v1.0

megatron/model/transformer.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Transformer."""
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from megatron import get_args, logging
+from megatron import mpu
+from .module import MegatronModule
+from megatron.enums import AttnMaskType, LayerType, AttnType, PositionEmbeddingType
+from megatron.model.fused_layer_norm import MixedFusedLayerNorm as 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
+
+import deepspeed
+
+from .glu_activations import GLU_ACTIVATIONS
+from .positional_embeddings import RotaryEmbedding, apply_rotary_pos_emb_torch, apply_rotary_pos_emb
+
+# flags required to enable jit fusion kernels
+torch._C._jit_set_profiling_mode(False)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_override_can_fuse_on_cpu(True)
+torch._C._jit_override_can_fuse_on_gpu(True)
+
+logger = logging.get_logger(__name__)
+
+""" 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 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. At the end, dropout is also
+    applied.
+    """
+
+    def __init__(self, init_method, output_layer_init_method):
+        super(ParallelMLP, self).__init__()
+        args = get_args()
+
+        # Project to ffn_hidden_size
+        self.dense_h_to_4h = mpu.ColumnParallelLinear(
+            args.hidden_size,
+            # GLU is a special activation that divides the dimension by a factor 2.
+            2 * args.ffn_hidden_size if args.glu_activation else args.ffn_hidden_size,
+            gather_output=False,
+            init_method=init_method,
+            skip_bias_add=True)
+
+        self.bias_gelu_fusion = args.bias_gelu_fusion
+        self.activation_func = F.gelu
+        if args.glu_activation:
+            self.activation_func = GLU_ACTIVATIONS[args.glu_activation]
+        elif args.openai_gelu:
+            self.activation_func = openai_gelu
+        elif args.onnx_safe:
+            self.activation_func = erf_gelu
+
+        # Project back to h.
+        self.dense_4h_to_h = mpu.RowParallelLinear(
+            args.ffn_hidden_size,
+            args.hidden_size,
+            input_is_parallel=True,
+            init_method=output_layer_init_method,
+            skip_bias_add=True)
+
+
+    def forward(self, hidden_states):
+
+        # [s, b, 4hp]
+        intermediate_parallel, bias_parallel = self.dense_h_to_4h(hidden_states)
+
+        if self.bias_gelu_fusion:
+             intermediate_parallel = \
+                     bias_gelu_impl(intermediate_parallel, bias_parallel)
+        else:
+            intermediate_parallel = \
+                self.activation_func(intermediate_parallel + bias_parallel)
+
+        # [s, b, h]
+        output, output_bias = self.dense_4h_to_h(intermediate_parallel)
+        return output, output_bias
+
+
+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.position_embedding_type = args.position_embedding_type
+
+        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
+
+        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 = mpu.divide(projection_size,
+                                                    world_size)
+        self.hidden_size_per_attention_head = mpu.divide(
+            projection_size, args.num_attention_heads)
+        self.num_attention_heads_per_partition = mpu.divide(
+            args.num_attention_heads, world_size)
+
+        # Strided linear layer.
+        if attention_type == AttnType.self_attn:
+            self.query_key_value = mpu.ColumnParallelLinear(
+                args.hidden_size,
+                3 * projection_size,
+                gather_output=False,
+                init_method=init_method)
+        else:
+            assert attention_type == AttnType.cross_attn
+            self.query = mpu.ColumnParallelLinear(
+                args.hidden_size,
+                projection_size,
+                gather_output=False,
+                init_method=init_method)
+
+            self.key_value = mpu.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 = mpu.RowParallelLinear(
+            projection_size,
+            args.hidden_size,
+            input_is_parallel=True,
+            init_method=output_layer_init_method,
+            skip_bias_add=True)
+
+        if deepspeed.checkpointing.is_configured():
+            global get_cuda_rng_tracker, checkpoint
+            get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker
+            checkpoint = deepspeed.checkpointing.checkpoint
+
+        if self.position_embedding_type == PositionEmbeddingType.rotary:
+            self.rotary_emb = RotaryEmbedding(self.hidden_size_per_attention_head, precision=args.params_dtype)
+
+    def forward(self, hidden_states, attention_mask, layer_past=None,
+                get_key_value=False, encoder_output=None, alibi=None):
+        # hidden_states: [sq, b, h]
+
+        # =====================
+        # 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) = mpu.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) = mpu.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 layer_past is not None:
+            past_key, past_value = layer_past
+            key_layer = torch.cat((past_key.type_as(key_layer),
+                                   key_layer), dim=0)
+            value_layer = torch.cat((past_value.type_as(value_layer),
+                                     value_layer), dim=0)
+        if get_key_value:
+            present = (key_layer, value_layer)
+
+        # ===================================
+        # 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]
+        if alibi is None:
+            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())
+        else:
+            matmul_result = alibi[:output_size[0]*output_size[1], :, :output_size[3]]
+
+        # Rotary embeddings
+        if self.position_embedding_type == PositionEmbeddingType.rotary:
+            apply_rotary_fn = apply_rotary_pos_emb_torch if self.bf16 else apply_rotary_pos_emb
+
+            seq_len = key_layer.shape[0]
+            offset = 0
+            if layer_past is not None and layer_past.numel() > 0:
+                offset = layer_past[0].shape[0]
+                seq_len += offset
+            cos, sin = self.rotary_emb(value_layer, seq_len=seq_len)
+            query_layer, key_layer = apply_rotary_fn(query_layer, key_layer, cos, sin, offset=offset)
+
+        # Raw attention scores. [b * np, sq, sk]
+        if alibi is None:
+            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))
+        else:
+            if not hasattr(self, "logged_alibi"):
+                logger.debug("Using Alibi.")
+                self.logged_alibi = True
+
+            if self.apply_query_key_layer_scaling:
+                beta = 1.0 / self.layer_number
+            else:
+                beta = 1.0
+
+            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=beta, alpha=(1.0 / self.norm_factor))
+
+        # change view to [b, np, sq, sk]
+        attention_scores = matmul_result.view(*output_size)
+        # ==================================================
+        # Update attention mask for inference. [b, np, sq, sk]
+        # ==================================================
+
+        if get_key_value:
+            with torch.no_grad():
+                # TODO @thomasw21 Handle case where `attention_mask` is None
+                if layer_past is not None:
+                    attention_mask = attention_mask[
+                        ...,
+                        attention_scores.size(3) - 1,
+                        :attention_scores.size(3)].unsqueeze(2)
+                else:
+                    attention_mask = attention_mask[
+                        ...,
+                        :attention_scores.size(3),
+                        :attention_scores.size(3)]
+
+        # ===========================
+        # 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 mpu.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)
+
+        if get_key_value:
+            output = [output, present]
+
+        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, bias, residual, prob):
+    # type: (Tensor, Tensor, Tensor, float) -> Tensor
+    return bias_dropout_add(x, bias, residual, prob, True)
+
+
+@torch.jit.script
+def bias_dropout_add_fused_inference(x, bias, residual, prob):
+    # type: (Tensor, Tensor, Tensor, float) -> Tensor
+    return bias_dropout_add(x, bias, residual, prob, False)
+
+
+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):
+        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)
+
+        # 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
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon)
+
+        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)
+
+        # MLP
+        self.mlp = ParallelMLP(init_method,
+                               output_layer_init_method)
+
+        # Alibi
+        if args.position_embedding_type == PositionEmbeddingType.alibi:
+            self.alibi = self._build_alibi_tensor(args.seq_length, args.num_attention_heads, args.micro_batch_size).to(torch.cuda.current_device())
+            if args.params_dtype == torch.float16:
+                self.alibi = self.alibi.to(torch.float16)
+            elif args.params_dtype == torch.bfloat16:
+                self.alibi = self.alibi.to(torch.bfloat16)
+        else:
+            self.alibi = None
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                layer_past=None, get_key_value=False):
+        # 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,
+                                layer_past=layer_past,
+                                get_key_value=get_key_value,
+                                alibi=self.alibi)
+
+        if get_key_value:
+            attention_output, presents = attention_output
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        # 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)
+
+        # 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
+
+        # 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)
+
+        if get_key_value:
+            output = [output, presents]
+
+        return output
+
+    @staticmethod
+    def _build_alibi_tensor(max_seq_len, num_attention_heads, batch_size):
+        # Based on https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
+        """Returns tensor shaped (batch_size * num_attention_heads, 1, max_seq_len)"""
+
+        def get_slopes(n):
+            def get_slopes_power_of_2(n):
+                start = (2 ** (-2 ** -(math.log2(n) - 3)))
+                ratio = start
+                return [start * ratio ** i for i in range(n)]
+
+            if math.log2(n).is_integer():
+                return get_slopes_power_of_2(n)
+            else:
+                closest_power_of_2 = 2 ** math.floor(math.log2(n))
+                return get_slopes_power_of_2(closest_power_of_2) + get_slopes(2 * closest_power_of_2)[0::2][
+                                                                   :n - closest_power_of_2]
+
+        slopes = torch.Tensor(get_slopes(num_attention_heads))
+        alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_seq_len).unsqueeze(0).unsqueeze(0).expand(
+            num_attention_heads, -1, -1)
+        
+        #Select the part of the tensor that corresponds to our tensor parallel index.
+        tp_world_size = mpu.get_tensor_model_parallel_world_size()
+        tp_index = mpu.get_tensor_model_parallel_rank()
+        alibi = alibi.reshape((tp_world_size, -1, *alibi.shape[1:]))[tp_index]
+        
+        alibi = alibi.repeat(batch_size, 1, 1)
+        return alibi
+
+class ParallelTransformerLayerPipe(ParallelTransformerLayer):
+    """Extends ParallelTransformerLayer to forward attention_mask through the pipeline.
+
+    Forward has two usages that affect attention mask communication:
+
+    1) forward((input, attn_mask) , **kwargs) -> (output, mask)
+       When the attention mask is provided as the second positional
+       argument, typical pipeline behavior is used and both the output
+       *and* mask are returned in a tuple. This tuple is then forwarded
+       to the next stage in the pipeline.
+
+       This version is useful if masks are dynamic.
+    
+    2) forward(input, **kwargs) -> output
+       When the mask is static over all samples, it is advantageous to
+       cache the mask and avoid communicating it.
+    """
+    def forward(self, inputs, **kwargs):
+        assert torch.is_tensor(inputs) or isinstance(inputs, tuple)
+        if torch.is_tensor(inputs) or len(inputs) == 1:
+            hidden_states, attention_mask = inputs, None
+            return super().forward(hidden_states, attention_mask, **kwargs)
+        elif len(inputs) == 2:
+            # Attention mask is an activation.
+            hidden_states, attention_mask = inputs[0], inputs[1]
+            return super().forward(*inputs, **kwargs), attention_mask
+        else:
+            raise RuntimeError('Received more inputs than understood.')
+
+
+class ParallelTransformer(MegatronModule):
+    """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):
+        super(ParallelTransformer, 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
+
+        # Store activation checkpoiting flag.
+        self.checkpoint_activations = args.checkpoint_activations
+        self.checkpoint_num_layers = args.checkpoint_num_layers
+
+        # Number of layers.
+        assert args.num_layers % mpu.get_pipeline_model_parallel_world_size() == 0, \
+            'num_layers must be divisible by pipeline_model_parallel_size'
+        self.num_layers = args.num_layers // mpu.get_pipeline_model_parallel_world_size()
+
+        # Transformer layers.
+        def build_layer(layer_number):
+            return ParallelTransformerLayer(
+                init_method,
+                output_layer_init_method,
+                layer_number,
+                layer_type=layer_type,
+                self_attn_mask_type=self_attn_mask_type)
+        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'
+            # 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.
+            offset = mpu.get_pipeline_model_parallel_rank() * self.num_layers
+
+        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)
+
+        if deepspeed.checkpointing.is_configured():
+            global get_cuda_rng_tracker, checkpoint
+            get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker
+            checkpoint = deepspeed.checkpointing.checkpoint
+
+    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
+
+        # Make sure memory is freed.
+        mpu.reset_checkpointed_activations_memory_buffer()
+        l = 0
+        while l < self.num_layers:
+            hidden_states = mpu.checkpoint(
+                custom(l, l + self.checkpoint_num_layers),
+                hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+            l += self.checkpoint_num_layers
+
+        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, layer_past=None,
+                get_key_value=False, encoder_output=None, enc_dec_attn_mask=None):
+
+        # Checks.
+        if layer_past is not None:
+            assert get_key_value, \
+                'for not None values in layer_past, ' \
+                'expected get_key_value to be set'
+        if get_key_value:
+            assert not self.checkpoint_activations, \
+                'get_key_value 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
+
+        if encoder_output is not None:
+             encoder_output = encoder_output.transpose(0, 1).contiguous()
+
+        if self.checkpoint_activations:
+            hidden_states = self._checkpointed_forward(hidden_states,
+                                                       attention_mask,
+                                                       encoder_output,
+                                                       enc_dec_attn_mask)
+        else:
+            if get_key_value:
+                presents = []
+            for index in range(self.num_layers):
+                layer = self._get_layer(index)
+                past = None
+                if layer_past is not None:
+                    past = layer_past[index]
+                hidden_states = layer(hidden_states,
+                                      attention_mask,
+                                      encoder_output=encoder_output,
+                                      enc_dec_attn_mask=enc_dec_attn_mask,
+                                      layer_past=past,
+                                      get_key_value=get_key_value)
+                if get_key_value:
+                    hidden_states, present = hidden_states
+                    presents.append(present)
+
+        # 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
+        if get_key_value:
+            output = [output, presents]
+
+        return output

megatron/model/utils.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Utilities for models."""
+
+import math
+from functools import wraps
+
+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):
+    args = get_args()
+    if args.curriculum_learning:
+        attention_mask_ = attention_mask
+        actual_seqlen = attention_scores.size()[2]
+        if actual_seqlen != attention_mask_.size()[2]:
+            # attention_mask has size [1, 1, seqlen, seqlen]
+            attention_mask_ = attention_mask_[:, :, :actual_seqlen, :actual_seqlen].contiguous()
+        attention_scores.masked_fill_(attention_mask_, torch.finfo(attention_scores.dtype).min)
+    else:
+        attention_scores.masked_fill_(attention_mask, torch.finfo(attention_scores.dtype).min)
+    return attention_scores
+
+
+def get_linear_layer(rows, columns, init_method):
+    """Simple linear layer with weight initialization."""
+    layer = torch.nn.Linear(rows, columns)
+    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))
+
+def log_debug_usage(logger, msg: str):
+    def log_debug_usage_(func):
+        """Helper function in order to log a message when using a function for the first time"""
+        func.__logged_message__ = False
+
+        @wraps(func)
+        def wrapped(*args, **kwargs):
+            if func.__logged_message__ is False:
+                logger.debug(msg)
+                func.__logged_message__ = True
+            return func(*args, **kwargs)
+
+        return wrapped
+    return log_debug_usage_

megatron/model/vit_model.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Vision Transformer(VIT) model."""
+
+import math
+import einops
+import torch
+import torch.nn.functional as F
+from megatron import get_args
+from megatron import mpu
+from megatron.model.transformer import ParallelTransformer
+from megatron.enums import AttnMaskType
+from .module import MegatronModule, fp32_to_float16
+from megatron.model.utils import (
+    get_linear_layer,
+    init_method_normal,
+    scaled_init_method_normal,
+)
+from .module import MegatronModule
+
+from deepspeed.pipe import PipelineModule, LayerSpec, TiedLayerSpec
+from megatron.model.fused_layer_norm import MixedFusedLayerNorm as LayerNorm
+from megatron.model.module import float16_to_fp32
+from .language_model import EmbeddingPipe
+from .transformer import ParallelTransformerLayerPipe
+from .language_model import parallel_lm_logits
+
+
+class VitMlpHead(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, num_classes):
+        super(VitMlpHead, self).__init__()
+        self.dense_in = torch.nn.Linear(hidden_size, hidden_size)
+        self.dense_out = torch.nn.Linear(hidden_size, num_classes)
+        torch.nn.init.constant_(self.dense_out.bias, -10)
+
+    def forward(self, hidden_states, sequence_index=0):
+        # hidden_states: [b, s, h]
+        # sequence_index: index of the token to pool.
+        x = hidden_states[:, sequence_index, :]
+        x = self.dense_in(x)
+        x = torch.tanh(x)
+        x = self.dense_out(x)
+        return x
+
+
+def twod_interpolate_position_embeddings_hook(
+    state_dict,
+    prefix,
+    local_metadata,
+    strict,
+    missing_keys,
+    unexpected_keys,
+    error_msgs,
+):
+
+    args = get_args()
+    num_patches_per_dim = args.img_dim // args.patch_dim
+    num_patches = num_patches_per_dim ** 2
+    seq_length = num_patches + 1
+    hidden_size = args.hidden_size
+
+    key = prefix + "weight"
+    # import pdb
+    # pdb.set_trace()
+    assert key in state_dict
+    if key in state_dict:
+        input_param = state_dict[key]
+
+        assert input_param.shape[1] == hidden_size
+        if input_param.shape[0] != seq_length:
+            # update input_param and load it to state_dict[key]
+
+            num_tok_input = input_param.shape[0] - 1
+            num_tok_new = seq_length - 1
+            input_param_tok, input_param_grid = (
+                input_param[:1, :],
+                input_param[1:, :],
+            )
+
+            gs_input = int(math.sqrt(num_tok_input))
+            gs_new = int(math.sqrt(num_tok_new))
+
+            input_param_grid = input_param_grid.transpose(0, 1).contiguous()
+            input_param_grid = input_param_grid.reshape(
+                (1, -1, gs_input, gs_input)
+            )
+            input_param_grid = input_param_grid.float()
+            scale_factor = gs_new / gs_input
+
+            input_param_grid = F.interpolate(
+                input_param_grid, scale_factor=scale_factor, mode="bilinear"
+            )
+
+            input_param_grid = input_param_grid.half()
+            input_param_grid = input_param_grid.reshape((-1, gs_new * gs_new))
+            input_param_grid = input_param_grid.transpose(0, 1).contiguous()
+
+            assert input_param_grid.shape[1] == hidden_size
+            input_param = torch.cat((input_param_tok, input_param_grid), dim=0)
+            assert (
+                input_param.shape[0] == seq_length
+                and input_param.shape[1] == hidden_size
+            )
+
+            state_dict[key] = input_param
+
+
+class VitModel(MegatronModule):
+    """Vision Transformer Model."""
+
+    def __init__(self, num_classes, finetune=False):
+        super(VitModel, self).__init__()
+        args = get_args()
+
+        self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
+        if args.init_method_xavier_uniform:
+            self.init_method = torch.nn.init.xavier_uniform_
+            self.scaled_init_method = torch.nn.init.xavier_uniform_
+        else:
+            self.init_method = init_method_normal(args.init_method_std)
+            self.scaled_init_method = scaled_init_method_normal(
+                args.init_method_std, args.num_layers
+            )
+
+        self.hidden_size = args.hidden_size
+        self.num_classes = num_classes
+        self.patch_dim = args.patch_dim
+        self.img_dim = args.img_dim
+        self.finetune = finetune
+
+        assert self.img_dim % self.patch_dim == 0
+        self.num_patches_per_dim = self.img_dim // self.patch_dim
+        self.num_patches = self.num_patches_per_dim ** 2
+        self.seq_length = self.num_patches + 1
+        self.flatten_dim = self.patch_dim * self.patch_dim * args.num_channels
+
+        # cls_token
+        self.cls_token = torch.nn.Parameter(torch.randn(1, 1, self.hidden_size))
+        torch.nn.init.zeros_(self.cls_token)
+
+        # Linear encoder
+        self.linear_encoder = torch.nn.Linear(
+            self.flatten_dim, self.hidden_size
+        )
+
+        # embedding
+        self.position_embeddings = torch.nn.Embedding(
+            self.seq_length, self.hidden_size
+        )
+        init_method_normal(args.init_method_std)(
+            self.position_embeddings.weight
+        )
+        self.position_ids = torch.arange(self.seq_length).expand(1, -1).cuda()
+
+        self.position_embeddings._register_load_state_dict_pre_hook(
+            twod_interpolate_position_embeddings_hook
+        )
+
+        self.embedding_dropout = torch.nn.Dropout(args.hidden_dropout)
+
+        # Transformer
+        self.transformer = ParallelTransformer(
+            self.init_method, self.scaled_init_method
+        )
+
+        # MLP head
+        if not self.finetune:
+            self.mlp_head = VitMlpHead(self.hidden_size, self.num_classes)
+        else:
+            self.class_head = get_linear_layer(
+                self.hidden_size, num_classes, torch.nn.init.zeros_
+            )
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.transformer.set_input_tensor(input_tensor)
+    
+    def forward(self, x):
+        x = einops.rearrange(
+            x,
+            "b c (h p1) (w p2) -> b (h w) (p1 p2 c)",
+            p1=self.patch_dim,
+            p2=self.patch_dim,
+        )
+
+        assert x.dtype == torch.half
+        x = self.linear_encoder(x)
+        cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        x = x + self.position_embeddings(self.position_ids)
+        x = self.embedding_dropout(x)
+        x = self.transformer(x, None)
+
+        if not self.finetune:
+            x = self.mlp_head(x)
+        else:
+            x = self.class_head(x[:, 0, :])
+
+        return x
+
+
+class PregrocessPipe(MegatronModule):
+    def __init__(self):
+        super(PregrocessPipe, self).__init__()
+        args = get_args()
+        self.hidden_size = args.hidden_size
+        self.patch_dim = args.patch_dim
+        self.img_dim = args.img_dim
+
+        assert self.img_dim % self.patch_dim == 0
+        self.num_patches_per_dim = self.img_dim // self.patch_dim
+        self.num_patches = self.num_patches_per_dim ** 2
+        self.seq_length = self.num_patches + 1
+        self.flatten_dim = self.patch_dim * self.patch_dim * args.num_channels
+
+        # cls_token
+        self.cls_token = torch.nn.Parameter(torch.randn(1, 1, self.hidden_size))
+        torch.nn.init.zeros_(self.cls_token)
+
+        # Linear encoder
+        self.linear_encoder = torch.nn.Linear(
+            self.flatten_dim, self.hidden_size
+        )
+
+        # embedding
+        self.position_embeddings = torch.nn.Embedding(
+            self.seq_length, self.hidden_size
+        )
+        # init_method_normal(args.init_method_std)(
+        #     self.position_embeddings.weight
+        # )
+        self.position_ids = torch.arange(self.seq_length).expand(1, -1).cuda()
+
+        # self.position_embeddings._register_load_state_dict_pre_hook(
+        #     twod_interpolate_position_embeddings_hook
+        # )
+
+        self.embedding_dropout = torch.nn.Dropout(args.hidden_dropout)
+    def forward(self, x):
+        x = einops.rearrange(
+            x,
+            "b c (h p1) (w p2) -> b (h w) (p1 p2 c)",
+            p1=self.patch_dim,
+            p2=self.patch_dim,
+        )
+
+        assert x.dtype == torch.half
+        x = self.linear_encoder(x)
+        cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        x = x + self.position_embeddings(self.position_ids)
+        x = self.embedding_dropout(x)
+        return x
+
+class VitModelPipe(PipelineModule,MegatronModule):
+    """Vision Transformer Model."""
+    def __init__(self, num_classes, finetune=False, attn_mask_type: AttnMaskType = AttnMaskType.causal):
+
+        args = get_args()
+        
+        if args.init_method_xavier_uniform:
+            init_method = torch.nn.init.xavier_uniform_
+            scaled_init_method = torch.nn.init.xavier_uniform_
+        else:
+            init_method = init_method_normal(args.init_method_std)
+            scaled_init_method = scaled_init_method_normal(
+                args.init_method_std, args.num_layers
+            )
+
+        self.specs = []
+        
+        def _to_float16(inputs):
+            if args.fp16:
+                return fp32_to_float16(inputs, lambda v: v.half())
+            elif args.bf16:
+                return fp32_to_float16(inputs, lambda v: v.bfloat16())
+            else:
+                return inputs
+        
+        self.specs.append(_to_float16)
+        
+        # Embedding layer
+        self.specs.append(TiedLayerSpec('embed', PregrocessPipe))
+        
+        if args.fp32_residual_connection:
+            if getattr(args, 'pretrain_causal_attention', False):
+                self.specs.append(lambda x: x.transpose(0, 1).contiguous().float())
+            else:
+                # EmbeddingPipe returns attention mask as well
+                self.specs.append(lambda x: (x[0].transpose(0, 1).contiguous().float(), *x[1:]))
+        else:
+            if getattr(args, 'pretrain_causal_attention', False):
+                self.specs.append(lambda x: x.transpose(0, 1).contiguous())
+            else:
+                # EmbeddingPipe returns attention mask as well
+                self.specs.append(lambda x: (x[0].transpose(0, 1).contiguous(), *x[1:]))
+        
+        for layer_idx in range(args.num_layers):
+            self.specs.append(
+                LayerSpec(ParallelTransformerLayerPipe,
+                    init_method=init_method,
+                    output_layer_init_method=scaled_init_method,
+                    layer_number=layer_idx,
+                    # TODO: Change naming of class from GPT to something that encapsulate prefix lm.
+                    self_attn_mask_type=attn_mask_type))
+         
+        # Undo data format change
+        def undo(x):
+            if not getattr(args, 'pretrain_causal_attention', False):
+                x = x[0]
+            return x.transpose(0, 1).contiguous()
+        self.specs.append(undo)
+
+        # Final layernorm after transformer layers
+        self.specs.append(
+            LayerSpec(LayerNorm,
+                      args.hidden_size,
+                      eps=args.layernorm_epsilon))
+        
+        # MLP head
+        if not finetune:
+            self.specs.append(VitMlpHead(args.hidden_size, num_classes))
+        else:
+            self.specs.append(lambda x: x[:, 0, :])
+            self.specs.append(get_linear_layer(args.hidden_size, num_classes, torch.nn.init.zeros_))
+        
+        # Convert to fp32 if needed
+        # if args.fp16 or args.bf16:
+        #     self.specs.append(float16_to_fp32)
+        
+        if args.checkpoint_activations:
+            interval = args.checkpoint_num_layers
+        else:
+            interval = 0
+        
+        from deepspeed.runtime.pipe.topology import PipeModelDataParallelTopology
+        topo = PipeModelDataParallelTopology(num_pp=mpu.get_pipeline_model_parallel_world_size(),
+                                             num_mp=mpu.get_tensor_model_parallel_world_size(),
+                                             num_dp=mpu.get_data_parallel_world_size())
+
+        # here one can extend the regex to include more layers to be counted towards partitioning,
+        # e.g. 'type:transformer|embedding' will add up all the transformer blocks and also the first
+        # and last embedding layers and then partition that transformers+2 layers - so to get a good
+        # balance you may want to use less transformer layers
+        #
+        # caveat emptor: the current implementation of PP fails unless each stage has at least one
+        # transformer layer
+        if args.pp_partition_method is not None:
+            partition_method = args.pp_partition_method
+        else:
+            partition_method = 'type:transformer'
+
+        super().__init__(layers=self.specs,
+                         loss_fn=get_cross_entropy(is_prefix=attn_mask_type is AttnMaskType.prefix),
+                         topology=topo,
+                         activation_checkpoint_interval=interval,
+                         partition_method=partition_method)
+
+
+def get_cross_entropy(is_prefix: bool):
+    def CrossEntropy(output, labels):
+        args = get_args()
+
+        losses = mpu.vocab_parallel_cross_entropy(output.contiguous().float(), labels)
+
+        loss = torch.sum(losses) / labels.shape[0]
+        return loss
+    return CrossEntropy
+    

megatron/mpu/__init__.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Model parallel utility interface."""
+
+from .cross_entropy import vocab_parallel_cross_entropy
+
+from .data import broadcast_data
+
+from .initialize import is_unitialized
+from .initialize import destroy_model_parallel
+from .initialize import get_data_parallel_group
+from .initialize import get_data_parallel_rank
+from .initialize import get_data_parallel_world_size
+from .initialize import get_embedding_group
+from .initialize import get_model_parallel_group
+from .initialize import get_tensor_model_parallel_group
+from .initialize import get_pipeline_model_parallel_group
+from .initialize import get_tensor_model_parallel_rank, set_tensor_model_parallel_rank
+from .initialize import get_pipeline_model_parallel_rank, set_pipeline_model_parallel_rank
+from .initialize import is_pipeline_first_stage, is_pipeline_last_stage
+from .initialize import get_tensor_model_parallel_src_rank
+from .initialize import get_pipeline_model_parallel_first_rank
+from .initialize import get_pipeline_model_parallel_last_rank
+from .initialize import get_pipeline_model_parallel_next_rank
+from .initialize import get_pipeline_model_parallel_prev_rank
+from .initialize import get_tensor_model_parallel_world_size, set_tensor_model_parallel_world_size
+from .initialize import get_pipeline_model_parallel_world_size, set_pipeline_model_parallel_world_size
+from .initialize import get_virtual_pipeline_model_parallel_rank, set_virtual_pipeline_model_parallel_rank
+from .initialize import initialize_model_parallel
+from .initialize import model_parallel_is_initialized
+from .initialize import get_model_parallel_world_size, get_model_parallel_rank
+
+from .layers import ColumnParallelLinear
+from .layers import RowParallelLinear
+from .layers import VocabParallelEmbedding
+from .layers import (set_tensor_model_parallel_attributes,
+                     set_defaults_if_not_set_tensor_model_parallel_attributes,
+                     copy_tensor_model_parallel_attributes)
+                     
+from .mappings import copy_to_tensor_model_parallel_region
+from .mappings import gather_from_tensor_model_parallel_region
+from .mappings import reduce_from_tensor_model_parallel_region
+from .mappings import scatter_to_tensor_model_parallel_region
+
+from .random import checkpoint
+from .random import get_cuda_rng_tracker
+from .random import init_checkpointed_activations_memory_buffer
+from .random import model_parallel_cuda_manual_seed
+from .random import reset_checkpointed_activations_memory_buffer
+from .random import gather_split_1d_tensor
+from .random import split_tensor_into_1d_equal_chunks
+
+from .utils import divide
+from .utils import split_tensor_along_last_dim

megatron/mpu/cross_entropy.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+
+from .initialize import get_tensor_model_parallel_group
+from .initialize import get_tensor_model_parallel_rank
+from .initialize import get_tensor_model_parallel_world_size
+from .utils import VocabUtility
+
+
+class _VocabParallelCrossEntropy(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, vocab_parallel_logits, target):
+
+        # Maximum value along vocab dimension across all GPUs.
+        logits_max = torch.max(vocab_parallel_logits, dim=-1)[0]
+        torch.distributed.all_reduce(logits_max,
+                                     op=torch.distributed.ReduceOp.MAX,
+                                     group=get_tensor_model_parallel_group())
+        # Subtract the maximum value.
+        vocab_parallel_logits.sub_(logits_max.unsqueeze(dim=-1))
+
+        # Get the partition's vocab indecies
+        get_vocab_range = VocabUtility.vocab_range_from_per_partition_vocab_size
+        partition_vocab_size = vocab_parallel_logits.size()[-1]
+        rank = get_tensor_model_parallel_rank()
+        world_size = get_tensor_model_parallel_world_size()
+        vocab_start_index, vocab_end_index = get_vocab_range(
+            partition_vocab_size, rank, world_size)
+
+        # Create a mask of valid vocab ids (1 means it needs to be masked).
+        target_mask = (target < vocab_start_index) | (target >= vocab_end_index)
+        masked_target = target.clone() - vocab_start_index
+        masked_target[target_mask] = 0
+
+        # Get predicted-logits = logits[target].
+        # For Simplicity, we convert logits to a 2-D tensor with size
+        # [*, partition-vocab-size] and target to a 1-D tensor of size [*].
+        logits_2d = vocab_parallel_logits.view(-1, partition_vocab_size)
+        masked_target_1d = masked_target.view(-1)
+        arange_1d = torch.arange(start=0, end=logits_2d.size()[0],
+                                 device=logits_2d.device)
+        predicted_logits_1d = logits_2d[arange_1d, masked_target_1d]
+        predicted_logits_1d = predicted_logits_1d.clone().contiguous()
+        predicted_logits = predicted_logits_1d.view_as(target)
+        predicted_logits[target_mask] = 0.0
+        # All reduce is needed to get the chunks from other GPUs.
+        torch.distributed.all_reduce(predicted_logits,
+                                     op=torch.distributed.ReduceOp.SUM,
+                                     group=get_tensor_model_parallel_group())
+
+        # Sum of exponential of logits along vocab dimension across all GPUs.
+        exp_logits = vocab_parallel_logits
+        torch.exp(vocab_parallel_logits, out=exp_logits)
+        sum_exp_logits = exp_logits.sum(dim=-1)
+        torch.distributed.all_reduce(sum_exp_logits,
+                                     op=torch.distributed.ReduceOp.SUM,
+                                     group=get_tensor_model_parallel_group())
+
+        # Loss = log(sum(exp(logits))) - predicted-logit.
+        loss = torch.log(sum_exp_logits) - predicted_logits
+
+        # Store softmax, target-mask and masked-target for backward pass.
+        exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1))
+        ctx.save_for_backward(exp_logits, target_mask, masked_target_1d)
+
+        return loss
+
+    @staticmethod
+    def backward(ctx, grad_output):
+
+        # Retreive tensors from the forward path.
+        softmax, target_mask, masked_target_1d = ctx.saved_tensors
+
+        # All the inputs have softmax as thier gradient.
+        grad_input = softmax
+        # For simplicity, work with the 2D gradient.
+        partition_vocab_size = softmax.size()[-1]
+        grad_2d = grad_input.view(-1, partition_vocab_size)
+
+        # Add the gradient from matching classes.
+        arange_1d = torch.arange(start=0, end=grad_2d.size()[0],
+                                 device=grad_2d.device)
+        grad_2d[arange_1d, masked_target_1d] -= (
+            1.0 - target_mask.view(-1).float())
+
+        # Finally elementwise multiplication with the output gradients.
+        grad_input.mul_(grad_output.unsqueeze(dim=-1))
+
+        return grad_input, None
+
+
+def vocab_parallel_cross_entropy(vocab_parallel_logits, target):
+    """Helper function for the cross entropy."""
+    return _VocabParallelCrossEntropy.apply(vocab_parallel_logits, target)

megatron/mpu/data.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from .initialize import get_tensor_model_parallel_group
+from .initialize import get_tensor_model_parallel_rank
+from .initialize import get_tensor_model_parallel_src_rank
+
+
+_MAX_DATA_DIM = 5
+
+
+def _check_data_types(keys, data, target_dtype):
+    """Check that all the keys have the same target data type."""
+    for key in keys:
+        assert data[key].dtype == target_dtype, '{} has data type {} which '\
+            'is different than {}'.format(key, data[key].dtype, target_dtype)
+
+
+def _build_key_size_numel_dictionaries(keys, data):
+    """Build the size on rank 0 and broadcast."""
+    max_dim = _MAX_DATA_DIM
+    sizes = [0 for _ in range(max_dim) for _ in keys]
+
+    # Pack the sizes on rank zero.
+    if get_tensor_model_parallel_rank() == 0:
+        offset = 0
+        for key in keys:
+            assert data[key].dim() < max_dim, 'you should increase MAX_DATA_DIM'
+            size = data[key].size()
+            for i, s in enumerate(size):
+                sizes[i + offset] = s
+            offset += max_dim
+
+    # Move to GPU and broadcast.
+    sizes_cuda = torch.cuda.LongTensor(sizes)
+    torch.distributed.broadcast(sizes_cuda, get_tensor_model_parallel_src_rank(),
+                                group=get_tensor_model_parallel_group())
+
+    # Move back to cpu and unpack.
+    sizes_cpu = sizes_cuda.cpu()
+    key_size = {}
+    key_numel = {}
+    total_numel = 0
+    offset = 0
+    for key in keys:
+        i = 0
+        size = []
+        numel = 1
+        while sizes_cpu[offset + i] > 0:
+            this_size = sizes_cpu[offset + i]
+            size.append(this_size)
+            numel *= this_size
+            i += 1
+        key_size[key] = size
+        key_numel[key] = numel
+        total_numel += numel
+        offset += max_dim
+
+    return key_size, key_numel, total_numel
+
+
+def broadcast_data(keys, data, datatype):
+    """Broadcast data from rank zero of each model parallel group to the
+    members of the same model parallel group.
+
+    Arguments:
+        keys: list of keys in the data disctionary to be broadcasted
+        data: data dictionary of string keys and cpu tensor values.
+        datatype: torch data type of all tensors in data associated
+                  with keys.
+    """
+    # Build (key, size) and (key, number of elements) dictionaries along
+    # with the total number of elements on all ranks.
+    key_size, key_numel, total_numel = _build_key_size_numel_dictionaries(keys,
+                                                                          data)
+
+    # Pack on rank zero.
+    if get_tensor_model_parallel_rank() == 0:
+        # Check that all keys have the same data type.
+        _check_data_types(keys, data, datatype)
+        # Flatten the data associated with the keys
+        flatten_data = torch.cat(
+            [data[key].contiguous().view(-1) for key in keys], dim=0).cuda()
+    else:
+        flatten_data = torch.empty(total_numel,
+                                   device=torch.cuda.current_device(),
+                                   dtype=datatype)
+
+    # Broadcast
+    torch.distributed.broadcast(flatten_data, get_tensor_model_parallel_src_rank(),
+                                group=get_tensor_model_parallel_group())
+
+    # Unpack
+    output = {}
+    offset = 0
+    for key in keys:
+        size = key_size[key]
+        numel = key_numel[key]
+        output[key] = flatten_data.narrow(0, offset, numel).view(size)
+        offset += numel
+
+    return output

megatron/mpu/initialize.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""Model and data parallel groups."""
+
+import torch
+
+from .utils import ensure_divisibility
+
+
+# Intra-layer model parallel group that the current rank belongs to.
+_TENSOR_MODEL_PARALLEL_GROUP = None
+# Inter-layer model parallel group that the current rank belongs to.
+_PIPELINE_MODEL_PARALLEL_GROUP = None
+# Model parallel group (both intra- and pipeline) that the current rank belongs to.
+_MODEL_PARALLEL_GROUP = None
+# Embedding group.
+_EMBEDDING_GROUP = None
+# Data parallel group that the current rank belongs to.
+_DATA_PARALLEL_GROUP = None
+
+_VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
+_VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+
+# These values enable us to change the mpu sizes on the fly.
+_MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
+_MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+_MPU_TENSOR_MODEL_PARALLEL_RANK = None
+_MPU_PIPELINE_MODEL_PARALLEL_RANK = None
+
+# A list of global ranks for each pipeline group to ease calculation of the source
+# rank when broadcasting from the first or last pipeline stage
+_PIPELINE_GLOBAL_RANKS = None
+
+def is_unitialized():
+    """Useful for code segments that may be accessed with or without mpu initialization"""
+    return _DATA_PARALLEL_GROUP is None
+
+
+def initialize_model_parallel(tensor_model_parallel_size_=1,
+                              pipeline_model_parallel_size_=1,
+                              virtual_pipeline_model_parallel_size_=None):
+    """
+    Initialize model data parallel groups.
+
+    Arguments:
+        tensor_model_parallel_size: number of GPUs used to parallelize model tensor.
+        pipeline_model_parallel_size: number of GPUs used to parallelize model pipeline.
+
+    Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
+    use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
+    the model pipeline. The present function will
+    create 8 tensor model-parallel groups, 4 pipeline model-parallel groups
+    and 8 data-parallel groups as:
+        8 data_parallel groups:
+            [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]
+        8 tensor model-parallel groups:
+            [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]
+        4 pipeline model-parallel groups:
+            [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]
+    Note that for efficiency, the caller should make sure adjacent ranks
+    are on the same DGX box. For example if we are using 2 DGX-1 boxes
+    with a total of 16 GPUs, rank 0 to 7 belong to the first box and
+    ranks 8 to 15 belong to the second box.
+    """
+    if torch.distributed.get_rank() == 0:
+        print('> initializing tensor model parallel with size {}'.format(
+            tensor_model_parallel_size_))
+        print('> initializing pipeline model parallel with size {}'.format(
+            pipeline_model_parallel_size_))
+    # Get world size and rank. Ensure some consistencies.
+    assert torch.distributed.is_initialized()
+    world_size = torch.distributed.get_world_size()
+    tensor_model_parallel_size = min(tensor_model_parallel_size_, world_size)
+    pipeline_model_parallel_size = min(pipeline_model_parallel_size_, world_size)
+    ensure_divisibility(world_size,
+                        tensor_model_parallel_size * pipeline_model_parallel_size)
+    data_parallel_size = world_size // (tensor_model_parallel_size *
+                                        pipeline_model_parallel_size)
+
+    num_tensor_model_parallel_groups = world_size // tensor_model_parallel_size
+    num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size
+    num_data_parallel_groups = world_size // data_parallel_size
+
+    if virtual_pipeline_model_parallel_size_ is not None:
+        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+        _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = 0
+        _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = virtual_pipeline_model_parallel_size_
+
+    rank = torch.distributed.get_rank()
+
+    # Build the data-parallel groups.
+    global _DATA_PARALLEL_GROUP
+    assert _DATA_PARALLEL_GROUP is None, \
+        'data parallel group is already initialized'
+    all_data_parallel_group_ranks = []
+    for i in range(pipeline_model_parallel_size):
+        start_rank = i * num_pipeline_model_parallel_groups
+        end_rank = (i + 1) * num_pipeline_model_parallel_groups
+        for j in range(tensor_model_parallel_size):
+            ranks = range(start_rank + j, end_rank,
+                          tensor_model_parallel_size)
+            all_data_parallel_group_ranks.append(list(ranks))
+            group = torch.distributed.new_group(ranks)
+            if rank in ranks:
+                _DATA_PARALLEL_GROUP = group
+
+    # Build the model-parallel groups.
+    global _MODEL_PARALLEL_GROUP
+    assert _MODEL_PARALLEL_GROUP is None, \
+        'model parallel group is already initialized'
+    for i in range(data_parallel_size):
+        ranks = [data_parallel_group_ranks[i]
+                 for data_parallel_group_ranks in all_data_parallel_group_ranks]
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _MODEL_PARALLEL_GROUP = group
+
+    # Build the tensor model-parallel groups.
+    global _TENSOR_MODEL_PARALLEL_GROUP
+    assert _TENSOR_MODEL_PARALLEL_GROUP is None, \
+        'tensor model parallel group is already initialized'
+    for i in range(num_tensor_model_parallel_groups):
+        ranks = range(i * tensor_model_parallel_size,
+                      (i + 1) * tensor_model_parallel_size)
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _TENSOR_MODEL_PARALLEL_GROUP = group
+
+    # Build the pipeline model-parallel groups and embedding groups
+    # (first and last rank in each pipeline model-parallel group).
+    global _PIPELINE_MODEL_PARALLEL_GROUP
+    global _PIPELINE_GLOBAL_RANKS
+    assert _PIPELINE_MODEL_PARALLEL_GROUP is None, \
+        'pipeline model parallel group is already initialized'
+    global _EMBEDDING_GROUP
+    assert _EMBEDDING_GROUP is None, \
+        'embedding group is already initialized'
+    for i in range(num_pipeline_model_parallel_groups):
+        ranks = range(i, world_size,
+                      num_pipeline_model_parallel_groups)
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _PIPELINE_MODEL_PARALLEL_GROUP = group
+            _PIPELINE_GLOBAL_RANKS = ranks
+        # Setup embedding group (to exchange gradients between
+        # first and last stages).
+        if len(ranks) > 1:
+            embedding_ranks = [ranks[0], ranks[-1]]
+        else:
+            embedding_ranks = ranks
+        group = torch.distributed.new_group(embedding_ranks)
+        if rank in embedding_ranks:
+            _EMBEDDING_GROUP = group
+
+
+def model_parallel_is_initialized():
+    """Check if model and data parallel groups are initialized."""
+    if _TENSOR_MODEL_PARALLEL_GROUP is None or \
+        _PIPELINE_MODEL_PARALLEL_GROUP is None or \
+        _DATA_PARALLEL_GROUP is None:
+        return False
+    return True
+
+
+def get_model_parallel_group():
+    """Get the model parallel group the caller rank belongs to."""
+    assert _MODEL_PARALLEL_GROUP is not None, \
+        'model parallel group is not initialized'
+    return _MODEL_PARALLEL_GROUP
+
+
+def get_tensor_model_parallel_group():
+    """Get the tensor model parallel group the caller rank belongs to."""
+    assert _TENSOR_MODEL_PARALLEL_GROUP is not None, \
+        'intra_layer_model parallel group is not initialized'
+    return _TENSOR_MODEL_PARALLEL_GROUP
+
+
+def get_pipeline_model_parallel_group():
+    """Get the pipeline model parallel group the caller rank belongs to."""
+    assert _PIPELINE_MODEL_PARALLEL_GROUP is not None, \
+        'pipeline_model parallel group is not initialized'
+    return _PIPELINE_MODEL_PARALLEL_GROUP
+
+
+def get_data_parallel_group():
+    """Get the data parallel group the caller rank belongs to."""
+    assert _DATA_PARALLEL_GROUP is not None, \
+        'data parallel group is not initialized'
+    return _DATA_PARALLEL_GROUP
+
+
+def get_embedding_group():
+    """Get the embedding group the caller rank belongs to."""
+    assert _EMBEDDING_GROUP is not None, \
+        'embedding group is not initialized'
+    return _EMBEDDING_GROUP
+
+
+def set_tensor_model_parallel_world_size(world_size):
+    """Set the tensor model parallel size"""
+    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = world_size
+
+
+def set_pipeline_model_parallel_world_size(world_size):
+    """Set the pipeline model parallel size"""
+    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = world_size
+
+
+def get_tensor_model_parallel_world_size():
+    """Return world size for the tensor model parallel group."""
+    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    if _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE is not None:
+        return _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    return torch.distributed.get_world_size(group=get_tensor_model_parallel_group())
+
+
+def get_model_parallel_world_size():
+    assert get_pipeline_model_parallel_world_size() == 1, "legacy get_model_parallel_world_size is only supported if PP is disabled"
+    return get_tensor_model_parallel_world_size()
+
+
+def get_pipeline_model_parallel_world_size():
+    """Return world size for the pipeline model parallel group."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    if _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE is not None:
+        return _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    return torch.distributed.get_world_size(group=get_pipeline_model_parallel_group())
+
+
+def set_tensor_model_parallel_rank(rank):
+    """Set tensor model parallel rank."""
+    global _MPU_TENSOR_MODEL_PARALLEL_RANK
+    _MPU_TENSOR_MODEL_PARALLEL_RANK = rank
+
+
+def set_pipeline_model_parallel_rank(rank):
+    """Set pipeline model parallel rank."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    _MPU_PIPELINE_MODEL_PARALLEL_RANK = rank
+
+
+def get_tensor_model_parallel_rank():
+    """Return my rank for the tensor model parallel group."""
+    global _MPU_TENSOR_MODEL_PARALLEL_RANK
+    if _MPU_TENSOR_MODEL_PARALLEL_RANK is not None:
+        return _MPU_TENSOR_MODEL_PARALLEL_RANK
+    return torch.distributed.get_rank(group=get_tensor_model_parallel_group())
+
+
+def get_model_parallel_rank():
+    assert get_pipeline_model_parallel_world_size() == 1, "legacy get_model_parallel_rank is only supported if PP is disabled"
+    return get_tensor_model_parallel_rank()
+
+
+def get_pipeline_model_parallel_rank():
+    """Return my rank for the pipeline model parallel group."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    if _MPU_PIPELINE_MODEL_PARALLEL_RANK is not None:
+        return _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    return torch.distributed.get_rank(group=get_pipeline_model_parallel_group())
+
+
+def is_pipeline_first_stage(ignore_virtual=False):
+    """Return True if in the first pipeline model-parallel stage, False otherwise."""
+    if not ignore_virtual:
+        if get_virtual_pipeline_model_parallel_world_size() is not None and \
+            get_virtual_pipeline_model_parallel_rank() != 0:
+            return False
+    return get_pipeline_model_parallel_rank() == 0
+
+
+def is_pipeline_last_stage(ignore_virtual=False):
+    """Return True if in the last pipeline model-parallel stage, False otherwise."""
+    if not ignore_virtual:
+        virtual_pipeline_model_parallel_world_size = \
+            get_virtual_pipeline_model_parallel_world_size()
+        if virtual_pipeline_model_parallel_world_size is not None and \
+            get_virtual_pipeline_model_parallel_rank() != (
+                virtual_pipeline_model_parallel_world_size - 1):
+            return False
+    return get_pipeline_model_parallel_rank() == (
+        get_pipeline_model_parallel_world_size() - 1)
+
+
+def get_virtual_pipeline_model_parallel_rank():
+    """Return the virtual pipeline-parallel rank."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+
+
+def set_virtual_pipeline_model_parallel_rank(rank):
+    """Set the virtual pipeline-parallel rank."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = rank
+
+
+def get_virtual_pipeline_model_parallel_world_size():
+    """Return the virtual pipeline-parallel world size."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+
+
+def get_tensor_model_parallel_src_rank():
+    """Calculate the global rank corresponding to the first local rank
+    in the tensor model parallel group."""
+    global_rank = torch.distributed.get_rank()
+    local_world_size = get_tensor_model_parallel_world_size()
+    return (global_rank // local_world_size) * local_world_size
+
+
+def get_pipeline_model_parallel_first_rank():
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    return _PIPELINE_GLOBAL_RANKS[0]
+
+
+def get_pipeline_model_parallel_last_rank():
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    last_rank_local = get_pipeline_model_parallel_world_size() - 1
+    return _PIPELINE_GLOBAL_RANKS[last_rank_local]
+
+def get_pipeline_model_parallel_next_rank():
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    rank_in_pipeline = get_pipeline_model_parallel_rank()
+    world_size = get_pipeline_model_parallel_world_size()
+    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline + 1) % world_size]
+
+
+def get_pipeline_model_parallel_prev_rank():
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    rank_in_pipeline = get_pipeline_model_parallel_rank()
+    world_size = get_pipeline_model_parallel_world_size()
+    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline - 1) % world_size]
+
+
+def get_data_parallel_world_size():
+    """Return world size for the data parallel group."""
+    return torch.distributed.get_world_size(group=get_data_parallel_group())
+
+
+def get_data_parallel_rank():
+    """Return my rank for the data parallel group."""
+    return torch.distributed.get_rank(group=get_data_parallel_group())
+
+
+def destroy_model_parallel():
+    """Set the groups to none."""
+    global _TENSOR_MODEL_PARALLEL_GROUP
+    _TENSOR_MODEL_PARALLEL_GROUP = None
+    global _PIPELINE_MODEL_PARALLEL_GROUP
+    _PIPELINE_MODEL_PARALLEL_GROUP = None
+    global _DATA_PARALLEL_GROUP
+    _DATA_PARALLEL_GROUP = None

megatron/mpu/layers.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parts of the code here are adapted from PyTorch
+# repo: https://github.com/pytorch/pytorch
+
+
+import math
+
+import torch
+import torch.nn.functional as F
+import torch.nn.init as init
+from torch.nn.parameter import Parameter
+from functools import partial
+
+from .initialize import get_tensor_model_parallel_rank
+from .initialize import get_tensor_model_parallel_world_size
+from .mappings import copy_to_tensor_model_parallel_region
+from .mappings import gather_from_tensor_model_parallel_region
+from .mappings import reduce_from_tensor_model_parallel_region
+from .mappings import scatter_to_tensor_model_parallel_region
+from .random import get_cuda_rng_tracker
+from .utils import divide
+from .utils import split_tensor_along_last_dim
+from .utils import VocabUtility
+from ..model.fused_layer_norm import MixedFusedLayerNorm as LayerNorm
+from megatron import get_args, mpu
+import deepspeed.runtime.activation_checkpointing.checkpointing as ds_checkpointing
+
+
+_MODEL_PARALLEL_ATTRIBUTE_DEFAULTS = {'tensor_model_parallel': False,
+                                      'partition_dim': -1,
+                                      'partition_stride': 1}
+
+
+def param_is_not_tensor_parallel_duplicate(param):
+    return (hasattr(param, 'tensor_model_parallel') and
+            param.tensor_model_parallel) or (
+                get_tensor_model_parallel_rank() == 0)
+
+
+def set_tensor_model_parallel_attributes(tensor, is_parallel, dim, stride):
+    # Make sure the attributes are not set.
+    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
+        assert not hasattr(tensor, attribute)
+    # Set the attributes.
+    setattr(tensor, 'tensor_model_parallel', is_parallel)
+    setattr(tensor, 'partition_dim', dim)
+    setattr(tensor, 'partition_stride', stride)
+
+
+def set_defaults_if_not_set_tensor_model_parallel_attributes(tensor):
+    def maybe_set(attribute, value):
+        if not hasattr(tensor, attribute):
+            setattr(tensor, attribute, value)
+    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
+        maybe_set(attribute, _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS[attribute])
+
+
+def copy_tensor_model_parallel_attributes(destination_tensor, source_tensor):
+    def maybe_copy(attribute):
+        if hasattr(source_tensor, attribute):
+            setattr(destination_tensor, attribute,
+                    getattr(source_tensor, attribute))
+    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
+        maybe_copy(attribute)
+
+
+def _initialize_affine_weight_gpu(weight, init_method,
+                                  partition_dim, stride=1):
+    """Initialize affine weight for model parallel on GPU."""
+
+    set_tensor_model_parallel_attributes(tensor=weight,
+                                         is_parallel=True,
+                                         dim=partition_dim,
+                                         stride=stride)
+
+    if ds_checkpointing.is_configured():
+        global get_cuda_rng_tracker
+        get_cuda_rng_tracker = ds_checkpointing.get_cuda_rng_tracker
+
+    with get_cuda_rng_tracker().fork():
+        init_method(weight)
+
+
+def _initialize_affine_weight_cpu(weight, output_size, input_size,
+                                  per_partition_size, partition_dim,
+                                  init_method, stride=1,
+                                  return_master_weight=False):
+    """Initialize affine weight for model parallel.
+
+    Build the master weight on all processes and scatter
+    the relevant chunk."""
+
+    set_tensor_model_parallel_attributes(tensor=weight,
+                                         is_parallel=True,
+                                         dim=partition_dim,
+                                         stride=stride)
+
+    # Initialize master weight
+    master_weight = torch.empty(output_size, input_size,
+                                dtype=torch.float,
+                                requires_grad=False)
+    init_method(master_weight)
+    args = get_args()
+    master_weight = master_weight.to(dtype=args.params_dtype)
+
+    # Split and copy
+    per_partition_per_stride_size = divide(per_partition_size, stride)
+    weight_list = torch.split(master_weight, per_partition_per_stride_size,
+                              dim=partition_dim)
+    rank = get_tensor_model_parallel_rank()
+    world_size = get_tensor_model_parallel_world_size()
+    my_weight_list = weight_list[rank::world_size]
+
+    with torch.no_grad():
+        torch.cat(my_weight_list, dim=partition_dim, out=weight)
+    if return_master_weight:
+        return master_weight
+    return None
+
+
+def xavier_uniform_tensor_parallel_(tensor, gain=1., tp_degree=1):
+    r"""
+    This is a modified torch.nn.init.xavier_uniform_ with changes to support
+    partitioned on the vocab size dim embedding with tensor parallel.
+
+    Additional args:
+    - tp_degree: degree of tensor parallel
+
+    Note: the code assumes all partitions are equal in size
+    """
+    # receptive_field_size=1 as dim==2, so we don't need init._calculate_fan_in_and_fan_out
+    fan_out, fan_in = tensor.shape
+    fan_out *= tp_degree # tp splits on num_embeddings dim
+
+    std = gain * math.sqrt(2.0 / float(fan_in + fan_out))
+    a = math.sqrt(3.0) * std  # Calculate uniform bounds from standard deviation
+
+    return torch.nn.init._no_grad_uniform_(tensor, -a, a)
+
+
+class VocabParallelEmbedding(torch.nn.Module):
+    """Embedding parallelized in the vocabulary dimension.
+
+    This is mainly adapted from torch.nn.Embedding and all the default
+    values are kept.
+    Arguments:
+        num_embeddings: vocabulary size.
+        embedding_dim: size of hidden state.
+        init_method: method to initialize weights.
+    """
+
+    def __init__(self, num_embeddings, embedding_dim,
+                 init_method=init.xavier_normal_):
+        super(VocabParallelEmbedding, self).__init__()
+        # Keep the input dimensions.
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        # Set the defaults for compatibility.
+        self.padding_idx = None
+        self.max_norm = None
+        self.norm_type = 2.
+        self.scale_grad_by_freq = False
+        self.sparse = False
+        self._weight = None
+        self.tensor_model_parallel_size = get_tensor_model_parallel_world_size()
+        # Divide the weight matrix along the vocabulary dimension.
+        self.vocab_start_index, self.vocab_end_index = \
+            VocabUtility.vocab_range_from_global_vocab_size(
+                self.num_embeddings, get_tensor_model_parallel_rank(),
+                self.tensor_model_parallel_size)
+        self.num_embeddings_per_partition = self.vocab_end_index - \
+            self.vocab_start_index
+
+        # Allocate weights and initialize.
+        args = get_args()
+
+        # only the first stage embedding runs this class' forward. The head's embedding does its own
+        # thing, so don't waste memory allocating LN weights.
+        if mpu.is_pipeline_first_stage() and (args.use_bnb_optimizer or args.embed_layernorm):
+            self.norm = LayerNorm(embedding_dim)
+
+        if args.use_bnb_optimizer:
+            # for BNB we ignore the passed init_method and use torch.nn.init.xavier_uniform_
+            # modified to calculate std on the unpartitioned embedding
+            init_method = partial(xavier_uniform_tensor_parallel_, tp_degree=self.tensor_model_parallel_size)
+
+        if args.use_cpu_initialization:
+            self.weight = Parameter(torch.empty(
+                self.num_embeddings_per_partition, self.embedding_dim,
+                dtype=args.params_dtype))
+            _initialize_affine_weight_cpu(
+                self.weight, self.num_embeddings, self.embedding_dim,
+                self.num_embeddings_per_partition, 0, init_method)
+        else:
+            self.weight = Parameter(torch.empty(
+                self.num_embeddings_per_partition, self.embedding_dim,
+                device=torch.cuda.current_device(), dtype=args.params_dtype))
+            _initialize_affine_weight_gpu(self.weight, init_method,
+                                          partition_dim=0, stride=1)
+
+        if args.use_bnb_optimizer:
+            from bitsandbytes.optim import GlobalOptimManager
+            GlobalOptimManager.get_instance().override_config(self.weight, 'optim_bits', 32)
+            GlobalOptimManager.get_instance().register_parameters(self.weight)
+
+
+    def forward(self, input_):
+        if torch.any(input_ >= self.num_embeddings):
+            raise ValueError(f"There is an input id in the input that is greater than the highest possible input id.\nInput: {input_}\nnum_embeddings: {self.num_embeddings}")
+
+        if self.tensor_model_parallel_size > 1:
+            # Build the mask.
+            input_mask = (input_ < self.vocab_start_index) | \
+                         (input_ >= self.vocab_end_index)
+            # Mask the input.
+            masked_input = input_.clone() - self.vocab_start_index
+            masked_input[input_mask] = 0
+        else:
+            # input_ is garanted to be in the range [0:self.vocab_end_index - self.vocab_start_index] thanks to the first check
+            masked_input = input_
+
+        # Get the embeddings.
+        output_parallel = F.embedding(masked_input, self.weight,
+                                      self.padding_idx, self.max_norm,
+                                      self.norm_type, self.scale_grad_by_freq,
+                                      self.sparse)
+        # Mask the output embedding.
+        if self.tensor_model_parallel_size > 1:
+            output_parallel[input_mask, :] = 0.0
+        # Reduce across all the model parallel GPUs.
+        output = reduce_from_tensor_model_parallel_region(output_parallel)
+
+        if hasattr(self, 'norm'):
+            output = self.norm(output)
+
+        return output
+
+
+class ColumnParallelLinear(torch.nn.Module):
+    """Linear layer with column parallelism.
+
+    The linear layer is defined as Y = XA + b. A is parallelized along
+    its second dimension as A = [A_1, ..., A_p].
+
+    Arguments:
+        input_size: first dimension of matrix A.
+        output_size: second dimension of matrix A.
+        bias: If true, add bias
+        gather_output: If true, call all-gether on output and make Y avaiable
+                       to all GPUs, otherwise, every GPU will have its output
+                       which is Y_i = XA_i
+        init_method: method to initialize weights. Note that bias is always set
+                     to zero.
+        stride: For the strided linear layers.
+        keep_master_weight_for_test: This was added for testing and should be
+                                     set to False. It returns the master weights
+                                     used for initialization.
+        skip_bias_add: This was added to enable performance optimations where bias
+                       can be fused with other elementwise operations. we skip
+                       adding bias but instead return it.
+    """
+
+    def __init__(self, input_size, output_size, bias=True, gather_output=True,
+                 init_method=init.xavier_normal_, stride=1,
+                 keep_master_weight_for_test=False,
+                 skip_bias_add=False):
+        super(ColumnParallelLinear, self).__init__()
+
+        # Keep input parameters
+        self.input_size = input_size
+        self.output_size = output_size
+        self.gather_output = gather_output
+        # Divide the weight matrix along the last dimension.
+        world_size = get_tensor_model_parallel_world_size()
+        self.output_size_per_partition = divide(output_size, world_size)
+        self.skip_bias_add = skip_bias_add
+
+        # Parameters.
+        # Note: torch.nn.functional.linear performs XA^T + b and as a result
+        # we allocate the transpose.
+        # Initialize weight.
+        args = get_args()
+        if args.use_cpu_initialization:
+            self.weight = Parameter(torch.empty(self.output_size_per_partition,
+                                                self.input_size,
+                                                dtype=args.params_dtype))
+            self.master_weight = _initialize_affine_weight_cpu(
+                self.weight, self.output_size, self.input_size,
+                self.output_size_per_partition, 0, init_method,
+                stride=stride, return_master_weight=keep_master_weight_for_test)
+        else:
+            self.weight = Parameter(torch.empty(
+                self.output_size_per_partition, self.input_size,
+                device=torch.cuda.current_device(), dtype=args.params_dtype))
+            _initialize_affine_weight_gpu(self.weight, init_method,
+                                          partition_dim=0, stride=stride)
+
+        if bias:
+            if args.use_cpu_initialization:
+                self.bias = Parameter(torch.empty(
+                    self.output_size_per_partition, dtype=args.params_dtype))
+            else:
+                self.bias = Parameter(torch.empty(
+                    self.output_size_per_partition,
+                    device=torch.cuda.current_device(),
+                    dtype=args.params_dtype))
+            set_tensor_model_parallel_attributes(self.bias, True, 0, stride)
+            # Always initialize bias to zero.
+            with torch.no_grad():
+                self.bias.zero_()
+        else:
+            self.register_parameter('bias', None)
+
+
+
+    def forward(self, input_):
+        # Set up backprop all-reduce.
+        input_parallel = copy_to_tensor_model_parallel_region(input_)
+        # Matrix multiply.
+
+        bias = self.bias if not self.skip_bias_add else None
+        output_parallel = F.linear(input_parallel, self.weight, bias)
+        if self.gather_output:
+            # All-gather across the partitions.
+            output = gather_from_tensor_model_parallel_region(output_parallel)
+        else:
+            output = output_parallel
+        output_bias = self.bias if self.skip_bias_add else None
+        return output, output_bias
+
+
+class RowParallelLinear(torch.nn.Module):
+    """Linear layer with row parallelism.
+
+    The linear layer is defined as Y = XA + b. A is parallelized along
+    its first dimension and X along its second dimension as:
+               -   -
+              | A_1 |
+              | .   |
+          A = | .   |        X = [X_1, ..., X_p]
+              | .   |
+              | A_p |
+               -   -
+    Arguments:
+        input_size: first dimension of matrix A.
+        output_size: second dimension of matrix A.
+        bias: If true, add bias. Note that bias is not parallelized.
+        input_is_parallel: If true, we assume that the input is already
+                           split across the GPUs and we do not split
+                           again.
+        init_method: method to initialize weights. Note that bias is always set
+                     to zero.
+        stride: For the strided linear layers.
+        keep_master_weight_for_test: This was added for testing and should be
+                                     set to False. It returns the master weights
+                                     used for initialization.
+        skip_bias_add: This was added to enable performance optimations where bias
+                       can be fused with other elementwise operations. we skip
+                       adding bias but instead return it.
+    """
+
+    def __init__(self, input_size, output_size, bias=True,
+                 input_is_parallel=False,
+                 init_method=init.xavier_normal_, stride=1,
+                 keep_master_weight_for_test=False,
+                 skip_bias_add=False):
+        super(RowParallelLinear, self).__init__()
+
+        # Keep input parameters
+        self.input_size = input_size
+        self.output_size = output_size
+        self.input_is_parallel = input_is_parallel
+        # Divide the weight matrix along the last dimension.
+        world_size = get_tensor_model_parallel_world_size()
+        self.input_size_per_partition = divide(input_size, world_size)
+        self.skip_bias_add = skip_bias_add
+
+        # Parameters.
+        # Note: torch.nn.functional.linear performs XA^T + b and as a result
+        # we allocate the transpose.
+        # Initialize weight.
+        args = get_args()
+        if args.use_cpu_initialization:
+            self.weight = Parameter(torch.empty(self.output_size,
+                                                self.input_size_per_partition,
+                                                dtype=args.params_dtype))
+            self.master_weight = _initialize_affine_weight_cpu(
+                self.weight, self.output_size, self.input_size,
+                self.input_size_per_partition, 1, init_method,
+                stride=stride, return_master_weight=keep_master_weight_for_test)
+        else:
+            self.weight = Parameter(torch.empty(
+                self.output_size, self.input_size_per_partition,
+                device=torch.cuda.current_device(), dtype=args.params_dtype))
+            _initialize_affine_weight_gpu(self.weight, init_method,
+                                          partition_dim=1, stride=stride)
+        if bias:
+            if args.use_cpu_initialization:
+                self.bias = Parameter(torch.empty(self.output_size,
+                                                  dtype=args.params_dtype))
+            else:
+                self.bias = Parameter(torch.empty(
+                    self.output_size, device=torch.cuda.current_device(),
+                    dtype=args.params_dtype))
+            # Always initialize bias to zero.
+            with torch.no_grad():
+                self.bias.zero_()
+        else:
+            self.register_parameter('bias', None)
+
+        self.bias_tp_auto_sync = args.sync_tp_duplicated_parameters
+
+
+    def forward(self, input_):
+        # Set up backprop all-reduce.
+        if self.input_is_parallel:
+            input_parallel = input_
+        else:
+            input_parallel = scatter_to_tensor_model_parallel_region(input_)
+        # Matrix multiply.
+        output_parallel = F.linear(input_parallel, self.weight)
+        # All-reduce across all the partitions.
+        output_ = reduce_from_tensor_model_parallel_region(output_parallel)
+
+        if self.bias_tp_auto_sync:
+            torch.distributed.all_reduce(self.bias, op=torch.distributed.ReduceOp.AVG, group=mpu.get_tensor_model_parallel_group())
+
+        if not self.skip_bias_add:
+            output = output_ + self.bias if self.bias is not None else output_
+            output_bias = None
+        else:
+            output = output_
+            output_bias = self.bias
+        return output, output_bias

megatron/mpu/mappings.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from .initialize import get_tensor_model_parallel_group, get_tensor_model_parallel_world_size, get_tensor_model_parallel_rank
+from .utils import split_tensor_along_last_dim
+
+
+def _reduce(input_):
+    """All-reduce the the input tensor across model parallel group."""
+
+    # Bypass the function if we are using only 1 GPU.
+    if get_tensor_model_parallel_world_size()==1:
+        return input_
+
+    # All-reduce.
+    torch.distributed.all_reduce(input_, group=get_tensor_model_parallel_group())
+
+    return input_
+
+
+def _split(input_):
+    """Split the tensor along its last dimension and keep the
+    corresponding slice."""
+
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size==1:
+        return input_
+
+    # Split along last dimension.
+    input_list = split_tensor_along_last_dim(input_, world_size)
+
+    # Note: torch.split does not create contiguous tensors by default.
+    rank = get_tensor_model_parallel_rank()
+    output = input_list[rank].contiguous()
+
+    return output
+
+
+def _gather(input_):
+    """Gather tensors and concatinate along the last dimension."""
+
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size==1:
+        return input_
+
+    # Size and dimension.
+    last_dim = input_.dim() - 1
+    rank = get_tensor_model_parallel_rank()
+
+    tensor_list = [torch.empty_like(input_) for _ in range(world_size)]
+    tensor_list[rank] = input_
+    torch.distributed.all_gather(tensor_list, input_, group=get_tensor_model_parallel_group())
+
+    # Note: torch.cat already creates a contiguous tensor.
+    output = torch.cat(tensor_list, dim=last_dim).contiguous()
+
+    return output
+
+
+class _CopyToModelParallelRegion(torch.autograd.Function):
+    """Pass the input to the model parallel region."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return input_
+    
+    @staticmethod
+    def forward(ctx, input_):
+        return input_
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _reduce(grad_output)
+
+
+class _ReduceFromModelParallelRegion(torch.autograd.Function):
+    """All-reduce the input from the model parallel region."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _reduce(input_)
+    
+    @staticmethod
+    def forward(ctx, input_):
+        return _reduce(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return grad_output
+
+
+class _ScatterToModelParallelRegion(torch.autograd.Function):
+    """Split the input and keep only the corresponding chuck to the rank."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _split(input_)
+
+    @staticmethod
+    def forward(ctx, input_):
+        return _split(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _gather(grad_output)
+
+
+class _GatherFromModelParallelRegion(torch.autograd.Function):
+    """Gather the input from model parallel region and concatinate."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _gather(input_)
+    
+    @staticmethod
+    def forward(ctx, input_):
+        return _gather(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _split(grad_output)
+
+
+# -----------------
+# Helper functions.
+# -----------------
+
+def copy_to_tensor_model_parallel_region(input_):
+    return _CopyToModelParallelRegion.apply(input_)
+
+
+def reduce_from_tensor_model_parallel_region(input_):
+    return _ReduceFromModelParallelRegion.apply(input_)
+
+
+def scatter_to_tensor_model_parallel_region(input_):
+    return _ScatterToModelParallelRegion.apply(input_)
+
+
+def gather_from_tensor_model_parallel_region(input_):
+    return _GatherFromModelParallelRegion.apply(input_)

megatron/mpu/random.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parts of the code here are adapted from PyTorch
+# repo: https://github.com/pytorch/pytorch
+
+import contextlib
+
+import torch
+from torch import _C
+from torch.cuda import _lazy_call, device as device_ctx_manager
+from torch.utils.checkpoint import detach_variable
+
+from megatron import get_args
+from megatron.memory import allocate_mem_buff
+
+from .initialize import get_data_parallel_rank
+from .initialize import get_tensor_model_parallel_group
+from .initialize import get_tensor_model_parallel_rank
+from .initialize import get_tensor_model_parallel_world_size
+
+
+# Default name for the model parallel rng tracker.
+_MODEL_PARALLEL_RNG_TRACKER_NAME = 'model-parallel-rng'
+
+
+# Whether apply model parallelsim to checkpointed hidden states.
+_CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER = None
+
+
+def init_checkpointed_activations_memory_buffer():
+    """Initialize the memory buffer for the checkpointed activations."""
+    args = get_args()
+
+    upper_bound_sequence_length = max(
+        args.seq_length if args.seq_length is not None else 0,
+        args.decoder_seq_length if args.decoder_seq_length is not None else 0
+    )
+    per_layer = args.micro_batch_size * upper_bound_sequence_length * \
+                args.hidden_size // args.tensor_model_parallel_size
+    assert args.num_layers % args.checkpoint_num_layers == 0, \
+        'number of layers is not divisible by checkpoint-num-layers'
+    num_checkpointer_layers = args.num_layers // args.checkpoint_num_layers
+    numel = per_layer * num_checkpointer_layers
+    dtype = torch.half
+    if not args.fp16:
+        dtype = torch.float
+
+    global _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER
+    assert _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER is None, \
+        'checkpointed activations memory buffer is already allocated.'
+    _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER = allocate_mem_buff(
+        'checkpointed activations', numel, dtype, track_usage=False)
+
+
+def reset_checkpointed_activations_memory_buffer():
+    """Reset the memory used for checkpointing."""
+    if _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER is not None:
+        _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER.reset()
+
+
+def _set_cuda_rng_state(new_state, device=-1):
+    """Sets the random number generator state of the current GPU.
+
+    Argumentss:
+        new_state (torch.ByteTensor): The desired state
+    This function is adapted from PyTorch repo (torch.cuda.set_rng_state)
+    with a single change: the input state is not cloned. Cloning caused
+    major performance issues for +4 GPU cases.
+    """
+    if hasattr(_C, '_cuda_setRNGState') and callable(_C._cuda_setRNGState):
+        # older PyTorch
+        def cb():
+            with device_ctx_manager(device):
+                _C._cuda_setRNGState(new_state)
+    else:
+        # newer PyTorch
+        if device == -1:
+            device = torch.device('cuda')
+        elif isinstance(device, str):
+            device = torch.device(device)
+        elif isinstance(device, int):
+            device = torch.device('cuda', device)
+
+        def cb():
+            idx = device.index
+            if idx is None:
+                idx = torch.cuda.current_device()
+            default_generator = torch.cuda.default_generators[idx]
+            default_generator.set_state(new_state)
+
+    _lazy_call(cb)
+
+
+def split_tensor_into_1d_equal_chunks(tensor):
+    """Break a tensor into equal 1D chunks."""
+    data = tensor.view(-1)
+    partition_size = torch.numel(data) // get_tensor_model_parallel_world_size()
+    start_index = partition_size * get_tensor_model_parallel_rank()
+    end_index = start_index + partition_size
+    return data[start_index:end_index]
+
+
+def gather_split_1d_tensor(tensor):
+    """Opposite of above function, gather values from model parallel ranks."""
+    world_size = get_tensor_model_parallel_world_size()
+    numel = torch.numel(tensor)
+    numel_gathered = world_size * numel
+    gathered = torch.empty(numel_gathered, dtype=tensor.dtype,
+                           device=torch.cuda.current_device(),
+                           requires_grad=False)
+    chunks = [gathered[i*numel:(i+1)*numel] for i in range(world_size)]
+    torch.distributed.all_gather(chunks, tensor,
+                                 group=get_tensor_model_parallel_group())
+    return gathered
+
+
+class CudaRNGStatesTracker:
+    """Tracker for the cuda RNG states.
+
+    Using the `add` method, a cuda rng state is initialized based on
+    the input `seed` and is assigned to `name`. Later, by forking the
+    rng state, we can perform operations and return to our starting
+    cuda state.
+    """
+
+    def __init__(self):
+        # Map from a string name to the cuda rng state.
+        self.states_ = {}
+        # Seeds are just for book keeping and ensure no seed is set twice.
+        self.seeds_ = set()
+
+    def reset(self):
+        """Set to the initial state (no tracker)."""
+        self.states_ = {}
+        self.seeds_ = set()
+
+    def get_states(self):
+        """Get rng states. Copy the dictionary so we have direct
+        pointers to the states, not just a pointer to the dictionary."""
+        states = {}
+        for name in self.states_:
+            states[name] = self.states_[name]
+        return states
+
+    def set_states(self, states):
+        """Set the rng states. For efficiency purposes, we do not check
+        the size of seed for compatibility."""
+        self.states_ = states
+
+    def add(self, name, seed):
+        """Track the rng state."""
+        # Check seed is not already used.
+        if seed in self.seeds_:
+            raise Exception('seed {} already exists'.format(seed))
+        self.seeds_.add(seed)
+        # Check that state is not already defined.
+        if name in self.states_:
+            raise Exception('cuda rng state {} already exists'.format(name))
+        # Get the current rng state.
+        orig_rng_state = torch.cuda.get_rng_state()
+        # Set the new state and store it.
+        torch.cuda.manual_seed(seed)
+        self.states_[name] = torch.cuda.get_rng_state()
+        # Reset rng state to what it was.
+        _set_cuda_rng_state(orig_rng_state)
+
+    @contextlib.contextmanager
+    def fork(self, name=_MODEL_PARALLEL_RNG_TRACKER_NAME):
+        """Fork the cuda rng state, perform operations, and exit with
+        the original state."""
+        # Check if we have added the state
+        if name not in self.states_:
+            print(name, self.states_)
+            raise Exception('cuda rng state {} is not added'.format(name))
+        # Store current rng state.
+        orig_cuda_rng_state = torch.cuda.get_rng_state()
+        # Set rng state to the desired one
+        _set_cuda_rng_state(self.states_[name])
+        # Do the stuff we wanted to do.
+        try:
+            yield
+        finally:
+            # Update the current rng state for later use.
+            self.states_[name] = torch.cuda.get_rng_state()
+            # And set the state to the original state we started with.
+            _set_cuda_rng_state(orig_cuda_rng_state)
+
+
+# RNG tracker object.
+_CUDA_RNG_STATE_TRACKER = CudaRNGStatesTracker()
+
+
+def get_cuda_rng_tracker():
+    """Get cuda rng tracker."""
+    return _CUDA_RNG_STATE_TRACKER
+
+
+def model_parallel_cuda_manual_seed(seed):
+    """Initialize model parallel cuda seed.
+
+    This function should be called after the model parallel is
+    initialized. Also, no torch.cuda.manual_seed should be called
+    after this function. Basically, this is replacement for that
+    function.
+    Two set of RNG states are tracked:
+        default state: This is for data parallelism and is the same among a
+                       set of model parallel GPUs but different across
+                       different model paralle groups. This is used for
+                       example for dropout in the non-tensor-model-parallel regions.
+        tensor-model-parallel state: This state is different among a set of model
+                              parallel GPUs, but the same across data parallel
+                              groups. This is used for example for dropout in
+                              model parallel regions.
+    """
+    # 2718 is just for fun and any POSITIVE value will work.
+    offset = seed + 2718
+    tensor_model_parallel_seed = offset + get_tensor_model_parallel_rank()
+    # Data parallel gets the original seed.
+    data_parallel_seed = seed
+
+    if torch.distributed.get_rank() == 0:
+        print('> initializing model parallel cuda seeds on global rank {}, '
+              'model parallel rank {}, and data parallel rank {} with '
+              'model parallel seed: {} and data parallel seed: {}'.format(
+                  torch.distributed.get_rank(), get_tensor_model_parallel_rank(),
+                  get_data_parallel_rank(), tensor_model_parallel_seed,
+                  data_parallel_seed), flush=True)
+    _CUDA_RNG_STATE_TRACKER.reset()
+    # Set the default state.
+    torch.cuda.manual_seed(data_parallel_seed)
+    # and model parallel state.
+    _CUDA_RNG_STATE_TRACKER.add(_MODEL_PARALLEL_RNG_TRACKER_NAME,
+                                tensor_model_parallel_seed)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    """This function is adapted from torch.utils.checkpoint with
+       two main changes:
+           1) torch.cuda.set_rng_state is replaced with `_set_cuda_rng_state`
+           2) the states in the model parallel tracker are also properly
+              tracked/set/reset.
+    """
+    @staticmethod
+    def forward(ctx, run_function, *args):
+        ctx.run_function = run_function
+
+        # Copy the rng states.
+        ctx.fwd_cpu_rng_state = torch.get_rng_state()
+        ctx.fwd_cuda_rng_state = torch.cuda.get_rng_state()
+        ctx.fwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states()
+
+        with torch.no_grad():
+            outputs = run_function(*args)
+
+        # Divide hidden states across model parallel group and only keep
+        # the chunk corresponding to the current rank.
+        if _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER is not None:
+            ctx.input_0_shape = args[0].data.shape
+            args[0].data = split_tensor_into_1d_equal_chunks(args[0].data)
+            args[0].data = _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER.add(
+                args[0].data)
+
+        # Store everything.
+        ctx.save_for_backward(*args)
+
+
+        return outputs
+
+    @staticmethod
+    def backward(ctx, *args):
+        if not torch.autograd._is_checkpoint_valid():
+            raise RuntimeError("Checkpointing is not compatible with .grad(), "
+                               "please use .backward() if possible")
+        inputs = ctx.saved_tensors
+        if _CHECKPOINTED_ACTIVATIONS_MEMORY_BUFFER is not None:
+            inputs[0].data = gather_split_1d_tensor(inputs[0].data)
+            inputs[0].data = inputs[0].data.view(ctx.input_0_shape)
+
+        # Store the current states.
+        bwd_cpu_rng_state = torch.get_rng_state()
+        bwd_cuda_rng_state = torch.cuda.get_rng_state()
+        bwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states()
+
+        # Set the states to what it used to be before the forward pass.
+        torch.set_rng_state(ctx.fwd_cpu_rng_state)
+        _set_cuda_rng_state(ctx.fwd_cuda_rng_state)
+        get_cuda_rng_tracker().set_states(ctx.fwd_cuda_rng_state_tracker)
+
+        # Compute the forward pass.
+        detached_inputs = detach_variable(inputs)
+        with torch.enable_grad():
+            outputs = ctx.run_function(*detached_inputs)
+
+        # Set the states back to what it was at the start of this function.
+        torch.set_rng_state(bwd_cpu_rng_state)
+        _set_cuda_rng_state(bwd_cuda_rng_state)
+        get_cuda_rng_tracker().set_states(bwd_cuda_rng_state_tracker)
+
+        if isinstance(outputs, torch.Tensor):
+            outputs = (outputs,)
+        torch.autograd.backward(outputs, args)
+        grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else inp
+                      for inp in detached_inputs)
+        return (None,) + grads
+
+
+def checkpoint(function, *args):
+    """Checkpoint a model or part of the model.
+    This has been directly copied from torch.utils.checkpoint."""
+    return CheckpointFunction.apply(function, *args)

megatron/mpu/tests/commons.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import os
+import random
+import numpy
+import torch
+
+import mpu
+
+
+class IdentityLayer(torch.nn.Module):
+    def __init__(self, size, scale=1.0):
+        super(IdentityLayer, self).__init__()
+        self.weight = torch.nn.Parameter(scale * torch.randn(size))
+
+    def forward(self):
+        return self.weight
+
+
+def set_random_seed(seed):
+    """Set random seed for reproducability."""
+    random.seed(seed)
+    numpy.random.seed(seed)
+    torch.manual_seed(seed)
+    mpu.model_parallel_cuda_manual_seed(seed)
+
+
+def initialize_distributed(backend='nccl'):
+    """Initialize torch.distributed."""
+    # Get local rank in case it is provided.
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--local_rank', type=int, default=None,
+                        help='local rank passed from distributed launcher')
+    args = parser.parse_args()
+    local_rank = args.local_rank
+
+    # Get rank and world size.
+    rank = int(os.getenv('RANK', '0'))
+    world_size = int(os.getenv("WORLD_SIZE", '1'))
+
+    print('> initializing torch.distributed with local rank: {}, '
+          'rank: {}, world size: {}'.format(local_rank, rank, world_size))
+
+    # Set the device id.
+    device = rank % torch.cuda.device_count()
+    if local_rank is not None:
+        device = local_rank
+    torch.cuda.set_device(device)
+
+    # Call the init process.
+    init_method = 'tcp://'
+    master_ip = os.getenv('MASTER_ADDR', 'localhost')
+    master_port = os.getenv('MASTER_PORT', '6000')
+    init_method += master_ip + ':' + master_port
+    torch.distributed.init_process_group(
+        backend=backend,
+        world_size=world_size,
+        rank=rank,
+        init_method=init_method)
+
+
+def print_separator(message):
+    torch.distributed.barrier()
+    filler_len = (78 - len(message)) // 2
+    filler = '-' * filler_len
+    string = '\n' + filler + ' {} '.format(message) + filler
+    if torch.distributed.get_rank() == 0:
+        print(string, flush=True)
+    torch.distributed.barrier()

megatron/mpu/tests/test_cross_entropy.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from commons import set_random_seed
+from commons import IdentityLayer
+from commons import print_separator
+from commons import initialize_distributed
+from mpu.cross_entropy import vocab_parallel_cross_entropy
+import mpu
+import torch.nn.functional as F
+import torch
+import random
+import sys
+sys.path.append("../..")
+
+
+def torch_cross_entropy(batch_size, seq_length, vocab_size,
+                        logits_scale, seed):
+    set_random_seed(seed)
+    identity = IdentityLayer((batch_size, seq_length, vocab_size),
+                             scale=logits_scale).cuda()
+    logits = identity()
+    target = torch.cuda.LongTensor(
+        size=(batch_size, seq_length)).random_(0, vocab_size)
+    loss = F.cross_entropy(logits.view(-1, logits.size()[-1]),
+                           target.view(-1),
+                           reduction='none').view_as(target).mean()
+    loss.backward()
+    return loss, identity.weight.grad
+
+
+def mpu_cross_entropy(batch_size, seq_length, vocab_size,
+                      logits_scale, seed):
+    set_random_seed(seed)
+    identity = IdentityLayer((batch_size, seq_length, vocab_size),
+                             scale=logits_scale).cuda()
+    logits = identity()
+    logits_parallel = mpu.scatter_to_tensor_model_parallel_region(logits)
+    target = torch.cuda.LongTensor(
+        size=(batch_size, seq_length)).random_(0, vocab_size)
+    loss = vocab_parallel_cross_entropy(logits_parallel, target).mean()
+    loss.backward()
+    return loss, identity.weight.grad
+
+
+def test_cross_entropy(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing cross entropy with model parallel size {} ...'.
+              format(tensor_model_parallel_size))
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    batch_size = 13
+    seq_length = 17
+    vocab_size_per_partition = 11
+    logits_scale = 1000.0
+    vocab_size = vocab_size_per_partition * tensor_model_parallel_size
+    seed = 1234
+
+    loss_torch, grad_torch = torch_cross_entropy(batch_size, seq_length,
+                                                 vocab_size, logits_scale,
+                                                 seed)
+    loss_mpu, grad_mpu = mpu_cross_entropy(batch_size, seq_length,
+                                           vocab_size, logits_scale,
+                                           seed)
+
+    error = loss_torch.sub_(loss_mpu).abs().max()
+    print('   max error in loss on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    error = grad_torch.sub_(grad_mpu).abs().max()
+    print('   max error in grad on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # Reset groups
+    mpu.destroy_tensor_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+if __name__ == '__main__':
+
+    initialize_distributed()
+    world_size = torch.distributed.get_world_size()
+
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test cross entropy')
+        test_cross_entropy(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2

megatron/mpu/tests/test_data.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from commons import print_separator
+from commons import initialize_distributed
+from mpu import data as data_utils
+import mpu
+import torch
+import functools
+import operator
+import sys
+sys.path.append("../..")
+
+
+def test_broadcast_data(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing broadcast_data with model parallel size {} ...'.
+              format(tensor_model_parallel_size))
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    torch.manual_seed(1234 + mpu.get_data_parallel_rank())
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    key_size_t = {'key1': [7, 11],
+                  'key2': [8, 2, 1],
+                  'key3': [13],
+                  'key4': [5, 1, 2],
+                  'key5': [5, 12]}
+    keys = list(key_size_t.keys())
+
+    data = {}
+    data_t = {}
+    for key in key_size_t:
+        data[key] = torch.LongTensor(size=key_size_t[key]).random_(0, 1000)
+        data_t[key] = data[key].clone()
+    data['keyX'] = torch.FloatTensor(size=(5, )).random_(0, 1000)
+    data_t['keyX'] = data['keyX'].clone()
+    if mpu.get_tensor_model_parallel_rank() != 0:
+        data = None
+
+    data_utils._check_data_types(keys, data_t, torch.int64)
+    key_size, key_numel, \
+        total_numel = data_utils._build_key_size_numel_dictionaries(keys, data)
+    for key in keys:
+        assert key_size[key] == key_size_t[key]
+    total_numel_t = 0
+    for key in keys:
+        target_size = functools.reduce(operator.mul, key_size_t[key], 1)
+        assert key_numel[key] == target_size
+        total_numel_t += target_size
+    assert total_numel == total_numel_t
+
+    data_b = data_utils.broadcast_data(keys, data, torch.int64)
+    for key in keys:
+        tensor = data_t[key].cuda()
+        assert data_b[key].sub(tensor).abs().max() == 0
+
+    # Reset groups
+    mpu.destroy_tensor_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+if __name__ == '__main__':
+
+    initialize_distributed()
+    world_size = torch.distributed.get_world_size()
+
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test test broadcast data')
+        test_broadcast_data(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2

megatron/mpu/tests/test_initialize.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from commons import print_separator
+from commons import initialize_distributed
+import mpu
+import torch
+import sys
+sys.path.append("../..")
+
+
+def test_initialize_model_parallel(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing initialize_model_parallel with size {} ...'.format(
+            tensor_model_parallel_size))
+    tensor_model_parallel_size_ = min(tensor_model_parallel_size,
+                               torch.distributed.get_world_size())
+    assert not mpu.model_parallel_is_initialized()
+    mpu.initialize_model_parallel(tensor_model_parallel_size_)
+    assert mpu.model_parallel_is_initialized()
+
+    # Checks.
+    def check(group, world_size, rank):
+        assert world_size == torch.distributed.get_world_size(group=group)
+        assert rank == torch.distributed.get_rank(group=group)
+
+    # Model parallel.
+    world_size = tensor_model_parallel_size_
+    rank = torch.distributed.get_rank() % tensor_model_parallel_size_
+    assert world_size == mpu.get_tensor_model_parallel_world_size()
+    assert rank == mpu.get_tensor_model_parallel_rank()
+    check(mpu.get_tensor_model_parallel_group(), world_size, rank)
+
+    # Data parallel.
+    world_size = torch.distributed.get_world_size() // tensor_model_parallel_size_
+    rank = torch.distributed.get_rank() // tensor_model_parallel_size
+    assert world_size == mpu.get_data_parallel_world_size()
+    assert rank == mpu.get_data_parallel_rank()
+    check(mpu.get_data_parallel_group(), world_size, rank)
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+def test_get_tensor_model_parallel_src_rank(tensor_model_parallel_size_):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing get_tensor_model_parallel_src_rank with size {} ...'.format(
+            tensor_model_parallel_size_))
+    tensor_model_parallel_size = min(tensor_model_parallel_size_,
+                              torch.distributed.get_world_size())
+    assert not mpu.model_parallel_is_initialized()
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    assert mpu.model_parallel_is_initialized()
+
+    # Checks
+    src_rank = torch.distributed.get_rank() - mpu.get_tensor_model_parallel_rank()
+    assert mpu.get_tensor_model_parallel_src_rank() == src_rank
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+if __name__ == '__main__':
+
+    initialize_distributed()
+    world_size = torch.distributed.get_world_size()
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test initialize model parallel')
+        test_initialize_model_parallel(tensor_model_parallel_size)
+        print_separator('test model parallel source rank')
+        test_get_tensor_model_parallel_src_rank(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2

megatron/mpu/tests/test_layers.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from mpu import layers
+from commons import set_random_seed
+from commons import print_separator
+from commons import initialize_distributed
+import mpu
+from torch.nn.parameter import Parameter
+import torch.nn.init as init
+import torch
+import random
+import sys
+sys.path.append("../..")
+
+
+def test_parallel_embedding(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing parallel embedding with model parallel size {} ...'.
+              format(tensor_model_parallel_size))
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    batch_size = 17
+    seq_length = 23
+    vocab_size = 48
+    hidden_size = 16
+    seed = 1236
+
+    set_random_seed(123)
+    input_data = torch.LongTensor(
+        size=(batch_size, seq_length)).random_(0, vocab_size).cuda()
+    loss_weight = torch.randn([batch_size, seq_length, hidden_size]).cuda()
+
+    set_random_seed(seed)
+    embedding_original = torch.nn.Embedding(vocab_size, hidden_size).cuda()
+
+    output = embedding_original(input_data)
+    loss_original = torch.mul(output, loss_weight).sum()
+    loss_original.backward()
+
+    set_random_seed(seed)
+    embedding_parallel = layers.ParallelEmbedding(
+        vocab_size, hidden_size, init_method=init.normal_).cuda()
+    output = embedding_parallel(input_data)
+    loss_parallel = torch.mul(output, loss_weight).sum()
+    loss_parallel.backward()
+
+    set_random_seed(seed)
+    embedding_vocab_parallel = layers.VocabParallelEmbedding(
+        vocab_size, hidden_size, init_method=init.normal_).cuda()
+    output = embedding_vocab_parallel(input_data)
+    loss_vocab_parallel = torch.mul(output, loss_weight).sum()
+    loss_vocab_parallel.backward()
+
+    torch.distributed.barrier()
+    error = loss_parallel.sub(loss_original).abs()
+    print('   error in loss (parallel) on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-12, 'error: {}'.format(error)
+
+    torch.distributed.barrier()
+    error = loss_vocab_parallel.sub(loss_original).abs()
+    print('   error in loss (vocab parallel) on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-12, 'error: {}'.format(error)
+
+    weight_grad_orig = torch.split(embedding_original.weight.grad,
+                                   hidden_size // tensor_model_parallel_size,
+                                   1)[mpu.get_tensor_model_parallel_rank()]
+    error = embedding_parallel.weight.grad.sub(weight_grad_orig).abs().max()
+    print('   error in grad (parallel) on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-12, 'error: {}'.format(error)
+
+    weight_grad_orig = torch.split(embedding_original.weight.grad,
+                                   vocab_size // tensor_model_parallel_size,
+                                   0)[mpu.get_tensor_model_parallel_rank()]
+    error = embedding_vocab_parallel.weight.grad.sub(
+        weight_grad_orig).abs().max()
+    print('   error in grad (vocab parallel) on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-12, 'error: {}'.format(error)
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+def test_initialize_affine_weight(tensor_model_parallel_size):
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    if torch.distributed.get_rank() == 0:
+        print('> testing initialize_affine_weight with model parallel '
+              'size: {}'.format(tensor_model_parallel_size))
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    seed = 12345
+    input_size_coeff = 13
+    input_size = input_size_coeff * tensor_model_parallel_size
+    output_size_coeff = 17
+    output_size = output_size_coeff * tensor_model_parallel_size
+
+    # ---------------
+    # Column parallel
+    # ---------------
+    weight = torch.empty(output_size_coeff, input_size)
+    set_random_seed(seed)
+    layers._initialize_affine_weight(weight, output_size, input_size,
+
+                                     output_size_coeff, 0,
+                                     torch.nn.init.normal_)
+    # Target.
+    set_random_seed(seed)
+    master_weight = torch.empty(output_size, input_size)
+    torch.nn.init.normal_(master_weight)
+    rank = mpu.get_tensor_model_parallel_rank()
+    my_weight = torch.split(master_weight, output_size_coeff,
+                            dim=0)[rank].contiguous().clone()
+
+    # Compare.
+    error = weight.sub(my_weight).abs().max()
+    torch.distributed.barrier()
+    print('   column parallel max error (should be zero) on global rank '
+          '{}: {}'.format(torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # ------------
+    # Row parallel
+    # ------------
+    weight = torch.empty(output_size, input_size_coeff)
+    set_random_seed(seed)
+    mpu.layers._initialize_affine_weight(weight, output_size, input_size,
+                                         input_size_coeff, 1,
+                                         torch.nn.init.normal_)
+    # Target.
+    set_random_seed(seed)
+    master_weight = torch.empty(output_size, input_size)
+    torch.nn.init.normal_(master_weight)
+    rank = mpu.get_tensor_model_parallel_rank()
+    my_weight = torch.split(master_weight, input_size_coeff,
+                            dim=1)[rank].contiguous().clone()
+
+    # Compare.
+    error = weight.sub(my_weight).abs().max()
+    torch.distributed.barrier()
+    print('   row parallel max error (should be zero) on global rank '
+          '{}: {}'.format(torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print(' >> passed the test :-)')
+
+
+class IdentityLayer2D(torch.nn.Module):
+    def __init__(self, m, n):
+        super(IdentityLayer2D, self).__init__()
+        self.weight = Parameter(torch.Tensor(m, n))
+        torch.nn.init.xavier_normal_(self.weight)
+
+    def forward(self):
+        return self.weight
+
+
+def test_column_parallel_linear(tensor_model_parallel_size):
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    if torch.distributed.get_rank() == 0:
+        print('> testing ColumnParallelLinear with model parallel '
+              'size: {}'.format(tensor_model_parallel_size))
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    seed = 12345
+    set_random_seed(seed)
+    input_size_coeff = 13
+    input_size = input_size_coeff * tensor_model_parallel_size
+    output_size_coeff = 17
+    output_size = output_size_coeff * tensor_model_parallel_size
+    batch_size = 7
+
+    # Network
+    identity_layer = IdentityLayer2D(batch_size, input_size).cuda()
+    linear_layer = mpu.ColumnParallelLinear(
+        input_size, output_size, keep_master_weight_for_test=True).cuda()
+    loss_weight = torch.randn([batch_size, output_size]).cuda()
+    # Forward
+    input_ = identity_layer()
+    output = linear_layer(input_)
+    loss = torch.mul(output, loss_weight).sum()
+    # Backward
+    loss.backward()
+
+    # Values.
+    dLdY = loss_weight
+    X = identity_layer.weight
+    A = linear_layer.master_weight.cuda()
+    dLdA = torch.matmul(dLdY.t(), X)
+    dLdb = torch.matmul(torch.ones(batch_size, 1).cuda().t(), dLdY).view(-1)
+    dLdX = torch.matmul(dLdY, A)
+
+    rank = mpu.get_tensor_model_parallel_rank()
+    my_dLdA = torch.split(dLdA, output_size_coeff,
+                          dim=0)[rank].contiguous().clone()
+    error = my_dLdA.sub(linear_layer.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   error in dLdA on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    my_dLdb = torch.split(dLdb, output_size_coeff,
+                          dim=0)[rank].contiguous().clone()
+    error = my_dLdb.sub(linear_layer.bias.grad).abs().max()
+    torch.distributed.barrier()
+    print('   error in dLdb on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    error = dLdX.sub(identity_layer.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   error in dLdX on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print(' >> passed the test :-)')
+
+
+def test_row_parallel_linear(tensor_model_parallel_size):
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    if torch.distributed.get_rank() == 0:
+        print('> testing RowParallelLinear with model parallel '
+              'size: {}'.format(tensor_model_parallel_size))
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    seed = 12345
+    set_random_seed(seed)
+    input_size_coeff = 13
+    input_size = input_size_coeff * tensor_model_parallel_size
+    output_size_coeff = 17
+    output_size = output_size_coeff * tensor_model_parallel_size
+    batch_size = 7
+
+    # Network
+    identity_layer = IdentityLayer2D(batch_size, input_size).cuda()
+    linear_layer = mpu.RowParallelLinear(
+        input_size, output_size, keep_master_weight_for_test=True).cuda()
+    loss_weight = torch.randn([batch_size, output_size]).cuda()
+    # Forward
+    input_ = identity_layer()
+    output = linear_layer(input_)
+    loss = torch.mul(output, loss_weight).sum()
+    # Backward
+    loss.backward()
+
+    # Values.
+    dLdY = loss_weight
+    X = identity_layer.weight
+    A = linear_layer.master_weight.cuda()
+    dLdA = torch.matmul(dLdY.t(), X)
+    dLdb = torch.matmul(torch.ones(batch_size, 1).cuda().t(), dLdY).view(-1)
+    dLdX = torch.matmul(dLdY, A)
+
+    rank = mpu.get_tensor_model_parallel_rank()
+    my_dLdA = torch.split(dLdA, input_size_coeff,
+                          dim=1)[rank].contiguous().clone()
+    error = my_dLdA.sub(linear_layer.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   error in dLdA on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    error = dLdb.sub(linear_layer.bias.grad).abs().max()
+    torch.distributed.barrier()
+    print('   error in dLdb on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    error = dLdX.sub(identity_layer.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   error in dLdX on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print(' >> passed the test :-)')
+
+
+class IdentityLayer3D(torch.nn.Module):
+    def __init__(self, m, n, k):
+        super(IdentityLayer3D, self).__init__()
+        self.weight = Parameter(torch.Tensor(m, n, k))
+        torch.nn.init.xavier_normal_(self.weight)
+
+    def forward(self):
+        return self.weight
+
+
+def parallel_self_attention(tensor_model_parallel_size, num_att_heads_per_partition,
+                            hidden_size_per_att_head, dropout_prob, batch_size,
+                            sequence_length):
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    seed = 12345
+    set_random_seed(seed)
+
+    num_att_heads = num_att_heads_per_partition * \
+        torch.distributed.get_world_size()
+    hidden_size = hidden_size_per_att_head * num_att_heads
+
+    # Network
+    identity_layer = IdentityLayer3D(batch_size, sequence_length,
+                                     hidden_size).cuda()
+    attention_layer = mpu.BertParallelSelfAttention(hidden_size, num_att_heads,
+                                                    dropout_prob).cuda()
+    loss_weight = torch.randn([batch_size, sequence_length, hidden_size]).cuda()
+    attention_mask = torch.randn([batch_size, 1, 1, sequence_length]).cuda()
+    # Forward
+    input_ = identity_layer()
+    output = attention_layer(input_, attention_mask)
+    loss = torch.mul(output, loss_weight).sum()
+    # Backward
+    loss.backward()
+
+    rank = mpu.get_tensor_model_parallel_rank()
+    mpu.destroy_model_parallel()
+    return rank, hidden_size, tensor_model_parallel_size, loss, \
+        attention_layer, identity_layer
+
+
+def test_parallel_self_attention(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing ParallelSelfAttention with model parallel '
+              'size: {}'.format(tensor_model_parallel_size))
+
+    num_att_heads_per_partition = 3
+    hidden_size_per_att_head = 7
+    dropout_prob = 0.0  # has to be zero
+    batch_size = 5
+    sequence_length = 13
+
+    rank_1, hideen_size_1, tensor_model_parallel_size_1, loss_1, \
+        attention_layer_1, identity_layer_1 = parallel_self_attention(
+            1, num_att_heads_per_partition,
+            hidden_size_per_att_head, dropout_prob, batch_size, sequence_length)
+
+    rank, hidden_size, tensor_model_parallel_size, loss, \
+        attention_layer, identity_layer = parallel_self_attention(
+            tensor_model_parallel_size, num_att_heads_per_partition,
+            hidden_size_per_att_head, dropout_prob, batch_size, sequence_length)
+    assert hideen_size_1 == hidden_size
+
+    error = loss_1.sub(loss).abs().max()
+    torch.distributed.barrier()
+    print('   loss error on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 5.0e-6
+
+    my_lin_grad_list = torch.split(
+        attention_layer_1.query_key_value.weight.grad,
+        hidden_size // tensor_model_parallel_size, 0)[rank::tensor_model_parallel_size]
+    my_lin_grad = torch.cat(my_lin_grad_list, dim=0)
+    error = my_lin_grad.sub(
+        attention_layer.query_key_value.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   weight gradient error on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 5.0e-6
+
+    error = identity_layer_1.weight.grad.sub(
+        identity_layer.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   input gradient error on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 5.0e-6
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print(' >> passed the test :-)')
+
+
+def parallel_transformer(tensor_model_parallel_size, num_att_heads_per_partition,
+                         hidden_size_per_att_head, batch_size, sequence_length):
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    seed = 12345
+    set_random_seed(seed)
+
+    num_att_heads = num_att_heads_per_partition * \
+        torch.distributed.get_world_size()
+    hidden_size = hidden_size_per_att_head * num_att_heads
+    intermediate_size = 4 * hidden_size
+
+    # Network
+    identity_layer = IdentityLayer3D(batch_size, sequence_length,
+                                     hidden_size).cuda()
+    transformer_layer = mpu.BertParallelTransformerLayer(
+        hidden_size, intermediate_size, num_att_heads, 0.0, 0.0,
+        torch.nn.functional.relu, 1.0e-5).cuda()
+
+    loss_weight = torch.randn([batch_size, sequence_length, hidden_size]).cuda()
+    attention_mask = torch.randn([batch_size, 1, 1, sequence_length]).cuda()
+    # Forward
+    input_ = identity_layer()
+    output = transformer_layer(input_, attention_mask)
+    loss = torch.mul(output, loss_weight).sum()
+    # Backward
+    loss.backward()
+
+    rank = mpu.get_tensor_model_parallel_rank()
+    mpu.destroy_model_parallel()
+    return rank, hidden_size, tensor_model_parallel_size, loss, \
+        transformer_layer, identity_layer
+
+
+def test_parallel_transformer_layer(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing ParallelTransformerLayer with model parallel '
+              'size: {}'.format(tensor_model_parallel_size))
+
+    num_att_heads_per_partition = 3
+    hidden_size_per_att_head = 7
+    batch_size = 5
+    sequence_length = 13
+
+    rank_1, hidden_size_1, tensor_model_parallel_size_1, loss_1, \
+        transformer_layer_1, identity_layer_1 = parallel_transformer(
+            1, num_att_heads_per_partition,
+            hidden_size_per_att_head, batch_size, sequence_length)
+
+    rank, hidden_size, tensor_model_parallel_size, loss, \
+        transformer_layer, identity_layer = parallel_transformer(
+            tensor_model_parallel_size, num_att_heads_per_partition,
+            hidden_size_per_att_head, batch_size, sequence_length)
+
+    error = loss_1.sub(loss).abs().max()
+    torch.distributed.barrier()
+    print('   loss error on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 5.0e-5, 'error: {}'.format(error)
+
+    error = identity_layer_1.weight.grad.sub(
+        identity_layer.weight.grad).abs().max()
+    torch.distributed.barrier()
+    print('   input gradient error on global rank {}: {}'.format(
+        torch.distributed.get_rank(), error))
+    assert error < 5.0e-5, 'error: {}'.format(error)
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print(' >> passed the test :-)')
+
+
+if __name__ == '__main__':
+
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+
+    initialize_distributed()
+    world_size = torch.distributed.get_world_size()
+
+    print_separator('test initialize affine weight')
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        test_initialize_affine_weight(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test parallel embedding')
+        test_parallel_embedding(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    print_separator('test column-parallel linear')
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        test_column_parallel_linear(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    print_separator('test row-parallel linear')
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        test_row_parallel_linear(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    print_separator('test parallel self-attention')
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        test_parallel_self_attention(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    print_separator('test parallel transformer')
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        test_parallel_transformer_layer(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2

megatron/mpu/tests/test_random.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from commons import print_separator
+from commons import initialize_distributed
+import mpu
+import torch
+import sys
+sys.path.append("../..")
+
+
+def test_set_cuda_rng_state(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing set_rng_state with size {} ...'.
+              format(tensor_model_parallel_size))
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    size = 123
+    seed = 1234
+    torch.cuda.manual_seed(1234)
+    tensor = torch.cuda.FloatTensor(size)
+
+    # Get the state
+    rng_state = torch.cuda.get_rng_state()
+    rng_state_copy = rng_state.clone()
+
+    # Do some stuff.
+    for _ in range(5):
+        torch.randn(size, out=tensor)
+    result_1 = tensor.clone()
+
+    assert rng_state.sub(rng_state_copy).max() == 0
+    assert torch.cuda.get_rng_state().sub(rng_state_copy).max() > 0
+
+    # State should be different.
+    new_rng_state = torch.cuda.get_rng_state()
+    max_diff = new_rng_state.sub(rng_state).max()
+    print('   max diff in rng state (should be non-zero) on global rank {}: {}'.
+          format(torch.distributed.get_rank(), max_diff))
+    assert max_diff > 0
+
+    # Reset the rng state and do the same stuff.
+    mpu.random._set_cuda_rng_state(rng_state)
+    for _ in range(5):
+        torch.randn(size, out=tensor)
+    mpu.random._set_cuda_rng_state(rng_state)
+    for _ in range(5):
+        torch.randn(size, out=tensor)
+    result_2 = tensor.clone()
+
+    # Results should be the same
+    error = result_2.sub(result_1).abs().max()
+    print('   max error in generated tensors (should be zero) on '
+          'global rank {}: {}'.format(torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # Input state should have remained intact.
+    error = rng_state.sub(rng_state_copy).max()
+    print('   max error in rng state (should be zero) on global rank {}: {}'.
+          format(torch.distributed.get_rank(), error))
+    assert error == 0
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+def test_cuda_rng_tracker(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing cuda rng tracker with size {} ...'.
+              format(tensor_model_parallel_size))
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    seed_1 = 1234
+    seed_2 = 4321
+    size = [12, 21]
+    tensor = torch.cuda.FloatTensor(size)
+
+    # Set to seed_1 and generate two tensors.
+    torch.cuda.manual_seed(seed_1)
+    torch.randn(size, out=tensor)
+    target_11 = tensor.clone()
+    torch.randn(size, out=tensor)
+    target_12 = tensor.clone()
+
+    # Set to seed_2 and generate two tensors.
+    torch.cuda.manual_seed(seed_2)
+    torch.randn(size, out=tensor)
+    target_21 = tensor.clone()
+    torch.randn(size, out=tensor)
+    target_22 = tensor.clone()
+
+    # Now if we interleave seed_1 and seed_2,
+    # we should still get the same tensors
+    torch.cuda.manual_seed(seed_1)
+    mpu.get_cuda_rng_tracker().add('test', seed_2)
+
+    torch.randn(size, out=tensor)
+    result_11 = tensor.clone()
+
+    with mpu.get_cuda_rng_tracker().fork('test'):
+        torch.randn(size, out=tensor)
+        result_21 = tensor.clone()
+
+    torch.randn(size, out=tensor)
+    result_12 = tensor.clone()
+
+    with mpu.get_cuda_rng_tracker().fork('test'):
+        torch.randn(size, out=tensor)
+        result_22 = tensor.clone()
+
+    diff = result_11.sub(result_21).abs().max()
+    diff = min(diff, result_12.sub(result_22).abs().max())
+    print('   max diff in generated tensors (should be non-zero) on '
+          'global rank {}: {}'.format(torch.distributed.get_rank(), diff))
+    assert diff > 1.0e-6
+    error = max(result_11.sub(target_11).abs().max(),
+                result_12.sub(target_12).abs().max())
+    error = max(error, result_21.sub(target_21).abs().max())
+    error = max(error, result_22.sub(target_22).abs().max())
+    print('   max error in generated tensors (should be zero) on '
+          'global rank {}: {}'.format(torch.distributed.get_rank(), error))
+    assert error < 1.0e-6
+
+    # Reset the tracker
+    mpu.get_cuda_rng_tracker().reset()
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+def test_model_parallel_cuda_manual_seed(tensor_model_parallel_size):
+
+    if torch.distributed.get_rank() == 0:
+        print('> testing model parallel cuda manual seed with size {} ...'.
+              format(tensor_model_parallel_size))
+
+    mpu.initialize_model_parallel(tensor_model_parallel_size)
+    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()
+
+    mpu.model_parallel_cuda_manual_seed(12345)
+    assert torch.cuda.initial_seed() == 12345
+    with mpu.get_cuda_rng_tracker().fork():
+        assert torch.cuda.initial_seed() == (12345 + 2718 +
+                                             mpu.get_tensor_model_parallel_rank())
+
+    # Reset the tracker
+    mpu.get_cuda_rng_tracker().reset()
+
+    # Reset groups
+    mpu.destroy_model_parallel()
+
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>> passed the test :-)')
+
+
+if __name__ == '__main__':
+
+    initialize_distributed()
+    world_size = torch.distributed.get_world_size()
+
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test set rng state')
+        test_set_cuda_rng_state(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test cuda rng tracker')
+        test_cuda_rng_tracker(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2
+
+    tensor_model_parallel_size = 1
+    while tensor_model_parallel_size <= world_size:
+        print_separator('test model parallel cuda manual seed')
+        test_model_parallel_cuda_manual_seed(tensor_model_parallel_size)
+        tensor_model_parallel_size *= 2

megatron/mpu/utils.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+
+
+def ensure_divisibility(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator."""
+    assert numerator % denominator == 0, '{} is not divisible by {}'.format(
+        numerator, denominator)
+
+
+def divide(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator and return
+    the division value."""
+    ensure_divisibility(numerator, denominator)
+    return numerator // denominator
+
+
+def split_tensor_along_last_dim(tensor, num_partitions,
+                                contiguous_split_chunks=False):
+    """Split a tensor along its last dimension.
+    Arguments:
+        tensor: input tensor.
+        num_partitions: number of partitions to split the tensor
+        contiguous_split_chunks: If True, make each chunk contiguous
+                                 in memory.
+    """
+    # Get the size and dimension.
+    last_dim = tensor.dim() - 1
+    last_dim_size = divide(tensor.size()[last_dim], num_partitions)
+    # Split.
+    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+    # Note: torch.split does not create contiguous tensors by default.
+    if contiguous_split_chunks:
+        return tuple(chunk.contiguous() for chunk in tensor_list)
+
+    return tensor_list
+
+
+class VocabUtility:
+    """Split the vocabulary into `world_size` chunks amd return the
+        first and last index of the vocabulary belonging to the `rank`
+        partition: Note that indecies in [fist, last)"""
+
+    @staticmethod
+    def vocab_range_from_per_partition_vocab_size(per_partition_vocab_size,
+                                                  rank, world_size):
+        index_f = rank * per_partition_vocab_size
+        index_l = index_f + per_partition_vocab_size
+        return index_f, index_l
+
+    @staticmethod
+    def vocab_range_from_global_vocab_size(global_vocab_size, rank, world_size):
+        per_partition_vocab_size = divide(global_vocab_size, world_size)
+        return VocabUtility.vocab_range_from_per_partition_vocab_size(
+            per_partition_vocab_size, rank, world_size)

megatron/optimizer/__init__.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from apex.optimizers import FusedAdam as Adam
+from apex.optimizers import FusedSGD as SGD
+
+from megatron import get_args
+from megatron.model.fused_layer_norm import MixedFusedLayerNorm as LayerNorm
+
+from .grad_scaler import ConstantGradScaler, DynamicGradScaler
+from .optimizer import Float16OptimizerWithFloat16Params, FP32Optimizer
+
+
+def _get_params_for_weight_decay_optimization(modules):
+    """Divide params into with-weight-decay and without-weight-decay groups.
+    Layernorms and baises will have no weight decay but the rest will.
+    """
+
+    weight_decay_params = {'params': []}
+    no_weight_decay_params = {'params': [], 'weight_decay': 0.0}
+    for module in modules:
+        for module_ in module.modules():
+            if isinstance(module_, LayerNorm):
+                no_weight_decay_params['params'].extend(
+                    [p for p in list(module_._parameters.values())
+                     if p is not None])
+            else:
+                weight_decay_params['params'].extend(
+                    [p for n, p in list(module_._parameters.items())
+                     if p is not None and n != 'bias'])
+                no_weight_decay_params['params'].extend(
+                    [p for n, p in list(module_._parameters.items())
+                     if p is not None and n == 'bias'])
+
+    # XXX: temp hack to workaround the crash in apex FusedAdam's multi_tensor_applier
+    #
+    # it crashes when the param count is larger than a certain size which we hit at 200B over 80
+    # A100 gpus - I think around 2.7B per gpu, so halving it works around the issue
+    param_count = len(weight_decay_params['params'])
+    first_half = weight_decay_params['params'][:param_count // 2]
+    second_half = weight_decay_params['params'][param_count // 2:]
+
+    first_half =  { 'params': first_half }
+    second_half = { 'params': second_half }
+
+    return first_half, second_half, no_weight_decay_params
+
+    #return weight_decay_params, no_weight_decay_params
+
+
+def get_megatron_optimizer(model):
+    args = get_args()
+
+    if args.cpu_optimizer:
+        raise NotImplementedError('need to add cpu adam')
+
+    # Base optimizer.
+    param_groups = _get_params_for_weight_decay_optimization(model)
+    if args.optimizer == 'adam':
+        if args.use_bnb_optimizer:
+            import bitsandbytes as bnb
+            adam_optimizer = bnb.optim.Adam8bit
+        else:
+            adam_optimizer = Adam
+        optimizer = adam_optimizer(param_groups,
+                                   lr=args.lr,
+                                   weight_decay=args.weight_decay,
+                                   betas=(args.adam_beta1, args.adam_beta2),
+                                   eps=args.adam_eps)
+    elif args.optimizer == 'sgd':
+        optimizer = SGD(param_groups,
+                        lr=args.lr,
+                        weight_decay=args.weight_decay,
+                        momentum=args.sgd_momentum)
+    else:
+        raise Exception('{} optimizer is not supported.'.format(
+            args.optimizer))
+
+    if args.deepspeed:
+        return optimizer
+
+    # Determine whether the params have main-grad field.
+    params_have_main_grad = False
+    if args.DDP_impl == 'local':
+        params_have_main_grad = True
+
+    if args.fp16 or args.bf16:
+
+        # Grad scaler:
+        #    if loss-scale is provided, instantiate the constant scaler.
+        #    if we are using fp16 and loss-scale is not present, use a
+        #       dynamic scaler.
+        #    otherwise we are running in bf16 with no loss-scale so
+        #       leave it as None.
+        grad_scaler = None
+        # Constant loss scale.
+        if args.loss_scale:
+            grad_scaler = ConstantGradScaler(args.loss_scale)
+        # Dynamic loss scale.
+        else:
+            if args.fp16:
+                grad_scaler = DynamicGradScaler(
+                    initial_scale=args.initial_loss_scale,
+                    min_scale=args.min_loss_scale,
+                    growth_factor=2.0,
+                    backoff_factor=0.5,
+                    growth_interval=args.loss_scale_window,
+                    hysteresis=args.hysteresis)
+
+        # Megatron optimizer.
+        return Float16OptimizerWithFloat16Params(optimizer,
+                                                 args.clip_grad,
+                                                 args.log_num_zeros_in_grad,
+                                                 params_have_main_grad,
+                                                 args.bf16,
+                                                 grad_scaler)
+
+    # FP32.
+    return FP32Optimizer(optimizer, args.clip_grad,
+                         args.log_num_zeros_in_grad,
+                         params_have_main_grad)

megatron/optimizer/clip_grads.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Gradient clipping."""
+
+import torch
+from torch._six import inf
+
+from apex.multi_tensor_apply import multi_tensor_applier
+import amp_C
+
+from megatron import mpu
+from megatron.model.module import param_is_not_shared
+from megatron.mpu.layers import param_is_not_tensor_parallel_duplicate
+
+
+def clip_grad_norm_fp32(parameters, max_norm, norm_type=2):
+    """Clips gradient norm of an iterable of parameters whose gradients
+       are in fp32.
+
+    This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and
+    added functionality to handle model parallel parameters. Note that
+    the gradients are modified in place.
+
+    Arguments:
+        parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a
+            single Tensor that will have gradients normalized
+        max_norm (float or int): max norm of the gradients
+        norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for
+            infinity norm.
+
+    Returns:
+        Total norm of the parameters (viewed as a single vector).
+    """
+
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+
+    # Filter parameters based on:
+    #   - grad should not be none
+    #   - parameter should not be shared
+    #   - should not be a replica due to tensor model parallelism
+    grads = []
+    grads_for_norm = []
+    for param in parameters:
+        grad_not_none = param.grad is not None
+        is_not_shared = param_is_not_shared(param)
+        is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
+        grad = param.grad.detach()
+        if grad_not_none:
+            # Make sure the grads are in fp32
+            assert param.grad.type() == 'torch.cuda.FloatTensor'
+            grads.append(grad)
+        if grad_not_none and is_not_shared and is_not_tp_duplicate:
+            grads_for_norm.append(grad)
+
+    # Norm parameters.
+    max_norm = float(max_norm)
+    norm_type = float(norm_type)
+    total_norm = 0.0
+
+    # Calculate norm.
+    if norm_type == inf:
+        total_norm = max(grad.abs().max() for grad in grads_for_norm)
+        total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)])
+        # Take max across all model-parallel GPUs.
+        torch.distributed.all_reduce(total_norm_cuda,
+                                     op=torch.distributed.ReduceOp.MAX,
+                                     group=mpu.get_model_parallel_group())
+        total_norm = total_norm_cuda[0].item()
+
+    else:
+        if norm_type == 2.0:
+            dummy_overflow_buf = torch.cuda.IntTensor([0])
+            # Use apex's multi-tensor applier for efficiency reasons.
+            # Multi-tensor applier takes a function and a list of list
+            # and performs the operation on that list all in one kernel.
+            grad_norm, _ = multi_tensor_applier(
+                amp_C.multi_tensor_l2norm,
+                dummy_overflow_buf,
+                [grads_for_norm],
+                False # no per-parameter norm
+            )
+            # Since we will be summing across data parallel groups,
+            # we need the pow(norm-type).
+            total_norm = grad_norm ** norm_type
+
+        else:
+            for grad in grads_for_norm:
+                grad_norm = torch.norm(grad, norm_type)
+                total_norm += grad_norm ** norm_type
+
+        # Sum across all model-parallel GPUs.
+        torch.distributed.all_reduce(total_norm,
+                                     op=torch.distributed.ReduceOp.SUM,
+                                     group=mpu.get_model_parallel_group())
+        total_norm = total_norm.item() ** (1.0 / norm_type)
+
+    # Scale.
+    clip_coeff = max_norm / (total_norm + 1.0e-6)
+    if clip_coeff < 1.0:
+        dummy_overflow_buf = torch.cuda.IntTensor([0])
+        multi_tensor_applier(amp_C.multi_tensor_scale,
+                             dummy_overflow_buf,
+                             [grads, grads],
+                             clip_coeff)
+
+    return total_norm
+
+
+def count_zeros_fp32(parameters):
+
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+
+    # Filter parameters based on:
+    #   - grad should not be none
+    #   - parameter should not be shared
+    #   - should not be a replica due to tensor model parallelism
+    total_num_zeros = 0.0
+    for param in parameters:
+        grad_not_none = param.grad is not None
+        is_not_shared = param_is_not_shared(param)
+        is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
+        if grad_not_none and is_not_shared and is_not_tp_duplicate:
+            grad = param.grad.detach()
+            num_zeros = grad.numel() - torch.count_nonzero(grad)
+            total_num_zeros = num_zeros + total_num_zeros
+
+    # Sum across all model-parallel GPUs.
+    torch.distributed.all_reduce(total_num_zeros,
+                                 op=torch.distributed.ReduceOp.SUM,
+                                 group=mpu.get_model_parallel_group())
+    total_num_zeros = total_num_zeros.item()
+
+    return total_num_zeros