# coding=utf-8 # Copyright (c) 2021-22, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # NOTE(mkozuki): This file is based on megatron-lm/mpu/random.py with some differences: # - Not using "viewless" tensor: # - _kernel_make_viewless_tensor # - MakeViewlessTensor # - make_viewless_tensor # - assert_viewless_tensor # - safely_set_viewless_tensor_data # Parts of the code here are adapted from PyTorch # repo: https://github.com/pytorch/pytorch import contextlib import torch from torch import _C from torch.cuda import _lazy_call, device as device_ctx_manager from torch.utils.checkpoint import detach_variable from apex.transformer.parallel_state import get_tensor_model_parallel_rank from apex.transformer.tensor_parallel.memory import allocate_mem_buff from apex.transformer.utils import split_tensor_into_1d_equal_chunks from apex.transformer.utils import gather_split_1d_tensor # Default name for the model parallel rng tracker. _MODEL_PARALLEL_RNG_TRACKER_NAME = "model-parallel-rng" # TODO(mkozuki): Remove `_CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER` as megatron-lm doesn't seem to use. # Whether apply model parallelism to checkpointed hidden states. _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER = None # TODO(mkozuki): Remove `init_checkpointed_activations_memory_buffer` as megatron-lm doesn't seem to use. def init_checkpointed_activations_memory_buffer( micro_batch_size, max_position_embeddings, hidden_size, num_layers, tensor_model_parallel_size, checkpoint_num_layers, fp16, ): """Initializ the memory buffer for the checkpointed activations.""" per_layer = ( micro_batch_size * max_position_embeddings * hidden_size // tensor_model_parallel_size ) assert ( num_layers % checkpoint_num_layers == 0 ), "number of layers is not divisible by checkpoint-num-layers" num_checkpointer_layers = num_layers // checkpoint_num_layers numel = per_layer * num_checkpointer_layers dtype = torch.half if not fp16: dtype = torch.float global _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER assert ( _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER is None ), "checkpointed activations memory buffer is already allocated." _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER = allocate_mem_buff( "checkpointed activations", numel, dtype, track_usage=False ) # TODO(mkozuki): Remove `reset_checkpointed_activations_memory_buffer` as megatron-lm doesn't seem to use. def reset_checkpointed_activations_memory_buffer(): """Reset the memory used for checkpointing.""" if _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER is not None: _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER.reset() def _set_cuda_rng_state(new_state, device=-1): """Sets the random number generator state of the current GPU. Arguments: new_state (torch.ByteTensor): The desired state This function is adapted from PyTorch repo (torch.cuda.set_rng_state) with a single change: the input state is not cloned. Cloning caused major performance issues for +4 GPU cases. """ if hasattr(_C, "_cuda_setRNGState") and callable(_C._cuda_setRNGState): # older PyTorch def cb(): with device_ctx_manager(device): _C._cuda_setRNGState(new_state) else: # newer PyTorch if device == -1: device = torch.device("cuda") elif isinstance(device, str): device = torch.device(device) elif isinstance(device, int): device = torch.device("cuda", device) def cb(): idx = device.index if idx is None: idx = torch.cuda.current_device() default_generator = torch.cuda.default_generators[idx] default_generator.set_state(new_state) _lazy_call(cb) class CudaRNGStatesTracker: """Tracker for the cuda RNG states. Using the `add` method, a cuda rng state is initialized based on the input `seed` and is assigned to `name`. Later, by forking the rng state, we can perform operations and return to our starting cuda state. """ def __init__(self): # Map from a string name to the cuda rng state. self.states_ = {} # Seeds are just for book keeping and ensure no seed is set twice. self.seeds_ = set() def reset(self): """Set to the initial state (no tracker).""" self.states_ = {} self.seeds_ = set() def get_states(self): """Get rng states. Copy the dictionary so we have direct pointers to the states, not just a pointer to the dictionary.""" states = {} for name in self.states_: states[name] = self.states_[name] return states def set_states(self, states): """Set the rng states. For efficiency purposes, we do not check the size of seed for compatibility.""" self.states_ = states def add(self, name, seed): """Track the rng state.""" # Check seed is not already used. if seed in self.seeds_: raise Exception("seed {} already exists".format(seed)) self.seeds_.add(seed) # Check that state is not already defined. if name in self.states_: raise Exception("cuda rng state {} already exists".format(name)) # Get the current rng state. orig_rng_state = torch.cuda.get_rng_state() # Set the new state and store it. torch.cuda.manual_seed(seed) self.states_[name] = torch.cuda.get_rng_state() # Reset rng state to what it was. _set_cuda_rng_state(orig_rng_state) @contextlib.contextmanager def fork(self, name=_MODEL_PARALLEL_RNG_TRACKER_NAME): """Fork the cuda rng state, perform operations, and exit with the original state.""" # Check if we have added the state if name not in self.states_: raise Exception("cuda rng state {} is not added".format(name)) # Store current rng state. orig_cuda_rng_state = torch.cuda.get_rng_state() # Set rng state to the desired one _set_cuda_rng_state(self.states_[name]) # Do the stuff we wanted to do. try: yield finally: # Update the current rng state for later use. self.states_[name] = torch.cuda.get_rng_state() # And set the state to the original state we started with. _set_cuda_rng_state(orig_cuda_rng_state) # RNG tracker object. _CUDA_RNG_STATE_TRACKER = CudaRNGStatesTracker() def get_cuda_rng_tracker(): """Get cuda rng tracker.""" return _CUDA_RNG_STATE_TRACKER def model_parallel_cuda_manual_seed(seed): """Initialize model parallel cuda seed. This function should be called after the model parallel is initialized. Also, no torch.cuda.manual_seed should be called after this function. Basically, this is replacement for that function. Two set of RNG states are tracked: default state: This is for data parallelism and is the same among a set of model parallel GPUs but different across different model paralle groups. This is used for example for dropout in the non-tensor-model-parallel regions. tensor-model-parallel state: This state is different among a set of model parallel GPUs, but the same across data parallel groups. This is used for example for dropout in model parallel regions. """ # 2718 is just for fun and any POSITIVE value will work. offset = seed + 2718 tensor_model_parallel_seed = offset + get_tensor_model_parallel_rank() # Data parallel gets the original seed. data_parallel_seed = seed _CUDA_RNG_STATE_TRACKER.reset() # Set the default state. torch.cuda.manual_seed(data_parallel_seed) # and model parallel state. _CUDA_RNG_STATE_TRACKER.add( _MODEL_PARALLEL_RNG_TRACKER_NAME, tensor_model_parallel_seed ) # TODO (mkozuki): Move the below gradient checkpoint related features to another (new) file. class CheckpointFunction(torch.autograd.Function): """This function is adapted from torch.utils.checkpoint with two main changes: 1) torch.cuda.set_rng_state is replaced with `_set_cuda_rng_state` 2) the states in the model parallel tracker are also properly tracked/set/reset. """ @staticmethod def forward(ctx, run_function, distribute_saved_activations, *args): ctx.run_function = run_function ctx.distribute_saved_activations = distribute_saved_activations # Copy the rng states. ctx.fwd_cpu_rng_state = torch.get_rng_state() ctx.fwd_cuda_rng_state = torch.cuda.get_rng_state() ctx.fwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states() with torch.no_grad(): outputs = run_function(*args) # Divide hidden states across model parallel group and only keep # the chunk corresponding to the current rank. if ctx.distribute_saved_activations: ctx.input_0_shape = args[0].shape # Store everything. ctx.save_for_backward(*args) return outputs @staticmethod def backward(ctx, *args): if not torch.autograd._is_checkpoint_valid(): raise RuntimeError( "Checkpointing is not compatible with .grad(), " "please use .backward() if possible" ) inputs = ctx.saved_tensors # Store the current states. bwd_cpu_rng_state = torch.get_rng_state() bwd_cuda_rng_state = torch.cuda.get_rng_state() bwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states() # Set the states to what it used to be before the forward pass. torch.set_rng_state(ctx.fwd_cpu_rng_state) _set_cuda_rng_state(ctx.fwd_cuda_rng_state) get_cuda_rng_tracker().set_states(ctx.fwd_cuda_rng_state_tracker) # Compute the forward pass. detached_inputs = detach_variable(inputs) with torch.enable_grad(): outputs = ctx.run_function(*detached_inputs) # Set the states back to what it was at the start of this function. torch.set_rng_state(bwd_cpu_rng_state) _set_cuda_rng_state(bwd_cuda_rng_state) get_cuda_rng_tracker().set_states(bwd_cuda_rng_state_tracker) if isinstance(outputs, torch.Tensor): outputs = (outputs,) torch.autograd.backward(outputs, args) grads = tuple( inp.grad if isinstance(inp, torch.Tensor) else inp for inp in detached_inputs ) return (None, None) + grads # NOTE(mkozuki): It doesn't look like `distribute_saved_activations` is used in apex.transformer # but I added this change to reduce the superficial difference from Megatron-LM. def checkpoint(function, distribute_saved_activations, *args): """Checkpoint a model or part of the model. This has been directly copied from torch.utils.checkpoint.""" return CheckpointFunction.apply(function, distribute_saved_activations, *args)