# Owner(s): ["oncall: distributed"] import functools import os import sys import warnings from collections import namedtuple from contextlib import nullcontext from copy import deepcopy from itertools import chain from typing import Any, Tuple import torch import torch.distributed as dist import torch.distributed.fsdp._traversal_utils as traversal_utils import torch.nn as nn from torch.distributed.fsdp import ( CPUOffload, FlatParameter, FullyShardedDataParallel as FSDP, ShardingStrategy, ) from torch.distributed.fsdp._flat_param import _FSDP_USE_UNSAFE_SETATTR from torch.distributed.fsdp._runtime_utils import HOMOGENEOUS_ATTR_NAMES from torch.distributed.fsdp.wrap import ( always_wrap_policy, ModuleWrapPolicy, transformer_auto_wrap_policy, ) from torch.distributed.optim import _apply_optimizer_in_backward from torch.nn import TransformerDecoderLayer, TransformerEncoderLayer from torch.nn.parallel import DistributedDataParallel as DDP from torch.testing._internal.common_distributed import skip_if_lt_x_gpu from torch.testing._internal.common_fsdp import ( _assert_module_states, CUDAInitMode, FSDPInitMode, FSDPTest, FSDPTestMultiThread, MLP, NestedWrappedModule, TransformerWithSharedParams, ) from torch.testing._internal.common_utils import ( instantiate_parametrized_tests, parametrize, run_tests, TEST_WITH_DEV_DBG_ASAN, ) if not dist.is_available(): print("Distributed not available, skipping tests", file=sys.stderr) sys.exit(0) if TEST_WITH_DEV_DBG_ASAN: print( "Skip dev-asan as torch + multiprocessing spawn have known issues", file=sys.stderr, ) sys.exit(0) class MyModel(nn.Module): def __init__(self) -> None: super().__init__() self.a = nn.Linear(2, 2) self.b = nn.Linear(2, 2) def forward(self, x, y): return self.b(self.a(x + y)) class TestFSDPMiscMultiProcess(FSDPTest): @property def world_size(self): return 2 @property def process_group(self): return dist.distributed_c10d._get_default_group() @skip_if_lt_x_gpu(2) @parametrize("use_index", [True, False]) def test_fsdp_device_id(self, use_index): """ Tests the FSDP ``device_id`` argument: - Wrapping a CPU module should move the module to the GPU matching ``device_id`` - Wrapping a GPU module already on the GPU matching ``device_id`` should not raise an error - Wrapping a GPU module already on GPU and passing a GPU device without specifying a device ID (i.e. ``torch.device("cuda")``) warns """ dev_id = ( torch.cuda.current_device() if use_index else torch.device("cuda", torch.cuda.current_device()) ) def _check_device_matches(module, device_id): """Checks that the ``FlatParameter``s in ``module`` have device matching ``device_id``.""" devices = { p.device for p in module.parameters() if isinstance(p, FlatParameter) } assert len(devices) > 0 self.assertEqual(1, len(devices)) found_device = devices.pop() if use_index and not isinstance(device_id, torch.device): device = torch.device("cuda", device_id) else: device = device_id self.assertEqual(found_device, device) # Check that FSDP parameters are moved to `device_id` for a CPU module nested_wrapped_module = NestedWrappedModule.init( self.process_group, FSDPInitMode.RECURSIVE, CUDAInitMode.CUDA_NEVER, fsdp_kwargs={"device_id": dev_id}, ) _check_device_matches(nested_wrapped_module, dev_id) # Check that specifying `device_id` for a GPU module already on that # device does not raise an error nested_wrapped_module = NestedWrappedModule.init( self.process_group, FSDPInitMode.RECURSIVE, CUDAInitMode.CUDA_BEFORE, fsdp_kwargs={"device_id": dev_id}, ) _check_device_matches(nested_wrapped_module, dev_id) # Check that passing in `torch.device("cuda")` for a GPU module warns regex = "does not have an explicit index" context = self.assertWarnsRegex( expected_warning=UserWarning, expected_regex=regex ) with context: nested_wrapped_module = NestedWrappedModule.init( self.process_group, FSDPInitMode.RECURSIVE, CUDAInitMode.CUDA_BEFORE, fsdp_kwargs={"device_id": torch.device("cuda")}, ) _check_device_matches( nested_wrapped_module, torch.device("cuda", torch.cuda.current_device()) ) @skip_if_lt_x_gpu(2) def test_fsdp_zero2_eval_with_prefetch(self): # Test FSDP validation with SHARD_GRAD_OP and forward_prefetch class Mnist(nn.Module): def __init__(self) -> None: super().__init__() self.conv1 = nn.Conv2d(1, 32, 3, 1) self.conv2 = nn.Conv2d(32, 64, 3, 1) self.dropout1 = nn.Dropout(0.25) self.dropout2 = nn.Dropout(0.5) self.fc1 = nn.Linear(9216, 128) self.fc2 = nn.Linear(128, 10) self.ln = nn.LayerNorm(9216) def forward(self, x, y): x = self.conv1(x) x = torch.nn.functional.relu(x) x = self.conv2(x) x = torch.nn.functional.relu(x) x = torch.nn.functional.max_pool2d(x, 2) x = self.dropout1(x) x = torch.flatten(x, 1) x = self.ln(x) x = self.fc1(x) x = torch.nn.functional.relu(x) x = self.dropout2(x) x = self.fc2(x) output = torch.nn.functional.log_softmax(x, dim=1) loss = torch.nn.functional.cross_entropy(output, y) return loss model = Mnist().cuda() model1 = Mnist().cuda() model1.load_state_dict(model.state_dict()) fsdp_model = FSDP( model, sharding_strategy=ShardingStrategy.SHARD_GRAD_OP, forward_prefetch=True, use_orig_params=True, auto_wrap_policy=ModuleWrapPolicy([nn.Linear, nn.Conv2d]), ) ddp_model = torch.nn.parallel.DistributedDataParallel( model1, ) fsdp_opt = torch.optim.SGD(fsdp_model.parameters(), lr=1e-4) ddp_opt = torch.optim.SGD(ddp_model.parameters(), lr=1e-4) seed = self.rank + 20231010 torch.manual_seed(seed) torch.cuda.manual_seed(seed) losses = [] grads = [] for i in range(5): x = torch.randn(8, 1, 28, 28, device="cuda").requires_grad_() y = torch.randint(low=0, high=9, size=(8,), device="cuda") for model, opt in ((fsdp_model, fsdp_opt), (ddp_model, ddp_opt)): seed = self.rank + i torch.manual_seed(seed) torch.cuda.manual_seed(seed) loss = model(x, y).sum() losses.append(loss) loss.backward() opt.step() grads.append(x.grad) opt.zero_grad() assert torch.allclose(losses[0], losses[1]) assert torch.allclose(grads[0], grads[1]) losses.clear() grads.clear() with torch.no_grad(): fsdp_model.eval() ddp_model.eval() for _ in range(5): x = torch.randn(8, 1, 28, 28, device="cuda").requires_grad_() y = torch.randint(low=0, high=9, size=(8,), device="cuda") fsdp_loss = fsdp_model(x, y) ddp_loss = ddp_model(x, y) assert torch.allclose(fsdp_loss, ddp_loss) fsdp_model.train() ddp_model.train() for i in range(5): x = torch.randn(8, 1, 28, 28, device="cuda").requires_grad_() y = torch.randint(low=0, high=9, size=(8,), device="cuda") for model, opt in ((fsdp_model, fsdp_opt), (ddp_model, ddp_opt)): seed = self.rank + i torch.manual_seed(seed) torch.cuda.manual_seed(seed) loss = model(x, y).sum() losses.append(loss) loss.backward() opt.step() grads.append(x.grad) opt.zero_grad() assert torch.allclose(losses[0], losses[1]) assert torch.allclose(grads[0], grads[1]) losses.clear() grads.clear() @skip_if_lt_x_gpu(2) @parametrize("use_second_layer", [True, False]) @parametrize("sharding_strategy", [ShardingStrategy.NO_SHARD, None]) def test_fsdp_module_no_compute_grad(self, use_second_layer, sharding_strategy): # When use_second_layer=True, b is involved in forward computation but does # not receive grad in backward. Otherwise, b is not involved in forward # computation. class MyModel(nn.Module): def __init__(self) -> None: super().__init__() self.a = nn.Linear(10, 10) self.b = nn.Linear(10, 10) def forward(self, x, y): out1 = self.a(x) if use_second_layer: out2 = self.b(y) return out1, out2 else: return out1 fsdp = FSDP( MyModel().cuda(), sharding_strategy=sharding_strategy, auto_wrap_policy=always_wrap_policy, ) x = torch.randn(10, 10, device="cuda") y = torch.randn(10, 10, device="cuda") for i in range(4): if use_second_layer: a, b = fsdp(x, y) else: a = fsdp(x, y) loss = a.sum() loss.backward() # self.a receives grad, self.b does not a_grad = fsdp.module.a._handle.flat_param.grad b_grad = fsdp.module.b._handle.flat_param.grad self.assertIsNotNone(a_grad) self.assertIsNone(b_grad) @skip_if_lt_x_gpu(2) def test_fsdp_not_all_outputs_used_in_loss(self): self.run_subtests( { "sharding_strategy": [ ShardingStrategy.FULL_SHARD, ShardingStrategy.SHARD_GRAD_OP, ShardingStrategy.NO_SHARD, ] }, self._test_fsdp_not_all_outputs_used_in_loss, ) def _test_fsdp_not_all_outputs_used_in_loss( self, sharding_strategy: ShardingStrategy ): class MyModule(nn.Module): def __init__(self) -> None: super().__init__() self.lin1 = nn.Linear(4, 4) self.lin2 = nn.Linear(4, 4) def forward(self, x): a = self.lin1(x) b = self.lin2(x) return (a, b) def _check_resharded(fsdp_module): handle = fsdp_module._handle if not handle: return param = handle.flat_param if handle.uses_sharded_strategy: full_param = param._full_param_padded self.assertEqual(full_param.storage().size(), 0) self.assertEqual(param.data_ptr(), param._local_shard.data_ptr()) def _check_equal(local, fsdp): with FSDP.summon_full_params(fsdp): for p1, p2 in zip(fsdp.parameters(), local.parameters()): torch.testing.assert_close(p1, p2) fsdp_ctor = functools.partial(FSDP, sharding_strategy=sharding_strategy) m = MyModule().cuda() m_local = deepcopy(m) local_m = m_local prev_params = [p.clone() for p in m_local.parameters()] m.lin1 = fsdp_ctor(m.lin1) m = fsdp_ctor(m) _check_equal(m_local, m) opt = torch.optim.SGD(m.parameters(), lr=1e-3) opt_local = torch.optim.SGD(local_m.parameters(), lr=1e-3) for i in range(6): t = torch.ones(4, device="cuda") a, b = m(t) local_a, local_b = local_m(t) if i < 2: # use both params in loss computation. Later, # b will go unused and we check grads are the # same as local training. loss = (a @ b).sum() loss_local = (local_a @ local_b).sum() else: loss = a.sum() loss_local = local_a.sum() loss.backward() loss_local.backward() _check_resharded(m) opt.step() opt_local.step() _check_equal(m_local, m) # Ensure at least some change from previous params, otherwise # above check would be vacuously true. self.assertTrue( any( not torch.equal(p1, p2) for p1, p2 in zip(prev_params, m_local.parameters()) ) ) prev_params = [p.clone() for p in local_m.parameters()] opt.zero_grad() opt_local.zero_grad() dist.barrier() @skip_if_lt_x_gpu(2) def test_fsdp_optim_overlap_no_use_orig_params_error(self): fsdp_overlap = FSDP( MyModel().cuda(), auto_wrap_policy=always_wrap_policy, use_orig_params=False, ) optim_cls = torch.optim.SGD optim_kwargs = {"lr": 0.03} _apply_optimizer_in_backward( optimizer_class=optim_cls, params=fsdp_overlap.parameters(), optimizer_kwargs=optim_kwargs, register_hook=False, ) inp = torch.randn(10, 10, device="cuda") with self.assertRaisesRegex( RuntimeError, "only supported with use_orig_params=True" ): fsdp_overlap(inp, inp) @skip_if_lt_x_gpu(2) def test_fsdp_optimizer_overlap(self): torch.manual_seed(0) for cpu_offload in [True, False]: offload = CPUOffload(offload_params=cpu_offload) model = MyModel().cuda() model_overlap = deepcopy(model) fsdp = FSDP( model.cuda(), auto_wrap_policy=always_wrap_policy, use_orig_params=True, cpu_offload=offload, ) fsdp_overlap = FSDP( model_overlap.cuda(), auto_wrap_policy=always_wrap_policy, use_orig_params=True, cpu_offload=offload, ) optim_cls = torch.optim.SGD optim_kwargs = {"lr": 0.03} _apply_optimizer_in_backward( optimizer_class=optim_cls, params=fsdp_overlap.parameters(), optimizer_kwargs=optim_kwargs, register_hook=False, ) for p in fsdp_overlap.parameters(): assert hasattr(p, "_in_backward_optimizers") optim = optim_cls(fsdp.parameters(), **optim_kwargs) # Verify params initially equal for p1, p2 in zip(fsdp.parameters(), fsdp_overlap.parameters()): self.assertEqual(p1, p2) with FSDP.summon_full_params(fsdp_overlap): fsdp_overlap_prev_params = [ (n, p.clone()) for n, p in fsdp_overlap.named_parameters() ] for i in range(6): inp = torch.randn(2, 2, device="cuda") with torch.no_grad(): inp_clone = inp.clone() fsdp(inp, inp).sum().backward() fsdp_overlap(inp_clone, inp_clone).sum().backward() optim.step() optim.zero_grad() # Overlapped optimizer FSDP module should have sharded_grad as None. for fsdp_unit in FSDP.fsdp_modules(fsdp_overlap): handle = fsdp_unit._handle if handle: handle_grad = handle.sharded_grad self.assertEqual( None, handle_grad, "Overlapped FSDP sharded_grad is not None!", ) # Note: FSDP without optimizer overlap won't set sharded_grad to None until the next # pre-forward since it needs to run FSDP specific logic that picks up that set_to_none=True # has been called (or that the gradients have been otherwise set to None) # Verify parameters are different than prev iteration with FSDP.summon_full_params(fsdp_overlap, with_grads=True): for (n, p), (n_prev, p_prev) in zip( fsdp_overlap.named_parameters(), fsdp_overlap_prev_params ): self.assertNotEqual( p, p_prev, f"{n_prev} Params at iter {i} same as previous iter!", ) # Verify overlap and non overlapped are the same with FSDP.summon_full_params(fsdp_overlap): with FSDP.summon_full_params(fsdp): for (n_overlap, p_overlap), (n, p) in zip( fsdp_overlap.named_parameters(), fsdp.named_parameters() ): self.assertEqual(n_overlap, n) self.assertEqual( p, p_overlap, f"Rank {self.rank}: Params not equal at iteration {i}: {n_overlap} - {p} vs {p_overlap}", ) self.assertEqual( None, p.grad, f"Expected param {n} grad to be None" ) self.assertEqual( None, p_overlap.grad, f"Expected param {n_overlap} grad to be None", ) fsdp_overlap_prev_params = [ (n, p.clone()) for n, p in fsdp_overlap.named_parameters() ] @skip_if_lt_x_gpu(2) def test_fsdp_cpu_training(self): """Tests FSDP training on CPU.""" gloo_pg = dist.new_group(backend="gloo") for ss in [ ShardingStrategy.NO_SHARD, ShardingStrategy.FULL_SHARD, ShardingStrategy.SHARD_GRAD_OP, ShardingStrategy.HYBRID_SHARD, ShardingStrategy._HYBRID_SHARD_ZERO2, ]: torch.manual_seed(42) model = MyModel() ref_model = DDP(deepcopy(model), process_group=gloo_pg) model = FSDP( model, auto_wrap_policy=always_wrap_policy, process_group=gloo_pg, device_id=torch.device("cpu"), ) ref_optim = torch.optim.Adam(ref_model.parameters(), lr=1e-2) optim = torch.optim.Adam(model.parameters(), lr=1e-2) torch.manual_seed(42 + self.rank) inp = torch.randn(2, 2) for _ in range(10): losses = [] for _model, _optim in ((ref_model, ref_optim), (model, optim)): loss = _model(inp, inp).sum() losses.append(loss) loss.backward() _optim.step() _optim.zero_grad() self.assertEqual(losses[0], losses[1]) @skip_if_lt_x_gpu(2) def test_fsdp_cpu_init_stays_on_cpu(self): # Move me to MT test once warning logging and backward collective issue # is resolved. """Tests that passing a CPU module to FSDP preserves that the wrapped module is on CPU after FSDP initialization, albeit after logging a warning, and that FSDP moves CPU input to GPU before the forward.""" torch.cuda.set_device(self.rank) regex = "passed-in `module` is on CPU" context = self.assertWarnsRegex( expected_warning=UserWarning, expected_regex=regex ) with context: nested_wrapped_module = NestedWrappedModule.init( self.process_group, FSDPInitMode.RECURSIVE, CUDAInitMode.CUDA_NEVER, ) fsdp_model = FSDP(nested_wrapped_module, self.process_group) devices = {p.device for p in fsdp_model.parameters()} self.assertEqual(1, len(devices)) self.assertEqual(torch.device("cpu"), devices.pop()) fsdp_model = fsdp_model.cuda() # Ensure fwd + backward can be performed after moving to CUDA. # CPU input also tests that input is correctly moved to appropriate # CUDA device. inp = fsdp_model.module.get_input(device=torch.device("cpu")) fsdp_model(*inp).sum().backward() @skip_if_lt_x_gpu(2) def test_cpu_init_with_sync_module_states(self): """ Tests that passing ``sync_module_states=True`` raises an error for a CPU module since the synchronization requires GPU communication, while additionally passing ``device_id`` does not raise an error, even when the model has CPU buffers. """ def init_nested_wrapped_module(): return NestedWrappedModule.init( self.process_group, FSDPInitMode.NO_FSDP, CUDAInitMode.CUDA_NEVER, ) with self.assertRaisesRegex( ValueError, "The module has CPU parameters or buffers when `sync_module_states=True`", ): FSDP( init_nested_wrapped_module(), self.process_group, sync_module_states=True, ) # Check that `device_id` with `sync_module_states=True` works nested_wrapped_module = init_nested_wrapped_module() nested_wrapped_module.buf = nn.Buffer( torch.ones((2, 2), device="cpu") * self.rank ) nested_wrapped_module.module[0].buf = nn.Buffer( torch.ones((3, 2), device="cpu") * self.rank ) nested_wrapped_module = FSDP( nested_wrapped_module, self.process_group, auto_wrap_policy=ModuleWrapPolicy({nn.Linear}), device_id=torch.cuda.current_device(), sync_module_states=True, ) # Each rank's buffers should be 0s since rank 0 is the source, and they # should be on GPU since we specified `device_id` self.assertEqual( nested_wrapped_module.buf.device, torch.device("cuda", torch.cuda.current_device()), ) self.assertEqual(nested_wrapped_module.buf, torch.zeros((2, 2))) self.assertEqual( nested_wrapped_module.module.module[0].buf.device, torch.device("cuda", torch.cuda.current_device()), ) self.assertEqual( nested_wrapped_module.module.module[0].buf, torch.zeros((3, 2)) ) class TestFSDPMiscMultiThread(FSDPTestMultiThread): @property def world_size(self): return 2 @property def process_group(self): return dist.distributed_c10d._get_default_group() @skip_if_lt_x_gpu(2) def test_fsdp_namedtuple(self): class MyModule(nn.Module): def __init__(self) -> None: super().__init__() self.lin = nn.Linear(100, 100) def forward(self, x): return x m = MyModule().cuda() m = FSDP(m) t = torch.ones(1, device="cuda", requires_grad=True) MyOutputType = namedtuple( "MyOutputType", ["a", "b", "c", "d"], defaults=(t, t, t, t) ) inp = MyOutputType() out = m(inp) # Ensure hooks are registered for x in out: self.assertNotEqual([], list(x._backward_hooks.values())) # TODO: we should check backward() and param is resharded # as well, but this is blocked by # https://github.com/pytorch/pytorch/issues/83107 and # https://github.com/pytorch/pytorch/issues/83129 @skip_if_lt_x_gpu(2) def test_device_id_auto_wrap(self): """Tests that ``auto_wrap_policy`` propagates ``device_id`` to all nested FSDP instances.""" self.run_subtests( {"use_callable": [False, True]}, self._test_device_id_auto_wrap, ) def _test_device_id_auto_wrap(self, use_callable: bool): module_classes = {TransformerEncoderLayer, TransformerDecoderLayer} if use_callable: auto_wrap_policy = functools.partial( transformer_auto_wrap_policy, transformer_layer_cls=module_classes, ) else: auto_wrap_policy = ModuleWrapPolicy(module_classes) fsdp_kwargs = { "auto_wrap_policy": auto_wrap_policy, "device_id": torch.cuda.current_device(), } fsdp_model = TransformerWithSharedParams.init( self.process_group, FSDPInitMode.RECURSIVE, CUDAInitMode.CUDA_BEFORE, fsdp_kwargs, ) for fsdp_module in FSDP.fsdp_modules(fsdp_model): self.assertEqual( fsdp_module.compute_device, torch.device("cuda", torch.cuda.current_device()), ) @skip_if_lt_x_gpu(2) def test_fsdp_device_id_cpu_offload(self): """ Tests FSDP when specifying both ``device_id`` and parameter CPU offloading. """ self.run_subtests( {"use_orig_params": [False, True]}, self._test_fsdp_device_id_cpu_offload, ) def _test_fsdp_device_id_cpu_offload(self, use_orig_params: bool): class MyModel(nn.Module): def __init__(self) -> None: super().__init__() self.seq = nn.Sequential( nn.Linear(10, 10), nn.Linear(10, 10), ) self.lin = nn.Linear(10, 10) def forward(self, x): return self.lin(self.seq(x)) model = MyModel() # Choose a wrapping policy such that there are (1) nested FSDP # instances and (2) the parent FSDP instance has managed parameters auto_wrap_policy = ModuleWrapPolicy({nn.Sequential}) fsdp_model = FSDP( model, auto_wrap_policy=auto_wrap_policy, cpu_offload=CPUOffload(offload_params=True), device_id=torch.cuda.current_device(), use_orig_params=use_orig_params, ) cpu_device = torch.device("cpu") for handle in traversal_utils._get_fsdp_handles(fsdp_model): self.assertEqual(handle.flat_param.device, cpu_device) @skip_if_lt_x_gpu(2) def test_module_device_mismatches_device_id(self): """Tests that specifying a ``device_id`` argument to FSDP for a GPU module that does not match the GPU device ID raises an error.""" # TODO: override FSDP MT Thread _run to set this instead of here for # every test. torch.cuda.set_device(self.rank) context = ( self.assertRaisesRegex(ValueError, f"cuda:{self.rank} vs cuda:0") if self.rank != 0 else nullcontext() ) with context: NestedWrappedModule.init( self.process_group, FSDPInitMode.RECURSIVE, # Move wrapped modules to CUDA before wrapping with FSDP cuda_init_mode=CUDAInitMode.CUDA_BEFORE, # Should raise error since rank 1 is given `device_id=0` when # the model is on cuda:1 fsdp_kwargs={"device_id": 0}, ) @skip_if_lt_x_gpu(2) def test_cpu_gpu_module(self): """Tests a CPU + GPU module supported if device_id is passed in, errors if device_id is not. """ torch.cuda.set_device(self.rank) class CPUGPUModule(nn.Module): def __init__(self) -> None: super().__init__() self.a = nn.Linear(1, 1).cuda() self.b = nn.Linear(1, 1) cpu_gpu = CPUGPUModule() fsdp = FSDP(cpu_gpu, device_id=torch.cuda.current_device()) for param in fsdp.parameters(): self.assertEqual(param.device, torch.device(torch.cuda.current_device())) # without device_id, we hit an error with self.assertRaisesRegex(RuntimeError, "please pass in device_id"): FSDP(CPUGPUModule()) @skip_if_lt_x_gpu(2) def test_fsdp_ignored_module_meta(self): torch.cuda.set_device(self.rank) class CPUGPUModule(nn.Module): def __init__(self) -> None: super().__init__() self.a = nn.Linear(1, 1) self.b = nn.Linear(1, 1) with torch.device("meta"): m = CPUGPUModule() m = FSDP(m, device_id=self.rank, ignored_modules=[m.a], use_orig_params=True) meta_device = torch.device("meta") self.assertEqual(meta_device, next(m.a.parameters()).device) # Test with param_init_fn with torch.device("meta"): m = CPUGPUModule() m = FSDP( m, device_id=torch.cuda.current_device(), ignored_modules=[m.a], use_orig_params=True, param_init_fn=lambda m: m.to_empty( device=torch.cuda.current_device(), recurse=False ), ) self.assertEqual(meta_device, next(m.a.parameters()).device) @skip_if_lt_x_gpu(2) def test_fsdp_device_id_no_move_ignored_params_and_bufs(self): class CPUGPUModule(nn.Module): def __init__(self) -> None: super().__init__() self.a = nn.Linear(1, 1) self.b = nn.Linear(1, 1) self.a.buf = torch.nn.Buffer(torch.ones(1)) m = CPUGPUModule() m = FSDP(m, device_id=self.rank, ignored_modules=[m.a], use_orig_params=True) ignored_params = m.a.parameters() ignored_bufs = m.a.buffers() for t in chain(ignored_params, ignored_bufs): self.assertEqual(torch.device("cpu"), t.device) @skip_if_lt_x_gpu(2) def test_multigpu_module(self): """ Module on multiple GPUs wrapped in FSDP should raise an error. """ class MultiGPUModule(nn.Module): def __init__(self, rank): super().__init__() self.rank = rank self.a = nn.Linear(1, 1).cuda(self.rank) self.b = nn.Linear(1, 1).cuda((self.rank + 1) % dist.get_world_size()) with self.assertRaisesRegex( RuntimeError, "FSDP only supports single device modules" ): FSDP(MultiGPUModule(self.rank)) @skip_if_lt_x_gpu(2) def test_no_params(self): """ Test that device_id and cpu init work if module has no params (they are effective noops, but ensure FSDP does not assume module has parameters during init) """ # TODO: override FSDP MT Thread _run to set this instead of here for # every test. torch.cuda.set_device(self.rank) # Test CPU no_params = nn.ReLU() module = FSDP(no_params) # Test CUDA no_params = nn.ReLU().cuda() module = FSDP(no_params) # Test CPU + device_id no_params = nn.ReLU() module = FSDP(no_params, device_id=torch.cuda.current_device()) # For modules with no params, wrong device_id will raise error about # inconsistency between compute_device and device_id, since compute_device # is computed as torch.cuda.current_device when there are no params. no_params = nn.ReLU().cuda() context = ( ( self.assertRaisesRegex( ValueError, f"Inconsistent.*cuda:{self.rank} vs cuda:0" ) ) if self.rank != 0 else nullcontext() ) with context: FSDP(no_params, device_id=0) @skip_if_lt_x_gpu(2) def test_fsdp_same_model_across_ranks(self): """ FSDP broadcasts model from rank 0 to ensure it starts off with the same values. """ class MyModel(nn.Module): def __init__(self, rank): super().__init__() # Seed via rank to make model different across ranks torch.manual_seed(rank) torch.cuda.manual_seed(rank) self.lin = nn.Linear(10, 10, bias=False) self.buffer = nn.Buffer(torch.ones(1) * rank) m = MyModel(self.rank).cuda() _assert_module_states( m, process_group=self.process_group, assert_fn=self.assertNotEqual ) # Passing sync_module_states into FSDP makes model the same during init. fsdp = FSDP(m, sync_module_states=True) with fsdp.summon_full_params(fsdp): _assert_module_states( fsdp, process_group=self.process_group, assert_fn=self.assertEqual ) # sync_module_states also works with CPU module with device_id passed in m = MyModel(self.rank) _assert_module_states( m, process_group=self.process_group, assert_fn=self.assertNotEqual ) # Passing sync_module_states into FSDP makes model the same during init. fsdp = FSDP(m, device_id=torch.cuda.current_device(), sync_module_states=True) with fsdp.summon_full_params(fsdp): _assert_module_states( fsdp, process_group=self.process_group, assert_fn=self.assertEqual ) @skip_if_lt_x_gpu(2) def test_homogeneous_attributes(self): """ Tests that passing heterogeneous values for attributes designated as homogeneous raises an error. """ # Manually construct this list but verify against the global list of # homogeneous attribute names all_attr_name_and_values = [ ("_use_orig_params", False, True), ("limit_all_gathers", False, True), ("_use_full_prec_in_eval", False, True), ] self.assertEqual( [ attr_name_and_values[0] for attr_name_and_values in all_attr_name_and_values ], HOMOGENEOUS_ATTR_NAMES, ) self.run_subtests( {"attr_name_and_values": all_attr_name_and_values}, self._test_homogeneous_attributes, ) def _test_homogeneous_attributes(self, attr_name_and_values: Tuple[str, Any, Any]): model = NestedWrappedModule.init( self.process_group, FSDPInitMode.NO_FSDP, CUDAInitMode.CUDA_BEFORE, {}, ) attr_name = attr_name_and_values[0] if "_use_full_prec_in_eval" == attr_name: model.module[1] = FSDP(model.module[1]) os.environ["FSDP_USE_FULL_PREC_IN_EVAL"] = "1" fsdp_model = FSDP(model) else: fsdp_kwargs_inner = {attr_name.lstrip("_"): attr_name_and_values[1]} fsdp_kwargs_outer = {attr_name.lstrip("_"): attr_name_and_values[2]} model.module[1] = FSDP(model.module[1], **fsdp_kwargs_inner) fsdp_model = FSDP(model, **fsdp_kwargs_outer) # Run a forward to trigger lazy initialization and the error with self.assertRaisesRegex( ValueError, f"Expects one homogeneous value for {attr_name}" ): inp = fsdp_model.module.get_input(torch.device("cuda")) fsdp_model(*inp) @skip_if_lt_x_gpu(2) def test_fsdp_unsupported_module_cls(self): regex = r"FSDP will not all-gather parameters for containers that do not implement forward" model = nn.ModuleList([MLP(8, torch.device("cpu")) for _ in range(3)]) with self.assertWarnsRegex(UserWarning, regex): FSDP(model, device_id="cuda") model = nn.ModuleDict( {"1": MLP(8, torch.device("cpu")), "2": MLP(8, torch.device("cpu"))} ) with self.assertWarnsRegex(UserWarning, regex): FSDP(model) class TestFSDPMiscWorldSize1(FSDPTestMultiThread): @property def world_size(self) -> int: return 1 @skip_if_lt_x_gpu(1) def test_world_size_1_sharding_strategy_warning(self): """ Tests that FSDP issues a warning when it switches to using ``NO_SHARD`` when the world size is 1. """ warning_prefix = "FSDP is switching to use `NO_SHARD` instead of" # If the user already passes `NO_SHARD`, then there should not be a # warning with warnings.catch_warnings(record=True) as w: warnings.simplefilter("always") # trigger all warnings FSDP(nn.Linear(3, 3).cuda(), sharding_strategy=ShardingStrategy.NO_SHARD) for warning in w: self.assertTrue( warning.category != UserWarning or not str(warning.message).startswith(warning_prefix) ) # Check that a warning is issued warning_suffix = " since the world size is 1." # - Pass `FULL_SHARD` or `None` expected_regex_full_shard = ( warning_prefix + " " + str(ShardingStrategy.FULL_SHARD) + warning_suffix ) with self.assertWarnsRegex(UserWarning, expected_regex_full_shard): FSDP(nn.Linear(3, 3).cuda(), sharding_strategy=ShardingStrategy.FULL_SHARD) with self.assertWarnsRegex(UserWarning, expected_regex_full_shard): FSDP(nn.Linear(3, 3).cuda()) # - Pass `SHARD_GRAD_OP` expected_regex_shard_grad_op = ( warning_prefix + " " + str(ShardingStrategy.SHARD_GRAD_OP) + warning_suffix ) with self.assertWarnsRegex(UserWarning, expected_regex_shard_grad_op): FSDP( nn.Linear(3, 3).cuda(), sharding_strategy=ShardingStrategy.SHARD_GRAD_OP ) @skip_if_lt_x_gpu(1) def test_training_device_mismatch_errors(self): """ Tests that, when training starts, if FSDP parameters are not on the expected device, then an informative error is raised. This applies for both no parameter CPU offloading and parameter CPU offloading. """ # Incorrectly not moving from CPU -> GPU model = torch.nn.Linear(10, 10) fsdp_model = FSDP(model) inp = torch.randn((2, 10)) with self.assertRaisesRegex( RuntimeError, "An FSDP-managed module unexpectedly has parameters on cpu. Make " "sure to move the module to cuda:0 before training.", ): fsdp_model(inp) # Incorrectly moving from CPU -> GPU model = torch.nn.Linear(10, 10) fsdp_model = FSDP(model, cpu_offload=CPUOffload(offload_params=True)) fsdp_model.to(torch.device("cuda")) inp = torch.randn((2, 10)) with self.assertRaisesRegex( RuntimeError, "An FSDP-managed module with parameter CPU offloading enabled has " "parameters on cuda:0. Make sure to not move the module from CPU " "when offloading parameters.", ): fsdp_model(inp) @skip_if_lt_x_gpu(2) def test_unsafe_setattr(self): """ Tests that the environment variable for using unsafe setattr gates as expected. """ self.run_subtests( {"use_orig_params": [False, True]}, self._test_unsafe_setattr, ) def _test_unsafe_setattr(self, use_orig_params: bool): called_setattr_override = False class SetattrLinear(nn.Module): def __init__(self, in_dim: int, out_dim: int, device: torch.device) -> None: super().__init__() self.weight = nn.Parameter( torch.randn((in_dim, out_dim), device=device) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return x @ self.weight def __setattr__(self, name: str, value: Any) -> None: nonlocal called_setattr_override called_setattr_override = True return super().__setattr__(name, value) # Construct FSDP module without changing any environment variables and # run forward, which triggers both unsharded and sharded view setting module = SetattrLinear(5, 5, torch.device("cuda")) fsdp_module = FSDP(module, use_orig_params=use_orig_params) inp = torch.randn((8, 5), device=torch.device("cuda")) called_setattr_override = False fsdp_module(inp) self.assertTrue(called_setattr_override) # Repeat with unsafe setattr explicitly enabled os.environ[_FSDP_USE_UNSAFE_SETATTR] = "1" module = SetattrLinear(5, 5, torch.device("cuda")) fsdp_module = FSDP(module, use_orig_params=use_orig_params) called_setattr_override = False fsdp_module(inp) self.assertFalse(called_setattr_override) # Repeat with unsafe setattr explicitly disabled os.environ[_FSDP_USE_UNSAFE_SETATTR] = "0" module = SetattrLinear(5, 5, torch.device("cuda")) fsdp_module = FSDP(module, use_orig_params=use_orig_params) called_setattr_override = False fsdp_module(inp) self.assertTrue(called_setattr_override) instantiate_parametrized_tests(TestFSDPMiscMultiThread) instantiate_parametrized_tests(TestFSDPMiscMultiProcess) if __name__ == "__main__": run_tests()