure model weights are still same after update. | ||
| 157 | + # Due to the trick we use for Partial aggregation, we only check the weight when local_rank = 0. | ||
| 158 | + self._check_module(model, model_tp) | ||
| 159 | + | ||
| 160 | + inp = torch.rand(*inp_size, device=self.device_type) | ||
| 161 | + output = model(inp) | ||
| 162 | + output_tp = model_tp(inp) | ||
| 163 | + self.assertEqual(output, output_tp) | ||
| 164 | + | ||
| 165 | + def _test_mlp_inference(self, device_mesh): | ||
| 166 | + inp_size = [8, 10] | ||
| 167 | + # Ensure all tp ranks have same input. | ||
| 168 | + torch.manual_seed(0) | ||
| 169 | + inp = torch.rand(*inp_size, device=self.device_type) | ||
| 170 | + model = MLPModule(self.device_type) | ||
| 171 | + model_tp = deepcopy(model) | ||
| 172 | + | ||
| 173 | + # Ensure model are initialized the same way. | ||
| 174 | + self._check_module(model, model_tp) | ||
| 175 | + | ||
| 176 | + # Shard module and initialize optimizer. | ||
| 177 | + parallelize_plan = { | ||
| 178 | + "net1": ColwiseParallel(), | ||
| 179 | + "net2": RowwiseParallel(), | ||
| 180 | + } | ||
| 181 | + model_tp = parallelize_module(model_tp, device_mesh, parallelize_plan) | ||
| 182 | + | ||
| 183 | + output = model(inp) | ||
| 184 | + output_tp = model_tp(inp) | ||
| 185 | + self.assertEqual(output, output_tp) | ||
| 186 | + | ||
| 187 | + | ||
| 188 | + | ||
| 189 | + | ||
| 190 | + # TODO: need to revisit input_reshard API about why it failed multi-gpu tests. | ||
| 191 | + # @parametrize("recompute_activation", [True, False]) | ||
| 192 | + | ||
| 193 | + def test_mlp_training(self, is_seq_parallel, recompute_activation): | ||
| 194 | + self._test_mlp_training_e2e( | ||
| 195 | + is_seq_parallel=is_seq_parallel, recompute_activation=recompute_activation | ||
| 196 | + ) | ||
| 197 | + | ||
| 198 | + | ||
| 199 | + | ||
| 200 | + def test_mlp_inference(self): | ||
| 201 | + device_mesh = DeviceMesh( | ||
| 202 | + self.device_type, | ||
| 203 | + torch.arange(0, self.world_size), | ||
| 204 | + ) | ||
| 205 | + with torch.inference_mode(): | ||
| 206 | + self._test_mlp_inference(device_mesh) | ||
| 207 | + | ||
| 208 | + def _setup_single_gpu_model(self, model_args, dtype): | ||
| 209 | + return Transformer(model_args).to(device=self.device_type, dtype=dtype) | ||
| 210 | + | ||
| 211 | + def _setup_tp_model(self, model, is_seq_parallel, dtype): | ||
| 212 | + model_tp = deepcopy(model) | ||
| 213 | + self._check_module(model, model_tp) | ||
| 214 | + device_mesh = DeviceMesh(self.device_type, torch.arange(0, self.world_size)) | ||
| 215 | + local_output_for_attn = dtype is torch.float64 | ||
| 216 | + return Transformer.parallelize( | ||
| 217 | + model_tp, | ||
| 218 | + device_mesh, | ||
| 219 | + is_seq_parallel, | ||
| 220 | + local_output_for_attn=local_output_for_attn, | ||
| 221 | + ) | ||
| 222 | + | ||
| 223 | + def _setup_optimizer(self, model, model_tp): | ||
| 224 | + # Step 3: Run test by comparing outputs from single-gpu and multi-gpu models. | ||
| 225 | + LR = 0.25 | ||
| 226 | + optim = torch.optim.Adam(model.parameters(), lr=LR) | ||
| 227 | + optim_tp = torch.optim.Adam(model_tp.parameters(), lr=LR) | ||
| 228 | + return optim, optim_tp | ||
| 229 | + | ||
| 230 | + def _validate_fwd( | ||
| 231 | + self, model, model_tp, inp, expected_comms_dict=None, check_comms=True | ||
| 232 | + ): | ||
| 233 | + # Compare outputs on the same input. | ||
| 234 | + output = model(inp) | ||
| 235 | + with CommDebugMode() as comm_mode: | ||
| 236 | + output_tp = model_tp(inp) | ||
| 237 | + self.assertEqual(output, output_tp) | ||
| 238 | + if check_comms: | ||
| 239 | + self.assertDictEqual(comm_mode.get_comm_counts(), expected_comms_dict or {}) | ||
| 240 | + return output, output_tp | ||
| 241 | + | ||
| 242 | + def _validate_bwd( | ||
| 243 | + self, | ||
| 244 | + model, | ||
| 245 | + model_tp, | ||
| 246 | + output, | ||
| 247 | + output_tp, | ||
| 248 | + expected_comms_dict=None, | ||
| 249 | + check_comms=True, | ||
| 250 | + ): | ||
| 251 | + # Ensure gradients are equal. | ||
| 252 | + output.sum().backward() | ||
| 253 | + with CommDebugMode() as comm_mode: | ||
| 254 | + output_tp.sum().backward() | ||
| 255 | + self._check_module(model, model_tp, check_grad=True) | ||
| 256 | + if check_comms: | ||
| 257 | + self.assertDictEqual(comm_mode.get_comm_counts(), expected_comms_dict or {}) | ||
| 258 | + | ||
| 259 | + def _validate_optim_step( | ||
| 260 | + self, | ||
| 261 | + model, | ||
| 262 | + model_tp, | ||
| 263 | + optim, | ||
| 264 | + optim_tp, | ||
| 265 | + expected_comms_dict=None, | ||
| 266 | + check_comms=True, | ||
| 267 | + ): | ||
| 268 | + optim.step() # Ensure model weights are still the same after update. | ||
| 269 | + from torch.distributed._tensor.experimental import implicit_replication | ||
| 270 | + | ||
| 271 | + with implicit_replication(): | ||
| 272 | + with CommDebugMode() as comm_mode: | ||
| 273 | + optim_tp.step() | ||
| 274 | + self._check_module(model, model_tp) | ||
| 275 | + if check_comms: | ||
| 276 | + self.assertDictEqual(comm_mode.get_comm_counts(), expected_comms_dict or {}) | ||
| 277 | + | ||
| 278 | + | ||
| 279 | + def _thaw_params(thaw_params, model, model_tp): | ||
| 280 | + if not thaw_params: | ||
| 281 | + return | ||
| 282 | + for target_model in [model, model_tp]: | ||
| 283 | + for n, p in target_model.named_parameters(): | ||
| 284 | + if n not in thaw_params: | ||
| 285 | + p.requires_grad_(False) | ||
| 286 | + | ||
| 287 | + | ||
| 288 | + | ||
| 289 | + | ||
| 290 | + | ||
| 291 | + def test_transformer_training(self, is_seq_parallel, dtype: torch.dtype): | ||
| 292 | + EXP_BASE_CC = ExpCommCounts( | ||
| 293 | + fwd={all_reduce: 6, all_gather: 1}, bwd={all_reduce: 9} | ||
| 294 | + ) | ||
| 295 | + EXP_SEQ_PARALLEL_CC = ExpCommCounts( | ||
| 296 | + fwd={reduce_scatter: 6, all_gather: 6}, | ||
| 297 | + bwd={reduce_scatter: 5, all_gather: 6}, | ||
| 298 | + optim={all_reduce: 30}, | ||
| 299 | + ) | ||
| 300 | + | ||
| 301 | + # Disable dropout in the test since we cannot reproduce the same random | ||
| 302 | + # behaviors when comparing single-gpu models with multi-gpu models. | ||
| 303 | + model_args = ModelArgs(dropout_p=0.0) | ||
| 304 | + model = self._setup_single_gpu_model( | ||
| 305 | + model_args, dtype | ||
| 306 | + ) # Step 1: Initialize single-gpu models. | ||
| 307 | + model_tp = self._setup_tp_model( | ||
| 308 | + model, is_seq_parallel, dtype | ||
| 309 | + ) # Step 2: Setup tp model, place onto device mesh. | ||
| 310 | + optim, optim_tp = self._setup_optimizer( | ||
| 311 | + model, model_tp | ||
| 312 | + ) # Step 3: Setup optimizers for both models | ||
| 313 | + | ||
| 314 | + # Initialize input and make sure all ranks have the same input. | ||
| 315 | + inp_size = [8, 8] # [batch_size, seq_len] | ||
| 316 | + if is_seq_parallel: | ||
| 317 | + assert inp_size[1] % self.world_size == 0 | ||
| 318 | + | ||
| 319 | + torch.manual_seed(0) | ||
| 320 | + steps = 10 if type(model) is torch.float64 else 1 | ||
| 321 | + for _ in range(steps): | ||
| 322 | + inp = torch.randint( | ||
| 323 | + model_args.vocab_size, inp_size, device=self.device_type | ||
| 324 | + ) | ||
| 325 | + expected_fwd_comms = ( | ||
| 326 | + EXP_SEQ_PARALLEL_CC.fwd if is_seq_parallel else EXP_BASE_CC.fwd | ||
| 327 | + ) | ||
| 328 | + output, output_tp = self._validate_fwd( | ||
| 329 | + model, model_tp, inp, expected_fwd_comms | ||
| 330 | + ) | ||
| 331 | + expected_bwd_comms = ( | ||
| 332 | + EXP_SEQ_PARALLEL_CC.bwd if is_seq_parallel else EXP_BASE_CC.bwd | ||
| 333 | + ) | ||
| 334 | + self._validate_bwd(model, model_tp, output, output_tp, expected_bwd_comms) | ||
| 335 | + expected_optim_comms = ( | ||
| 336 | + EXP_SEQ_PARALLEL_CC.optim if is_seq_parallel else EXP_BASE_CC.optim | ||
| 337 | + ) | ||
| 338 | + self._validate_optim_step( | ||
| 339 | + model, model_tp, optim, optim_tp, expected_optim_comms | ||
| 340 | + ) | ||
| 341 | + | ||
| 342 | + | ||
| 343 | + | ||
| 344 | + | ||
| 345 | + "thaw_params, is_seq_parallel, dtype, exp_cnts", | ||
| 346 | + [ | ||
| 347 | + ( | ||
| 348 | + None, # all require grad seq_parallel float32 baseline | ||
| 349 | + True, | ||
| 350 | + torch.float32, | ||
| 351 | + ExpCommCounts( | ||
| 352 | + bwd={reduce_scatter: 5, all_gather: 6}, optim={all_reduce: 30} | ||
| 353 | + ), | ||
| 354 | + ), | ||
| 355 | + ( | ||
| 356 | + None, # all require grad no seq_parallel float64 baseline | ||
| 357 | + False, | ||
| 358 | + torch.float64, | ||
| 359 | + ExpCommCounts(bwd={all_reduce: 9}), | ||
| 360 | + ), | ||
| 361 | + # test a subset of LayerNorm bwd output_masks | ||
| 362 | + ( | ||
| 363 | + ("output.weight", "norm.weight", "norm.bias"), # [False, True, True] | ||
| 364 | + True, | ||
| 365 | + torch.float32, | ||
| 366 | + ExpCommCounts(bwd={reduce_scatter: 1}, optim={all_reduce: 6}), | ||
| 367 | + ), | ||
| 368 | + ( | ||
| 369 | + ("tok_embeddings.weight", "output.weight"), # [True, False, False] | ||
| 370 | + True, | ||
| 371 | + torch.float32, | ||
| 372 | + ExpCommCounts(bwd={reduce_scatter: 5, all_gather: 5}), | ||
| 373 | + ), | ||
| 374 | + ( | ||
| 375 | + ( | ||
| 376 | + "tok_embeddings.weight", | ||
| 377 | + "output.weight", | ||
| 378 | + "norm.weight", | ||
| 379 | + "norm.bias", | ||
| 380 | + ), # [True, True, True] | ||
| 381 | + True, | ||
| 382 | + torch.float32, | ||
| 383 | + ExpCommCounts( | ||
| 384 | + bwd={reduce_scatter: 5, all_gather: 5}, optim={all_reduce: 6} | ||
| 385 | + ), | ||
| 386 | + ), | ||
| 387 | + ( | ||
| 388 | + ( | ||
| 389 | + "tok_embeddings.weight", | ||
| 390 | + "output.weight", | ||
| 391 | + "norm.weight", | ||
| 392 | + "norm.bias", | ||
| 393 | + "layers.1.ffn_norm.weight", | ||
| 394 | + "layers.1.ffn_norm.bias", | ||
| 395 | + ), # a single transformerblock layernorm | ||
| 396 | + True, | ||
| 397 | + torch.float32, | ||
| 398 | + ExpCommCounts( | ||
| 399 | + bwd={reduce_scatter: 5, all_gather: 5}, optim={all_reduce: 12} | ||
| 400 | + ), | ||
| 401 | + ), | ||
| 402 | + ( | ||
| 403 | + ( | ||
| 404 | + "tok_embeddings.weight", | ||
| 405 | + "layers.0.attention.wv.weight", | ||
| 406 | + "layers.0.feed_forward.w1.bias", | ||
| 407 | + "layers.1.ffn_norm.bias", | ||
| 408 | + "layers.1.feed_forward.w2.weight", | ||
| 409 | + "output.weight", | ||
| 410 | + ), # varied layer/param types | ||
| 411 | + True, | ||
| 412 | + torch.float32, | ||
| 413 | + ExpCommCounts( | ||
| 414 | + bwd={reduce_scatter: 5, all_gather: 5}, optim={all_reduce: 3} | ||
| 415 | + ), | ||
| 416 | + ), | ||
| 417 | + ], | ||
| 418 | + name_fn=lambda thaw, seq, dtype, *_: f"{'seq_parallel_' if seq else ''}" | ||
| 419 | + + f"{str(dtype).split('.')[-1]}_" | ||
| 420 | + + f"thaw_{'__'.join(sorted({n.rpartition('.')[0].replace('.', '_') for n in thaw})) if thaw else 'all'}", | ||
| 421 | + ) | ||
| 422 | + def test_transformer_req_grad(self, thaw_params, is_seq_parallel, dtype, exp_cnts): | ||
| 423 | + # Sample a subset of `requires_grad` patterns | ||
| 424 | + | ||
| 425 | + # disabling dropout to facilitate single gpu to multi-device comparison | ||
| 426 | + # disable weight-tying to enable more fine-tuning configurations | ||
| 427 | + model_args = ModelArgs(dropout_p=0.0, weight_tying=False) | ||
| 428 | + model = self._setup_single_gpu_model( | ||
| 429 | + model_args, dtype | ||
| 430 | + ) # Step 1: Initialize single-gpu models. | ||
| 431 | + model_tp = self._setup_tp_model( | ||
| 432 | + model, is_seq_parallel, dtype | ||
| 433 | + ) # Step 2: Setup tp model, place onto device mesh. | ||
| 434 | + optim, optim_tp = self._setup_optimizer( | ||
| 435 | + model, model_tp | ||
| 436 | + ) # Step 3: Setup optimizers for both models | ||
| 437 | + DistTensorParallelExampleTest._thaw_params( | ||
| 438 | + thaw_params, model, model_tp | ||
| 439 | + ) # Step 4: set `requires_grad` patterns | ||
| 440 | + | ||
| 441 | + # Initialize input and make sure all ranks have the same input. | ||
| 442 | + inp_size = [8, 8] # [batch_size, seq_len] | ||
| 443 | + if is_seq_parallel: | ||
| 444 | + assert inp_size[1] % self.world_size == 0 | ||
| 445 | + | ||
| 446 | + torch.manual_seed(0) | ||
| 447 | + inp = torch.randint(model_args.vocab_size, inp_size, device=self.device_type) | ||
| 448 | + output, output_tp = self._validate_fwd(model, model_tp, inp, check_comms=False) | ||
| 449 | + self._validate_bwd( | ||
| 450 | + model, model_tp, output, output_tp, exp_cnts.bwd, check_comms=True | ||
| 451 | + ) | ||
| 452 | + self._validate_optim_step( | ||
| 453 | + model, model_tp, optim, optim_tp, exp_cnts.optim, check_comms=True | ||
| 454 | + ) | ||
| 455 | + | ||
| 456 | + | ||
| 457 | + | ||
| 458 | + def test_weight_tying(self): | ||
| 459 | + class TestModule(torch.nn.Module): | ||
| 460 | + def __init__(self) -> None: | ||
| 461 | + super().__init__() | ||
| 462 | + # Initialize different weights for embedding and fc. | ||
| 463 | + torch.manual_seed(1) | ||
| 464 | + self.embedding = torch.nn.Embedding(16, 8) | ||
| 465 | + torch.manual_seed(2) | ||
| 466 | + self.fc = torch.nn.Linear(8, 16) | ||
| 467 | + | ||
| 468 | + def forward(self, x): | ||
| 469 | + return self.fc(self.embedding(x)) | ||
| 470 | + | ||
| 471 | + model = TestModule().to(self.device_type) | ||
| 472 | + parallelize_plan = { | ||
| 473 | + "embedding": ColwiseParallel(), | ||
| 474 | + "fc": RowwiseParallel(), | ||
| 475 | + } | ||
| 476 | + device_mesh = DeviceMesh(self.device_type, list(range(self.world_size))) | ||
| 477 | + parallelize_module(model, device_mesh, parallelize_plan) | ||
| 478 | + | ||
| 479 | + input_size = [5] | ||
| 480 | + torch.manual_seed(0) | ||
| 481 | + inp = torch.randint(16, input_size, device=self.device_type) | ||
| 482 | + | ||
| 483 | + # Without weight tying. | ||
| 484 | + self.assertNotEqual( | ||
| 485 | + model.embedding.weight.to_local(), model.fc.weight.to_local() | ||
| 486 | + ) | ||
| 487 | + output = model(inp) | ||
| 488 | + output.sum().backward() | ||
| 489 | + self.assertNotEqual( | ||
| 490 | + model.embedding.weight.grad.to_local(), model.fc.weight.grad.to_local() | ||
| 491 | + ) | ||
| 492 | + model.zero_grad() | ||
| 493 | + | ||
| 494 | + # With weight tying. | ||
| 495 | + model.fc.weight = model.embedding.weight | ||
| 496 | + | ||
| 497 | + self.assertEqual(model.embedding.weight, model.fc.weight) | ||
| 498 | + self.assertEqual(id(model.embedding.weight), id(model.fc.weight)) | ||
| 499 | + output = model(inp) | ||
| 500 | + output.sum().backward() | ||
| 501 | + self.assertEqual(model.embedding.weight.grad, model.fc.weight.grad) | ||
| 502 | + self.assertEqual(id(model.embedding.weight.grad), id(model.fc.weight.grad)) | ||
| 503 | + | ||
| 504 | + | ||
| 505 | + | ||
| 506 | + def test_loss_parallel(self): | ||
| 507 | + device_mesh = self.build_device_mesh() | ||
| 508 | + comm_mode = CommDebugMode() | ||
| 509 | + | ||
| 510 | + channel_size, channel_dim = 16, 1 | ||
| 511 | + test_setup = [ | ||
| 512 | + (2, (8, channel_size), (8,)), # calling aten.nll_loss_forward | ||
| 513 | + (3, (8, channel_size, 12), (8, 12)), # calling aten.nll_loss2d_forward | ||
| 514 | + ] | ||
| 515 | + weight = torch.rand(channel_size, device=self.device_type) | ||
| 516 | + for input_ndim, input_size, target_size in test_setup: | ||
| 517 | + x = torch.rand(*input_size, device=self.device_type, requires_grad=True) | ||
| 518 | + target = torch.randint(channel_size, target_size, device=self.device_type) | ||
| 519 | + | ||
| 520 | + shard_dims = list(range(input_ndim)) | ||
| 521 | + reductions = ["none", "mean", "sum"] | ||
| 522 | + for shard_dim, reduction in itertools.product(shard_dims, reductions): | ||
| 523 | + dist_x = distribute_tensor(x, device_mesh, [Shard(shard_dim)]) | ||
| 524 | + y = F.cross_entropy(x, target, weight, reduction=reduction) | ||
| 525 | + with loss_parallel(): | ||
| 526 | + if shard_dim == channel_dim: | ||
| 527 | + with comm_mode: | ||
| 528 | + dist_y = F.cross_entropy( | ||
| 529 | + dist_x, target, weight, reduction=reduction | ||
| 530 | + ) | ||
| 531 | + self.assertEqual(comm_mode.get_total_counts(), 3) | ||
| 532 | + self.assertEqual( | ||
| 533 | + comm_mode.get_comm_counts()[c10d_functional.all_reduce], | ||
| 534 | + 3, | ||
| 535 | + ) | ||
| 536 | + self.assertTrue(dist_y.placements[0].is_replicate()) | ||
| 537 | + self.assertEqual(dist_y.to_local(), y) | ||
| 538 | + | ||
| 539 | + with comm_mode: | ||
| 540 | + if reduction == "none": | ||
| 541 | + y.sum().backward() | ||
| 542 | + dist_y.sum().backward() | ||
| 543 | + else: | ||
| 544 | + y.backward() | ||
| 545 | + dist_y.backward() | ||
| 546 | + self.assertEqual(comm_mode.get_total_counts(), 0) | ||
| 547 | + self.assertTrue( | ||
| 548 | + dist_x.grad.placements[0].is_shard(shard_dim) | ||
| 549 | + ) | ||
| 550 | + self.assertEqual(dist_x.grad.full_tensor(), x.grad) | ||
| 551 | + x.grad.zero_() | ||
| 552 | + else: | ||
| 553 | + with self.assertRaisesRegex( | ||
| 554 | + ValueError, | ||
| 555 | + "loss_parallel", | ||
| 556 | + ): | ||
| 557 | + dist_y = F.cross_entropy( | ||
| 558 | + dist_x, target, reduction=reduction | ||
| 559 | + ) | ||
| 560 | + | ||
| 561 | + | ||
| 562 | +instantiate_parametrized_tests(DistTensorParallelExampleTest) | ||
| 563 | + | ||
| 564 | +if __name__ == "__main__": | ||
| 565 | + run_tests() | ||