Merge branch 'main' of https://github.com/NVIDIA/Megatron-LM

megatron/model/distributed.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+from abc import ABC
+from abc import abstractmethod
+import math
+
+import torch
+from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
+
+from megatron import get_args
+from megatron.core import mpu
+from .module import MegatronModule
+
+
+class MemoryBuffer:
+
+    def __init__(self, numel, numel_padded, dtype):
+        self.numel = numel
+        self.numel_padded = numel_padded
+        self.dtype = dtype
+        self.data = torch.zeros(self.numel_padded,
+                                dtype=self.dtype,
+                                device=torch.cuda.current_device(),
+                                requires_grad=False)
+
+    def zero(self):
+        """Reset the buffer to zero."""
+        self.data.zero_()
+
+
+    def get(self, shape, start_index):
+        """Return a tensor with the input `shape` as a view into the
+        1-D data starting at `start_index`."""
+        end_index = start_index + shape.numel()
+        assert end_index <= self.numel, \
+            'requested tensor is out of the buffer range.'
+        buffer_tensor = self.data[start_index:end_index]
+        buffer_tensor = buffer_tensor.view(shape)
+        return buffer_tensor
+
+
+
+class DistributedDataParallelBase(MegatronModule, ABC):
+    """Abstract class for DDP."""
+
+    def __init__(self, module):
+        super(DistributedDataParallelBase, self).__init__()
+        # Keep a pointer to the model.
+        self.module = module
+
+
+    @abstractmethod
+    def allreduce_gradients(self):
+        pass
+
+
+    def forward(self, *inputs, **kwargs):
+        return self.module(*inputs, **kwargs)
+
+
+    def state_dict(self, prefix='', keep_vars=False):
+        return self.module.state_dict(prefix=prefix, keep_vars=keep_vars)
+
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        return self.module.state_dict_for_save_checkpoint(prefix=prefix,
+                                                          keep_vars=keep_vars)
+
+
+    def load_state_dict(self, state_dict, strict=True):
+        self.module.load_state_dict(state_dict, strict=strict)
+
+
+
+class DistributedDataParallel(DistributedDataParallelBase):
+    """DDP with contiguous buffers options to storre and accumulate gradients.
+    This class:
+        - has the potential to reduce memory fragmentation.
+        - provides the option to do the gradient accumulation
+          in a type other than the params type (for example fp32)
+
+    Arguments:
+        module: input model.
+        accumulate_allreduce_grads_in_fp32: if true do the gradient accumulation
+            and the gradient all-reduce all in in float32. If this option is
+            true, we require `use_contiguous_buffers` to be true too.
+        use_contiguous_buffers: if true, use a contiguous buffer to store the
+            gradients.
+    """
+
+    def __init__(self, module,
+                 accumulate_allreduce_grads_in_fp32,
+                 use_contiguous_buffers):
+
+        super(DistributedDataParallel, self).__init__(module)
+
+        self.accumulate_allreduce_grads_in_fp32 \
+            = accumulate_allreduce_grads_in_fp32
+        self.use_contiguous_buffers = use_contiguous_buffers
+        # If we are using fp32-accumulate-allreduce explicitly
+        # this means we need main grads in a continous buffer.
+        if self.accumulate_allreduce_grads_in_fp32:
+            assert self.use_contiguous_buffers
+
+        # ===================================
+        # Rest of this part applies only to
+        # the case we use continuous buffers.
+        # ===================================
+        self._grad_buffers = None
+        self._grad_buffer_param_index_map = None
+        if self.use_contiguous_buffers:
+            self._grad_buffers = {}
+            self._grad_buffer_param_index_map = {}
+            data_parallel_world_size = mpu.get_data_parallel_world_size()
+
+            # Simple function to define buffer type.
+            def _get_buffer_type(param):
+                return torch.float if \
+                    self.accumulate_allreduce_grads_in_fp32 else param.dtype
+
+            # First calculate total number of elements per type.
+            type_num_elements = {}
+            for param in self.module.parameters():
+                if param.requires_grad:
+                    dtype = _get_buffer_type(param)
+                    type_num_elements[dtype] = type_num_elements.get(dtype, 0) \
+                                               + param.data.nelement()
+
+            # Allocate the buffer.
+            for dtype, num_elements in type_num_elements.items():
+
+                # If using distributed optimizer, pad memory buffer to be
+                # multiple of data_parallel_world_size. (This padding is done
+                # due to a constraint with the reduce_scatter op, which requires
+                # all tensors have equal size. See: optimizer.py.)
+                num_elements_padded = data_parallel_world_size * \
+                    int(math.ceil(num_elements / data_parallel_world_size))
+
+                # Allocate grad buffer.
+                self._grad_buffers[dtype] = MemoryBuffer(num_elements,
+                                                         num_elements_padded,
+                                                         dtype)
+
+            # Assume the back prop order is reverse the params order,
+            # store the start index for the gradients.
+            for param in self.module.parameters():
+                if param.requires_grad:
+                    dtype = _get_buffer_type(param)
+                    type_num_elements[dtype] -= param.data.nelement()
+                    param.main_grad = self._grad_buffers[dtype].get(
+                        param.data.shape, type_num_elements[dtype])
+                    if dtype not in self._grad_buffer_param_index_map:
+                        self._grad_buffer_param_index_map[dtype] = {}
+                    self._grad_buffer_param_index_map[dtype][param] = (
+                        type_num_elements[dtype],
+                        type_num_elements[dtype] + param.data.nelement(),
+                    )
+
+            # Backward hook.
+            # Accumalation function for the gradients. We need
+            # to store them so they don't go out of scope.
+            self.grad_accs = []
+            # Loop over all the parameters in the model.
+            for param in self.module.parameters():
+                if param.requires_grad:
+                    # Expand so we get access to grad_fn.
+                    param_tmp = param.expand_as(param)
+                    # Get the gradient accumulator functtion.
+                    grad_acc = param_tmp.grad_fn.next_functions[0][0]
+                    grad_acc.register_hook(self._make_param_hook(param))
+                    self.grad_accs.append(grad_acc)
+
+
+    def _make_param_hook(self, param):
+        """Create the all-reduce hook for backprop."""
+        # Hook used for back-prop.
+        def param_hook(*unused):
+            # Add the gradient to the buffer.
+            if param.grad is not None:
+                # The gradient function of linear layers is fused with GEMMs
+                param.main_grad.add_(param.grad.data)
+                # Now we can deallocate grad memory.
+                param.grad = None
+        return param_hook
+
+
+    def zero_grad_buffer(self):
+        """Set the grad buffer data to zero. Needs to be called at the
+        begining of each iteration."""
+        assert self._grad_buffers is not None, 'buffers are not initialized.'
+        for _, buffer_ in self._grad_buffers.items():
+            buffer_.zero()
+
+
+    def broadcast_params(self):
+        for param in self.module.parameters():
+            torch.distributed.broadcast(param.data,
+                                        src=mpu.get_data_parallel_src_rank(),
+                                        group=mpu.get_data_parallel_group())
+
+
+    def allreduce_gradients(self):
+        """Reduce gradients across data parallel ranks."""
+        # If we have buffers, simply reduce the data in the buffer.
+        if self._grad_buffers is not None:
+            for _, buffer_ in self._grad_buffers.items():
+                buffer_.data /= mpu.get_data_parallel_world_size()
+                torch.distributed.all_reduce(
+                    buffer_.data, group=mpu.get_data_parallel_group())
+        else:
+            # Otherwise, bucketize and all-reduce
+            buckets = {}
+            # Pack the buckets.
+            for param in self.module.parameters():
+                if param.requires_grad and param.grad is not None:
+                    tp = param.data.type()
+                    if tp not in buckets:
+                        buckets[tp] = []
+                    buckets[tp].append(param)
+                    param.main_grad = param.grad
+
+            # For each bucket, all-reduce and copy all-reduced grads.
+            for tp in buckets:
+                bucket = buckets[tp]
+                grads = [param.grad.data for param in bucket]
+                coalesced = _flatten_dense_tensors(grads)
+                coalesced /= mpu.get_data_parallel_world_size()
+                torch.distributed.all_reduce(
+                    coalesced, group=mpu.get_data_parallel_group())
+                for buf, synced in zip(grads, _unflatten_dense_tensors(
+                        coalesced, grads)):
+                    buf.copy_(synced)

megatron/model/enums.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+import enum
+
+class LayerType(enum.Enum):
+    encoder = 1
+    decoder = 2
+ 
+class AttnType(enum.Enum):
+    self_attn = 1
+    cross_attn = 2
+
+class AttnMaskType(enum.Enum):
+    padding = 1
+    causal = 2
+
+# For backward compatibility with old model checkpoints
+from megatron.core.enums import ModelType

megatron/model/fused_bias_gelu.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+import torch
+
+
+###### BIAS GELU FUSION/ NO AUTOGRAD ################
+# 1/sqrt(2*pi)-> 0.3989423
+# 1/sqrt(2)   -> 0.70710678
+# sqrt(2/pi)  -> 0.79788456
+# this function is tanh approximation of gelu
+# actual gelu is:
+# x * 0.5 * (1.0 + torch.erf(x * 0.70710678))
+
+@torch.jit.script
+def bias_gelu(bias, y):
+    x = bias + y
+    return  x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))
+
+# gradient of tanh approximation of gelu
+# gradient of actual gelu is:
+# 0.5 * (1. + torch.erf(x * 0.70710678)) + 0.3989423 * x * torch.exp(-0.5 * x * x)
+@torch.jit.script
+def bias_gelu_back(g, bias, y):
+    x = bias + y
+    tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
+    # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
+    ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
+    return ff*g
+
+class GeLUFunction(torch.autograd.Function):
+    @staticmethod
+    # bias is an optional argument
+    def forward(ctx, input, bias):
+        ctx.save_for_backward(input, bias)
+        return bias_gelu(bias, input)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, bias = ctx.saved_tensors
+        tmp = bias_gelu_back(grad_output, bias, input)
+        return tmp, tmp
+
+bias_gelu_impl = GeLUFunction.apply

megatron/model/fused_layer_norm.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""This code is copied fron NVIDIA apex:
+      https://github.com/NVIDIA/apex
+   with some changes. """
+
+import numbers
+import torch
+from torch.nn.parameter import Parameter
+from torch.nn import init
+import importlib
+
+from megatron.core.utils import make_viewless_tensor
+
+try:
+    from apex.contrib.layer_norm.layer_norm import FastLayerNormFN
+    HAVE_PERSIST_LAYER_NORM = True
+except:
+    HAVE_PERSIST_LAYER_NORM = False
+
+from apex.normalization.fused_layer_norm import FusedLayerNormAffineFunction
+
+
+global fused_layer_norm_cuda
+fused_layer_norm_cuda = None
+
+
+class MixedFusedLayerNorm(torch.nn.Module):
+
+  def __init__(self, normalized_shape, eps=1e-5,
+               no_persist_layer_norm=True,
+               sequence_parallel=False,
+               apply_layernorm_1p=False):
+        super(MixedFusedLayerNorm, self).__init__()
+
+        self.apply_layernorm_1p = apply_layernorm_1p
+
+        global fused_layer_norm_cuda
+        fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda")
+
+        # List of hiddens sizes supported in the persistent layer norm kernel
+        # If the hidden size is not supported, fall back to the non-persistent
+        # kernel.
+        persist_ln_hidden_sizes = [1024, 1536, 2048, 2304, 3072, 3840, 4096,
+            5120, 6144, 8192, 10240, 12288, 12800, 15360, 16384, 18432, 20480,
+            24576, 25600, 30720, 32768, 40960, 49152, 65536]
+        if normalized_shape not in persist_ln_hidden_sizes or \
+                not HAVE_PERSIST_LAYER_NORM:
+            no_persist_layer_norm = True
+
+        if isinstance(normalized_shape, numbers.Integral):
+            normalized_shape = (normalized_shape,)
+        self.normalized_shape = torch.Size(normalized_shape)
+        self.eps = eps
+        self.weight = Parameter(torch.Tensor(*normalized_shape))
+        self.bias = Parameter(torch.Tensor(*normalized_shape))
+        self.reset_parameters()
+        self.no_persist_layer_norm = no_persist_layer_norm
+        self.sequence_parallel = sequence_parallel
+
+        # set sequence parallelism flag on weight and bias parameters
+        setattr(self.weight, 'sequence_parallel', self.sequence_parallel)
+        setattr(self.bias, 'sequence_parallel', self.sequence_parallel)
+
+
+  def reset_parameters(self):
+
+    if self.apply_layernorm_1p:
+        init.zeros_(self.weight)
+        init.zeros_(self.bias)
+    else:
+        init.ones_(self.weight)
+        init.zeros_(self.bias)
+
+  def forward(self, input):
+
+    weight = self.weight + 1 if self.apply_layernorm_1p else self.weight
+
+    if self.no_persist_layer_norm:
+        return FusedLayerNormAffineFunction.apply(input, weight, self.bias, self.normalized_shape, self.eps)
+    else:
+        output = FastLayerNormFN.apply(input, weight, self.bias, self.eps)
+
+        # Apex's fast layer norm function outputs a 'view' tensor (i.e., has
+        # a populated '_base' field). This will result in schedule.py's
+        # deallocate_output_tensor() throwing an error, so a viewless tensor is
+        # created to prevent this.
+        output = make_viewless_tensor(inp = output,
+                                      requires_grad = input.requires_grad,
+                                      keep_graph = True)
+
+        return output

megatron/model/fused_softmax.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+
+import torch
+import torch.nn as nn
+from megatron.model.enums import AttnMaskType
+
+
+class ScaledUpperTriangMaskedSoftmax(torch.autograd.Function):
+    """
+    Fused operation which performs following three operations in sequence
+    1. Scale the tensor.
+    2. Apply upper triangular mask (typically used in gpt models).
+    3. Perform softmax.
+    """
+
+    @staticmethod
+    def forward(ctx, inputs, scale):
+        import scaled_upper_triang_masked_softmax_cuda
+
+        scale_t = torch.tensor([scale])
+        softmax_results = scaled_upper_triang_masked_softmax_cuda.forward(
+            inputs, scale_t[0]
+        )
+
+        ctx.save_for_backward(softmax_results, scale_t)
+        return softmax_results
+
+    @staticmethod
+    def backward(ctx, output_grads):
+        import scaled_upper_triang_masked_softmax_cuda
+
+        softmax_results, scale_t = ctx.saved_tensors
+        input_grads = scaled_upper_triang_masked_softmax_cuda.backward(
+            output_grads, softmax_results, scale_t[0]
+        )
+
+        return input_grads, None
+
+
+class ScaledMaskedSoftmax(torch.autograd.Function):
+    """
+    Fused operation which performs following three operations in sequence
+    1. Scale the tensor.
+    2. Apply the mask.
+    3. Perform softmax.
+    """
+
+    @staticmethod
+    def forward(ctx, inputs, mask, scale):
+        import scaled_masked_softmax_cuda
+
+        scale_t = torch.tensor([scale])
+
+        softmax_results = scaled_masked_softmax_cuda.forward(inputs, mask, scale_t[0])
+        ctx.save_for_backward(softmax_results, scale_t)
+        return softmax_results
+
+    @staticmethod
+    def backward(ctx, output_grads):
+        import scaled_masked_softmax_cuda
+
+        softmax_results, scale_t = ctx.saved_tensors
+
+        input_grads = scaled_masked_softmax_cuda.backward(
+            output_grads, softmax_results, scale_t[0]
+        )
+        return input_grads, None, None
+
+
+class ScaledSoftmax(torch.autograd.Function):
+    """
+    Fused operation which performs following two operations in sequence
+    1. Scale the tensor.
+    2. Perform softmax.
+    """
+
+    @staticmethod
+    def forward(ctx, inputs, scale):
+        import scaled_softmax_cuda
+
+        scale_t = torch.tensor([scale])
+
+        softmax_results = scaled_softmax_cuda.forward(
+            inputs, scale_t[0]
+        )
+        ctx.save_for_backward(softmax_results, scale_t)
+        return softmax_results
+
+    @staticmethod
+    def backward(ctx, output_grads):
+        import scaled_softmax_cuda
+
+        softmax_results, scale_t = ctx.saved_tensors
+
+        input_grads = scaled_softmax_cuda.backward(
+            output_grads, softmax_results, scale_t[0]
+        )
+        return input_grads, None, None
+
+
+class FusedScaleMaskSoftmax(nn.Module):
+    """
+    fused operation: scaling + mask + softmax
+
+    Arguments:
+        input_in_fp16: flag to indicate if input in fp16 data format.
+        input_in_bf16: flag to indicate if input in bf16 data format.
+        attn_mask_type: attention mask type (pad or causal)
+        scaled_masked_softmax_fusion: flag to indicate user want to use softmax fusion
+        mask_func: mask function to be applied.
+        softmax_in_fp32: if true, softmax in performed at fp32 precision.
+        scale: scaling factor used in input tensor scaling.
+    """
+
+    def __init__(
+        self,
+        input_in_fp16,
+        input_in_bf16,
+        attn_mask_type,
+        scaled_masked_softmax_fusion,
+        mask_func,
+        softmax_in_fp32,
+        scale,
+    ):
+        super(FusedScaleMaskSoftmax, self).__init__()
+        self.input_in_fp16 = input_in_fp16
+        self.input_in_bf16 = input_in_bf16
+        assert not (
+            self.input_in_fp16 and self.input_in_bf16
+        ), "both fp16 and bf16 flags cannot be active at the same time."
+        self.input_in_float16 = self.input_in_fp16 or self.input_in_bf16
+        self.attn_mask_type = attn_mask_type
+        self.scaled_masked_softmax_fusion = scaled_masked_softmax_fusion
+        self.mask_func = mask_func
+        self.softmax_in_fp32 = softmax_in_fp32
+        self.scale = scale
+
+        assert (
+            self.scale is None or softmax_in_fp32
+        ), "softmax should be in fp32 when scaled"
+
+    def forward(self, input, mask):
+        # [b, np, sq, sk]
+        assert input.dim() == 4
+
+        if self.is_kernel_available(mask, *input.size()):
+            return self.forward_fused_softmax(input, mask)
+        else:
+            return self.forward_torch_softmax(input, mask)
+
+    def is_kernel_available(self, mask, b, np, sq, sk):
+        attn_batches = b * np
+
+        if (
+            self.scaled_masked_softmax_fusion  # user want to fuse
+            and self.input_in_float16  # input must be fp16
+            and 16 < sk <= 4096  # sk must be 16 ~ 2048
+            and sq % 4 == 0  # sq must be divisor of 4
+            and sk % 4 == 0  # sk must be divisor of 4 
+            and attn_batches % 4 == 0  # np * b must be divisor of 4
+        ):
+            if 0 <= sk <= 4096:
+                batch_per_block = self.get_batch_per_block(sq, sk, b, np)
+
+                if self.attn_mask_type == AttnMaskType.causal:
+                    if attn_batches % batch_per_block == 0:
+                        return True
+                else:
+                    if sq % batch_per_block == 0:
+                        return True
+        return False
+
+    def forward_fused_softmax(self, input, mask):
+        b, np, sq, sk = input.size()
+        scale = self.scale if self.scale is not None else 1.0
+
+        if self.attn_mask_type == AttnMaskType.causal:
+            assert sq == sk, "causal mask is only for self attention"
+
+            # input is 3D tensor (attn_batches, sq, sk)
+            input = input.view(-1, sq, sk)
+            probs = ScaledUpperTriangMaskedSoftmax.apply(input, scale)
+            return probs.view(b, np, sq, sk)
+        else:
+            # input is 4D tensor (b, np, sq, sk)
+            if mask is not None:
+                return ScaledMaskedSoftmax.apply(input, mask, scale)
+            else:
+                return ScaledSoftmax.apply(input, scale)
+
+    def forward_torch_softmax(self, input, mask):
+        if self.input_in_float16 and self.softmax_in_fp32:
+            input = input.float()
+
+        if self.scale is not None:
+            input = input * self.scale
+        mask_output = self.mask_func(input, mask) if mask is not None else input
+        probs = torch.nn.Softmax(dim=-1)(mask_output)
+
+        if self.input_in_float16 and self.softmax_in_fp32:
+            if self.input_in_fp16:
+                probs = probs.half()
+            else:
+                probs = probs.bfloat16()
+
+        return probs
+
+    @staticmethod
+    def get_batch_per_block(sq, sk, b, np):
+        import scaled_masked_softmax_cuda
+
+        return scaled_masked_softmax_cuda.get_batch_per_block(sq, sk, b, np)

megatron/model/gpt_model.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""GPT-2 model."""
+
+import torch
+
+from megatron import get_args
+from megatron.core import tensor_parallel
+from .module import MegatronModule
+
+from .enums import AttnMaskType
+from .language_model import parallel_lm_logits
+from .language_model import get_language_model
+from .utils import init_method_normal
+from .utils import scaled_init_method_normal
+
+
+def post_language_model_processing(lm_output, labels, logit_weights,
+                                   parallel_output,
+                                   fp16_lm_cross_entropy):
+
+    # Output. Format [s b h]
+    output = parallel_lm_logits(
+        lm_output,
+        logit_weights,
+        parallel_output)
+
+    if labels is None:
+        # [s b h] => [b s h]
+        return output.transpose(0,1).contiguous()
+    else:
+        # [b s] => [s b]
+        labels = labels.transpose(0,1).contiguous()
+        if fp16_lm_cross_entropy:
+            assert output.dtype == torch.half
+            loss = tensor_parallel.vocab_parallel_cross_entropy(output, labels)
+        else:
+            loss = tensor_parallel.vocab_parallel_cross_entropy(output.float(), labels)
+        
+        # [s b] => [b, s]
+        loss = loss.transpose(0,1).contiguous()
+        return loss
+
+
+class GPTModel(MegatronModule):
+    """GPT-2 Language model."""
+
+    def __init__(self,
+                 num_tokentypes=0,
+                 parallel_output=True,
+                 pre_process=True,
+                 post_process=True):
+        args = get_args()
+        super(GPTModel, self).__init__(share_word_embeddings=not args.untie_embeddings_and_output_weights)
+
+        self.parallel_output = parallel_output
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
+        self.untie_embeddings_and_output_weights = args.untie_embeddings_and_output_weights
+
+        self.language_model, self._language_model_key = get_language_model(
+            num_tokentypes=num_tokentypes,
+            add_pooler=False,
+            encoder_attn_mask_type=AttnMaskType.causal,
+            init_method=init_method_normal(args.init_method_std),
+            scaled_init_method=scaled_init_method_normal(args.init_method_std,
+                                                         args.num_layers),
+            pre_process=self.pre_process,
+            post_process=self.post_process)
+        
+        if not args.untie_embeddings_and_output_weights:
+            self.initialize_word_embeddings(init_method_normal)
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.language_model.set_input_tensor(input_tensor)
+
+    def forward(self, input_ids, position_ids, attention_mask,
+                ret_input_ids=None, ret_position_ids=None, ret_attn_mask=None,
+                labels=None, tokentype_ids=None, inference_params=None):
+
+        lm_output = self.language_model(
+            input_ids,
+            position_ids,
+            attention_mask,
+            ret_input_ids=ret_input_ids,
+            ret_position_ids=ret_position_ids,
+            ret_attn_mask=ret_attn_mask,
+            inference_params=inference_params)
+
+        if self.post_process:
+            return post_language_model_processing(
+                lm_output, labels,
+                self.language_model.output_layer.weight if self.untie_embeddings_and_output_weights else self.word_embeddings_weight(),
+                self.parallel_output,
+                self.fp16_lm_cross_entropy)
+        else:
+            return lm_output
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+
+        state_dict_ = {}
+        state_dict_[self._language_model_key] \
+            = self.language_model.state_dict_for_save_checkpoint(
+                prefix=prefix, keep_vars=keep_vars)
+        # Save word_embeddings.
+        if self.post_process and not self.pre_process and not self.untie_embeddings_and_output_weights:
+            state_dict_[self._word_embeddings_for_head_key] \
+                = self.word_embeddings.state_dict(prefix=prefix,
+                                                  keep_vars=keep_vars)
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+
+        # Load word_embeddings.
+        if self.post_process and not self.pre_process and not self.untie_embeddings_and_output_weights:
+            self.word_embeddings.load_state_dict(
+                state_dict[self._word_embeddings_for_head_key], strict=strict)
+        if self._language_model_key in state_dict:
+            state_dict = state_dict[self._language_model_key]
+        self.language_model.load_state_dict(state_dict, strict=strict)

megatron/model/language_model.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Transformer based language model."""
+
+import torch
+import torch.nn.functional as F
+
+from megatron import get_args
+from megatron.core import mpu, tensor_parallel
+
+from .enums import LayerType, AttnMaskType
+from .module import MegatronModule
+from .retro_transformer import ParallelRetroEncoder, ParallelRetroTransformer
+from .rotary_pos_embedding import apply_rotary_pos_emb, RotaryEmbedding
+from .transformer import ParallelTransformer
+from .utils import get_linear_layer
+from .utils import init_method_normal, scaled_init_method_normal
+
+
+def parallel_lm_logits(input_, word_embeddings_weight, parallel_output,
+                       bias=None):
+    """LM logits using word embedding weights."""
+    args = get_args()
+    # Parallel logits.
+    if args.async_tensor_model_parallel_allreduce or\
+            args.sequence_parallel:
+        input_parallel = input_
+        model_parallel = mpu.get_tensor_model_parallel_world_size() > 1
+        async_grad_allreduce = args.async_tensor_model_parallel_allreduce and \
+            model_parallel and not args.sequence_parallel
+    else:
+        input_parallel = tensor_parallel.copy_to_tensor_model_parallel_region(input_)
+        async_grad_allreduce = False
+
+    # Matrix multiply.
+    logits_parallel = tensor_parallel.linear_with_grad_accumulation_and_async_allreduce(
+        input=input_parallel,
+        weight=word_embeddings_weight,
+        bias=bias,
+        gradient_accumulation_fusion=args.gradient_accumulation_fusion,
+        async_grad_allreduce=async_grad_allreduce,
+        sequence_parallel_enabled=args.sequence_parallel)
+    # Gather if needed.
+
+    if parallel_output:
+        return logits_parallel
+
+    return tensor_parallel.gather_from_tensor_model_parallel_region(logits_parallel)
+
+
+def get_language_model(num_tokentypes, add_pooler,
+                       encoder_attn_mask_type, init_method=None,
+                       scaled_init_method=None, add_encoder=True,
+                       add_decoder=False,
+                       decoder_attn_mask_type=AttnMaskType.causal,
+                       pre_process=True, post_process=True):
+    """Build language model and return along with the key to save."""
+    args = get_args()
+
+    if init_method is None:
+        init_method = init_method_normal(args.init_method_std)
+
+    if scaled_init_method is None:
+        scaled_init_method = scaled_init_method_normal(args.init_method_std,
+                                                       args.num_layers)
+
+    # Language model.
+    language_model = TransformerLanguageModel(
+        init_method,
+        scaled_init_method,
+        encoder_attn_mask_type,
+        num_tokentypes=num_tokentypes,
+        add_encoder=add_encoder,
+        add_decoder=add_decoder,
+        decoder_attn_mask_type=decoder_attn_mask_type,
+        add_pooler=add_pooler,
+        pre_process=pre_process,
+        post_process=post_process
+    )
+    # key used for checkpoints.
+    language_model_key = 'language_model'
+
+    return language_model, language_model_key
+
+
+class Pooler(MegatronModule):
+    """Pooler layer.
+
+    Pool hidden states of a specific token (for example start of the
+    sequence) and add a linear transformation followed by a tanh.
+
+    Arguments:
+        hidden_size: hidden size
+        init_method: weight initialization method for the linear layer.
+            bias is set to zero.
+    """
+
+    def __init__(self, hidden_size, init_method):
+        super(Pooler, self).__init__()
+        args = get_args()
+        self.dense = get_linear_layer(hidden_size, hidden_size, init_method)
+        self.sequence_parallel = args.sequence_parallel
+
+
+    def forward(self, hidden_states, sequence_index=0):
+        # hidden_states: [s, b, h]
+        # sequence_index: index of the token to pool.
+
+        # gather data along sequence dimensions
+        # same pooler is run on all tensor parallel nodes
+        if self.sequence_parallel:
+            hidden_states = tensor_parallel.gather_from_sequence_parallel_region(
+                hidden_states,
+                tensor_parallel_output_grad=False)
+
+        pooled = hidden_states[sequence_index, :, :]
+        pooled = self.dense(pooled)
+        pooled = torch.tanh(pooled)
+        return pooled
+
+
+class Embedding(MegatronModule):
+    """Language model embeddings.
+
+    Arguments:
+        hidden_size: hidden size
+        vocab_size: vocabulary size
+        max_sequence_length: maximum size of sequence. This
+                             is used for positional embedding
+        embedding_dropout_prob: dropout probability for embeddings
+        init_method: weight initialization method
+        num_tokentypes: size of the token-type embeddings. 0 value
+                        will ignore this embedding
+    """
+
+    def __init__(self,
+                 hidden_size,
+                 vocab_size,
+                 max_sequence_length,
+                 embedding_dropout_prob,
+                 init_method,
+                 num_tokentypes=0):
+        super(Embedding, self).__init__()
+
+        self.hidden_size = hidden_size
+        self.init_method = init_method
+        self.num_tokentypes = num_tokentypes
+
+        args = get_args()
+
+        # Word embeddings (parallel).
+        self.word_embeddings = tensor_parallel.VocabParallelEmbedding(
+            vocab_size, self.hidden_size,
+            init_method=self.init_method,
+            params_dtype=args.params_dtype,
+            use_cpu_initialization=args.use_cpu_initialization,
+            perform_initialization=args.perform_initialization
+        )
+        self._word_embeddings_key = 'word_embeddings'
+
+        # Position embedding (serial).
+        self.add_position_embedding = args.add_position_embedding
+        if self.add_position_embedding:
+            self.position_embeddings = torch.nn.Embedding(
+                max_sequence_length, self.hidden_size)
+            self._position_embeddings_key = 'position_embeddings'
+            # Initialize the position embeddings.
+            if args.perform_initialization:
+                self.init_method(self.position_embeddings.weight)
+
+        # Token type embedding.
+        # Add this as an optional field that can be added through
+        # method call so we can load a pretrain model without
+        # token types and add them as needed.
+        self._tokentype_embeddings_key = 'tokentype_embeddings'
+        if self.num_tokentypes > 0:
+            self.tokentype_embeddings = torch.nn.Embedding(self.num_tokentypes,
+                                                           self.hidden_size)
+            # Initialize the token-type embeddings.
+            if args.perform_initialization:
+                self.init_method(self.tokentype_embeddings.weight)
+        else:
+            self.tokentype_embeddings = None
+
+        self.fp32_residual_connection = args.fp32_residual_connection 
+        self.sequence_parallel = args.sequence_parallel
+        # Embeddings dropout
+        self.embedding_dropout = torch.nn.Dropout(embedding_dropout_prob)
+
+    def zero_parameters(self):
+        """Zero out all parameters in embedding."""
+        self.word_embeddings.weight.data.fill_(0)
+        self.word_embeddings.weight.shared = True
+        if self.add_position_embedding:
+            self.position_embeddings.weight.data.fill_(0)
+            self.position_embeddings.weight.shared = True
+        if self.num_tokentypes > 0:
+            self.tokentype_embeddings.weight.data.fill_(0)
+            self.tokentype_embeddings.weight.shared = True
+
+    def add_tokentype_embeddings(self, num_tokentypes):
+        """Add token-type embedding. This function is provided so we can add
+        token-type embeddings in case the pretrained model does not have it.
+        This allows us to load the model normally and then add this embedding.
+        """
+        if self.tokentype_embeddings is not None:
+            raise Exception('tokentype embeddings is already initialized')
+        if torch.distributed.get_rank() == 0:
+            print('adding embedding for {} tokentypes'.format(num_tokentypes),
+                  flush=True)
+        self.num_tokentypes = num_tokentypes
+        self.tokentype_embeddings = torch.nn.Embedding(num_tokentypes,
+                                                       self.hidden_size)
+        # Initialize the token-type embeddings.
+        args = get_args()
+        self.init_method(self.tokentype_embeddings.weight)
+
+    def forward(self, input_ids, position_ids, tokentype_ids=None):
+        # Embeddings.
+        words_embeddings = self.word_embeddings(input_ids)
+        if self.add_position_embedding:
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings = words_embeddings + position_embeddings
+        else:
+            embeddings = words_embeddings
+
+        if tokentype_ids is not None:
+            assert self.tokentype_embeddings is not None
+            embeddings = embeddings + self.tokentype_embeddings(tokentype_ids)
+        else:
+            assert self.tokentype_embeddings is None
+
+        # Data format change to avoid explicit tranposes : [b s h] --> [s b h].
+        embeddings = embeddings.transpose(0, 1).contiguous()
+
+        # If the input flag for fp32 residual connection is set, convert for float.
+        if self.fp32_residual_connection:
+            embeddings = embeddings.float()
+
+        # Dropout.
+        if self.sequence_parallel:
+            embeddings = tensor_parallel.scatter_to_sequence_parallel_region(embeddings)
+            with tensor_parallel.get_cuda_rng_tracker().fork():
+                embeddings = self.embedding_dropout(embeddings)
+        else:
+            embeddings = self.embedding_dropout(embeddings)
+
+        return embeddings
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """For easy load."""
+
+        state_dict_ = {}
+        state_dict_[self._word_embeddings_key] \
+            = self.word_embeddings.state_dict(prefix=prefix,
+                                              keep_vars=keep_vars)
+        if self.add_position_embedding:
+            state_dict_[self._position_embeddings_key] \
+                = self.position_embeddings.state_dict(prefix=prefix,
+                                                  keep_vars=keep_vars)
+        if self.num_tokentypes > 0:
+            state_dict_[self._tokentype_embeddings_key] \
+                = self.tokentype_embeddings.state_dict(prefix=prefix,
+                                                       keep_vars=keep_vars)
+
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+
+        # Word embedding.
+        if self._word_embeddings_key in state_dict:
+            state_dict_ = state_dict[self._word_embeddings_key]
+        else:
+            # for backward compatibility.
+            state_dict_ = {}
+            for key in state_dict.keys():
+                if 'word_embeddings' in key:
+                    state_dict_[key.split('word_embeddings.')[1]] \
+                        = state_dict[key]
+        self.word_embeddings.load_state_dict(state_dict_, strict=strict)
+
+        # Position embedding.
+        if self.add_position_embedding:
+            if self._position_embeddings_key in state_dict:
+                state_dict_ = state_dict[self._position_embeddings_key]
+            else:
+                # for backward compatibility.
+                state_dict_ = {}
+                for key in state_dict.keys():
+                    if 'position_embeddings' in key:
+                        state_dict_[key.split('position_embeddings.')[1]] \
+                            = state_dict[key]
+            self.position_embeddings.load_state_dict(state_dict_, strict=strict)
+
+        # Tokentype embedding.
+        if self.num_tokentypes > 0:
+            state_dict_ = {}
+            if self._tokentype_embeddings_key in state_dict:
+                state_dict_ = state_dict[self._tokentype_embeddings_key]
+            else:
+                # for backward compatibility.
+                for key in state_dict.keys():
+                    if 'tokentype_embeddings' in key:
+                        state_dict_[key.split('tokentype_embeddings.')[1]] \
+                            = state_dict[key]
+            if len(state_dict_.keys()) > 0:
+                self.tokentype_embeddings.load_state_dict(state_dict_,
+                                                          strict=strict)
+            else:
+                print('***WARNING*** expected tokentype embeddings in the '
+                      'checkpoint but could not find it', flush=True)
+
+
+class TransformerLanguageModel(MegatronModule):
+    """Transformer language model.
+
+    Arguments:
+        transformer_hparams: transformer hyperparameters
+        vocab_size: vocabulary size
+        max_sequence_length: maximum size of sequence. This
+                             is used for positional embedding
+        embedding_dropout_prob: dropout probability for embeddings
+        num_tokentypes: size of the token-type embeddings. 0 value
+                        will ignore this embedding
+    """
+
+    def __init__(self,
+                 init_method,
+                 output_layer_init_method,
+                 encoder_attn_mask_type,
+                 num_tokentypes=0,
+                 add_encoder=True,
+                 add_decoder=False,
+                 decoder_attn_mask_type=AttnMaskType.causal,
+                 add_pooler=False,
+                 pre_process=True,
+                 post_process=True):
+        args = get_args()
+        # TODO: passing share_word_embeddings=False will not work correctly for T5 and embeddings will not be synced. Fix later for T5.
+        if args.untie_embeddings_and_output_weights: assert not add_decoder
+        super(TransformerLanguageModel, self).__init__(share_word_embeddings=not args.untie_embeddings_and_output_weights)
+
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.hidden_size = args.hidden_size
+        self.num_tokentypes = num_tokentypes
+        self.init_method = init_method
+        self.add_encoder = add_encoder
+        self.encoder_attn_mask_type = encoder_attn_mask_type
+        self.add_decoder = add_decoder
+        self.decoder_attn_mask_type = decoder_attn_mask_type
+        self.add_pooler = add_pooler
+        self.encoder_hidden_state = None
+        self.untie_embeddings_and_output_weights = args.untie_embeddings_and_output_weights
+
+        # Embeddings.
+        if self.pre_process:
+            self.embedding = Embedding(self.hidden_size,
+                                       args.padded_vocab_size,
+                                       args.max_position_embeddings,
+                                       args.hidden_dropout,
+                                       self.init_method,
+                                       self.num_tokentypes)
+            self._embedding_key = 'embedding'
+
+        # Rotary positional embeddings
+        self.use_rotary_position_embeddings = \
+            args.use_rotary_position_embeddings
+        if args.use_rotary_position_embeddings:
+            self.seq_length = args.seq_length
+            rotary_dim = args.hidden_size // args.num_attention_heads \
+                if args.kv_channels is None else args.kv_channels
+
+            if args.rotary_percent < 1.0:
+                rotary_dim = int(rotary_dim * args.rotary_percent)
+
+            # partial rotary embeddings, which is better than full rotary
+            # Wang and Komatsuzaki et al
+            # https://github.com/kingoflolz/mesh-transformer-jax/
+            self.rotary_pos_emb = RotaryEmbedding(rotary_dim)
+
+        # Retriever (bi-directional transformer with cross attention)
+        if args.retro_add_retriever:
+            self.retriever = ParallelRetroEncoder(
+                self.init_method,
+                output_layer_init_method,
+                self_attn_mask_type=AttnMaskType.padding,
+                pre_process=self.pre_process,
+                post_process=False,
+            )
+            self._retriever_key = 'retriever'
+        else:
+            self.retriever = None
+
+        # Encoder (usually set to True, False if part of an encoder-decoder
+        # architecture and in encoder-only stage).
+        if self.add_encoder:
+            if args.retro_add_retriever:
+                self.encoder = ParallelRetroTransformer(
+                    self.init_method,
+                    output_layer_init_method,
+                    self_attn_mask_type=self.encoder_attn_mask_type,
+                    pre_process=self.pre_process,
+                    post_process=self.post_process,
+                    retriever=self.retriever,
+                )
+            else:
+                self.encoder = ParallelTransformer(
+                    self.init_method,
+                    output_layer_init_method,
+                    self_attn_mask_type=self.encoder_attn_mask_type,
+                    pre_process=self.pre_process,
+                    post_process=self.post_process,
+                )
+            self._encoder_key = 'encoder'
+        else:
+            self.encoder = None
+
+        # Decoder (usually set to False, True if part of an encoder-decoder
+        # architecture and in decoder-only stage).
+        if self.add_decoder:
+            self.decoder = ParallelTransformer(
+                self.init_method,
+                output_layer_init_method,
+                layer_type=LayerType.decoder,
+                self_attn_mask_type=self.decoder_attn_mask_type,
+                pre_process=self.pre_process,
+                post_process=self.post_process)
+            self._decoder_key = 'decoder'
+        else:
+            self.decoder = None
+
+        if self.post_process:
+            # Pooler.
+            if self.add_pooler:
+                self.pooler = Pooler(self.hidden_size, self.init_method)
+                self._pooler_key = 'pooler'
+
+            if self.untie_embeddings_and_output_weights:
+                self.output_layer = tensor_parallel.ColumnParallelLinear(
+                    args.hidden_size,
+                    args.padded_vocab_size,
+                    bias=False, # Setting bias to False always to keep it consistent with embedding tying that also does not have a bias.
+                    init_method=self.init_method)
+                self._output_layer_key = 'output_layer'
+
+    def set_input_tensor(self, input_tensor):
+        """ See megatron.model.transformer.set_input_tensor()"""
+
+        # This is usually handled in schedules.py but some inference code still
+        # gives us non-lists or None
+        if not isinstance(input_tensor, list):
+            input_tensor = [input_tensor]
+
+        if self.add_encoder and self.add_decoder:
+            assert len(input_tensor) == 1, \
+                'input_tensor should only be length 1 for stage with both encoder and decoder'
+            self.encoder.set_input_tensor(input_tensor[0])
+        elif self.add_encoder:
+            assert len(input_tensor) == 1, \
+                'input_tensor should only be length 1 for stage with only encoder'
+            self.encoder.set_input_tensor(input_tensor[0])
+        elif self.add_decoder:
+            if len(input_tensor) == 2:
+                self.decoder.set_input_tensor(input_tensor[0])
+                self.encoder_hidden_state = input_tensor[1]
+            elif len(input_tensor) == 1:
+                self.decoder.set_input_tensor(None)
+                self.encoder_hidden_state = input_tensor[0]
+            else:
+                raise Exception('input_tensor must have either length 1 or 2')
+        else:
+            raise Exception('Stage must have at least either encoder or decoder')
+
+    def forward(self, enc_input_ids, enc_position_ids, enc_attn_mask,
+                dec_input_ids=None, dec_position_ids=None, dec_attn_mask=None,
+                ret_input_ids=None, ret_position_ids=None, ret_attn_mask=None,
+                enc_dec_attn_mask=None, tokentype_ids=None,
+                inference_params=None,
+                pooling_sequence_index=0,
+                enc_hidden_states=None, output_enc_hidden=False):
+
+        # Retriever embedding.
+        if self.retriever and self.pre_process:
+            retriever_input = self.embedding(ret_input_ids, ret_position_ids,
+                                             tokentype_ids=tokentype_ids)
+        else:
+            retriever_input = None
+
+        # Encoder embedding.
+        if self.pre_process:
+            encoder_input = self.embedding(enc_input_ids, enc_position_ids,
+                                           tokentype_ids=tokentype_ids)
+        else:
+            encoder_input = None
+
+        # Rotary positional embeddings
+        rotary_pos_emb = None
+        if self.use_rotary_position_embeddings:
+            if inference_params is not None:
+                rotary_pos_emb = \
+                    self.rotary_pos_emb(inference_params.max_sequence_len)
+            else:
+                rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
+
+        # Run encoder.
+        if enc_hidden_states is None:
+            if self.encoder is not None:
+                if self.retriever:
+                    encoder_output = self.encoder(
+                        encoder_input,
+                        enc_attn_mask,
+                        retriever_output=retriever_input,
+                        retriever_attn_mask=ret_attn_mask,
+                        inference_params=inference_params)
+                else:
+                    encoder_output = self.encoder(
+                        encoder_input,
+                        enc_attn_mask,
+                        inference_params=inference_params,
+                        rotary_pos_emb=rotary_pos_emb)
+            else:
+                encoder_output = self.encoder_hidden_state
+        else:
+            encoder_output = enc_hidden_states.to(encoder_input.dtype)
+
+        if self.post_process:
+            if self.add_pooler:
+                pooled_output = self.pooler(encoder_output,
+                                            pooling_sequence_index)
+
+        # output_enc_hidden refers to when we just need the encoder's
+        # output. For example, it is helpful to compute
+        # similarity between two sequences by average pooling
+        if not self.add_decoder or output_enc_hidden:
+            if self.add_pooler and self.post_process:
+                return encoder_output, pooled_output
+            else:
+                return encoder_output
+
+        # Decoder embedding.
+        if self.pre_process:
+            decoder_input = self.embedding(dec_input_ids,
+                                           dec_position_ids)
+        else:
+            decoder_input = None
+
+        # Run decoder.
+        decoder_output = self.decoder(
+            decoder_input,
+            dec_attn_mask,
+            encoder_output=encoder_output,
+            enc_dec_attn_mask=enc_dec_attn_mask,
+            inference_params=inference_params,
+            rotary_pos_emb=rotary_pos_emb)
+
+        if self.add_pooler and self.post_process:
+            return decoder_output, encoder_output, pooled_output
+        else:
+            return decoder_output, encoder_output
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """For easy load."""
+
+        state_dict_ = {}
+        if self.pre_process:
+            state_dict_[self._embedding_key] \
+                = self.embedding.state_dict_for_save_checkpoint(prefix=prefix,
+                                                                keep_vars=keep_vars)
+        if self.add_encoder:
+            state_dict_[self._encoder_key] \
+                = self.encoder.state_dict_for_save_checkpoint(prefix=prefix,
+                                                              keep_vars=keep_vars)
+        if self.post_process:
+            if self.add_pooler:
+                state_dict_[self._pooler_key] \
+                    = self.pooler.state_dict_for_save_checkpoint(prefix=prefix,
+                                                                 keep_vars=keep_vars)
+            if self.untie_embeddings_and_output_weights:
+                state_dict_[self._output_layer_key] \
+                    = self.output_layer.state_dict(prefix=prefix, keep_vars=keep_vars)
+
+        if self.add_decoder:
+            state_dict_[self._decoder_key] \
+                = self.decoder.state_dict_for_save_checkpoint(prefix=prefix,
+                                                              keep_vars=keep_vars)
+
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+
+        # Embedding.
+        if self.pre_process:
+            if self._embedding_key in state_dict:
+                state_dict_ = state_dict[self._embedding_key]
+            else:
+                # for backward compatibility.
+                state_dict_ = {}
+                for key in state_dict.keys():
+                    if '_embeddings' in key:
+                        state_dict_[key] = state_dict[key]
+            self.embedding.load_state_dict(state_dict_, strict=strict)
+
+        # Encoder.
+        if self.add_encoder:
+            if self._encoder_key in state_dict:
+                state_dict_ = state_dict[self._encoder_key]
+            # For backward compatibility.
+            elif 'transformer' in state_dict:
+                state_dict_ = state_dict['transformer']
+            else:
+                # For backward compatibility.
+                state_dict_ = {}
+                for key in state_dict.keys():
+                    if 'transformer.' in key:
+                        state_dict_[key.split('transformer.')[1]] = state_dict[key]
+
+            # For backward compatibility.
+            state_dict_self_attention = {}
+            for key in state_dict_.keys():
+                if '.attention.' in key:
+                    state_dict_self_attention[key.replace(".attention.",
+                        ".self_attention.")] = state_dict_[key]
+                else:
+                    state_dict_self_attention[key] = state_dict_[key]
+            state_dict_ = state_dict_self_attention
+
+            self.encoder.load_state_dict(state_dict_, strict=strict)
+
+        # Pooler.
+        if self.post_process:
+            if self.add_pooler:
+                assert 'pooler' in state_dict, \
+                    'could not find data for pooler in the checkpoint'
+                self.pooler.load_state_dict(state_dict[self._pooler_key],
+                                            strict=strict)
+            if self.untie_embeddings_and_output_weights:
+                assert 'output_layer' in state_dict, \
+                    'could not find data for output_layer in the checkpoint'
+                self.output_layer.load_state_dict(state_dict[self._output_layer_key],
+                                                  strict=strict)
+        # Decoder.
+        if self.add_decoder:
+            assert 'decoder' in state_dict, \
+                'could not find data for pooler in the checkpoint'
+            self.decoder.load_state_dict(state_dict[self._decoder_key],
+                                         strict=strict)

megatron/model/module.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Megatron Module"""
+
+import torch
+from torch.autograd import Variable
+from torch.nn.parameter import Parameter
+
+from megatron import get_args
+from megatron.core import mpu, tensor_parallel
+
+
+_FLOAT_TYPES = (torch.FloatTensor, torch.cuda.FloatTensor)
+_HALF_TYPES = (torch.HalfTensor, torch.cuda.HalfTensor)
+_BF16_TYPES = (torch.BFloat16Tensor, torch.cuda.BFloat16Tensor)
+
+
+
+def param_is_not_shared(param):
+    return not hasattr(param, 'shared') or not param.shared
+
+
+
+class MegatronModule(torch.nn.Module):
+    """Megatron specific extensions of torch Module with support
+    for pipelining."""
+
+    def __init__(self, share_word_embeddings=True):
+        super(MegatronModule, self).__init__()
+        self.share_word_embeddings = share_word_embeddings
+
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """Use this function to override the state dict for
+        saving checkpoints."""
+        return self.state_dict(prefix=prefix, keep_vars=keep_vars)
+
+
+    def word_embeddings_weight(self):
+        if self.pre_process:
+            return self.language_model.embedding.word_embeddings.weight
+        else:
+            if not self.share_word_embeddings:
+                raise Exception('word_embeddings_weight() called for last '
+                                'stage, but share_word_embeddings is false')
+            return self.word_embeddings.weight
+
+
+    def initialize_word_embeddings(self, init_method_normal):
+        args = get_args()
+        if not self.share_word_embeddings:
+            raise Exception('initialize_word_embeddings() was called but '
+                            'share_word_embeddings is false')
+
+        # This function just initializes the word embeddings in the final stage
+        # when we are using pipeline parallelism. Nothing to do if we aren't
+        # using pipeline parallelism.
+        if args.pipeline_model_parallel_size == 1:
+            return
+
+        # Parameters are shared between the word embeddings layers, and the
+        # heads at the end of the model. In a pipelined setup with more than
+        # one stage, the initial embedding layer and the head are on different
+        # workers, so we do the following:
+        # 1. Create a second copy of word_embeddings on the last stage, with
+        #    initial parameters of 0.0.
+        # 2. Do an all-reduce between the first and last stage to ensure that
+        #    the two copies of word_embeddings start off with the same
+        #    parameter values.
+        # 3. In the training loop, before an all-reduce between the grads of
+        #    the two word_embeddings layers to ensure that every applied weight
+        #    update is the same on both stages.
+        if mpu.is_pipeline_last_stage() and not self.pre_process:
+            assert not mpu.is_pipeline_first_stage()
+            self._word_embeddings_for_head_key = 'word_embeddings_for_head'
+            # set word_embeddings weights to 0 here, then copy first
+            # stage's weights using all_reduce below.
+            self.word_embeddings = tensor_parallel.VocabParallelEmbedding(
+                args.padded_vocab_size, args.hidden_size,
+                init_method=init_method_normal(args.init_method_std),
+                params_dtype=args.params_dtype,
+                use_cpu_initialization=args.use_cpu_initialization,
+                perform_initialization=args.perform_initialization)
+            self.word_embeddings.weight.data.fill_(0)
+            self.word_embeddings.weight.shared = True
+
+        # Zero out initial weights for decoder embedding.
+        # NOTE: We don't currently support T5 with the interleaved schedule.
+        if not mpu.is_pipeline_first_stage(ignore_virtual=True) and \
+                self.pre_process:
+            self.language_model.embedding.zero_parameters()
+
+        if not torch.distributed.is_initialized():
+            if not getattr(MegatronModule, "embedding_warning_printed", False):
+                print("WARNING! Distributed processes aren't initialized, so "
+                      "word embeddings in the last layer are not initialized. "
+                      "If you are just manipulating a model this is fine, but "
+                      "this needs to be handled manually. If you are training "
+                      "something is definitely wrong.")
+                MegatronModule.embedding_warning_printed = True
+            return
+
+        # Ensure that first and last stages have the same initial parameter
+        # values.
+        if mpu.is_rank_in_embedding_group():
+            torch.distributed.all_reduce(self.word_embeddings_weight().data,
+                                         group=mpu.get_embedding_group())
+
+        # Ensure that encoder(first stage) and decoder(split stage) position
+        # embeddings have the same initial parameter values
+        # NOTE: We don't currently support T5 with the interleaved schedule.
+        if mpu.is_rank_in_position_embedding_group() and \
+                args.pipeline_model_parallel_split_rank is not None:
+            # TODO: Support tokentype embedding.
+            self.language_model.embedding.cuda()
+            position_embeddings = self.language_model.embedding.position_embeddings
+            torch.distributed.all_reduce(position_embeddings.weight.data,
+                                         group=mpu.get_position_embedding_group())
+
+
+def conversion_helper(val, conversion):
+    """Apply conversion to val. Recursively apply conversion if `val`
+    #is a nested tuple/list structure."""
+    if not isinstance(val, (tuple, list)):
+        return conversion(val)
+    rtn = [conversion_helper(v, conversion) for v in val]
+    if isinstance(val, tuple):
+        rtn = tuple(rtn)
+    return rtn
+
+
+def fp32_to_float16(val, float16_convertor):
+    """Convert fp32 `val` to fp16/bf16"""
+    def half_conversion(val):
+        val_typecheck = val
+        if isinstance(val_typecheck, (Parameter, Variable)):
+            val_typecheck = val.data
+        if isinstance(val_typecheck, _FLOAT_TYPES):
+            val = float16_convertor(val)
+        return val
+    return conversion_helper(val, half_conversion)
+
+
+def float16_to_fp32(val):
+    """Convert fp16/bf16 `val` to fp32"""
+    def float_conversion(val):
+        val_typecheck = val
+        if isinstance(val_typecheck, (Parameter, Variable)):
+            val_typecheck = val.data
+        if isinstance(val_typecheck, (_BF16_TYPES, _HALF_TYPES)):
+            val = val.float()
+        return val
+    return conversion_helper(val, float_conversion)
+
+
+
+class Float16Module(MegatronModule):
+
+    def __init__(self, module, args):
+        super(Float16Module, self).__init__()
+
+        if args.fp16:
+            self.add_module('module', module.half())
+            def float16_convertor(val):
+                return val.half()
+        elif args.bf16:
+            self.add_module('module', module.bfloat16())
+            def float16_convertor(val):
+                return val.bfloat16()
+        else:
+            raise Exception('should not be here')
+
+        self.float16_convertor = float16_convertor
+
+
+    def set_input_tensor(self, input_tensor):
+        return self.module.set_input_tensor(input_tensor)
+
+
+    def forward(self, *inputs, **kwargs):
+        if mpu.is_pipeline_first_stage():
+            inputs = fp32_to_float16(inputs, self.float16_convertor)
+        outputs = self.module(*inputs, **kwargs)
+        if mpu.is_pipeline_last_stage():
+            outputs = float16_to_fp32(outputs)
+        return outputs
+
+
+    def state_dict(self, prefix='', keep_vars=False):
+        return self.module.state_dict(prefix=prefix, keep_vars=keep_vars)
+
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        return self.module.state_dict_for_save_checkpoint(prefix=prefix,
+                                                          keep_vars=keep_vars)
+
+
+    def load_state_dict(self, state_dict, strict=True):
+        self.module.load_state_dict(state_dict, strict=strict)

megatron/model/multiple_choice.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Multiple choice model."""
+
+import torch
+
+from megatron import get_args, print_rank_last
+from megatron.model.enums import AttnMaskType
+from megatron.model.bert_model import bert_extended_attention_mask, bert_position_ids
+from megatron.model.language_model import get_language_model
+from megatron.model.utils import get_linear_layer
+from megatron.model.utils import init_method_normal
+from megatron.model.utils import scaled_init_method_normal
+from .module import MegatronModule
+
+
+class MultipleChoice(MegatronModule):
+
+    def __init__(self,
+                 num_tokentypes=2,
+                 pre_process=True,
+                 post_process=True):
+        super(MultipleChoice, self).__init__(share_word_embeddings=False)
+        args = get_args()
+
+        init_method = init_method_normal(args.init_method_std)
+        self.pre_process = pre_process
+        self.post_process = post_process
+
+        self.language_model, self._language_model_key = get_language_model(
+            num_tokentypes=num_tokentypes,
+            add_pooler=True,
+            encoder_attn_mask_type=AttnMaskType.padding,
+            init_method=init_method,
+            scaled_init_method=scaled_init_method_normal(args.init_method_std,
+                                                         args.num_layers),
+            pre_process=self.pre_process,
+            post_process=self.post_process)
+
+        # Multi-choice head.
+        if self.post_process:
+            self.multichoice_dropout = torch.nn.Dropout(args.hidden_dropout)
+            self.multichoice_head = get_linear_layer(args.hidden_size, 1,
+                                                     init_method)
+            self._multichoice_head_key = 'multichoice_head'
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.language_model.set_input_tensor(input_tensor)
+
+    def forward(self, model_input, attention_mask, tokentype_ids=None):
+
+        # [batch, choices, sequence] --> [batch * choices, sequence] -->
+        #    transformer --> [batch, choices] --> softmax
+
+        # Ensure the shape is [batch-size, choices, sequence]
+        assert len(attention_mask.shape) == 3
+        num_choices = attention_mask.shape[1]
+
+        # Reshape and treat choice dimension the same as batch.
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1))
+        extended_attention_mask = bert_extended_attention_mask(attention_mask)
+
+        input_ids = model_input
+        # Do the same as attention_mask for input_ids, tokentype_ids
+        assert len(input_ids.shape) == 3
+        assert len(tokentype_ids.shape) == 3
+        input_ids = input_ids.view(-1, input_ids.size(-1))
+        tokentype_ids = tokentype_ids.view(-1, tokentype_ids.size(-1))
+        position_ids = bert_position_ids(input_ids)
+
+        lm_output = self.language_model(
+            input_ids,
+            position_ids,
+            extended_attention_mask,
+            tokentype_ids=tokentype_ids
+        )
+        if self.post_process:
+            _, pooled_output = lm_output
+            multichoice_output = self.multichoice_dropout(pooled_output)
+            multichoice_logits = self.multichoice_head(multichoice_output)
+
+            # Reshape back to separate choices.
+            multichoice_logits = multichoice_logits.view(-1, num_choices)
+
+            return multichoice_logits
+        return lm_output
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """For easy load when model is combined with other heads,
+        add an extra key."""
+
+        state_dict_ = {}
+        state_dict_[self._language_model_key] \
+            = self.language_model.state_dict_for_save_checkpoint(prefix=prefix,
+                                                                 keep_vars=keep_vars)
+        if self.post_process:
+            state_dict_[self._multichoice_head_key] \
+                = self.multichoice_head.state_dict(prefix=prefix, keep_vars=keep_vars)
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+
+        self.language_model.load_state_dict(
+            state_dict[self._language_model_key], strict=strict)
+        if self.post_process:
+            if self._multichoice_head_key in state_dict:
+                self.multichoice_head.load_state_dict(
+                    state_dict[self._multichoice_head_key], strict=strict)
+            else:
+                print_rank_last('***WARNING*** could not find {} in the checkpoint, '
+                                'initializing to random'.format(
+                                    self._multichoice_head_key))

megatron/model/realm_model.py

+import os
+import torch
+
+from megatron import get_args, print_rank_0
+from megatron.checkpointing import get_checkpoint_tracker_filename, get_checkpoint_name
+from megatron.model import BertModel
+from .module import MegatronModule
+from megatron.core import mpu
+from megatron.model.enums import AttnMaskType
+from megatron.model.utils import get_linear_layer
+from megatron.model.utils import init_method_normal
+from megatron.model.language_model import get_language_model
+from megatron.model.utils import scaled_init_method_normal
+from megatron.model.bert_model import bert_extended_attention_mask, bert_position_ids
+
+
+def general_ict_model_provider(only_query_model=False, only_block_model=False):
+    """Build the model."""
+    args = get_args()
+    assert args.ict_head_size is not None, \
+        "Need to specify --ict-head-size to provide an ICTBertModel"
+    assert mpu.get_tensor_model_parallel_world_size() == 1 and mpu.get_pipeline_model_parallel_world_size() == 1, \
+        "Model parallel size > 1 not supported for ICT"
+
+    print_rank_0('building ICTBertModel...')
+
+    # simpler to just keep using 2 tokentypes since the LM we initialize with has 2 tokentypes
+    model = ICTBertModel(
+        ict_head_size=args.ict_head_size,
+        num_tokentypes=2,
+        parallel_output=True,
+        only_query_model=only_query_model,
+        only_block_model=only_block_model)
+
+    return model
+
+
+class ICTBertModel(MegatronModule):
+    """Bert-based module for Inverse Cloze task."""
+    def __init__(self,
+                 ict_head_size,
+                 num_tokentypes=1,
+                 parallel_output=True,
+                 only_query_model=False,
+                 only_block_model=False):
+        super(ICTBertModel, self).__init__()
+        bert_kwargs = dict(
+            ict_head_size=ict_head_size,
+            num_tokentypes=num_tokentypes,
+            parallel_output=parallel_output
+        )
+        assert not (only_block_model and only_query_model)
+        self.use_block_model = not only_query_model
+        self.use_query_model = not only_block_model
+
+        if self.use_query_model:
+            # this model embeds (pseudo-)queries - Embed_input in the paper
+            self.query_model = IREncoderBertModel(**bert_kwargs)
+            self._query_key = 'question_model'
+
+        if self.use_block_model:
+            # this model embeds evidence blocks - Embed_doc in the paper
+            self.block_model = IREncoderBertModel(**bert_kwargs)
+            self._block_key = 'context_model'
+
+    def forward(self, query_tokens, query_attention_mask, block_tokens, block_attention_mask):
+        """Run a forward pass for each of the models and return the respective embeddings."""
+        query_logits = self.embed_query(query_tokens, query_attention_mask)
+        block_logits = self.embed_block(block_tokens, block_attention_mask)
+        return query_logits, block_logits
+
+    def embed_query(self, query_tokens, query_attention_mask):
+        """Embed a batch of tokens using the query model"""
+        if self.use_query_model:
+            query_types = torch.cuda.LongTensor(*query_tokens.shape).fill_(0)
+            query_ict_logits, _ = self.query_model.forward(query_tokens, query_attention_mask, query_types)
+            return query_ict_logits
+        else:
+            raise ValueError("Cannot embed query without query model.")
+
+    def embed_block(self, block_tokens, block_attention_mask):
+        """Embed a batch of tokens using the block model"""
+        if self.use_block_model:
+            block_types = torch.cuda.LongTensor(*block_tokens.shape).fill_(0)
+            block_ict_logits, _ = self.block_model.forward(block_tokens, block_attention_mask, block_types)
+            return block_ict_logits
+        else:
+            raise ValueError("Cannot embed block without block model.")
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """Save dict with state dicts of each of the models."""
+        state_dict_ = {}
+        if self.use_query_model:
+            state_dict_[self._query_key] \
+                = self.query_model.state_dict_for_save_checkpoint(
+                    prefix=prefix, keep_vars=keep_vars)
+
+        if self.use_block_model:
+            state_dict_[self._block_key] \
+                = self.block_model.state_dict_for_save_checkpoint(
+                    prefix=prefix, keep_vars=keep_vars)
+
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Load the state dicts of each of the models"""
+        if self.use_query_model:
+            print("Loading ICT query model", flush=True)
+            self.query_model.load_state_dict(
+                state_dict[self._query_key], strict=strict)
+
+        if self.use_block_model:
+            print("Loading ICT block model", flush=True)
+            self.block_model.load_state_dict(
+                state_dict[self._block_key], strict=strict)
+
+    def init_state_dict_from_bert(self):
+        """Initialize the state from a pretrained BERT model on iteration zero of ICT pretraining"""
+        args = get_args()
+        tracker_filename = get_checkpoint_tracker_filename(args.bert_load)
+        if not os.path.isfile(tracker_filename):
+            raise FileNotFoundError("Could not find BERT load for ICT")
+        with open(tracker_filename, 'r') as f:
+            iteration = int(f.read().strip())
+            assert iteration > 0
+
+        checkpoint_name = get_checkpoint_name(args.bert_load, iteration, False)
+        if mpu.get_data_parallel_rank() == 0:
+            print('global rank {} is loading checkpoint {}'.format(
+                torch.distributed.get_rank(), checkpoint_name))
+
+        try:
+            state_dict = torch.load(checkpoint_name, map_location='cpu')
+        except BaseException:
+            raise ValueError("Could not load checkpoint")
+
+        # load the LM state dict into each model
+        model_dict = state_dict['model']['language_model']
+        self.query_model.language_model.load_state_dict(model_dict)
+        self.block_model.language_model.load_state_dict(model_dict)
+
+        # give each model the same ict_head to begin with as well
+        query_ict_head_state_dict = self.state_dict_for_save_checkpoint()[self._query_key]['ict_head']
+        self.block_model.ict_head.load_state_dict(query_ict_head_state_dict)
+
+
+class IREncoderBertModel(MegatronModule):
+    """BERT-based encoder for queries or blocks used for learned information retrieval."""
+    def __init__(self, ict_head_size, num_tokentypes=2, parallel_output=True):
+        super(IREncoderBertModel, self).__init__()
+        args = get_args()
+
+        self.ict_head_size = ict_head_size
+        self.parallel_output = parallel_output
+        init_method = init_method_normal(args.init_method_std)
+        scaled_init_method = scaled_init_method_normal(args.init_method_std,
+                                                       args.num_layers)
+
+        self.language_model, self._language_model_key = get_language_model(
+            num_tokentypes=num_tokentypes,
+            add_pooler=True,
+            encoder_attn_mask_type=AttnMaskType.padding,
+            init_method=init_method,
+            scaled_init_method=scaled_init_method)
+
+        self.ict_head = get_linear_layer(args.hidden_size, ict_head_size, init_method)
+        self._ict_head_key = 'ict_head'
+
+    def forward(self, input_ids, attention_mask, tokentype_ids=None):
+        extended_attention_mask = bert_extended_attention_mask(
+            attention_mask, next(self.language_model.parameters()).dtype)
+        position_ids = bert_position_ids(input_ids)
+
+        lm_output, pooled_output = self.language_model(
+            input_ids,
+            position_ids,
+            extended_attention_mask,
+            tokentype_ids=tokentype_ids)
+
+        # Output.
+        ict_logits = self.ict_head(pooled_output)
+        return ict_logits, None
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """For easy load when model is combined with other heads,
+        add an extra key."""
+
+        state_dict_ = {}
+        state_dict_[self._language_model_key] \
+            = self.language_model.state_dict_for_save_checkpoint(prefix=prefix,
+                                                                 keep_vars=keep_vars)
+        state_dict_[self._ict_head_key] \
+            = self.ict_head.state_dict(prefix=prefix,
+                                       keep_vars=keep_vars)
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+        self.language_model.load_state_dict(
+            state_dict[self._language_model_key], strict=strict)
+        self.ict_head.load_state_dict(
+            state_dict[self._ict_head_key], strict=strict)
+
+

megatron/model/retro_transformer.py

+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+
+"""Retro Transformer.
+
+** Special note about this file **
+Many classes and methods in this file directly parallel those in transformer.py
+in name and utility. However, due to 1) subtle changes in the code over time
+(i.e., transposes and contexts), and 2) other code that is soon to be merged,
+this file will *temporarily* remain as is, until a larger integration is
+complete.
+"""
+
+import math
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from megatron import get_args, get_retro_args, get_tensorboard_writer
+from megatron.core import parallel_state
+from megatron.core import tensor_parallel
+from megatron.core import utils as core_utils
+from megatron.core.enums import ModelType
+from megatron.model.enums import AttnMaskType, LayerType, AttnType
+from megatron.model import LayerNorm
+from megatron.model.fused_softmax import FusedScaleMaskSoftmax
+from megatron.model.fused_bias_gelu import bias_gelu_impl
+from megatron.model.utils import attention_mask_func, openai_gelu, erf_gelu, init_method_normal
+
+from .module import MegatronModule
+from .transformer import _get_num_layers, ParallelMLP, NoopTransformerLayer
+
+""" We use the following notation throughout this file:
+     h: hidden size
+     n: number of attention heads
+     p: number of model parallel partitions
+     np: n/p
+     hp: h/p
+     hn: h/n
+     b: batch size
+     s: sequence length
+     l: number of layers
+    Transformer takes input of size [s, b, h] and returns a
+    tensor of the same size. We use the following arguments:
+        hyperparameters: transformer hyperparameters
+"""
+
+
+class DropPath(MegatronModule):
+    """Drop paths (Stochastic Depth) per sample
+    (when applied in main path of residual blocks).
+
+    *Note: differs from transformer.py/DropPath in hidden_state transpose.
+    """
+
+    def __init__(self, drop_prob=0.):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, hidden_state):
+        if self.drop_prob == 0. or not self.training:
+            return hidden_state
+        keep_prob = 1 - self.drop_prob
+        # work with diff dim tensors, not just 2D ConvNets
+        shape = (hidden_state.shape[0],) + (1,) * (hidden_state.ndim - 1)
+        random_tensor = keep_prob + \
+            torch.rand(shape, dtype=hidden_state.dtype, device=hidden_state.device)
+        random_tensor.floor_() # binarize
+        output = hidden_state.div(keep_prob) * random_tensor
+        return output
+
+
+class SwitchMLP(MegatronModule):
+    """
+    Routes input to one of N MLP "experts"
+    """
+    def __init__(self, init_method, output_layer_init_method):
+        super(SwitchMLP, self).__init__()
+        args = get_args()
+        self.router = torch.nn.Linear(args.hidden_size, args.num_experts)
+        self.experts = torch.nn.ModuleList()
+        for i in range(args.num_experts):
+            self.experts.append(ParallelMLP(init_method, output_layer_init_method))
+
+    def forward(self, hidden_states):
+        # hidden_states: [b, s, h]
+        b = hidden_states.size(0)
+        s = hidden_states.size(1)
+        h = hidden_states.size(2)
+        route = self.router(hidden_states)
+        route = torch.nn.functional.softmax(route, dim=2)
+        max_prob, max_ind = torch.max(route, dim=2)
+        max_prob = torch.unsqueeze(max_prob, 2) # [b s 1]
+
+        # TODO (rprenger) TODO this could be made easier to read
+        # Converting [b, s, h] to [b*s, h].
+        # Each vector could be routed differently
+        hidden_states = hidden_states.view(-1, hidden_states.size(2)) # [b*s h]
+        max_prob = max_prob.view(-1, max_prob.size(2)) # [b*s 1]
+        max_ind = max_ind.view(-1) # [b*s]
+
+        output_total = torch.empty_like(hidden_states)
+        output_bias_total = torch.empty_like(hidden_states)
+        #TODO (rprenger) This does each expert in serial, but it could be parallelized
+
+        for expert_num, expert in enumerate(self.experts):
+            local_indices = (max_ind == expert_num).nonzero()
+            hidden = hidden_states[local_indices,:]
+            output, output_bias = expert(hidden)
+            output_bias = output_bias.expand_as(output)
+            output_total[local_indices,:] = output
+            output_bias_total[local_indices,:] = output_bias
+
+        output_total = output_total*max_prob
+        output_bias_total = output_bias_total*max_prob
+        output_total = output_total.view(b, s, h)
+        output_bias_total = output_bias_total.view(b, s, h)
+
+        return output_total, output_bias_total
+
+class ParallelAttention(MegatronModule):
+    """Parallel self-attention layer abstract class.
+
+    Self-attention layer takes input with size [b, s, h]
+    and returns output of the same size.
+    """
+
+    def __init__(self, init_method,
+                 output_layer_init_method, layer_number,
+                 attention_type=AttnType.self_attn,
+                 attn_mask_type=AttnMaskType.padding):
+        super(ParallelAttention, self).__init__()
+        args = get_args()
+        self.fp16 = args.fp16
+        self.bf16 = args.bf16
+
+        self.apply_query_key_layer_scaling = args.apply_query_key_layer_scaling
+        self.attention_softmax_in_fp32 = args.attention_softmax_in_fp32
+        if self.apply_query_key_layer_scaling:
+            self.attention_softmax_in_fp32 = True
+        self.layer_number = max(1, layer_number)
+        self.attention_type = attention_type
+        self.attn_mask_type = attn_mask_type
+        self.params_dtype = args.params_dtype
+
+        projection_size = args.kv_channels * args.num_attention_heads
+
+        # Per attention head and per partition values.
+        world_size = parallel_state.get_tensor_model_parallel_world_size()
+        self.hidden_size_per_partition = core_utils.divide(projection_size,
+                                                           world_size)
+        self.hidden_size_per_attention_head = core_utils.divide(
+            projection_size, args.num_attention_heads)
+        self.num_attention_heads_per_partition = core_utils.divide(
+            args.num_attention_heads, world_size)
+
+        # Strided linear layer.
+        if attention_type == AttnType.self_attn:
+            self.query_key_value = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                3 * projection_size,
+                gather_output=False,
+                init_method=init_method)
+        else:
+            assert attention_type == AttnType.cross_attn
+            self.query = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                projection_size,
+                gather_output=False,
+                init_method=init_method)
+
+            self.key_value = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                2 * projection_size,
+                gather_output=False,
+                init_method=init_method)
+
+        coeff = None
+        self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
+        if self.apply_query_key_layer_scaling:
+            coeff = self.layer_number
+            self.norm_factor *= coeff
+
+        self.scale_mask_softmax = FusedScaleMaskSoftmax(
+            self.fp16, self.bf16,
+            self.attn_mask_type,
+            args.masked_softmax_fusion,
+            attention_mask_func,
+            self.attention_softmax_in_fp32,
+            coeff)
+
+        # Dropout. Note that for a single iteration, this layer will generate
+        # different outputs on different number of parallel partitions but
+        # on average it should not be partition dependent.
+        self.attention_dropout = torch.nn.Dropout(args.attention_dropout)
+
+        # Output.
+        self.dense = tensor_parallel.RowParallelLinear(
+            projection_size,
+            args.hidden_size,
+            input_is_parallel=True,
+            init_method=output_layer_init_method,
+            skip_bias_add=True)
+
+    def _allocate_memory(self, inference_max_sequence_len, batch_size):
+        return torch.empty(
+            inference_max_sequence_len,
+            batch_size,
+            self.num_attention_heads_per_partition,
+            self.hidden_size_per_attention_head,
+            dtype=self.params_dtype,
+            device=torch.cuda.current_device())
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, inference_params=None):
+        # hidden_states: [sq, b, h]
+
+        # =================================================
+        # Pre-allocate memory for key-values for inference.
+        # =================================================
+        if inference_params:
+            if self.layer_number not in inference_params.key_value_memory_dict:
+                inf_max_seq_len = inference_params.max_sequence_len
+                inf_max_batch_size = inference_params.max_batch_size
+                inference_key_memory = self._allocate_memory(
+                    inf_max_seq_len, inf_max_batch_size)
+                inference_value_memory = self._allocate_memory(
+                    inf_max_seq_len, inf_max_batch_size)
+                inference_params.key_value_memory_dict[self.layer_number] = (
+                    inference_key_memory, inference_value_memory)
+            else:
+                inference_key_memory, inference_value_memory = \
+                    inference_params.key_value_memory_dict[self.layer_number]
+
+        # =====================
+        # Query, Key, and Value
+        # =====================
+
+        if self.attention_type == AttnType.self_attn:
+            # Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
+            mixed_x_layer, _ = self.query_key_value(hidden_states)
+
+            # [sq, b, (np * 3 * hn)] --> [sq, b, np, 3 * hn]
+            new_tensor_shape = mixed_x_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 3 * self.hidden_size_per_attention_head)
+            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
+
+            # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
+            (query_layer,
+             key_layer,
+             value_layer) = tensor_parallel \
+                 .split_tensor_along_last_dim(mixed_x_layer, 3)
+        else:
+            # Attention heads [sk, b, h] --> [sk, b, (np * 2 * hn)]
+            mixed_kv_layer, _ = self.key_value(encoder_output)
+
+            # [sk, b, (np * 2 * hn)] --> [sk, b, np, 2 * hn]
+            new_tensor_shape = mixed_kv_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 2 * self.hidden_size_per_attention_head)
+            mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape)
+
+            # [sk, b, np, 2 * hn] --> 2 [sk, b, np, hn]
+            (key_layer,
+             value_layer) = tensor_parallel \
+                 .split_tensor_along_last_dim(mixed_kv_layer, 2)
+
+            # Attention head [sq, b, h] --> [sq, b, hp]
+            query_layer, _ = self.query(hidden_states)
+            # [sq, b, hp] --> [sq, b, np, hn]
+            new_tensor_shape = query_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 self.hidden_size_per_attention_head)
+            query_layer = query_layer.view(*new_tensor_shape)
+
+        # ==================================
+        # Adjust key and value for inference
+        # ==================================
+
+        if inference_params:
+            batch_start = inference_params.batch_size_offset
+            batch_end = batch_start + key_layer.size(1)
+            assert batch_end <= inference_key_memory.size(1)
+            sequence_start = inference_params.sequence_len_offset
+            sequence_end = sequence_start + key_layer.size(0)
+            assert sequence_end <= inference_key_memory.size(0)
+            # Copy key and values.
+            inference_key_memory[sequence_start:sequence_end,
+                                 batch_start:batch_end, ...] = key_layer
+            inference_value_memory[sequence_start:sequence_end,
+                                   batch_start:batch_end, ...] = value_layer
+            key_layer = inference_key_memory[
+                :sequence_end, batch_start:batch_end, ...]
+            value_layer = inference_value_memory[
+                :sequence_end, batch_start:batch_end, ...]
+
+        # ===================================
+        # Raw attention scores. [b, np, s, s]
+        # ===================================
+
+        # [b, np, sq, sk]
+        output_size = (query_layer.size(1),
+                       query_layer.size(2),
+                       query_layer.size(0),
+                       key_layer.size(0))
+
+        # [sq, b, np, hn] -> [sq, b * np, hn]
+        query_layer = query_layer.view(output_size[2],
+                                       output_size[0] * output_size[1], -1)
+        # [sk, b, np, hn] -> [sk, b * np, hn]
+        key_layer = key_layer.view(output_size[3],
+                                   output_size[0] * output_size[1], -1)
+
+        # preallocting result tensor: [b * np, sq, sk]
+        matmul_result = torch.empty(
+            output_size[0]*output_size[1],
+            output_size[2],
+            output_size[3],
+            dtype=query_layer.dtype,
+            device=torch.cuda.current_device())
+
+        # Raw attention scores. [b * np, sq, sk]
+        matmul_result = torch.baddbmm(
+            matmul_result,
+            query_layer.transpose(0, 1),   # [b * np, sq, hn]
+            key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
+            beta=0.0, alpha=(1.0/self.norm_factor))
+
+        # change view to [b, np, sq, sk]
+        attention_scores = matmul_result.view(*output_size)
+
+        # ===========================
+        # Attention probs and dropout
+        # ===========================
+
+        # attention scores and attention mask [b, np, sq, sk]
+        attention_probs = self.scale_mask_softmax(attention_scores,
+                                                  attention_mask)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        with tensor_parallel.get_cuda_rng_tracker().fork():
+            attention_probs = self.attention_dropout(attention_probs)
+
+        # =========================
+        # Context layer. [sq, b, hp]
+        # =========================
+
+        # value_layer -> context layer.
+        # [sk, b, np, hn] --> [b, np, sq, hn]
+
+        # context layer shape: [b, np, sq, hn]
+        output_size = (value_layer.size(1),
+                       value_layer.size(2),
+                       query_layer.size(0),
+                       value_layer.size(3))
+
+        # change view [sk, b * np, hn]
+        value_layer = value_layer.view(value_layer.size(0),
+                                       output_size[0] * output_size[1], -1)
+
+        # change view [b * np, sq, sk]
+        attention_probs = attention_probs.view(output_size[0] * output_size[1],
+                                               output_size[2], -1)
+
+        # matmul: [b * np, sq, hn]
+        context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
+
+        # change view [b, np, sq, hn]
+        context_layer = context_layer.view(*output_size)
+
+        # [b, np, sq, hn] --> [sq, b, np, hn]
+        context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+        # [sq, b, np, hn] --> [sq, b, hp]
+        new_context_layer_shape = context_layer.size()[:-2] + \
+            (self.hidden_size_per_partition,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        # =================
+        # Output. [sq, b, h]
+        # =================
+
+        output, bias = self.dense(context_layer)
+
+        return output, bias
+
+
+def bias_dropout_add(x, bias, residual, prob, training):
+    # type: (Tensor, Tensor, Tensor, float, bool) -> Tensor
+    out = torch.nn.functional.dropout(x + bias, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+def get_bias_dropout_add(training):
+    def _bias_dropout_add(x, bias, residual, prob):
+        return bias_dropout_add(x, bias, residual, prob, training)
+    return _bias_dropout_add
+
+
+@torch.jit.script
+def bias_dropout_add_fused_train(x: torch.Tensor,
+                                 bias: torch.Tensor,
+                                 residual: torch.Tensor,
+                                 prob: float) -> torch.Tensor:
+    return bias_dropout_add(x, bias, residual, prob, True)
+
+
+@torch.jit.script
+def bias_dropout_add_fused_inference(x: torch.Tensor,
+                                     bias: torch.Tensor,
+                                     residual: torch.Tensor,
+                                     prob: float) -> torch.Tensor:
+    return bias_dropout_add(x, bias, residual, prob, False)
+
+
+class ParallelRetroTransformerEncoderLayer(MegatronModule):
+    """A single transformer layer for Retro Decoder with an retriever encoder inside and cross attention.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0., retriever=None):
+        args = get_args()
+
+        super(ParallelRetroTransformerEncoderLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        self.apply_residual_connection_post_layernorm \
+            = args.apply_residual_connection_post_layernorm
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
+        # Retro Encoder
+        self.retriever = retriever
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        self.inter_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.cross_attn)
+
+        # Layernorm on the attention output.
+        self.post_inter_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [s, b, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.  # [ns, bs, d]
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        """
+        notations:
+            l: number of chunks
+            m: number of token per chunk
+            bs: batch size
+            d: hidden size
+            k: number of neighbors
+            r: number of tokens per neighbors (neighbors + continuation)
+        """
+
+        args = get_args()
+        retro_args = get_retro_args()
+
+        chunk_length = retro_args.retro_gpt_chunk_length
+        retrieved_length = retro_args.retro_gpt_retrieved_length
+        num_neighbors = args.retro_num_neighbors
+
+        ns, bs, d = layernorm_output.shape
+        l = int(np.ceil(ns / chunk_length))
+        first_ns = ns % chunk_length
+        if first_ns > 0:
+            first_chunk, rest_chunk = \
+                layernorm_output[:first_ns], layernorm_output[first_ns:]
+            first_chunk = torch.nn.functional.pad(
+                first_chunk,
+                (0, 0, 0, 0, 0, chunk_length - first_ns),
+                'constant',
+                0)
+            chunked_output = \
+                torch.cat((first_chunk, rest_chunk), dim=0) # [l * m, bs, d]
+        else:
+            chunked_output = layernorm_output # [l * m, bs, d]
+        chunked_output = chunked_output \
+            .reshape(l, chunk_length, bs, d) \
+            .permute(1, 2, 0, 3) \
+            .reshape(chunk_length, bs * l, d) \
+            .contiguous()
+
+        # Get Encoder Output
+        retriever_output = self.retriever(
+            retriever_output,
+            retriever_attn_mask,
+            retriever_output=chunked_output,
+            retriever_attn_mask=retriever_attn_mask,
+            inference_params=inference_params) # [r, k * bs * l , d]
+        retriever_output = retriever_output.reshape(
+            retrieved_length * num_neighbors, bs * l, d) # [r * k, bs * l, d]
+
+        # Chunked Cross attention with Retriever Encoder
+        pad = (ns - 1) % chunk_length
+        attending_chunks = layernorm_output[pad:] # [ns - m + 1, bs, d]
+        padded_chunks = torch.nn.functional.pad(
+            attending_chunks,
+            (0, 0, 0, 0, 0, chunk_length-1),
+            'constant', 0) # [ns, bs, d]
+        padded_chunked_output = padded_chunks \
+            .reshape(l, chunk_length, bs, d) \
+            .permute(1, 2, 0, 3)
+        padded_chunked_output = padded_chunked_output.reshape(
+            chunk_length, bs * l, d).contiguous() # [m, bs * l, d]
+
+        # attention_output: [m, bs * l, d]
+        # attention_bias: [d]
+        attention_output, attention_bias = \
+            self.inter_attention(
+                padded_chunked_output, # Q: main model embedding
+                None,
+                encoder_output=retriever_output) # KV: retriever output embedding
+
+        # Residual connection
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        # Re-enable torch grad to enable fused optimization.
+        with torch.enable_grad():
+            layernorm_input = bias_dropout_add_func(
+                attention_output,
+                attention_bias.expand_as(attention_output),
+                torch.zeros_like(attention_output),
+                self.hidden_dropout)
+            layernorm_input = layernorm_input \
+                .reshape(chunk_length, bs, l, d) \
+                .permute(2, 0, 1, 3)  # [l, m, bs, d]
+            layernorm_input = layernorm_input.reshape(chunk_length * l, bs, d)
+            layernorm_input = torch.nn.functional.pad(
+                layernorm_input,
+                (0, 0, 0, 0, pad, 0),
+                'constant', 0)[:ns] # [ns, b, d]
+            layernorm_input = layernorm_input + residual
+
+        # Layer norm post the decoder attention
+        layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output, retriever_output
+
+
+class ParallelRetroTransformerLayer(MegatronModule):
+    """A single transformer layer for Retro Decoder with cross attention.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelRetroTransformerLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        self.apply_residual_connection_post_layernorm \
+            = args.apply_residual_connection_post_layernorm
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        self.inter_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.cross_attn)
+        # Layernorm on the attention output.
+        self.post_inter_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [b, s, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        args = get_args()
+        retro_args = get_retro_args()
+        chunk_length = retro_args.retro_gpt_chunk_length
+
+        ns, bs, d = layernorm_output.shape
+        l = int(np.ceil(ns / chunk_length))
+        pad = (ns - 1) % chunk_length
+        attending_chunks = layernorm_output[pad:]
+        padded_chunks = torch.nn.functional.pad(
+            attending_chunks,
+            (0, 0, 0, 0, 0, chunk_length - 1),
+            'constant', 0)
+        padded_chunked_output = padded_chunks \
+            .reshape(l, chunk_length, bs, d) \
+            .permute(1, 2, 0, 3)
+        padded_chunked_output = padded_chunked_output.reshape(
+            chunk_length, bs * l, d).contiguous()
+
+        # Encoder output.
+        attention_output, attention_bias = \
+            self.inter_attention(padded_chunked_output,
+                                 None,
+                                 encoder_output=retriever_output)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        # Re-enable torch grad to enable fused optimization.
+        with torch.enable_grad():
+            layernorm_input = bias_dropout_add_func(
+                attention_output,
+                attention_bias.expand_as(attention_output),
+                torch.zeros_like(attention_output),
+                self.hidden_dropout)
+            layernorm_input = layernorm_input \
+                .reshape(chunk_length, bs, l, d) \
+                .permute(2, 0, 1, 3) # [l, m, bs, d]
+            layernorm_input = layernorm_input.reshape(chunk_length * l, bs, d)
+            layernorm_input = torch.nn.functional.pad(
+                layernorm_input,
+                (0, 0, 0, 0, pad, 0),
+                'constant', 0)[:ns]
+            layernorm_input = layernorm_input + residual
+
+        # Layer norm post the decoder attention
+        layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class ParallelRetroEncoderTransformerCALayer(MegatronModule):
+    """A single transformer layer for Retro Encoder with cross attention.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelRetroEncoderTransformerCALayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        self.apply_residual_connection_post_layernorm \
+            = args.apply_residual_connection_post_layernorm
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.self_attention.attention_dropout = \
+            torch.nn.Dropout(args.retro_encoder_attention_dropout)
+        self.hidden_dropout = args.retro_encoder_hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        self.inter_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.cross_attn)
+        # Layernorm on the attention output.
+        self.post_inter_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [s, b, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        # Neighbors.
+        args = get_args()
+        retro_args = get_retro_args()
+
+        retrieved_length = retro_args.retro_gpt_retrieved_length
+        num_neighbors = args.retro_num_neighbors
+
+        ns, bs, d = layernorm_output.shape # [r, bs * l * k, d]
+        chunked_outputs = layernorm_output.reshape(retrieved_length, -1,
+                                                   num_neighbors, d)
+        chunked_outputs_before_layer_norm = \
+            layernorm_input.reshape(retrieved_length, -1,
+                                    num_neighbors, d) # [r, bs * l, k, d]
+
+        layernorm_inputs = []
+        layernorm_outputs = []
+        for k in range(num_neighbors):
+            chunked_output = chunked_outputs[:,:,k].contiguous()
+            attention_output, attention_bias = \
+                self.inter_attention(
+                    chunked_output, # Q (neighbor embedding)
+                    None,
+                    encoder_output=retriever_output) # K, V (hidden act)
+
+            # Residual connection.
+            if self.apply_residual_connection_post_layernorm:
+                residual = chunked_output
+            else:
+                residual = chunked_outputs_before_layer_norm[:,:,k]
+
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+
+                layernorm_inputs.append(layernorm_input)
+
+            # Layer norm post the decoder attention
+            layernorm_output = \
+                self.post_inter_attention_layernorm(layernorm_input)
+            layernorm_outputs.append(layernorm_output)
+
+        # layernorm_input : [r, k * bs * l, d]
+        # layernorm_output : [r, k * bs * l, d]
+        layernorm_input = \
+            torch.stack(layernorm_inputs, dim=1).reshape(ns, bs, d)
+        layernorm_output = \
+            torch.stack(layernorm_outputs, dim=1).reshape(ns, bs, d)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class ParallelTransformerLayer(MegatronModule):
+    """A single transformer layer.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelTransformerLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        self.apply_residual_connection_post_layernorm \
+            = args.apply_residual_connection_post_layernorm
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        if self.layer_type == LayerType.decoder:
+            self.inter_attention = ParallelAttention(
+                init_method,
+                output_layer_init_method,
+                layer_number,
+                attention_type=AttnType.cross_attn)
+            # Layernorm on the attention output.
+            self.post_inter_attention_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [b, s, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        if self.layer_type == LayerType.decoder:
+            attention_output, attention_bias = \
+                self.inter_attention(layernorm_output,
+                                     enc_dec_attn_mask,
+                                     encoder_output=encoder_output)
+            # residual connection
+            if self.apply_residual_connection_post_layernorm:
+                residual = layernorm_output
+            else:
+                residual = layernorm_input
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+
+            # Layer norm post the decoder attention
+            layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class ParallelRetroEncoder(MegatronModule):
+    """ Retro Transformer class for encoder ."""
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 pre_process=True, post_process=True,
+                 drop_path_rate=0.0):
+        super(ParallelRetroEncoder, self).__init__()
+        args = get_args()
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.input_tensor = None
+        self.drop_path_rate = drop_path_rate
+
+        # Store activation checkpoiting flag.
+        self.recompute_granularity = args.recompute_granularity
+        self.recompute_method = args.recompute_method
+        self.recompute_num_layers = args.recompute_num_layers
+        self.distribute_saved_activations = \
+            args.distribute_saved_activations and not args.sequence_parallel
+
+        self.sequence_parallel = args.sequence_parallel
+
+        # Number of layers.
+        self.num_layers = args.retro_encoder_layers
+
+        self.drop_path_rates = [rate.item() for rate in torch.linspace(0, self.drop_path_rate, args.num_layers)]
+
+        if args.retro_add_retriever:
+            self.P = [1]
+
+        # Transformer layers.
+        assert args.retro_add_retriever
+        def build_layer(layer_number):
+            if layer_number in self.P:
+                return ParallelRetroEncoderTransformerCALayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+            else:
+                layer = ParallelTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+                layer.self_attention.attention_dropout = \
+                    torch.nn.Dropout(args.retro_encoder_attention_dropout)
+                layer.hidden_dropout = args.retro_encoder_hidden_dropout
+                return layer
+
+        if args.virtual_pipeline_model_parallel_size is not None:
+            assert args.num_layers % args.virtual_pipeline_model_parallel_size == 0, \
+                'num_layers_per_stage must be divisible by ' \
+                'virtual_pipeline_model_parallel_size'
+            assert args.model_type != ModelType.encoder_and_decoder
+
+            # Number of layers in each model chunk is the number of layers in
+            # the stage, divided by the number of model chunks in a stage.
+            self.num_layers = self.num_layers // args.virtual_pipeline_model_parallel_size
+
+            # With 8 layers, 2 stages, and 4 model chunks, we want an
+            # assignment of layers to stages like (each list is a model chunk):
+            #   Stage 0: [0]  [2]  [4]  [6]
+            #   Stage 1: [1]  [3]  [5]  [7]
+            # With 8 layers, 2 stages, and 2 virtual stages, we want an
+            # assignment of layers to stages like (each list is a model chunk):
+            #   Stage 0: [0, 1]  [4, 5]
+            #   Stage 1: [2, 3]  [6, 7]
+            offset = parallel_state.get_virtual_pipeline_model_parallel_rank() * (
+                args.num_layers // args.virtual_pipeline_model_parallel_size) + \
+                (parallel_state.get_pipeline_model_parallel_rank() * self.num_layers)
+        else:
+            # Each stage gets a contiguous set of layers.
+            if args.model_type == ModelType.encoder_and_decoder and \
+                    parallel_state.get_pipeline_model_parallel_world_size() > 1:
+                pipeline_rank = parallel_state.get_pipeline_model_parallel_rank()
+                if layer_type == LayerType.encoder:
+                    offset = pipeline_rank * self.num_layers
+                else:
+                    num_ranks_in_enc = args.pipeline_model_parallel_split_rank
+                    offset = (pipeline_rank - num_ranks_in_enc) * self.num_layers
+            else:
+                offset = parallel_state.get_pipeline_model_parallel_rank() * self.num_layers
+
+        if self.num_layers == 0:
+            # When a standalone embedding stage is used (e.g.,
+            # args.standalone_embedding_stage == True), virtual pipeline ranks
+            # on pipeline rank 0 will have zero transformer layers assigned to
+            # them. This results in the model's input and output tensors to be
+            # the same, which will cause failure for certain output tensor
+            # optimizations (e.g., pipeline output deallocation). To remedy
+            # this, we assign a 'no-op' layer on these ranks, which will
+            # disconnect the input tensor from the output tensor.
+            self.num_layers = 1
+            self.layers = torch.nn.ModuleList([ NoopTransformerLayer(1) ])
+        else:
+            self.layers = torch.nn.ModuleList(
+                [build_layer(i + 1 + offset) for i in range(self.num_layers)])
+
+        if self.post_process:
+            # Final layer norm before output.
+            self.final_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm)
+
+    def _get_layer(self, layer_number):
+        return self.layers[layer_number]
+
+    def _checkpointed_forward(self, hidden_states, attention_mask,
+                              encoder_output, enc_dec_attn_mask):
+        """Forward method with activation checkpointing."""
+        def custom(start, end):
+            def custom_forward(*inputs):
+                x_ = inputs[0]
+                attention_mask = inputs[1]
+                encoder_output = inputs[2]
+                enc_dec_attn_mask = inputs[3]
+                for index in range(start, end):
+                    layer = self._get_layer(index)
+                    x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask)
+                return x_
+            return custom_forward
+
+        if self.activations_checkpoint_method == 'uniform':
+            # Uniformly divide the total number of Transformer layers and
+            # checkpoint the input activation of each divided chunk.
+            # A method to further reduce memory usage reducing checkpoints.
+            l = 0
+            while l < self.num_layers:
+                hidden_states = parallel_state.checkpoint(
+                    custom(l, l + self.activations_checkpoint_num_layers),
+                    self.distribute_checkpointed_activations,
+                    hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                l += self.activations_checkpoint_num_layers
+        elif self.activations_checkpoint_method == 'block':
+            # Checkpoint the input activation of only a set number of individual
+            # Transformer layers and skip the rest.
+            # A method fully use the device memory removing redundant re-computation.
+            for l in range(self.num_layers):
+                if l < self.activations_checkpoint_num_layers:
+                    hidden_states = parallel_state.checkpoint(
+                        custom(l, l + 1),
+                        self.distribute_checkpointed_activations,
+                        hidden_states, attention_mask,
+                        encoder_output, enc_dec_attn_mask)
+                else:
+                    hidden_states = custom(l, l + 1)(
+                        hidden_states, attention_mask,
+                        encoder_output, enc_dec_attn_mask)
+        else:
+            raise ValueError("Invalid activation checkpoint method.")
+
+        return hidden_states
+
+    def set_input_tensor(self, input_tensor):
+        """Set input tensor to be used instead of forward()'s input.
+
+        When doing pipeline parallelism the input from the previous
+        stage comes from communication, not from the input, so the
+        model's forward_step_func won't have it. This function is thus
+        used by internal code to bypass the input provided by the
+        forward_step_func"""
+        self.input_tensor = input_tensor
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+
+        # Checks.
+        if inference_params:
+            assert self.activations_checkpoint_method is None, \
+                'inference does not work with activation checkpointing'
+
+        if self.pre_process:
+            # Data format change to avoid explicit tranposes : [b s h] --> [s b h].
+            # If the input flag for fp32 residual connection is set, convert for float.
+            if self.fp32_residual_connection:
+                hidden_states = hidden_states.transpose(0, 1).contiguous().float()
+            # Otherwise, leave it as is.
+            else:
+                hidden_states = hidden_states.transpose(0, 1).contiguous()
+        else:
+            # See set_input_tensor()
+            hidden_states = self.input_tensor
+
+        # Viewless tensor.
+        # - We only need to create a viewless tensor in the case of micro batch
+        #   size (mbs) == 1, since in this case, 'hidden_states.transpose()'
+        #   above creates a view tensor, and '.contiguous()' is a pass-through.
+        #   For mbs >= 2, '.contiguous()' creates a new tensor, eliminating
+        #   the need to make it viewless.
+        #
+        #   However, we don't explicitly check mbs == 1 here because
+        #   make_viewless_tensor() has negligible overhead when its input
+        #   is already viewless.
+        #
+        # - For the 'else' case above, calling make_viewless_tensor() here is
+        #   likely redundant, since p2p_communication.py (likely originator)
+        #   already creates viewless tensors. That said, make_viewless_tensor()
+        #   is called here to be future-proof and corner-case-proof.
+        hidden_states = core_utils.make_viewless_tensor(
+            hidden_states,
+            requires_grad = True,
+            keep_graph = True,
+        )
+
+        # Transpose encoder output.
+        if encoder_output is not None:
+            encoder_output = encoder_output.transpose(0, 1).contiguous()
+
+        args = get_args()
+        assert not args.sequence_parallel, "if SP, need rng context."
+
+        # Forward pass.
+        if self.recompute_granularity == 'full':
+            hidden_states = self._checkpointed_forward(hidden_states,
+                                                       attention_mask,
+                                                       encoder_output,
+                                                       enc_dec_attn_mask)
+        else:
+            for index in range(self.num_layers):
+                layer = self._get_layer(index)
+                if index + 1 in self.P:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask,
+                        retriever_output=retriever_output,
+                        retriever_attn_mask=retriever_attn_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+                else:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+
+        # Final layer norm.
+        if self.post_process:
+            # Reverting data format change [s b h] --> [b s h].
+            hidden_states = hidden_states.transpose(0, 1).contiguous()
+            output = self.final_layernorm(hidden_states)
+        else:
+            output = hidden_states
+
+        return output
+
+
+class ParallelRetroTransformer(MegatronModule):
+    """Standard GPT Transformer class."""
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 pre_process=True, post_process=True,
+                 drop_path_rate=0.0, retriever=None):
+        super(ParallelRetroTransformer, self).__init__()
+        args = get_args()
+
+        assert pre_process and post_process, \
+            "pipeline parallelism un-supported."
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.input_tensor = None
+        self.drop_path_rate = drop_path_rate
+
+        # Store activation checkpoiting flag.
+        self.recompute_granularity = args.recompute_granularity
+        self.recompute_method = args.recompute_method
+        self.recompute_num_layers = args.recompute_num_layers
+        self.distribute_saved_activations = \
+            args.distribute_saved_activations and not args.sequence_parallel
+
+        self.sequence_parallel = args.sequence_parallel
+
+        # Number of layers.
+        self.num_layers = _get_num_layers(
+            args, args.model_type == ModelType.encoder_and_decoder)
+
+        self.drop_path_rates = [rate.item() for rate in torch.linspace(0, self.drop_path_rate, args.num_layers)]
+
+        if args.retro_add_retriever:
+            if args.num_layers == 12:
+                self.P = [6, 9, 12]
+            elif args.num_layers == 24:
+                self.P = np.arange(9, 25, 3).tolist()
+            elif args.num_layers == 40:
+                self.P = np.arange(9, 41, 3).tolist()
+                self.P.append(40)
+            self.retriever = retriever
+
+        # Transformer layers.
+        assert args.retro_add_retriever
+        def build_layer(layer_number):
+            if layer_number == min(self.P):
+                return ParallelRetroTransformerEncoderLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1],
+                    retriever=retriever
+                )
+            elif layer_number in self.P:
+                return ParallelRetroTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+            else:
+                return ParallelTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+
+        if args.virtual_pipeline_model_parallel_size is not None:
+            assert args.num_layers % args.virtual_pipeline_model_parallel_size == 0, \
+                'num_layers_per_stage must be divisible by ' \
+                'virtual_pipeline_model_parallel_size'
+            assert args.model_type != ModelType.encoder_and_decoder
+            # Number of layers in each model chunk is the number of layers in the stage,
+            # divided by the number of model chunks in a stage.
+            self.num_layers = self.num_layers // args.virtual_pipeline_model_parallel_size
+            # With 8 layers, 2 stages, and 4 model chunks, we want an assignment of
+            # layers to stages like (each list is a model chunk):
+            # Stage 0: [0]  [2]  [4]  [6]
+            # Stage 1: [1]  [3]  [5]  [7]
+            # With 8 layers, 2 stages, and 2 virtual stages, we want an assignment of
+            # layers to stages like (each list is a model chunk):
+            # Stage 0: [0, 1]  [4, 5]
+            # Stage 1: [2, 3]  [6, 7]
+            offset = parallel_state.get_virtual_pipeline_model_parallel_rank() * (
+                args.num_layers // args.virtual_pipeline_model_parallel_size) + \
+                (parallel_state.get_pipeline_model_parallel_rank() * self.num_layers)
+        else:
+            # Each stage gets a contiguous set of layers.
+            if args.model_type == ModelType.encoder_and_decoder and \
+                    parallel_state.get_pipeline_model_parallel_world_size() > 1:
+                pipeline_rank = parallel_state.get_pipeline_model_parallel_rank()
+                if layer_type == LayerType.encoder:
+                    offset = pipeline_rank * self.num_layers
+                else:
+                    num_ranks_in_enc = args.pipeline_model_parallel_split_rank
+                    offset = (pipeline_rank - num_ranks_in_enc) * self.num_layers
+            else:
+                offset = parallel_state.get_pipeline_model_parallel_rank() * self.num_layers
+
+        if self.num_layers == 0:
+            # When a standalone embedding stage is used (e.g.,
+            # args.standalone_embedding_stage == True), virtual pipeline ranks
+            # on pipeline rank 0 will have zero transformer layers assigned to
+            # them. This results in the model's input and output tensors to be
+            # the same, which will cause failure for certain output tensor
+            # optimizations (e.g., pipeline output deallocation). To remedy
+            # this, we assign a 'no-op' layer on these ranks, which will
+            # disconnect the input tensor from the output tensor.
+            self.num_layers = 1
+            self.layers = torch.nn.ModuleList([ NoopTransformerLayer(1) ])
+        else:
+            self.layers = torch.nn.ModuleList(
+                [build_layer(i + 1 + offset) for i in range(self.num_layers)])
+
+        if self.post_process:
+            # Final layer norm before output.
+            self.final_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm)
+
+    def _get_layer(self, layer_number):
+        return self.layers[layer_number]
+
+    def _checkpointed_forward(self, hidden_states, attention_mask,
+                              encoder_output, enc_dec_attn_mask):
+        """Forward method with activation checkpointing."""
+        def custom(start, end):
+            def custom_forward(*inputs):
+                x_ = inputs[0]
+                attention_mask = inputs[1]
+                encoder_output = inputs[2]
+                enc_dec_attn_mask = inputs[3]
+                for index in range(start, end):
+                    layer = self._get_layer(index)
+                    x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask)
+                return x_
+            return custom_forward
+
+        if self.activations_checkpoint_method == 'uniform':
+            # Uniformly divide the total number of Transformer layers and checkpoint
+            # the input activation of each divided chunk.
+            # A method to further reduce memory usage reducing checkpoints.
+            l = 0
+            while l < self.num_layers:
+                hidden_states = parallel_state.checkpoint(
+                    custom(l, l + self.activations_checkpoint_num_layers),
+                    self.distribute_checkpointed_activations,
+                    hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                l += self.activations_checkpoint_num_layers
+        elif self.activations_checkpoint_method == 'block':
+            # Checkpoint the input activation of only a set number of individual
+            # Transformer layers and skip the rest.
+            # A method fully use the device memory removing redundant re-computation.
+            for l in range(self.num_layers):
+                if l < self.activations_checkpoint_num_layers:
+                    hidden_states = parallel_state.checkpoint(
+                        custom(l, l + 1),
+                        self.distribute_checkpointed_activations,
+                        hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                else:
+                    hidden_states = custom(l, l + 1)(
+                        hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+        else:
+            raise ValueError("Invalid activation checkpoint method.")
+
+        return hidden_states
+
+    def set_input_tensor(self, input_tensor):
+        """Set input tensor to be used instead of forward()'s input.
+
+        When doing pipeline parallelism the input from the previous
+        stage comes from communication, not from the input, so the
+        model's forward_step_func won't have it. This function is thus
+        used by internal code to bypass the input provided by the
+        forward_step_func"""
+        self.input_tensor = input_tensor
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output=None, retriever_attn_mask=None,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+
+        # Checks.
+        if inference_params:
+            assert self.recompute_granularity is None, \
+                'inference does not work with activation checkpointing'
+
+        args = get_args()
+
+        # Transpose retriever output, to match hidden_states shape.
+        retriever_output = retriever_output.transpose(0, 1).contiguous()
+
+        # Viewless tensor.
+        # - We only need to create a viewless tensor in the case of micro batch
+        #   size (mbs) == 1, since in this case, 'hidden_states.transpose()'
+        #   above creates a view tensor, and '.contiguous()' is a pass-through.
+        #   For mbs >= 2, '.contiguous()' creates a new tensor, eliminating
+        #   the need to make it viewless.
+        #
+        #   However, we don't explicitly check mbs == 1 here because
+        #   make_viewless_tensor() has negligible overhead when its input
+        #   is already viewless.
+        #
+        # - For the 'else' case above, calling make_viewless_tensor() here is
+        #   likely redundant, since p2p_communication.py (likely originator)
+        #   already creates viewless tensors. That said, make_viewless_tensor()
+        #   is called here to be future-proof and corner-case-proof.
+        hidden_states = core_utils.make_viewless_tensor(
+            hidden_states,
+            requires_grad=True,
+            keep_graph=True,
+        )
+
+        # Transpose encoder output.
+        if encoder_output is not None:
+            encoder_output = encoder_output.transpose(0, 1).contiguous()
+
+        # Forward pass.
+        assert not args.sequence_parallel, "if SP, need rng context."
+        if self.recompute_granularity == 'full':
+            hidden_states = self._checkpointed_forward(hidden_states,
+                                                       attention_mask,
+                                                       encoder_output,
+                                                       enc_dec_attn_mask)
+        else:
+            for index in range(self.num_layers):
+                layer = self._get_layer(index)
+                if args.retro_add_retriever and index + 1 == min(self.P):
+                    hidden_states, E = layer(
+                        hidden_states,
+                        attention_mask,
+                        retriever_output=retriever_output,
+                        retriever_attn_mask=retriever_attn_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+                elif args.retro_add_retriever and index + 1 in self.P:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask,
+                        retriever_output=E,
+                        retriever_attn_mask=retriever_attn_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+                else:
+                    hidden_states = layer(
+                    hidden_states,
+                    attention_mask,
+                    encoder_output=encoder_output,
+                    enc_dec_attn_mask=enc_dec_attn_mask,
+                    inference_params=inference_params)
+
+        # Final layer norm.
+        output = self.final_layernorm(hidden_states)
+
+        return output

megatron/model/rotary_pos_embedding.py

+# coding=utf-8
+
+# The following code has been taken from https://github.com/NVIDIA/NeMo/blob/ \
+# 782b4e1652aaa43c8be390d9db0dc89544afa080/nemo/collections/nlp/modules/ \
+# common/megatron/rotary_pos_embedding.py
+
+import importlib.util
+import torch
+
+from torch import einsum, nn
+
+__all__ = ['RotaryEmbedding', 'apply_rotary_pos_emb']
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer('inv_freq', inv_freq)
+        if importlib.util.find_spec('einops') is None:
+            raise RuntimeError("einops is required for Rotary Embedding")
+
+    def forward(self, max_seq_len, offset=0):
+        seq = torch.arange(max_seq_len, device=self.inv_freq.device) + offset
+        freqs = einsum('i , j -> i j', seq.type_as(self.inv_freq), self.inv_freq)
+        # first part even vector components, second part odd vector components,
+        #  2 * dim in dimension size
+        emb = torch.cat((freqs, freqs), dim=-1)
+        # emb [seq_length, .., dim]
+        from einops import rearrange
+        return rearrange(emb, 'n d -> n 1 1 d')
+
+
+def _rotate_half(x):
+    """
+    change sign so the last dimension becomes [-odd, +even]
+    """
+    from einops import rearrange
+    x = rearrange(x, '... (j d) -> ... j d', j=2)
+    x1, x2 = x.unbind(dim=-2)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(t, freqs):
+    """
+    input tensor t is of shape [seq_length, ..., dim]
+    rotary positional embeding tensor freqs is of shape [seq_length, ..., dim]
+    check https://kexue.fm/archives/8265 for detailed formulas
+    """
+    rot_dim = freqs.shape[-1]
+    # ideally t_pass is empty so rotary pos embedding is applied to all tensor t
+    t, t_pass = t[..., :rot_dim], t[..., rot_dim:]
+
+    # first part is cosine component
+    # second part is sine component, need to change signs with _rotate_half method
+    t = (t * freqs.cos()) + (_rotate_half(t) * freqs.sin())
+    return torch.cat((t, t_pass), dim=-1)

megatron/model/t5_model.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""T5 model."""
+
+import torch
+
+from megatron import get_args
+from megatron.core import tensor_parallel
+from megatron.model.enums import AttnMaskType
+from megatron.model.language_model import parallel_lm_logits, get_language_model
+from megatron.model import LayerNorm
+from megatron.model.utils import (
+    openai_gelu,
+    get_linear_layer,
+    init_method_normal,
+    scaled_init_method_normal
+)
+from .module import MegatronModule
+
+
+def t5_extended_attention_mask(attention_mask_list):
+
+    def attn_mask_postprocess(attn_mask):
+        # [b, 1, s, s]
+        extended_attention_mask = attn_mask.unsqueeze(1)
+        return extended_attention_mask
+
+    return [attn_mask_postprocess(attn_mask) for attn_mask in attention_mask_list]
+
+
+def t5_position_ids(token_ids):
+    # Create position ids
+    seq_length = token_ids.size(1)
+    position_ids = torch.arange(seq_length, dtype=torch.long,
+                                device=token_ids.device)
+    position_ids = position_ids.unsqueeze(0).expand_as(token_ids)
+
+    return position_ids
+
+
+class T5LMHead(MegatronModule):
+    """Masked LM head for T5
+
+    Arguments:
+        mpu_vocab_size: model parallel size of vocabulary.
+        hidden_size: hidden size
+        init_method: init method for weight initialization
+        layernorm_epsilon: tolerance for layer norm divisions
+        parallel_output: wether output logits being distributed or not.
+    """
+
+    def __init__(self, mpu_vocab_size, parallel_output):
+        super(T5LMHead, self).__init__()
+
+        args = get_args()
+
+        self.bias = torch.nn.Parameter(torch.zeros(mpu_vocab_size))
+        self.bias.model_parallel = True
+        self.bias.partition_dim = 0
+        self.bias.stride = 1
+        self.parallel_output = parallel_output
+
+    def forward(self, hidden_states, word_embeddings_weight):
+        output = parallel_lm_logits(hidden_states,
+                                    word_embeddings_weight,
+                                    self.parallel_output,
+                                    bias=self.bias)
+        return output
+
+
+class T5Model(MegatronModule):
+    """T5 Language model."""
+
+    def __init__(self,
+                 num_tokentypes=0,
+                 parallel_output=True,
+                 pre_process=True,
+                 post_process=True,
+                 add_encoder=True,
+                 add_decoder=True):
+        super(T5Model, self).__init__()
+        args = get_args()
+
+        self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
+        self.parallel_output = parallel_output
+        init_method = init_method_normal(args.init_method_std)
+        scaled_init_method = scaled_init_method_normal(args.init_method_std,
+                                                       args.num_layers)
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.add_encoder = add_encoder
+        self.add_decoder = add_decoder
+
+        self.language_model, self._language_model_key = get_language_model(
+            num_tokentypes=num_tokentypes,
+            add_pooler=False,
+            add_encoder=add_encoder,
+            add_decoder=add_decoder,
+            encoder_attn_mask_type=AttnMaskType.padding,
+            init_method=init_method,
+            scaled_init_method=scaled_init_method,
+            pre_process=self.pre_process,
+            post_process=self.post_process)
+
+        self.initialize_word_embeddings(init_method_normal)
+
+        if self.post_process and self.add_decoder:
+            self.lm_head = T5LMHead(
+                self.word_embeddings_weight().size(0),
+                parallel_output)
+            self._lm_head_key = 'lm_head'
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.language_model.set_input_tensor(input_tensor)
+
+    def forward(self, encoder_input_ids, decoder_input_ids, encoder_attn_mask,
+                decoder_attn_mask, encoder_decoder_attn_mask,
+                tokentype_ids=None, lm_labels=None, enc_hidden_states=None):
+
+        # Converting the attention masks to proper parameter settings
+        encoder_attn_mask, decoder_attn_mask, encoder_decoder_attn_mask = t5_extended_attention_mask(
+            [encoder_attn_mask, decoder_attn_mask, encoder_decoder_attn_mask])
+
+        encoder_position_ids = t5_position_ids(encoder_input_ids)
+        decoder_position_ids = t5_position_ids(decoder_input_ids)
+
+        lm_output = self.language_model(encoder_input_ids,
+                                        encoder_position_ids,
+                                        encoder_attn_mask,
+                                        decoder_input_ids,
+                                        decoder_position_ids,
+                                        decoder_attn_mask,
+                                        encoder_decoder_attn_mask,
+                                        tokentype_ids=tokentype_ids,
+                                        enc_hidden_states=enc_hidden_states)
+
+        if self.post_process and self.add_decoder:
+            decoder_output, encoder_output = lm_output
+            # Output. [s, b, h]
+            lm_logits = self.lm_head(decoder_output,
+                                     self.word_embeddings_weight())
+
+            if lm_labels is None:
+                # [s b h] => [b s h]
+                return lm_logits.transpose(0,1).contiguous()
+            else:
+                # [b s] => [s b]
+                lm_labels = lm_labels.transpose(0,1).contiguous()
+                if self.fp16_lm_cross_entropy:
+                    assert lm_logits.dtype == torch.half
+                    lm_loss = tensor_parallel.vocab_parallel_cross_entropy(lm_logits, lm_labels)
+                else:
+                    lm_loss = tensor_parallel.vocab_parallel_cross_entropy(lm_logits.float(),
+                                                                                lm_labels)
+                # [s b] => [b s]
+                lm_loss = lm_loss.transpose(0,1).contiguous()
+            return lm_loss
+        elif self.add_decoder and not self.add_encoder:
+            decoder_output, encoder_output = lm_output
+            return decoder_output
+        else:
+            encoder_output = lm_output
+            return encoder_output
+
+    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):
+        """For easy load when model is combined with other heads,
+        add an extra key."""
+
+        state_dict_ = {}
+        state_dict_[self._language_model_key] \
+            = self.language_model.state_dict_for_save_checkpoint(prefix=prefix,
+                                                                 keep_vars=keep_vars)
+        if self.post_process and self.add_decoder:
+            state_dict_[self._lm_head_key] \
+                = self.lm_head.state_dict_for_save_checkpoint(prefix=prefix,
+                                                              keep_vars=keep_vars)
+         # Save word_embeddings.
+        if self.post_process and not self.pre_process and self.add_decoder:
+            state_dict_[self._word_embeddings_for_head_key] \
+                = self.word_embeddings.state_dict(prefix=prefix,
+                                                  keep_vars=keep_vars)
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+
+        self.language_model.load_state_dict(
+            state_dict[self._language_model_key], strict=strict)
+        if self.post_process and self.add_decoder:
+            self.lm_head.load_state_dict(state_dict[self._lm_head_key],
+                                         strict=strict)
+        # Load word embeddings.
+        if self.post_process and not self.pre_process and self.add_decoder:
+            self.word_embeddings.load_state_dict(
+                state_dict[self._word_embeddings_for_head_key], strict=strict)

megatron/model/transformer.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Transformer."""
+import math
+from contextlib import nullcontext
+import torch
+import torch.nn.functional as F
+from typing import Optional
+
+from megatron import get_timers, get_args, core, get_num_microbatches
+from .module import MegatronModule
+from megatron.core import mpu, tensor_parallel
+from megatron.core.enums import ModelType
+from megatron.model import LayerNorm
+from megatron.model.enums import AttnMaskType, LayerType, AttnType
+from megatron.model.fused_softmax import FusedScaleMaskSoftmax
+from megatron.model.fused_bias_gelu import bias_gelu_impl
+from megatron.model.rotary_pos_embedding import apply_rotary_pos_emb
+from megatron.model.utils import attention_mask_func, openai_gelu, erf_gelu
+
+try:
+    from einops import rearrange
+except ImportError:
+    rearrange = None
+
+try:
+    from flash_attn.flash_attn_interface import flash_attn_unpadded_func
+except ImportError:
+    flash_attn_unpadded_func = None
+
+""" We use the following notation throughout this file:
+     h: hidden size
+     n: number of attention heads
+     p: number of model parallel partitions
+     np: n/p
+     hp: h/p
+     hn: h/n
+     b: batch size
+     s: sequence length
+     l: number of layers
+    Transformer takes input of size [s, b, h] and returns a
+    tensor of the same size. We use the following arguments:
+        hyperparameters: transformer hyperparameters
+"""
+
+class DropPath(MegatronModule):
+    """Drop paths (Stochastic Depth) per sample
+    (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob=0.):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, hidden_state):
+        if self.drop_prob == 0. or not self.training:
+            return hidden_state
+        keep_prob = 1 - self.drop_prob
+        # work with diff dim tensors, not just 2D ConvNets
+        # hidden_state: [s, b, h]
+        shape = (1,) + (hidden_state.shape[1],) + (1,) * (hidden_state.ndim - 2)
+        random_tensor = keep_prob + \
+            torch.rand(shape, dtype=hidden_state.dtype, device=hidden_state.device)
+        random_tensor.floor_()  # binarize
+        output = hidden_state.div(keep_prob) * random_tensor
+        return output
+
+def _args_to_kwargs():
+    args = get_args()
+
+    common_kwargs = {
+        "params_dtype": args.params_dtype,
+        "use_cpu_initialization": args.use_cpu_initialization,
+        "perform_initialization": args.perform_initialization,
+        "gradient_accumulation_fusion": args.gradient_accumulation_fusion,
+        "sequence_parallel_enabled": args.sequence_parallel,
+    }
+    return common_kwargs
+
+class ParallelMLP(MegatronModule):
+    """MLP.
+
+    MLP will take the input with h hidden state, project it to 4*h
+    hidden dimension, perform nonlinear transformation, and project the
+    state back into h hidden dimension.
+    """
+
+    def __init__(self, init_method, output_layer_init_method):
+        super(ParallelMLP, self).__init__()
+        args = get_args()
+
+        self.add_bias = args.add_bias_linear
+
+        # Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
+        self.dense_h_to_4h = tensor_parallel.ColumnParallelLinear(
+            args.hidden_size,
+            args.ffn_hidden_size * 2 if args.swiglu else args.ffn_hidden_size,
+            bias=self.add_bias,
+            gather_output=False,
+            init_method=init_method,
+            skip_bias_add=True,
+            async_tensor_model_parallel_allreduce=args.async_tensor_model_parallel_allreduce,
+            **_args_to_kwargs())
+
+        self.bias_gelu_fusion = False
+        self.activation_func = None
+        self.swiglu = args.swiglu
+
+        if args.openai_gelu:
+            self.activation_func = openai_gelu
+        elif args.onnx_safe:
+            self.activation_func = erf_gelu
+        elif args.swiglu:
+            def swiglu(x):
+                x = torch.chunk(x, 2, dim=-1)
+                return F.silu(x[0]) * x[1]
+            self.activation_func = swiglu
+        elif args.squared_relu:
+            def squared_relu(x):
+                return torch.pow(F.relu(x), 2)
+            self.activation_func = squared_relu
+        else:
+            self.bias_gelu_fusion = args.bias_gelu_fusion
+            self.activation_func = F.gelu
+
+        # Project back to h.
+        self.dense_4h_to_h = tensor_parallel.RowParallelLinear(
+            args.ffn_hidden_size,
+            args.hidden_size,
+            bias=self.add_bias,
+            input_is_parallel=True,
+            init_method=output_layer_init_method,
+            skip_bias_add=True,
+            **_args_to_kwargs())
+
+    def forward(self, hidden_states):
+
+        # [s, b, 4hp]
+        intermediate_parallel, bias_parallel = self.dense_h_to_4h(hidden_states)
+
+        if self.bias_gelu_fusion:
+            assert self.add_bias is True
+            assert self.activation_func == F.gelu
+            intermediate_parallel = bias_gelu_impl(intermediate_parallel, bias_parallel)
+        else:
+            if bias_parallel is not None:
+                intermediate_parallel = intermediate_parallel + bias_parallel
+            intermediate_parallel = self.activation_func(intermediate_parallel)
+
+        # [s, b, h]
+        output, output_bias = self.dense_4h_to_h(intermediate_parallel)
+        return output, output_bias
+
+class SwitchMLP(MegatronModule):
+    """
+    Routes input to one of N MLP "experts"
+    """
+    def __init__(self, init_method, output_layer_init_method):
+        super(SwitchMLP, self).__init__()
+        args = get_args()
+        self.router = torch.nn.Linear(args.hidden_size, args.num_experts)
+        self.experts = torch.nn.ModuleList()
+        for i in range(args.num_experts):
+            self.experts.append(ParallelMLP(init_method, output_layer_init_method))
+
+    def forward(self, hidden_states):
+        # hidden_states: [s, b, h]
+        s = hidden_states.size(0)
+        b = hidden_states.size(1)
+        h = hidden_states.size(2)
+        route = self.router(hidden_states)
+        route = torch.nn.functional.softmax(route, dim=2)
+        max_prob, max_ind = torch.max(route, dim=2)
+        max_prob = torch.unsqueeze(max_prob, 2) # [s b 1]
+
+        # TODO (rprenger) TODO this could be made easier to read
+        # Converting [s, b, h] to [s*b, h].
+        # Each vector could be routed differently
+        hidden_states = hidden_states.view(-1, hidden_states.size(2)) # [s*b h]
+        max_prob = max_prob.view(-1, max_prob.size(2)) # [s*b 1]
+        max_ind = max_ind.view(-1) # [s*b]
+
+        output_total = torch.empty_like(hidden_states)
+        output_bias_total = torch.empty_like(hidden_states)
+        #TODO (rprenger) This does each expert in serial, but it could be parallelized
+
+        for expert_num, expert in enumerate(self.experts):
+            local_indices = (max_ind == expert_num).nonzero()
+            hidden = hidden_states[local_indices,:]
+            output, output_bias = expert(hidden)
+            output_bias = output_bias.expand_as(output)
+            output_total[local_indices,:] = output
+            output_bias_total[local_indices,:] = output_bias
+
+        output_total = output_total*max_prob
+        output_bias_total = output_bias_total*max_prob
+        output_total = output_total.view(s, b, h)
+        output_bias_total = output_bias_total.view(s, b, h)
+
+        return output_total, output_bias_total
+
+
+class CoreAttention(MegatronModule):
+
+    def __init__(self, layer_number,
+                 attn_mask_type=AttnMaskType.padding):
+        super(CoreAttention, self).__init__()
+        args = get_args()
+        self.fp16 = args.fp16
+        self.bf16 = args.bf16
+
+        self.apply_query_key_layer_scaling = args.apply_query_key_layer_scaling
+        self.attention_softmax_in_fp32 = args.attention_softmax_in_fp32
+        if self.apply_query_key_layer_scaling:
+            self.attention_softmax_in_fp32 = True
+        self.layer_number = max(1, layer_number)
+        self.attn_mask_type = attn_mask_type
+        self.sequence_parallel = args.sequence_parallel
+
+        projection_size = args.kv_channels * args.num_attention_heads
+
+        # Per attention head and per partition values.
+        world_size = mpu.get_tensor_model_parallel_world_size()
+        self.hidden_size_per_partition = core.utils.divide(projection_size,
+                                                           world_size)
+        self.hidden_size_per_attention_head = core.utils.divide(
+            projection_size, args.num_attention_heads)
+        self.num_attention_heads_per_partition = core.utils.divide(
+            args.num_attention_heads, world_size)
+
+        coeff = None
+        self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
+        if self.apply_query_key_layer_scaling:
+            coeff = self.layer_number
+            self.norm_factor *= coeff
+
+        self.scale_mask_softmax = FusedScaleMaskSoftmax(
+            self.fp16, self.bf16,
+            self.attn_mask_type,
+            args.masked_softmax_fusion,
+            attention_mask_func,
+            self.attention_softmax_in_fp32,
+            coeff)
+
+        # Dropout. Note that for a single iteration, this layer will generate
+        # different outputs on different number of parallel partitions but
+        # on average it should not be partition dependent.
+        self.attention_dropout = torch.nn.Dropout(args.attention_dropout)
+
+    def forward(self, query_layer, key_layer,
+                value_layer, attention_mask):
+
+        # ===================================
+        # Raw attention scores. [b, np, s, s]
+        # ===================================
+
+        # [b, np, sq, sk]
+        output_size = (query_layer.size(1),
+                       query_layer.size(2),
+                       query_layer.size(0),
+                       key_layer.size(0))
+
+        # [sq, b, np, hn] -> [sq, b * np, hn]
+        query_layer = query_layer.view(output_size[2],
+                                       output_size[0] * output_size[1], -1)
+        # [sk, b, np, hn] -> [sk, b * np, hn]
+        key_layer = key_layer.view(output_size[3],
+                                   output_size[0] * output_size[1], -1)
+
+        # preallocting input tensor: [b * np, sq, sk]
+        matmul_input_buffer = mpu.get_global_memory_buffer().get_tensor(
+            (output_size[0]*output_size[1], output_size[2], output_size[3]),
+            query_layer.dtype, "mpu")
+
+        # Raw attention scores. [b * np, sq, sk]
+        matmul_result = torch.baddbmm(
+            matmul_input_buffer,
+            query_layer.transpose(0, 1),   # [b * np, sq, hn]
+            key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
+            beta=0.0, alpha=(1.0/self.norm_factor))
+
+        # change view to [b, np, sq, sk]
+        attention_scores = matmul_result.view(*output_size)
+
+        # ===========================
+        # Attention probs and dropout
+        # ===========================
+
+        # attention scores and attention mask [b, np, sq, sk]
+        attention_probs = self.scale_mask_softmax(attention_scores,
+                                                  attention_mask)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        if not self.sequence_parallel:
+            with tensor_parallel.get_cuda_rng_tracker().fork():
+                attention_probs = self.attention_dropout(attention_probs)
+        else:
+            attention_probs = self.attention_dropout(attention_probs)
+
+        # =========================
+        # Context layer. [sq, b, hp]
+        # =========================
+
+        # value_layer -> context layer.
+        # [sk, b, np, hn] --> [b, np, sq, hn]
+
+        # context layer shape: [b, np, sq, hn]
+        output_size = (value_layer.size(1),
+                       value_layer.size(2),
+                       query_layer.size(0),
+                       value_layer.size(3))
+
+        # change view [sk, b * np, hn]
+        value_layer = value_layer.view(value_layer.size(0),
+                                       output_size[0] * output_size[1], -1)
+
+        # change view [b * np, sq, sk]
+        attention_probs = attention_probs.view(output_size[0] * output_size[1],
+                                               output_size[2], -1)
+
+        # matmul: [b * np, sq, hn]
+        context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
+
+        # change view [b, np, sq, hn]
+        context_layer = context_layer.view(*output_size)
+
+        # [b, np, sq, hn] --> [sq, b, np, hn]
+        context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+        # [sq, b, np, hn] --> [sq, b, hp]
+        new_context_layer_shape = context_layer.size()[:-2] + \
+            (self.hidden_size_per_partition,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        return context_layer
+
+
+class FlashSelfAttention(torch.nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0,
+                 device=None, dtype=None):
+        super().__init__()
+        assert flash_attn_unpadded_func is not None, ('Please install FlashAttention first, '
+                                                      'e.g., with pip install flash-attn')
+        assert rearrange is not None, 'Please install einops first, e.g., with pip install einops'
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout_p = attention_dropout
+
+    def forward(self, q, k, v):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q, k, v: The tensor containing the query, key, and value. (B, S, H, D)
+        """
+
+        assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q,k,v)))
+        assert all((i.is_cuda for i in (q,k,v)))
+
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        seqlen_k = k.shape[1]
+
+        q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
+        cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32,
+                                    device=q.device)
+
+        if self.training:
+            # during training q,k,v always have same seqlen
+            assert seqlen_k == seqlen_q
+
+            is_causal = self.causal
+            cu_seqlens_k = cu_seqlens_q
+        else:
+            # turn off FA causal mask after first inference autoregressive iteration
+            # only on first autoregressive step q,k,v have same seqlen
+            is_causal = seqlen_q == seqlen_k
+            cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32,
+                        device=q.device)
+            self.dropout_p = 0
+
+        output = flash_attn_unpadded_func(
+            q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k,
+            self.dropout_p,
+            softmax_scale=self.softmax_scale, causal=is_causal
+        )
+
+        output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
+        return output
+
+
+class ParallelAttention(MegatronModule):
+    """Parallel self-attention layer abstract class.
+
+    Self-attention layer takes input with size [s, b, h]
+    and returns output of the same size.
+    """
+
+    def __init__(self, init_method,
+                 output_layer_init_method, layer_number,
+                 attention_type=AttnType.self_attn,
+                 attn_mask_type=AttnMaskType.padding):
+        super(ParallelAttention, self).__init__()
+        args = get_args()
+        self.layer_number = max(1, layer_number)
+        self.attention_type = attention_type
+        self.attn_mask_type = attn_mask_type
+        self.params_dtype = args.params_dtype
+        self.sequence_parallel = args.sequence_parallel
+
+        self.use_flash_attn = args.use_flash_attn
+        if self.use_flash_attn:
+            if flash_attn_unpadded_func is None:
+                raise ImportError('FlashAttention is not installed, please install with '
+                                  'pip install flash-attn')
+            assert attention_type == AttnType.self_attn, ('FlashAttention code path only supports '
+                                                          'self-attention for now')
+            assert self.attn_mask_type == AttnMaskType.causal, ('FlashAttention code path only '
+                                                                'supports causal mask for now')
+            if rearrange is None:
+                raise ImportError('einops is not installed, please install with pip install einops')
+
+        projection_size = args.kv_channels * args.num_attention_heads
+
+        # Per attention head and per partition values.
+        world_size = mpu.get_tensor_model_parallel_world_size()
+        self.hidden_size_per_attention_head = core.utils.divide(
+            projection_size, args.num_attention_heads)
+        self.num_attention_heads_per_partition = core.utils.divide(
+            args.num_attention_heads, world_size)
+
+        # Strided linear layer.
+        if attention_type == AttnType.self_attn:
+            self.query_key_value = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                3 * projection_size,
+                bias=args.add_bias_linear,
+                gather_output=False,
+                init_method=init_method,
+                async_tensor_model_parallel_allreduce=args.async_tensor_model_parallel_allreduce,
+                **_args_to_kwargs())
+        else:
+            assert attention_type == AttnType.cross_attn
+            self.query = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                projection_size,
+                bias=args.add_bias_linear,
+                gather_output=False,
+                init_method=init_method,
+                async_tensor_model_parallel_allreduce=args.async_tensor_model_parallel_allreduce,
+                **_args_to_kwargs())
+
+
+            self.key_value = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                2 * projection_size,
+                bias=args.add_bias_linear,
+                gather_output=False,
+                init_method=init_method,
+                async_tensor_model_parallel_allreduce=args.async_tensor_model_parallel_allreduce,
+                **_args_to_kwargs())
+
+        self.core_attention = CoreAttention(self.layer_number,
+                                            self.attn_mask_type)
+        self.checkpoint_core_attention = args.recompute_granularity == 'selective'
+
+        if self.use_flash_attn:
+            self.core_attention_flash = FlashSelfAttention(
+                causal=True, attention_dropout=args.attention_dropout
+            )
+
+        # Output.
+        self.dense = tensor_parallel.RowParallelLinear(
+            projection_size,
+            args.hidden_size,
+            bias=args.add_bias_linear,
+            input_is_parallel=True,
+            init_method=output_layer_init_method,
+            skip_bias_add=True,
+            **_args_to_kwargs())
+
+    def _checkpointed_attention_forward(self, query_layer, key_layer,
+                                        value_layer, attention_mask,
+                                        rotary_pos_emb=None):
+        """Forward method with activation checkpointing."""
+        def custom_forward(*inputs):
+            query_layer = inputs[0]
+            key_layer = inputs[1]
+            value_layer = inputs[2]
+            attention_mask = inputs[3]
+            output_ = self.core_attention(query_layer, key_layer,
+                                          value_layer, attention_mask)
+            return output_
+
+        q_pos_emb, k_pos_emb = (None, None) if rotary_pos_emb is None \
+            else rotary_pos_emb
+
+        hidden_states = tensor_parallel.checkpoint(
+            custom_forward,
+            False, query_layer, key_layer, value_layer, attention_mask,
+            q_pos_emb, k_pos_emb)
+
+        return hidden_states
+
+    def _allocate_memory(self, inference_max_sequence_len, batch_size):
+        return torch.empty(
+            inference_max_sequence_len,
+            batch_size,
+            self.num_attention_heads_per_partition,
+            self.hidden_size_per_attention_head,
+            dtype=self.params_dtype,
+            device=torch.cuda.current_device())
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, inference_params=None,
+                rotary_pos_emb=None):
+        # hidden_states: [sq, b, h]
+
+        # =================================================
+        # Pre-allocate memory for key-values for inference.
+        # =================================================
+        is_first_step = False
+        if inference_params:
+            if self.layer_number not in inference_params.key_value_memory_dict:
+                inf_max_seq_len = inference_params.max_sequence_len
+                inf_max_batch_size = inference_params.max_batch_size
+                inference_key_memory = self._allocate_memory(
+                    inf_max_seq_len, inf_max_batch_size)
+                inference_value_memory = self._allocate_memory(
+                    inf_max_seq_len, inf_max_batch_size)
+                inference_params.key_value_memory_dict[self.layer_number] = (
+                    inference_key_memory, inference_value_memory)
+                is_first_step = True
+            else:
+                inference_key_memory, inference_value_memory = \
+                    inference_params.key_value_memory_dict[self.layer_number]
+
+        # =====================
+        # Query, Key, and Value
+        # =====================
+
+        if self.attention_type == AttnType.self_attn:
+            # Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
+            mixed_x_layer, _ = self.query_key_value(hidden_states)
+
+            # [sq, b, (np * 3 * hn)] --> [sq, b, np, 3 * hn]
+            new_tensor_shape = mixed_x_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 3 * self.hidden_size_per_attention_head)
+            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
+
+            # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
+            (query_layer,
+             key_layer,
+             value_layer) = tensor_parallel.split_tensor_along_last_dim(mixed_x_layer, 3)
+        else:
+            # Attention heads [sk, b, h] --> [sk, b, (np * 2 * hn)]
+            mixed_kv_layer, _ = self.key_value(encoder_output)
+
+            # [sk, b, (np * 2 * hn)] --> [sk, b, np, 2 * hn]
+            new_tensor_shape = mixed_kv_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 2 * self.hidden_size_per_attention_head)
+            mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape)
+
+            # [sk, b, np, 2 * hn] --> 2 [sk, b, np, hn]
+            (key_layer,
+             value_layer) = tensor_parallel.split_tensor_along_last_dim(mixed_kv_layer, 2)
+
+            # Attention head [sq, b, h] --> [sq, b, hp]
+            query_layer, _ = self.query(hidden_states)
+            # [sq, b, hp] --> [sq, b, np, hn]
+            new_tensor_shape = query_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 self.hidden_size_per_attention_head)
+            query_layer = query_layer.view(*new_tensor_shape)
+
+        # ==================================
+        # Adjust key and value for inference
+        # ==================================
+
+        # duplicate the pos_emb for self attention
+        if rotary_pos_emb is not None:
+            if isinstance(rotary_pos_emb, tuple):
+                rotary_pos_emb = rotary_pos_emb
+            else:
+                rotary_pos_emb = ((rotary_pos_emb,) * 2)
+
+        if inference_params:
+            batch_start = inference_params.batch_size_offset
+            batch_end = batch_start + key_layer.size(1)
+            assert batch_end <= inference_key_memory.size(1)
+            sequence_start = inference_params.sequence_len_offset
+            sequence_end = sequence_start + key_layer.size(0)
+            assert sequence_end <= inference_key_memory.size(0)
+            # Copy key and values.
+            inference_key_memory[sequence_start:sequence_end,
+                                 batch_start:batch_end, ...] = key_layer
+            inference_value_memory[sequence_start:sequence_end,
+                                   batch_start:batch_end, ...] = value_layer
+            key_layer = inference_key_memory[
+                :sequence_end, batch_start:batch_end, ...]
+            value_layer = inference_value_memory[
+                :sequence_end, batch_start:batch_end, ...]
+
+
+            # adjust the key rotary positional embedding
+            if rotary_pos_emb is not None:
+                q_pos_emb, k_pos_emb = rotary_pos_emb
+                # need to cross check this condition during inference
+                # if not set_inference_key_value_memory:
+                if not is_first_step:
+                    # In inference, we compute one token at a time.
+                    # Select the correct positional embedding
+                    # (only the last token in the sequence)
+                    q_pos_emb = q_pos_emb[sequence_end - 1 : sequence_end]
+                else:
+                    # In the first forward pass of inference,
+                    # we use the entire provided prefix.
+                    # q_pos_emb here has the rope embeddings of the entire
+                    # prefix + to-be-generated output so
+                    # we slice to just the prefix.
+                    q_pos_emb = q_pos_emb[:sequence_end, :, :, :]
+                k_pos_emb = k_pos_emb[:sequence_end, :, :, :]
+                rotary_pos_emb = (q_pos_emb, k_pos_emb)
+
+
+        # ==================================
+        # core attention computation
+        # ==================================
+
+        # apply relative positional encoding (rotary embedding)
+        if rotary_pos_emb is not None:
+            q_pos_emb, k_pos_emb = rotary_pos_emb
+            query_layer = apply_rotary_pos_emb(query_layer, q_pos_emb)
+            key_layer = apply_rotary_pos_emb(key_layer, k_pos_emb)
+            # TODO, can apply positional embedding to value_layer so it has
+            # absolute positional embedding.
+            # otherwise, only relative positional embedding takes effect
+            # value_layer = apply_rotary_pos_emb(value_layer, k_pos_emb)
+
+        if not self.use_flash_attn:
+            if self.checkpoint_core_attention:
+                context_layer = self._checkpointed_attention_forward(
+                    query_layer, key_layer, value_layer, attention_mask)
+            else:
+                context_layer = self.core_attention(
+                    query_layer, key_layer, value_layer, attention_mask)
+        else:
+            q, k, v = [rearrange(x, 's b ... -> b s ...').contiguous()
+                       for x in (query_layer, key_layer, value_layer)]
+            if not self.sequence_parallel:
+                with tensor_parallel.get_cuda_rng_tracker().fork():
+                    context_layer = self.core_attention_flash(q, k, v)
+            else:
+                context_layer = self.core_attention_flash(q, k, v)
+            context_layer = rearrange(context_layer, 'b s h d -> s b (h d)').contiguous()
+
+        # =================
+        # Output. [sq, b, h]
+        # =================
+
+        output, bias = self.dense(context_layer)
+
+        return output, bias
+
+
+def bias_dropout_add(x, bias, residual, prob, training):
+    # type: (Tensor, Optional[Tensor], Tensor, float, bool) -> Tensor
+    if bias is not None:
+        x = x + bias
+    out = torch.nn.functional.dropout(x, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+def get_bias_dropout_add(training):
+    def _bias_dropout_add(x, bias, residual, prob):
+        return bias_dropout_add(x, bias, residual, prob, training)
+    return _bias_dropout_add
+
+
+@torch.jit.script
+def bias_dropout_add_fused_train(x: torch.Tensor,
+                                 bias: Optional[torch.Tensor],
+                                 residual: torch.Tensor,
+                                 prob: float) -> torch.Tensor:
+    return bias_dropout_add(x, bias, residual, prob, True)
+
+
+@torch.jit.script
+def bias_dropout_add_fused_inference(x: torch.Tensor,
+                                     bias: Optional[torch.Tensor],
+                                     residual: torch.Tensor,
+                                     prob: float) -> torch.Tensor:
+    return bias_dropout_add(x, bias, residual, prob, False)
+
+
+class ParallelTransformerLayer(MegatronModule):
+    """A single transformer layer.
+
+    Transformer layer takes input with size [s, b, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelTransformerLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        self.apply_residual_connection_post_layernorm \
+            = args.apply_residual_connection_post_layernorm
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm,
+            sequence_parallel=args.sequence_parallel,
+            apply_layernorm_1p=args.apply_layernorm_1p)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm,
+            sequence_parallel=args.sequence_parallel,
+            apply_layernorm_1p=args.apply_layernorm_1p)
+
+        if self.layer_type == LayerType.decoder:
+            self.inter_attention = ParallelAttention(
+                init_method,
+                output_layer_init_method,
+                layer_number,
+                attention_type=AttnType.cross_attn)
+            # Layernorm on the attention output.
+            self.post_inter_attention_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm,
+                sequence_parallel=args.sequence_parallel,
+                apply_layernorm_1p=args.apply_layernorm_1p)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+        # Set bias+dropout+add fusion grad_enable execution handler.
+        TORCH_MAJOR = int(torch.__version__.split('.')[0])
+        TORCH_MINOR = int(torch.__version__.split('.')[1])
+        use_nvfuser = TORCH_MAJOR > 1 or (TORCH_MAJOR == 1 and TORCH_MINOR >= 10)
+        self.bias_dropout_add_exec_handler = \
+                nullcontext if use_nvfuser else torch.enable_grad
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None, rotary_pos_emb=None):
+        # hidden_states: [s, b, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params,
+                rotary_pos_emb=rotary_pos_emb)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            if attention_bias is not None:
+                attention_bias = attention_bias.expand_as(residual)
+            with self.bias_dropout_add_exec_handler():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias,
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        if self.layer_type == LayerType.decoder:
+            attention_output, attention_bias = \
+                self.inter_attention(layernorm_output,
+                                     enc_dec_attn_mask,
+                                     encoder_output=encoder_output)
+            # residual connection
+            if self.apply_residual_connection_post_layernorm:
+                residual = layernorm_output
+            else:
+                residual = layernorm_input
+
+            if attention_bias is not None:
+                attention_bias = attention_bias.expand_as(residual)
+
+            with self.bias_dropout_add_exec_handler():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias,
+                    residual,
+                    self.hidden_dropout)
+
+            # Layer norm post the decoder attention
+            layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            if mlp_bias is not None:
+                mlp_bias = mlp_bias.expand_as(residual)
+            with self.bias_dropout_add_exec_handler():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias,
+                    residual,
+                    self.hidden_dropout)
+
+            # Jit compiled function creates 'view' tensor. This tensor
+            # potentially gets saved in the MPU checkpoint function context,
+            # which rejects view tensors. While making a viewless tensor here
+            # won't result in memory savings (like the data loader, or
+            # p2p_communication), it serves to document the origin of this
+            # 'view' tensor.
+            output = core.utils.make_viewless_tensor(inp = output,
+                                                     requires_grad = output.requires_grad,
+                                                     keep_graph = True)
+
+        else:
+            if mlp_bias is not None:
+                mlp_output = mlp_output + mlp_bias
+            out = torch.nn.functional.dropout(mlp_output,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class NoopTransformerLayer(MegatronModule):
+    """A single 'no-op' transformer layer.
+
+    The sole purpose of this layer is for when a standalone embedding layer
+    is used (i.e., args.standalone_embedding_stage == True). In this case,
+    zero transformer layers are assigned when pipeline rank == 0. Additionally,
+    when virtual pipeline rank >= 1, zero total model parameters are created
+    (virtual rank 0 contains the input embedding). This results in the model's
+    input and output tensors being the same, which causes an error when
+    performing certain memory optimiations on the output tensor (e.g.,
+    deallocating it). Thus, this layer disconnects the input from the output
+    via a clone. Since ranks containing a no-op layer are generally under-
+    utilized (both compute and memory), there's no worry of any performance
+    degredation.
+    """
+
+    def __init__(self, layer_number):
+        super().__init__()
+        self.layer_number = layer_number
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        return hidden_states.clone()
+
+
+def _get_num_layers(args, is_encoder_and_decoder_model, is_decoder=False):
+    """Compute the number of transformer layers resident on the current rank."""
+    if mpu.get_pipeline_model_parallel_world_size() > 1:
+        if is_encoder_and_decoder_model:
+            assert args.pipeline_model_parallel_split_rank is not None
+
+            # When a standalone embedding stage is used, a rank is taken from
+            # the encoder's ranks, to be used for the encoder's embedding
+            # layer. This way, the rank referenced by the 'split rank' remains
+            # the same whether or not a standalone embedding stage is used.
+            num_ranks_in_encoder = (
+                args.pipeline_model_parallel_split_rank - 1
+                if args.standalone_embedding_stage else
+                args.pipeline_model_parallel_split_rank
+            )
+            num_ranks_in_decoder = args.transformer_pipeline_model_parallel_size - num_ranks_in_encoder
+            assert args.encoder_num_layers % num_ranks_in_encoder == 0, \
+                    'encoder_num_layers (%d) must be divisible by number of ranks given to encoder (%d)' % (args.encoder_num_layers, num_ranks_in_encoder)
+            assert args.decoder_num_layers % num_ranks_in_decoder == 0, \
+                    'decoder_num_layers (%d) must be divisible by number of ranks given to decoder (%d)' % (args.decoder_num_layers, num_ranks_in_decoder)
+            if mpu.is_pipeline_stage_before_split():
+                num_layers = (
+                    0
+                    if args.standalone_embedding_stage
+                    and mpu.get_pipeline_model_parallel_rank() == 0 else
+                    args.encoder_num_layers // num_ranks_in_encoder
+                )
+            else:
+                num_layers = args.decoder_num_layers // num_ranks_in_decoder
+        else:
+            assert args.num_layers == args.encoder_num_layers
+            assert args.num_layers % args.transformer_pipeline_model_parallel_size == 0, \
+                'num_layers must be divisible by transformer_pipeline_model_parallel_size'
+
+            # When a standalone embedding stage is used, all transformer layers
+            # are divided among pipeline rank >= 1, while on pipeline rank 0,
+            # ranks either contain the input embedding layer (virtual pp rank 0),
+            # or no layers at all (virtual pp rank >= 1).
+            num_layers = (
+                0
+                if args.standalone_embedding_stage
+                and mpu.get_pipeline_model_parallel_rank() == 0 else
+                args.num_layers // args.transformer_pipeline_model_parallel_size
+            )
+    else:
+        if not is_decoder:
+            num_layers = args.encoder_num_layers
+        else:
+            num_layers = args.decoder_num_layers
+    return num_layers
+
+
+class ParallelTransformer(MegatronModule):
+    """Transformer class."""
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 post_layer_norm=True,
+                 pre_process=True, post_process=True,
+                 drop_path_rate=0.0):
+        super(ParallelTransformer, self).__init__()
+        args = get_args()
+
+        self.layer_type = layer_type
+        self.model_type = args.model_type
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+        self.post_layer_norm = post_layer_norm
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.input_tensor = None
+        self.drop_path_rate = drop_path_rate
+        self.transformer_impl = args.transformer_impl
+
+        # Store activation checkpoiting flag.
+        self.recompute_granularity = args.recompute_granularity
+        self.recompute_method = args.recompute_method
+        self.recompute_num_layers = args.recompute_num_layers
+        self.distribute_saved_activations = \
+            args.distribute_saved_activations and not args.sequence_parallel
+
+        self.sequence_parallel = args.sequence_parallel
+
+        # Transformer Engine Init.
+        if self.transformer_impl == 'transformer_engine':
+            global transformer_engine
+            import transformer_engine
+        self.use_fp8 = args.fp8_e4m3 or args.fp8_hybrid
+        self.fp8_recipe = None
+        self.fp8_group = None
+        if self.use_fp8:
+            self.fp8_group = mpu.get_data_parallel_group()
+            if args.fp8_e4m3:
+                fp8_format = transformer_engine.common.recipe.Format.E4M3
+            elif args.fp8_hybrid:
+                fp8_format = transformer_engine.common.recipe.Format.HYBRID
+            self.fp8_recipe = transformer_engine.common.recipe.DelayedScaling(
+                margin=args.fp8_margin,
+                interval=args.fp8_interval,
+                fp8_format=fp8_format,
+                amax_history_len=args.fp8_amax_history_len,
+                amax_compute_algo=args.fp8_amax_compute_algo,
+                override_linear_precision=(False, False, not args.fp8_wgrad),
+            )
+
+        self.num_microbatches_in_previous_step = -1
+        self.microbatch_count = 0
+        self.checkpoint_core_attention = args.recompute_granularity == 'selective'
+
+        # Number of layers.
+        self.num_layers = _get_num_layers(
+            args,
+            args.model_type == ModelType.encoder_and_decoder,
+            layer_type == LayerType.decoder)
+
+        self.drop_path_rates = [rate.item() for rate in torch.linspace(0, self.drop_path_rate, args.num_layers)]
+
+        # Transformer layers.
+        def build_layer(layer_number):
+            if args.transformer_impl == 'local':
+                return ParallelTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+            else:
+                return transformer_engine.pytorch.TransformerLayer(
+                    args.hidden_size,
+                    args.ffn_hidden_size,
+                    args.num_attention_heads,
+                    layernorm_epsilon=args.layernorm_epsilon,
+                    hidden_dropout=args.hidden_dropout,
+                    attention_dropout=args.attention_dropout,
+                    init_method=init_method,
+                    output_layer_init_method=output_layer_init_method,
+                    layer_number=layer_number,
+                    kv_channels=args.kv_channels,
+                    self_attn_mask_type=self_attn_mask_type.name,
+                    tp_group=mpu.get_tensor_model_parallel_group(),
+                    get_rng_state_tracker=tensor_parallel.get_cuda_rng_tracker,
+                    fuse_wgrad_accumulation=args.gradient_accumulation_fusion,
+                    apply_query_key_layer_scaling=args.apply_query_key_layer_scaling,
+                    attention_softmax_in_fp32=args.attention_softmax_in_fp32,
+                    seq_length=args.seq_length,
+                    micro_batch_size=args.micro_batch_size,
+                    sequence_parallel=args.sequence_parallel,
+                    params_dtype=args.params_dtype,
+                    apply_residual_connection_post_layernorm=args.apply_residual_connection_post_layernorm,
+                    output_layernorm=False,
+                    layer_type="encoder",
+                    drop_path_rate=self.drop_path_rates[layer_number - 1],
+                    set_parallel_mode=True,
+                    fuse_qkv_params=True)
+
+        if args.virtual_pipeline_model_parallel_size is not None:
+            assert args.num_layers % args.virtual_pipeline_model_parallel_size == 0, \
+                'num_layers_per_stage must be divisible by ' \
+                'virtual_pipeline_model_parallel_size'
+            assert args.model_type != ModelType.encoder_and_decoder
+            # Number of layers in each model chunk is the number of layers in the stage,
+            # divided by the number of model chunks in a stage.
+            self.num_layers = self.num_layers // args.virtual_pipeline_model_parallel_size
+            # With 8 layers, 2 stages, and 4 model chunks, we want an assignment of
+            # layers to stages like (each list is a model chunk):
+            # Stage 0: [0]  [2]  [4]  [6]
+            # Stage 1: [1]  [3]  [5]  [7]
+            # With 8 layers, 2 stages, and 2 virtual stages, we want an assignment of
+            # layers to stages like (each list is a model chunk):
+            # Stage 0: [0, 1]  [4, 5]
+            # Stage 1: [2, 3]  [6, 7]
+            offset = mpu.get_virtual_pipeline_model_parallel_rank() * (
+                args.num_layers // args.virtual_pipeline_model_parallel_size) + \
+                (mpu.get_pipeline_model_parallel_rank() * self.num_layers)
+        else:
+            # Each stage gets a contiguous set of layers.
+            if args.model_type == ModelType.encoder_and_decoder and \
+                    mpu.get_pipeline_model_parallel_world_size() > 1:
+                pipeline_rank = mpu.get_pipeline_model_parallel_rank()
+                if layer_type == LayerType.encoder:
+                    offset = pipeline_rank * self.num_layers
+                else:
+                    num_ranks_in_enc = args.pipeline_model_parallel_split_rank
+                    offset = (pipeline_rank - num_ranks_in_enc) * self.num_layers
+            else:
+                offset = mpu.get_pipeline_model_parallel_rank() * self.num_layers
+
+        if self.num_layers == 0:
+            # When a standalone embedding stage is used (e.g.,
+            # args.standalone_embedding_stage == True), virtual pipeline ranks
+            # on pipeline rank 0 will have zero transformer layers assigned to
+            # them. This results in the model's input and output tensors to be
+            # the same, which will cause failure for certain output tensor
+            # optimizations (e.g., pipeline output deallocation). To remedy
+            # this, we assign a 'no-op' layer on these ranks, which will
+            # disconnect the input tensor from the output tensor.
+            self.num_layers = 1
+            self.layers = torch.nn.ModuleList([ NoopTransformerLayer(1) ])
+        else:
+            self.layers = torch.nn.ModuleList(
+                [build_layer(i + 1 + offset) for i in range(self.num_layers)])
+
+        if self.post_process and self.post_layer_norm:
+            # Final layer norm before output.
+            self.final_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm,
+                sequence_parallel=args.sequence_parallel,
+                apply_layernorm_1p=args.apply_layernorm_1p)
+
+    def _get_layer(self, layer_number):
+        return self.layers[layer_number]
+
+    def _checkpointed_forward(self, hidden_states, attention_mask,
+                              encoder_output, enc_dec_attn_mask,
+                              rotary_pos_emb, is_first_microbatch):
+        """Forward method with activation checkpointing."""
+        def custom(start, end, is_transformer_engine=False):
+            def custom_forward(*args, **kwargs):
+                x_, *args = args
+                for index in range(start, end):
+                    layer = self._get_layer(index)
+                    x_ = layer(x_, *args, **kwargs)
+                return x_
+            def custom_forward_transformer_engine(*args, **kwargs):
+                return custom_forward(*args, is_first_microbatch=is_first_microbatch, **kwargs)
+            if not is_transformer_engine:
+                return custom_forward
+            else:
+                return custom_forward_transformer_engine
+
+        if self.recompute_method == 'uniform':
+            # Uniformly divide the total number of Transformer layers and checkpoint
+            # the input activation of each divided chunk.
+            # A method to further reduce memory usage reducing checkpoints.
+            l = 0
+            while l < self.num_layers:
+                if self.transformer_impl == 'transformer_engine':
+                    hidden_states = transformer_engine.pytorch.distributed.checkpoint(
+                        custom(l, l + self.recompute_num_layers, is_transformer_engine=True),
+                        self.distribute_saved_activations,
+                        tensor_parallel.get_cuda_rng_tracker,
+                        mpu.get_tensor_model_parallel_group(),
+                        hidden_states, attention_mask, encoder_output,
+                        enc_dec_attn_mask, rotary_pos_emb)
+                else:
+                    hidden_states = tensor_parallel.checkpoint(
+                        custom(l, l + self.recompute_num_layers),
+                        self.distribute_saved_activations,
+                        hidden_states, attention_mask, encoder_output,
+                        enc_dec_attn_mask, rotary_pos_emb)
+
+                l += self.recompute_num_layers
+
+        elif self.recompute_method == 'block':
+            # Checkpoint the input activation of only a set number of individual
+            # Transformer layers and skip the rest.
+            # A method fully use the device memory removing redundant re-computation.
+            for l in range(self.num_layers):
+                if l < self.recompute_num_layers:
+                    if self.transformer_impl == 'transformer_engine':
+                        hidden_states = transformer_engine.pytorch.distributed.checkpoint(
+                            custom(l, l + 1, is_transformer_engine=True),
+                            self.distribute_saved_activations,
+                            tensor_parallel.get_cuda_rng_tracker,
+                            mpu.get_tensor_model_parallel_group(),
+                            hidden_states, attention_mask, encoder_output,
+                            enc_dec_attn_mask, rotary_pos_emb)
+                    else:
+                        hidden_states = tensor_parallel.checkpoint(
+                            custom(l, l + 1),
+                            self.distribute_saved_activations,
+                            hidden_states, attention_mask, encoder_output,
+                            enc_dec_attn_mask, rotary_pos_emb)
+                else:
+                    if self.transformer_impl == 'transformer_engine':
+                        hidden_states = custom(l, l + 1, is_transformer_engine=True)(
+                            hidden_states, attention_mask, encoder_output,
+                            enc_dec_attn_mask, rotary_pos_emb)
+                    else:
+                        hidden_states = custom(l, l + 1)(
+                            hidden_states, attention_mask, encoder_output,
+                            enc_dec_attn_mask, rotary_pos_emb)
+        else:
+            raise ValueError("Invalid activation recompute method.")
+
+        return hidden_states
+
+    def set_input_tensor(self, input_tensor):
+        """Set input tensor to be used instead of forward()'s input.
+
+        When doing pipeline parallelism the input from the previous
+        stage comes from communication, not from the input, so the
+        model's forward_step_func won't have it. This function is thus
+        used by internal code to bypass the input provided by the
+        forward_step_func"""
+        self.input_tensor = input_tensor
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None, rotary_pos_emb=None):
+        # hidden_states: [s, b, h]
+
+        # Checks.
+        if inference_params:
+            assert self.recompute_granularity is None, \
+                'inference does not work with activation checkpointing'
+
+        if not self.pre_process:
+            # See set_input_tensor()
+            hidden_states = self.input_tensor
+
+        # Viewless tensor.
+        # - We only need to create a viewless tensor in the case of micro batch
+        #   size (mbs) == 1, since in this case, 'hidden_states.transpose()'
+        #   above creates a view tensor, and '.contiguous()' is a pass-through.
+        #   For mbs >= 2, '.contiguous()' creates a new tensor, eliminating
+        #   the need to make it viewless.
+        #
+        #   However, we don't explicitly check mbs == 1 here because
+        #   make_viewless_tensor() has negligible overhead when its input
+        #   is already viewless.
+        #
+        # - For the 'else' case above, calling make_viewless_tensor() here is
+        #   likely redundant, since p2p_communication.py (likely originator)
+        #   already creates viewless tensors. That said, make_viewless_tensor()
+        #   is called here to be future-proof and corner-case-proof.
+        hidden_states = core.utils.make_viewless_tensor(
+            hidden_states,
+            requires_grad=True,
+            keep_graph=True,
+        )
+
+        if self.sequence_parallel:
+            rng_context = tensor_parallel.get_cuda_rng_tracker().fork()
+        else:
+            rng_context = nullcontext()
+
+        with rng_context:
+            # The fp8_autocast context manager is a no-op when enabled=True
+            # The if...else serves to short circuit name resolution for fp8_autocast
+            with transformer_engine.pytorch.fp8_autocast(
+                enabled=self.use_fp8,
+                fp8_recipe=self.fp8_recipe,
+                fp8_group=self.fp8_group
+            ) if self.use_fp8 else nullcontext():
+                # Determine if the current iteration is first microbatch
+                if self.num_microbatches_in_previous_step != get_num_microbatches():
+                    self.microbatch_count = 0 # Reset count on new batch size rampup interval
+                self.num_microbatches_in_previous_step = get_num_microbatches()
+                is_first_microbatch = self.microbatch_count % get_num_microbatches() == 0
+
+                # Forward pass.
+                if self.recompute_granularity == 'full':
+                    hidden_states = self._checkpointed_forward(hidden_states,
+                                                               attention_mask,
+                                                               encoder_output,
+                                                               enc_dec_attn_mask,
+                                                               rotary_pos_emb,
+                                                               is_first_microbatch)
+                else:
+                    forward_kwargs = {
+                        'encoder_output': encoder_output,
+                        'enc_dec_attn_mask': enc_dec_attn_mask,
+                        'inference_params': inference_params,
+                        'rotary_pos_emb': rotary_pos_emb,
+                    }
+
+                    if self.transformer_impl == 'transformer_engine':
+                        forward_kwargs['is_first_microbatch'] = is_first_microbatch
+                        forward_kwargs['checkpoint_core_attention'] = self.checkpoint_core_attention
+
+                    for index in range(self.num_layers):
+                        layer = self._get_layer(index)
+
+                        hidden_states = layer(
+                            hidden_states,
+                            attention_mask,
+                            **forward_kwargs)
+
+                # Skip counter update for eval and activation checkpointing
+                if torch.is_grad_enabled() and self.training:
+                    self.microbatch_count += 1
+
+        # Final layer norm.
+        if self.post_process and self.post_layer_norm:
+            hidden_states = self.final_layernorm(hidden_states)
+
+        return hidden_states

megatron/model/utils.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Utilities for models."""
+
+import math
+
+import torch
+
+from megatron import get_args
+
+def init_method_normal(sigma):
+    """Init method based on N(0, sigma)."""
+    def init_(tensor):
+        return torch.nn.init.normal_(tensor, mean=0.0, std=sigma)
+
+    return init_
+
+
+def scaled_init_method_normal(sigma, num_layers):
+    """Init method based on N(0, sigma/sqrt(2*num_layers)."""
+    std = sigma / math.sqrt(2.0 * num_layers)
+
+    def init_(tensor):
+        return torch.nn.init.normal_(tensor, mean=0.0, std=std)
+
+    return init_
+
+
+def attention_mask_func(attention_scores, attention_mask):
+    attention_scores.masked_fill_(attention_mask, -10000.0)
+    return attention_scores
+
+
+def get_linear_layer(rows, columns, init_method):
+    """Simple linear layer with weight initialization."""
+    layer = torch.nn.Linear(rows, columns)
+    if get_args().perform_initialization:
+        init_method(layer.weight)
+    with torch.no_grad():
+        layer.bias.zero_()
+    return layer
+
+@torch.jit.script
+def gelu_impl(x):
+    """OpenAI's gelu implementation."""
+    return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x *
+                                       (1.0 + 0.044715 * x * x)))
+def openai_gelu(x):
+    return gelu_impl(x)
+
+#This is actually Python equivalent of torch.nn.functional.gelu(), also with type hints for ONNX exporter
+@torch.jit.script
+def erf_gelu(x):
+    return x * 0.5 * (torch.erf(x / 1.41421).to(dtype=x.dtype)+torch.ones_like(x).to(dtype=x.dtype))

megatron/model/vision/classification.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Vision Transformer(VIT) model."""
+
+import torch
+from torch.nn.init import trunc_normal_
+from megatron import get_args
+from megatron.model.utils import get_linear_layer
+from megatron.model.vision.vit_backbone import VitBackbone, VitMlpHead
+from megatron.model.vision.mit_backbone import mit_b3_avg
+from megatron.model.module import MegatronModule
+
+class VitClassificationModel(MegatronModule):
+    """Vision Transformer Model."""
+
+    def __init__(self, num_classes, finetune=False,
+                 pre_process=True, post_process=True):
+        super(VitClassificationModel, self).__init__()
+        args = get_args()
+
+        self.hidden_size = args.hidden_size
+        self.num_classes = num_classes
+        self.finetune = finetune
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.backbone = VitBackbone(
+            pre_process=self.pre_process,
+            post_process=self.post_process,
+            single_token_output=True
+        )
+        
+        if self.post_process:
+            if not self.finetune:
+                self.head = VitMlpHead(self.hidden_size, self.num_classes)
+            else:
+                self.head = get_linear_layer(
+                    self.hidden_size,
+                    self.num_classes,
+                    torch.nn.init.zeros_
+                )
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.backbone.set_input_tensor(input_tensor)
+
+    def forward(self, input):
+        hidden_states = self.backbone(input)
+
+        if self.post_process:
+            hidden_states = self.head(hidden_states)
+
+        return hidden_states
+
+
+class MitClassificationModel(MegatronModule):
+    """Mix vision Transformer Model."""
+
+    def __init__(self, num_classes,
+                 pre_process=True, post_process=True):
+        super(MitClassificationModel, self).__init__()
+        args = get_args()
+
+        self.hidden_size = args.hidden_size
+        self.num_classes = num_classes
+
+        self.backbone = mit_b3_avg()
+        self.head = torch.nn.Linear(512, num_classes)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, torch.nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, torch.nn.Linear) and m.bias is not None:
+                torch.nn.init.constant_(m.bias, 0)
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        pass
+
+    def forward(self, input):
+        hidden_states = self.backbone(input)
+        hidden_states = self.head(hidden_states)
+
+        return hidden_states

megatron/model/vision/dino.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the Apache license found in the
+# LICENSE file in the root directory of this source tree.
+
+# copied from https://github.com/facebookresearch/dino/blob/main/main_dino.py
+# reworked/refactored some parts to make it run in Megatron.
+import math
+import apex
+import einops
+import torch
+import numpy as np
+import torch.nn.functional as F
+from torch.nn.init import trunc_normal_
+from megatron import get_args, print_rank_0
+from megatron.model.utils import get_linear_layer
+from megatron.model.vision.vit_backbone import VitBackbone
+from megatron.model.module import MegatronModule
+from megatron.model.vision.mit_backbone import mit_b5_avg
+from megatron.model.vision.esvit_swin_backbone import get_swin
+
+
+class DINOLoss(torch.nn.Module):
+    def __init__(self, out_dim, ncrops, warmup_teacher_temp, teacher_temp,
+                 warmup_teacher_temp_epochs, nepochs, student_temp=0.1,
+                 center_momentum=0.9):
+        super().__init__()
+        self.student_temp = student_temp
+        self.center_momentum = center_momentum
+        self.ncrops = ncrops
+        self.register_buffer("center", torch.zeros(1, out_dim))
+        # we apply a warm up for the teacher temperature because
+        # a too high temperature makes the training instable at the beginning
+        self.teacher_temp_schedule = np.concatenate((
+            np.linspace(warmup_teacher_temp,
+                        teacher_temp, warmup_teacher_temp_epochs),
+            np.ones(nepochs - warmup_teacher_temp_epochs) * teacher_temp
+        ))
+        self.teacher_temp = teacher_temp
+
+    def forward(self, student_output, teacher_output, iteration):
+        """
+        Cross-entropy between softmax outputs of the teacher
+        and student network.
+        """
+        args = get_args()
+        student_out = student_output / self.student_temp
+        student_out = student_out.chunk(self.ncrops)
+
+        epoch = iteration // args.iter_per_epoch
+
+        # teacher centering and sharpening
+        temp = self.teacher_temp_schedule[epoch]
+        teacher_out = F.softmax((teacher_output - self.center) / temp, dim=-1)
+
+        teacher_out = teacher_out.detach().chunk(2)
+
+        total_loss = 0
+        n_loss_terms = 0
+        for iq, q in enumerate(teacher_out):
+            for v in range(len(student_out)):
+                if v == iq:
+                    # we skip cases where student and teacher operate on the same view
+                    continue
+                loss = torch.sum(-q * F.log_softmax(student_out[v], dim=-1), dim=-1)
+                total_loss += loss.mean()
+                n_loss_terms += 1
+        total_loss /= n_loss_terms
+        self.update_center(teacher_output)
+        return total_loss
+
+    @torch.no_grad()
+    def update_center(self, teacher_output):
+        """
+        Update center used for teacher output.
+        """
+        batch_center = torch.sum(teacher_output, dim=0, keepdim=True)
+        torch.distributed.all_reduce(batch_center)
+        batch_center = batch_center / (len(teacher_output) * torch.distributed.get_world_size())
+        self.center = self.center * self.center_momentum + batch_center * (1 - self.center_momentum)
+
+class DINOHead(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, norm_last_layer=True, nlayers=3):
+        super().__init__()
+        args = get_args()
+        hidden_dim = args.dino_head_hidden_size
+        bottleneck_dim = args.dino_bottleneck_size
+        nlayers = max(nlayers, 1)
+        if nlayers == 1:
+            self.mlp = torch.nn.Linear(in_dim, bottleneck_dim)
+        else:
+            layers = [torch.nn.Linear(in_dim, hidden_dim)]
+            layers.append(torch.nn.GELU())
+            for _ in range(nlayers - 2):
+                layers.append(torch.nn.Linear(hidden_dim, hidden_dim))
+                layers.append(torch.nn.GELU())
+            layers.append(torch.nn.Linear(hidden_dim, bottleneck_dim))
+            self.mlp = torch.nn.Sequential(*layers)
+        self.apply(self._init_weights)
+        self.last_layer = torch.nn.utils.weight_norm(torch.nn.Linear(bottleneck_dim, out_dim, bias=False))
+        self.last_layer.weight_g.data.fill_(1)
+        if norm_last_layer:
+            self.last_layer.weight_g.requires_grad = False
+
+    def _init_weights(self, m):
+        if isinstance(m, torch.nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, torch.nn.Linear) and m.bias is not None:
+                torch.nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        x = self.mlp(x)
+        x = torch.nn.functional.normalize(x, dim=-1, p=2)
+        x = self.last_layer(x)
+        return x
+
+
+class MultiCropWrapper(MegatronModule):
+
+    """
+    Perform forward pass separately on each resolution input.
+    The inputs corresponding to a single resolution are clubbed and single
+    forward is run on the same resolution inputs. Hence we do several
+    forward passes = number of different resolutions used. We then
+    concatenate all the output features and run the head forward on these
+    concatenated features.
+    """
+    def __init__(self, backbone, head):
+        super(MultiCropWrapper, self).__init__()
+        # disable layers dedicated to ImageNet labels classification
+        #backbone.fc, backbone.head = torch.nn.Identity(), torch.nn.Identity()
+        self.backbone = backbone
+        self.head = head
+
+    def forward(self, x):
+        # convert to list
+        if not isinstance(x, list):
+            x = [x]
+        idx_crops = torch.cumsum(torch.unique_consecutive(
+            torch.tensor([inp.shape[-1] for inp in x]),
+            return_counts=True,
+        )[1], 0)
+
+        start_idx = 0
+        for end_idx in idx_crops:
+            _out = self.backbone(torch.cat(x[start_idx: end_idx]))
+            if start_idx == 0:
+                output = _out
+            else:
+                output = torch.cat((output, _out))
+            start_idx = end_idx
+        # Run the head forward on the concatenated features.
+        if self.training:
+            return self.head(output)
+        else:
+            return output
+
+
+def cosine_scheduler(base_value, final_value, epochs, niter_per_ep,
+                     warmup_epochs=0, start_warmup_value=0):
+    warmup_schedule = np.array([])
+    warmup_iters = warmup_epochs * niter_per_ep
+    if warmup_epochs > 0:
+        warmup_schedule = \
+                np.linspace(start_warmup_value, base_value, warmup_iters)
+
+    iters = np.arange(epochs * niter_per_ep - warmup_iters)
+    schedule = final_value + 0.5 * (base_value - final_value) \
+        * (1 + np.cos(np.pi * iters / len(iters)))
+
+    schedule = np.concatenate((warmup_schedule, schedule))
+    assert len(schedule) == epochs * niter_per_ep
+    return schedule
+
+
+def get_student_backbone_and_num_features(pre_process=True, post_process=True):
+    args = get_args()
+
+    if args.vision_backbone_type == 'vit':
+        student = VitBackbone(pre_process=pre_process,
+                              post_process=post_process,
+                              drop_path_rate=0.1,
+                              single_token_output=True)
+        num_features = args.hidden_size
+    elif args.vision_backbone_type == 'mit':
+        student = mit_b5_avg(drop_path_rate=0.1)
+        num_features = 512
+    elif args.vision_backbone_type == 'swin':
+        student = get_swin()
+        num_features = student.num_features
+    else:
+        raise Exception('{} vision backbone is not supported.'.format(
+                              args.vision_backbone_type))
+ 
+    return student, num_features
+
+def get_teacher_backbone_and_num_features(pre_process=True, post_process=True):
+    args = get_args()
+
+    if args.vision_backbone_type == 'vit':
+        teacher = VitBackbone(pre_process=pre_process,
+                              post_process=post_process,
+                              single_token_output=True)
+        num_features = args.hidden_size
+    elif args.vision_backbone_type == 'mit':
+        teacher = mit_b5_avg(drop_path_rate=0.0)
+        num_features = 512
+    elif args.vision_backbone_type == 'swin':
+        teacher = get_swin(is_teacher=True)
+        num_features = teacher.num_features
+    else:
+        raise Exception('{} vision backbone is not supported.'.format(
+                              args.vision_backbone_type))
+    return teacher, num_features
+
+
+class DINOPretrainModel(MegatronModule):
+    def __init__(self, pre_process=True, post_process=True):
+        super(DINOPretrainModel, self).__init__()
+        args = get_args()
+        self.out_dim = 65536
+
+        self.dino_loss = DINOLoss(
+            self.out_dim,
+            args.dino_local_crops_number + 2,
+            args.dino_warmup_teacher_temp,
+            args.dino_teacher_temp,
+            args.dino_warmup_teacher_temp_epochs,
+            300,
+        )
+
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.momentum_teacher = 0.996
+
+        student_backbone, num_features = \
+            get_student_backbone_and_num_features(pre_process, post_process)
+
+        self.student = MultiCropWrapper(
+            student_backbone,
+            DINOHead(num_features, self.out_dim,
+                     norm_last_layer=args.dino_norm_last_layer)
+        )
+
+        self.momentum_schedule = cosine_scheduler(
+            self.momentum_teacher, 1,
+            args.train_iters // args.iter_per_epoch,
+            args.iter_per_epoch
+        )
+
+        teacher_backbone, num_features = \
+            get_teacher_backbone_and_num_features(pre_process, post_process)
+        self.teacher = MultiCropWrapper(
+            teacher_backbone,
+            DINOHead(num_features, self.out_dim)
+        )
+        self.teacher.load_state_dict(self.student.state_dict())
+
+        for p in self.teacher.parameters():
+            if hasattr(p, "requires_grad") and p.requires_grad is not None:
+                p.requires_grad = False
+
+    def set_input_tensor(self, tensor):
+        pass
+
+    def forward(self, input):
+        student_output = None
+        if self.training:
+            student_output = self.student(input)
+            teacher_output = self.teacher(input[:2])
+        else:
+            teacher_output = self.teacher(input)
+        return student_output, teacher_output
+
+    def cancel_gradients_last_layer(self, iteration):
+        args = get_args()
+        epoch = iteration // args.iter_per_epoch
+        if epoch < args.dino_freeze_last_layer:
+            for n, p in self.student.named_parameters():
+                if "last_layer" in n:
+                    p.grad = None
+
+    def update_momentum(self, iteration):
+        with torch.no_grad():
+            m = self.momentum_schedule[iteration]
+            for param_q, param_k in zip(self.student.parameters(), self.teacher.parameters()):
+                param_k.data.mul_(m).add_((1 - m) * param_q.detach().data)
+

megatron/model/vision/esvit_swin_backbone.py

+# Copyright (c) 2021 Microsoft
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+# --------------------------------------------------------
+# Modified by Chunyuan Li (chunyl@microsoft.com)
+# Swin Transformer
+# --------------------------------------------------------
+
+import os
+import logging
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+import torch.distributed as dist
+from torch.nn.init import trunc_normal_
+from megatron.model.transformer import DropPath
+from megatron import get_args
+from megatron.model import LayerNorm
+import numpy as np
+from math import sqrt
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None,
+                 out_features=None, act_layer=nn.GELU, drop=0.):
+        super(Mlp, self).__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r"""Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super(WindowAttention, self).__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2 Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0).type(attn.type())
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn_out = attn
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn_out
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+    @staticmethod
+    def compute_macs(module, input, output):
+        B, N, C = input[0].shape
+
+        module.__flops__ += module.flops(N) * B
+
+
+class SwinTransformerBlock(nn.Module):
+    r"""Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = input_resolution[0]
+        self.W = input_resolution[1]
+
+        self.attn_mask_dict = {}
+
+
+    def create_attn_mask(self, H, W):
+        # calculate attention mask for SW-MSA
+
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1))  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+
+    def forward(self, x):
+        B, L, C = x.shape
+        H = int(sqrt(L))
+        W = H
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+
+            if H in self.attn_mask_dict.keys():
+                attn_mask = self.attn_mask_dict[H]
+            else:
+                self.attn_mask_dict[H] = self.create_attn_mask(self.H, self.W).to(x.device)
+                attn_mask = self.attn_mask_dict[H]
+
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows, attn = self.attn(x_windows, attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x, attn
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size} mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+
+class PatchMerging(nn.Module):
+    r"""Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """ Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        H = int(sqrt(L))
+        W = H
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+    def extra_repr(self) -> str:
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.dim
+        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
+        return flops
+
+
+class BasicLayer(nn.Module):
+    """A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer)
+            for i in range(depth)])
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        for blk in self.blocks:
+            x, _ = blk(x)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def forward_with_features(self, x):
+        fea = []
+        for blk in self.blocks:
+            x, _ = blk(x)
+            fea.append(x)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x, fea
+
+    def forward_with_attention(self, x):
+        attns = []
+        for blk in self.blocks:
+            x, attn = blk(x)
+            attns.append(attn)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x, attns
+
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None):
+        super().__init__()
+        img_size = (img_size, img_size)
+        patch_size = (patch_size, patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+
+        x = self.proj(x).flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+
+    def flops(self):
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops
+
+class SwinTransformer(nn.Module):
+    r""" Swin Transformer
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+    Args:
+        img_size (int | tuple(int)): Input image size.
+        patch_size (int | tuple(int)): Patch size.
+        in_chans (int): Number of input channels.
+        num_classes (int): Number of classes for classification head.
+        embed_dim (int): Embedding dimension.
+        depths (tuple(int)): Depth of Swin Transformer layers.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: Truee
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate.
+        drop_path_rate (float): Stochastic depth rate.
+        norm_layer (nn.Module): normalization layer.
+        ape (bool): If True, add absolute position embedding to the patch embedding.
+        patch_norm (bool): If True, add normalization after patch embedding.
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000,
+                 embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24],
+                 window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm, ape=False, patch_norm=True, **kwargs):
+        super().__init__()
+
+        self.num_classes = num_classes
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
+        self.mlp_ratio = mlp_ratio
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(dim=int(embed_dim * 2 ** i_layer),
+                               input_resolution=(patches_resolution[0] // (2 ** i_layer),
+                                                 patches_resolution[1] // (2 ** i_layer)),
+                               depth=depths[i_layer],
+                               num_heads=num_heads[i_layer],
+                               window_size=window_size,
+                               mlp_ratio=self.mlp_ratio,
+                               qkv_bias=qkv_bias, qk_scale=qk_scale,
+                               drop=drop_rate, attn_drop=attn_drop_rate,
+                               drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                               norm_layer=norm_layer,
+                               downsample=PatchMerging if (i_layer < self.num_layers - 1) else None)
+            self.layers.append(layer)
+
+        self.norm = norm_layer(self.num_features)
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        # todo: to be implemented
+        return {'relative_position_bias_table'}
+
+    def forward(self, x):
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x)
+
+        x_region = self.norm(x)  # B L C
+        x = self.avgpool(x_region.transpose(1, 2))  # B C 1
+        x = torch.flatten(x, 1)
+
+        return x
+
+
+    def forward_feature_maps(self, x):
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x)
+
+        x_grid = self.norm(x)  # B L C
+        x = self.avgpool(x_grid.transpose(1, 2))  # B C 1
+        x = torch.flatten(x, 1)
+
+        return x, x_grid
+
+
+    def forward_selfattention(self, x, n=1):
+        # n=1 return the last layer attn map; otherwise return attn maps in all layers
+
+        
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        if n==1:
+            return self.forward_last_selfattention(x)
+        else:
+            return self.forward_all_selfattention(x)
+
+    def forward_last_selfattention(self, x):
+
+        for i, layer in enumerate(self.layers):
+            if i < len(self.layers) - 1:
+                x = layer(x)
+            else:
+                x, attns = layer.forward_with_attention(x)
+                return attns[-1]
+
+    def forward_all_selfattention(self, x):
+        attn_out = []
+
+        for layer in self.layers:
+            x, attns = layer.forward_with_attention(x)
+            attn_out += attns
+
+        return attn_out
+
+
+    def forward_return_n_last_blocks(self, x, n=1, return_patch_avgpool=False, depth=[]):
+
+        num_blks = sum(depth)
+        start_idx = num_blks - n
+
+        sum_cur = 0
+        for i, d in enumerate(depth):
+            sum_cur_new = sum_cur + d
+            if start_idx >= sum_cur and start_idx < sum_cur_new:
+                start_stage = i
+                start_blk = start_idx - sum_cur
+            sum_cur = sum_cur_new
+
+
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        # we will return the averaged token features from the `n` last blocks
+        # note: there is no [CLS] token in Swin Transformer
+        output = []
+        s = 0
+        for i, layer in enumerate(self.layers):
+            x, fea = layer.forward_with_features(x)
+
+            if i >= start_stage:
+                for x_ in fea[start_blk:]:
+
+                    if i == len(self.layers)-1: # use the norm in the last stage
+                        x_ = self.norm(x_)
+
+                    x_avg = torch.flatten(self.avgpool(x_.transpose(1, 2)), 1)  # B C     
+                    # print(f'Stage {i},  x_avg {x_avg.shape}')          
+                    output.append(x_avg)
+
+                start_blk = 0
+
+        return torch.cat(output, dim=-1)
+
+
+
+    def flops(self):
+        flops = 0
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+            if dist.get_rank() == 0:
+                print(f"GFLOPs layer_{i}: {layer.flops() / 1e9}")
+        flops += self.num_features * self.patches_resolution[0] * self.patches_resolution[1] // (2 ** self.num_layers)
+        flops += self.num_features * self.num_classes
+        return flops
+
+    def init_weights(self, pretrained='', pretrained_layers=[], verbose=True):
+        if os.path.isfile(pretrained):
+            pretrained_dict = torch.load(pretrained, map_location='cpu')
+            logging.info(f'=> loading pretrained model {pretrained}')
+            model_dict = self.state_dict()
+            pretrained_dict = {
+                k: v for k, v in pretrained_dict.items()
+                if k in model_dict.keys()
+            }
+            need_init_state_dict = {}
+            for k, v in pretrained_dict.items():
+                need_init = (
+                        k.split('.')[0] in pretrained_layers
+                        or pretrained_layers[0] is '*'
+                        or 'relative_position_index' not in k
+                        or 'attn_mask' not in k
+                )
+
+                if need_init:
+                    if verbose:
+                        logging.info(f'=> init {k} from {pretrained}')
+
+                    if 'relative_position_bias_table' in k and v.size() != model_dict[k].size():
+                        relative_position_bias_table_pretrained = v
+                        relative_position_bias_table_current = model_dict[k]
+                        L1, nH1 = relative_position_bias_table_pretrained.size()
+                        L2, nH2 = relative_position_bias_table_current.size()
+                        if nH1 != nH2:
+                            logging.info(f"Error in loading {k}, passing")
+                        else:
+                            if L1 != L2:
+                                logging.info(
+                                    '=> load_pretrained: resized variant: {} to {}'
+                                        .format((L1, nH1), (L2, nH2))
+                                )
+                                S1 = int(L1 ** 0.5)
+                                S2 = int(L2 ** 0.5)
+                                relative_position_bias_table_pretrained_resized = torch.nn.functional.interpolate(
+                                    relative_position_bias_table_pretrained.permute(1, 0).view(1, nH1, S1, S1),
+                                    size=(S2, S2),
+                                    mode='bicubic')
+                                v = relative_position_bias_table_pretrained_resized.view(nH2, L2).permute(1, 0)
+
+                    if 'absolute_pos_embed' in k and v.size() != model_dict[k].size():
+                        absolute_pos_embed_pretrained = v
+                        absolute_pos_embed_current = model_dict[k]
+                        _, L1, C1 = absolute_pos_embed_pretrained.size()
+                        _, L2, C2 = absolute_pos_embed_current.size()
+                        if C1 != C1:
+                            logging.info(f"Error in loading {k}, passing")
+                        else:
+                            if L1 != L2:
+                                logging.info(
+                                    '=> load_pretrained: resized variant: {} to {}'
+                                        .format((1, L1, C1), (1, L2, C2))
+                                )
+                                S1 = int(L1 ** 0.5)
+                                S2 = int(L2 ** 0.5)
+                                absolute_pos_embed_pretrained = absolute_pos_embed_pretrained.reshape(-1, S1, S1, C1)
+                                absolute_pos_embed_pretrained = absolute_pos_embed_pretrained.permute(0, 3, 1, 2)
+                                absolute_pos_embed_pretrained_resized = torch.nn.functional.interpolate(
+                                    absolute_pos_embed_pretrained, size=(S2, S2), mode='bicubic')
+                                v = absolute_pos_embed_pretrained_resized.permute(0, 2, 3, 1).flatten(1, 2)
+
+                    need_init_state_dict[k] = v
+            self.load_state_dict(need_init_state_dict, strict=False)
+
+    def freeze_pretrained_layers(self, frozen_layers=[]):
+        for name, module in self.named_modules():
+            if (
+                    name.split('.')[0] in frozen_layers
+                    or '.'.join(name.split('.')[0:2]) in frozen_layers
+                    or (len(frozen_layers) > 0 and frozen_layers[0] is '*')
+            ):
+                for _name, param in module.named_parameters():
+                    param.requires_grad = False
+                logging.info(
+                    '=> set param {} requires grad to False'
+                        .format(name)
+                )
+        for name, param in self.named_parameters():
+            if (
+                    name.split('.')[0] in frozen_layers
+                    or (len(frozen_layers) > 0 and frozen_layers[0] is '*')
+                    and param.requires_grad is True
+            ):
+                param.requires_grad = False
+                logging.info(
+                    '=> set param {} requires grad to False'
+                        .format(name)
+                )
+        return self
+
+
+def get_swin(is_teacher=False):
+    args = get_args()
+
+    if args.swin_backbone_type == "tiny":
+        embed_dim = 96
+        depths = [2, 2, 6, 2]
+        num_heads = [3, 6, 12, 24]
+        drop_path_rate = 0.1
+    elif args.swin_backbone_type == 'h3':
+        embed_dim = 384
+        depths = [2, 2, 18, 2]
+        num_heads = [6, 12, 24, 48]
+        drop_path_rate = 0.2
+    else:
+        embed_dim = 128
+        depths = [2, 2, 18, 2]
+        num_heads = [4, 8, 16, 32]
+        drop_path_rate = 0.2
+
+    swin = SwinTransformer(
+        img_size=224,
+        in_chans=3,
+        num_classes=1000,
+        patch_size=4,
+        embed_dim=embed_dim,
+        depths=depths,
+        num_heads=num_heads,
+        window_size=7,
+        mlp_ratio=4,
+        qkv_bias=True,
+        drop_rate=0,
+        attn_drop_rate=0,
+        drop_path_rate=(0.0 if is_teacher else drop_path_rate),
+        norm_layer=partial(LayerNorm, eps=1e-6),
+        ape=False,
+        patch_norm=True,
+    )
+
+    return swin
+

megatron/model/vision/inpainting.py

+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# This source code is licensed under the BSD license found in the
+# LICENSE file in the root directory of this source tree.
+i
+import math
+import apex
+import einops
+import torch
+import torch.nn.functional as F
+from megatron import get_args, print_rank_0
+from megatron.model.utils import get_linear_layer
+from megatron.model.vision.vit_backbone import VitBackbone
+from megatron.model.module import MegatronModule
+from megatron.model.vision.mit_backbone import mit_b3
+from megatron.model.vision.utils import resize_
+
+
+class VitInpaintingModel(MegatronModule):
+
+    def __init__(self, pre_process=True, post_process=True):
+        super(VitInpaintingModel, self).__init__()
+        args = get_args()
+
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.hidden_size = args.hidden_size
+        self.backbone = VitBackbone(
+            pre_process=self.pre_process,
+            post_process=self.post_process,
+            class_token=False,
+        )
+        self.patch_dim = args.patch_dim
+        self.img_h = args.img_h
+        self.img_w = args.img_w
+        self.seq_length = args.seq_length
+        # full mask
+
+        if self.post_process:
+            self.linear_decoder = get_linear_layer(
+                self.hidden_size,
+                self.backbone.flatten_dim,
+                torch.nn.init.zeros_
+            )
+
+    def set_input_tensor(self, input_tensor):
+        self.backbone.set_input_tensor(input_tensor)
+
+    def forward(self, input):
+
+        hidden_states = self.backbone(input)
+
+        if not self.post_process:
+            return hidden_states
+        decoded_output = self.linear_decoder(hidden_states)
+        output = einops.rearrange(
+                decoded_output,
+                "b (h w) (p1 p2 c) -> b c (h p1) (w p2)",
+                p1=self.patch_dim,
+                p2=self.patch_dim,
+                h=self.img_h//self.patch_dim,
+                w=self.img_w//self.patch_dim,
+            )
+
+        return output
+
+
+class MLP(torch.nn.Module):
+    """
+    Linear Embedding
+    """
+    def __init__(self, input_dim=2048, embed_dim=768):
+        super().__init__()
+        self.proj = torch.nn.Linear(input_dim, embed_dim)
+
+    def forward(self, x):
+        x = x.flatten(2).transpose(1, 2)
+        x = self.proj(x)
+        return x
+
+
+class MitInpaintingModel(MegatronModule):
+    """Mix vision Transformer Model."""
+
+    def __init__(self, pre_process=True, post_process=True):
+        super(MitInpaintingModel, self).__init__()
+        self.pre_process = pre_process
+        self.post_process = post_process
+
+        args = get_args()
+        self.patch_dim = args.patch_dim
+        self.img_h = args.img_h
+        self.img_w = args.img_w
+        self.flatten_dim = self.patch_dim * self.patch_dim * 3
+        self.backbone = mit_b3()
+
+        self.in_channels = [64, 128, 320, 512]
+        self.embedding_dim = 768
+
+        c1_in_channels, c2_in_channels, c3_in_channels, c4_in_channels = self.in_channels
+
+        self.linear_c4 = MLP(input_dim=c4_in_channels, embed_dim=self.embedding_dim)
+        self.linear_c3 = MLP(input_dim=c3_in_channels, embed_dim=self.embedding_dim)
+        self.linear_c2 = MLP(input_dim=c2_in_channels, embed_dim=self.embedding_dim)
+        self.linear_c1 = MLP(input_dim=c1_in_channels, embed_dim=self.embedding_dim)
+
+        self.conv_fuse = torch.nn.Conv2d(self.embedding_dim*4, self.embedding_dim, 1, 1, bias=False)
+        self.norm = apex.parallel.SyncBatchNorm(self.embedding_dim)
+        self.dropout = torch.nn.Dropout2d(0.1)
+        
+        self.linear_pred = torch.nn.Conv2d(self.embedding_dim, self.flatten_dim, kernel_size=1)
+    
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        pass
+
+    def forward(self, input):
+        c1, c2, c3, c4 = self.backbone(input)
+
+        n, _, h, w = c4.shape
+        _c4 = self.linear_c4(c4).permute(0, 2, 1).reshape(n, -1, c4.shape[2], c4.shape[3])
+        _c4 = resize(_c4, size=c1.size()[2:], mode='bilinear', align_corners=False)
+    
+        _c3 = self.linear_c3(c3).permute(0, 2, 1).reshape(n, -1, c3.shape[2], c3.shape[3])
+        _c3 = resize(_c3, size=c1.size()[2:], mode='bilinear', align_corners=False)
+
+        _c2 = self.linear_c2(c2).permute(0, 2, 1).reshape(n, -1, c2.shape[2], c2.shape[3])
+        _c2 = resize(_c2, size=c1.size()[2:], mode='bilinear', align_corners=False)
+
+        _c1 = self.linear_c1(c1).permute(0, 2, 1).reshape(n, -1, c1.shape[2], c1.shape[3])
+
+        _c = torch.cat([_c4, _c3, _c2, _c1], dim=1)
+        _c = self.conv_fuse(_c)
+ 
+        x = self.norm(_c)
+        x = F.relu(x, inplace=True)
+        x = self.dropout(x)
+
+        x = self.linear_pred(x)
+
+        output = einops.rearrange(
+            x,
+            "b (c p1 p2) h w -> b c (h p1) (w p2)",
+            p1=self.patch_dim,
+            p2=self.patch_dim,
+            h=self.img_h//self.patch_dim,
+            w=self.img_w//self.patch_dim,
+        )
+
+        return output

megatron/model/vision/knn_monitor.py

+import torch.nn.functional as F
+import torch
+from megatron import print_rank_0, get_args
+from megatron.core import mpu
+from megatron.data.vit_dataset import ClassificationTransform
+from megatron.data.image_folder import ImageFolder
+
+_FEATURE_BANK = None
+
+
+def build_data_loader(dataset, drop_last=True, shuffle=False):
+    """Data loader. Note that batch-size is the local (per GPU) batch-size."""
+    # Sampler.
+    args = get_args()
+    micro_batch_size = 16
+    num_workers = args.num_workers
+    world_size = mpu.get_data_parallel_world_size()
+    rank = mpu.get_data_parallel_rank()
+    sampler = torch.utils.data.distributed.DistributedSampler(
+        dataset, num_replicas=world_size, rank=rank,
+        drop_last=drop_last, shuffle=shuffle
+    )
+
+    # Data loader. Note that batch size is the per GPU batch size.
+    data_loader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=micro_batch_size,
+        sampler=sampler,
+        shuffle=False,
+        num_workers=num_workers,
+        drop_last=not drop_last,
+        pin_memory=True,
+    )
+    return data_loader
+
+
+def compute_feature_bank(model):
+    args = get_args()
+    global _FEATURE_BANK
+    feature_bank = []
+    feature_label = []
+
+    train_ds = ImageFolder(
+        root=args.data_path[0],
+        transform=ClassificationTransform((args.img_h, args.img_w), train=False),
+        data_per_class_fraction=1.0
+    )
+    classes = len(train_ds.classes)
+    dataloader = build_data_loader(train_ds)
+     
+    for m in model:
+        m.eval()
+
+    with torch.no_grad():
+        for i, batch in enumerate(dataloader):
+            images = batch[0].cuda().contiguous()
+            labels = batch[1].cuda().contiguous()
+            student_feature, teacher_feature = model[0](images)
+            feature = F.normalize(teacher_feature.float(), dim=1)
+            feature_bank.append(feature)
+            feature_label.append(labels)
+    
+    for m in model:
+        m.train()
+
+    # [N', D]
+    feature_bank = torch.cat(feature_bank, dim=0).contiguous()
+    feature_label = torch.cat(feature_label, dim=0).contiguous()
+
+    feature_banks = [torch.zeros_like(feature_bank)
+                     for i in range(mpu.get_data_parallel_world_size())]
+    torch.distributed.all_gather(feature_banks,
+                                 feature_bank,
+                                 group=mpu.get_data_parallel_group())
+
+    assert torch.all(torch.eq(feature_banks[mpu.get_data_parallel_rank()],
+                              feature_bank))
+
+    feature_labels = [torch.zeros_like(feature_label)
+                      for i in range(mpu.get_data_parallel_world_size())]
+    torch.distributed.all_gather(feature_labels,
+                                 feature_label,
+                                 group=mpu.get_data_parallel_group())
+
+    # [D, N]
+    feature_banks = torch.cat(feature_banks, dim=0).t().contiguous()
+    # [N]
+    feature_labels = torch.cat(feature_labels, dim=0).contiguous()
+    print_rank_0("feature_banks size is {}".format(feature_banks.size()))
+    print_rank_0("feature labels size is {}".format(feature_labels.size()))
+
+    _FEATURE_BANK = (feature_banks, feature_labels, classes)
+
+
+def get_feature_bank():
+    global _FEATURE_BANK
+    assert _FEATURE_BANK is not None
+    return _FEATURE_BANK
+
+
+# knn monitor as in InstDisc https://arxiv.org/abs/1805.01978
+# implementation follows http://github.com/zhirongw/lemniscate.pytorch and
+# https://github.com/leftthomas/SimCLR
+def knn_predict(feature, feature_bank, feature_labels, classes, knn_k, knn_t):
+    # compute cos similarity between each feature vector and feature bank ---> [B, N]
+    sim_matrix = torch.mm(feature, feature_bank)
+    # [B, K]
+    sim_weight, sim_indices = sim_matrix.topk(k=knn_k, dim=-1)
+    # [B, K]
+    sim_labels = torch.gather(feature_labels.expand(feature.size(0), -1),
+                              dim=-1,
+                              index=sim_indices)
+    sim_weight = (sim_weight / knn_t).exp()
+
+    # counts for each class
+    one_hot_label = torch.zeros(feature.size(0) * knn_k,
+                                classes,
+                                device=sim_labels.device)
+    # [B*K, C]
+    one_hot_label = one_hot_label.scatter(dim=-1,
+                                          index=sim_labels.view(-1, 1),
+                                          value=1.0)
+    # weighted score ---> [B, C]
+    pred_scores = torch.sum(
+            one_hot_label.view(feature.size(0), -1, classes) * sim_weight.unsqueeze(dim=-1),
+            dim=1)
+
+    pred_labels = pred_scores.argsort(dim=-1, descending=True)
+    return pred_labels