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

megatron/model/vision/mit_backbone.py

+# ---------------------------------------------------------------
+# Copyright (c) 2021, NVIDIA Corporation. All rights reserved.
+#
+# This work is licensed under the NVIDIA Source Code License
+# found in the LICENSE file in the root directory of this 
+# source tree.
+# ---------------------------------------------------------------
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+from torch.nn.init import trunc_normal_
+from megatron.model.transformer import DropPath
+from megatron.model import LayerNorm
+
+
+class Mlp(nn.Module):
+    def __init__(self,
+                 in_features,
+                 hidden_features=None,
+                 out_features=None,
+                 act_layer=nn.GELU,
+                 drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self,
+                 dim,
+                 num_heads=8,
+                 qkv_bias=False,
+                 qk_scale=None,
+                 attn_drop=0.,
+                 proj_drop=0.,
+                 sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = (img_size, img_size)
+        patch_size = (patch_size, patch_size)
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class MixVisionTransformer(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], output_avg=False):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+        self.output_avg = output_avg
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_chans=in_chans,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_chans=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_chans=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_chans=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        if not self.output_avg:
+            x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+    def forward(self, x):
+        x = self.forward_features(x)
+    
+        if self.output_avg:
+            x = x[3].mean(dim=1)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W)
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+class mit_b0(MixVisionTransformer):
+    def __init__(self, **kwargs):
+        super(mit_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+class mit_b1(MixVisionTransformer):
+    def __init__(self, **kwargs):
+        super(mit_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+class mit_b2(MixVisionTransformer):
+    def __init__(self, **kwargs):
+        super(mit_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+ 
+class mit_b3(MixVisionTransformer):
+    def __init__(self, **kwargs):
+        super(mit_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+class mit_b3_avg(MixVisionTransformer):
+    def __init__(self, drop_path_rate=0.1, **kwargs):
+        super(mit_b3_avg, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=drop_path_rate, output_avg=True)
+
+class mit_b4(MixVisionTransformer):
+    def __init__(self, **kwargs):
+        super(mit_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+class mit_b5(MixVisionTransformer):
+    def __init__(self, **kwargs):
+        super(mit_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+class mit_b5_avg(MixVisionTransformer):
+    def __init__(self, drop_path_rate=0.1, **kwargs):
+        super(mit_b5_avg, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=drop_path_rate, output_avg=True)
+

megatron/model/vision/swin_backbone.py

+# Copyright (c) 2021 Microsoft
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+# --------------------------------------------------------
+# Swin Transformer
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from math import sqrt
+
+from megatron import get_args
+from functools import partial
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None,
+                 out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = input_resolution[0]
+        self.W = input_resolution[1]
+
+        self.attn_mask_dict = {} 
+
+    def create_attn_mask(self, H, W):
+        # calculate attention mask for SW-MSA
+
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1))  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+
+    def forward(self, x):
+        B, L, C = x.shape
+        H = int(sqrt(L))
+        W = H
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.dim
+        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
+        return flops
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+        x_b4_ds = x
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x_b4_ds, x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        assert H == self.img_size[0] and W == self.img_size[1], \
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def flops(self):
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops
+
+
+class SwinTransformer(nn.Module):
+    r""" Swin Transformer
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 224
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3,
+                 embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24],
+                 window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.3,
+                 norm_layer=partial(nn.LayerNorm, eps=1e-6), ape=False, patch_norm=True,
+                 use_checkpoint=False, output_avg=False, **kwargs):
+        super().__init__()
+
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
+        self.mlp_ratio = mlp_ratio
+        self.img_size = to_2tuple(img_size)
+        self.patch_size = to_2tuple(patch_size)
+        self.output_avg = output_avg
+        
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(dim=int(embed_dim * 2 ** i_layer),
+                               input_resolution=(patches_resolution[0] // (2 ** i_layer),
+                                                 patches_resolution[1] // (2 ** i_layer)),
+                               depth=depths[i_layer],
+                               num_heads=num_heads[i_layer],
+                               window_size=window_size,
+                               mlp_ratio=self.mlp_ratio,
+                               qkv_bias=qkv_bias, qk_scale=qk_scale,
+                               drop=drop_rate, attn_drop=attn_drop_rate,
+                               drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                               norm_layer=norm_layer,
+                               downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                               use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+    def forward(self, x):
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        h = self.img_size[0] // self.patch_size[0]
+        w = self.img_size[1] // self.patch_size[1]
+        outs = []
+
+        for i, layer in enumerate(self.layers):
+            px, x = layer(x)
+            b, n, c = px.shape
+
+            if i != len(self.layers) - 1 or not self.output_avg:
+                px = px.permute(0, 2, 1).contiguous()
+                px = px.reshape(b, c, h, w)
+            # is this a fair assumption ?? i think it's baked into the architecture
+            h, w = h//2, w//2
+            outs.append(px)
+
+        if self.output_avg:
+            return outs[-1].mean(dim=1)
+
+        return outs
+
+    def flops(self):
+        flops = 0
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+        flops += self.num_features * self.patches_resolution[0] * self.patches_resolution[1] // (2 ** self.num_layers)
+        flops += self.num_features * self.num_classes
+        return flops
+
+
+def get_swin(drop_path_rate=0.3, output_avg=False):
+    args = get_args()
+
+    window_size = 7
+    embed_dim = 128
+    depths = [2, 2, 18, 2]
+    num_heads = [4, 8, 16, 32]
+    swin = SwinTransformer(
+        img_size=(args.img_h, args.img_w,),
+        in_chans=3,
+        patch_size=args.patch_dim,
+        embed_dim=embed_dim,
+        depths=depths,
+        num_heads=num_heads,
+        window_size=window_size,
+        drop_path_rate=drop_path_rate,
+        output_avg=output_avg,
+    )
+
+    return swin
+

megatron/model/vision/utils.py

+import warnings
+import torch
+import torch.nn.functional as F
+
+
+def resize(input,
+           size=None,
+           scale_factor=None,
+           mode='nearest',
+           align_corners=None,
+           warning=True):
+    if warning:
+        if size is not None and align_corners:
+            input_h, input_w = tuple(int(x) for x in input.shape[2:])
+            output_h, output_w = tuple(int(x) for x in size)
+            if output_h > input_h or output_w > output_h:
+                if ((output_h > 1 and output_w > 1 and input_h > 1
+                     and input_w > 1) and (output_h - 1) % (input_h - 1)
+                        and (output_w - 1) % (input_w - 1)):
+                    warnings.warn(
+                        f'When align_corners={align_corners}, '
+                        'the output would more aligned if '
+                        f'input size {(input_h, input_w)} is `x+1` and '
+                        f'out size {(output_h, output_w)} is `nx+1`')
+    if isinstance(size, torch.Size):
+        size = tuple(int(x) for x in size)
+    return F.interpolate(input, size, scale_factor, mode, align_corners)

megatron/model/vision/vit_backbone.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Vision Transformer(VIT) model."""
+
+import math
+import einops
+import torch
+import apex
+import torch.nn.functional as F
+from megatron import get_args
+from megatron.model.transformer import ParallelTransformer
+from megatron.model.utils import (
+    get_linear_layer,
+    init_method_normal,
+    scaled_init_method_normal,
+)
+from megatron.model.module import MegatronModule
+
+CLASS_TOKEN_LENGTH = 8
+
+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.relu = torch.nn.ReLU()
+        self.dense_out = torch.nn.Linear(hidden_size, num_classes)
+        torch.nn.init.constant_(self.dense_out.bias, -10)
+
+    def forward(self, hidden_states):
+        # hidden_states: [b, 1, h]
+        # sequence_index: index of the token to pool.
+        dense_in_result = self.dense_in(hidden_states)
+        tanh_result = torch.tanh(dense_in_result)
+        dense_out_result = self.dense_out(tanh_result)
+        return dense_out_result
+
+
+def isPerfectSquare(x):
+    if(x >= 0):
+        sr = math.sqrt(x)
+        return (int(sr) * int(sr) == x)
+    return False
+
+
+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_h = args.img_h // args.patch_dim
+    num_patches_per_dim_w = args.img_w // args.patch_dim
+    num_patches = num_patches_per_dim_h * num_patches_per_dim_w
+    hidden_size = args.hidden_size
+
+    key = prefix + "weight"
+
+    assert key in state_dict
+    if key in state_dict:
+        input_param = state_dict[key]
+
+        input_seq_len = input_param.shape[0]
+        assert(isPerfectSquare(input_seq_len) or isPerfectSquare(input_seq_len - CLASS_TOKEN_LENGTH))
+        input_has_class_token = not isPerfectSquare(input_seq_len)
+        num_tok_input = input_seq_len - CLASS_TOKEN_LENGTH if input_has_class_token else input_seq_len
+        num_tok_output = num_patches
+        output_has_class_token = args.class_token_present
+
+        # update input_param and load it to state_dict[key]
+        if input_has_class_token:
+            input_param_tok = input_param[:CLASS_TOKEN_LENGTH, :]
+            input_param_grid = input_param[CLASS_TOKEN_LENGTH:, :]
+        else:
+            input_param_tok = torch.zeros(CLASS_TOKEN_LENGTH, hidden_size)
+            input_param_grid = input_param
+
+        assert input_param.shape[1] == hidden_size
+
+        if num_tok_input != num_tok_output:
+
+            gs_input = int(math.sqrt(num_tok_input))
+            gs_new = (num_patches_per_dim_h, num_patches_per_dim_w)
+
+            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[0] / gs_input, gs_new[1] / 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, num_tok_output))
+            input_param_grid = input_param_grid.transpose(0, 1).contiguous()
+
+            assert input_param_grid.shape[1] == hidden_size
+
+        input_param = input_param_grid
+        assert (
+            input_param.shape[0] == num_tok_output
+            and input_param.shape[1] == hidden_size
+        )
+
+        if output_has_class_token:
+            input_param = torch.cat((input_param_tok, input_param), dim=0)
+
+        state_dict[key] = input_param
+
+
+class VitBackbone(MegatronModule):
+    """Vision Transformer Model."""
+
+    def __init__(self,
+                 pre_process=True,
+                 post_process=True,
+                 class_token=True,
+                 single_token_output=False,
+                 post_layer_norm=True,
+                 drop_path_rate=0.0):
+        super(VitBackbone, self).__init__(share_word_embeddings=False)
+        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.pre_process = pre_process
+        self.post_process = post_process
+        self.class_token = class_token
+        self.post_layer_norm = post_layer_norm
+        self.hidden_size = args.hidden_size
+        self.patch_dim = args.patch_dim
+        self.img_h = args.img_h
+        self.img_w = args.img_w
+        self.micro_batch_size = args.micro_batch_size
+        self.single_token_output = single_token_output
+        self.drop_path_rate = drop_path_rate
+
+        assert self.img_h % self.patch_dim == 0
+        assert self.img_w % self.patch_dim == 0
+        self.num_patches_per_dim_h = self.img_h // self.patch_dim
+        self.num_patches_per_dim_w = self.img_w // self.patch_dim
+        self.num_patches = self.num_patches_per_dim_h * self.num_patches_per_dim_w
+        self.seq_length = self.num_patches + (CLASS_TOKEN_LENGTH if self.class_token else 0)
+        self.flatten_dim = self.patch_dim * self.patch_dim * args.num_channels
+        self.input_tensor = None
+        self.position_ids = None
+
+        if self.pre_process:
+            # cls_token
+            if self.class_token:
+                self.cls_token = torch.nn.Parameter(
+                    torch.randn(1, CLASS_TOKEN_LENGTH, self.hidden_size)
+                )
+                torch.nn.init.zeros_(self.cls_token)
+            self.position_ids = torch.arange(self.seq_length).expand(1, -1).cuda()
+            
+            # 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
+            )
+
+            args.class_token_present = self.class_token
+            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,
+            pre_process=self.pre_process,
+            post_process=self.post_process,
+            post_layer_norm=self.post_layer_norm,
+            drop_path_rate=self.drop_path_rate
+        )
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.transformer.set_input_tensor(input_tensor)
+
+    def forward(self, input):
+
+        if self.pre_process:
+            rearranged_input = einops.rearrange(
+                input,
+                "b c (h p1) (w p2) -> b (h w) (p1 p2 c)",
+                p1=self.patch_dim,
+                p2=self.patch_dim,
+            )
+
+            assert rearranged_input.dtype == torch.half
+            encoder_output = self.linear_encoder(rearranged_input)
+
+            concatenated_tokens = encoder_output
+            if self.class_token:
+                cls_tokens = self.cls_token.expand(encoder_output.shape[0], -1, -1)
+                concatenated_tokens = torch.cat((cls_tokens, encoder_output), dim=1)
+
+            token_embeddings = concatenated_tokens + \
+                    self.position_embeddings(self.position_ids[:, :concatenated_tokens.shape[1]])
+            # [b, s, h] => [s, b, h]
+            token_embeddings = token_embeddings.transpose(0, 1).contiguous()
+            hidden_states = self.embedding_dropout(token_embeddings)
+        else:
+            hidden_states = input
+
+        hidden_states = self.transformer(hidden_states, None)
+
+        if self.post_process:
+            # [s b h] => [b s h]
+            if self.single_token_output:
+                hidden_states = hidden_states[0]
+            else:
+                hidden_states = hidden_states.transpose(0, 1).contiguous()
+
+        return hidden_states
+

megatron/mpu/tests/commons.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+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

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+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

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+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

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+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

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+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

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+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/optimizer/__init__.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+from apex.optimizers import FusedAdam as Adam
+from apex.optimizers import FusedSGD as SGD
+
+from megatron import get_args
+
+from .distrib_optimizer import DistributedOptimizer
+from .grad_scaler import ConstantGradScaler, DynamicGradScaler
+from .optimizer import Float16OptimizerWithFloat16Params, FP32Optimizer
+
+
+def get_param_groups(modules,
+                     no_weight_decay_cond,
+                     scale_lr_cond,
+                     lr_mult):
+    """creates param groups based on weight decay condition (regularized vs non regularized)
+       and learning rate scale condition (args.lr vs lr_mult * args.lr)
+       scale_lr_cond is used during finetuning where head of the network requires a scaled
+       version of the base learning rate. 
+    """
+    wd_no_scale_lr = []
+    wd_scale_lr = []
+    no_wd_no_scale_lr = []
+    no_wd_scale_lr = []
+    for module in modules:
+        for name, param in module.named_parameters():
+            if not param.requires_grad:
+                continue
+
+            if no_weight_decay_cond is not None:
+                no_wd = no_weight_decay_cond(name, param)
+            else:
+                # do not regularize biases nor Norm parameters
+                no_wd = name.endswith(".bias") or len(param.shape) == 1
+
+            if scale_lr_cond is not None:
+                scale_lr = scale_lr_cond(name, param)
+            else:
+                scale_lr = False
+
+            if not no_wd and not scale_lr:
+                wd_no_scale_lr.append(param)
+            elif not no_wd and scale_lr:
+                wd_scale_lr.append(param)
+            elif no_wd and not scale_lr:
+                no_wd_no_scale_lr.append(param)
+            else:
+                no_wd_scale_lr.append(param)
+
+    param_groups = []
+    if len(wd_no_scale_lr):
+        param_groups.append({'params': wd_no_scale_lr, 'wd_mult': 1.0, 'lr_mult': 1.0})
+    if len(wd_scale_lr):
+        param_groups.append({'params': wd_scale_lr, 'wd_mult': 1.0, 'lr_mult': lr_mult})
+    if len(no_wd_no_scale_lr):
+        param_groups.append({'params': no_wd_no_scale_lr, 'wd_mult': 0.0, 'lr_mult': 1.0})
+    if len(no_wd_scale_lr):
+        param_groups.append({'params': no_wd_scale_lr, 'wd_mult': 0.0, 'lr_mult': lr_mult})
+
+    return param_groups
+
+def get_megatron_optimizer(model,
+                           no_weight_decay_cond=None,
+                           scale_lr_cond=None,
+                           lr_mult=1.0):
+    args = get_args()
+
+    # Base optimizer.
+    param_groups = get_param_groups(model,
+                                    no_weight_decay_cond,
+                                    scale_lr_cond,
+                                    lr_mult)
+
+    if args.optimizer == 'adam':
+        optimizer = Adam(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))
+
+    # Determine whether the params have main-grad field.
+    params_have_main_grad = False
+    if args.DDP_impl == 'local':
+        params_have_main_grad = True
+
+    # Mixed precision optimizer.
+    # - Note: both the Float16Optimizer and the DistributedOptimizer inherit
+    #   from the MixedPrecisionOptimizer, which manages any optimizer where
+    #   the model params and main params are distinct.
+    if args.fp16 or args.bf16 or args.use_distributed_optimizer:
+
+        # 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.
+        opt_ty = DistributedOptimizer \
+            if args.use_distributed_optimizer else \
+            Float16OptimizerWithFloat16Params
+        return opt_ty(optimizer,
+                      args.clip_grad,
+                      args.log_num_zeros_in_grad,
+                      params_have_main_grad,
+                      args.use_contiguous_buffers_in_local_ddp,
+                      args.fp16,
+                      args.bf16,
+                      args.params_dtype,
+                      grad_scaler,
+                      model)
+
+    # FP32.
+    return FP32Optimizer(optimizer, args.clip_grad,
+                         args.log_num_zeros_in_grad,
+                         params_have_main_grad,
+                         args.use_contiguous_buffers_in_local_ddp,
+                         model)

megatron/optimizer/clip_grads.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Gradient clipping."""
+
+import torch
+from torch._six import inf
+
+from apex.multi_tensor_apply import multi_tensor_applier
+import amp_C
+
+from megatron.model.module import param_is_not_shared
+from megatron.core.tensor_parallel import param_is_not_tensor_parallel_duplicate
+
+
+def clip_grad_norm_fp32(parameters, grads_for_norm,
+                        max_norm, norm_type=2,
+                        model_parallel_group=None):
+    """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
+        grads_for_norm (Iterable[Tensor]): an iterable of Tensors or a single
+            Tensor that will be used for calculating the grad norm.
+        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.
+        model_parallel_group (group): given the nature of the distributed
+            optimizer, this is passed as an argument.
+
+    Returns:
+        Total norm of the parameters (viewed as a single vector).
+    """
+
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    if isinstance(grads_for_norm, torch.Tensor):
+        grads_for_norm = [grads_for_norm]
+
+    # Grads.
+    grads = []
+    for param in parameters:
+        if param.grad is not None:
+            assert param.grad.type() == 'torch.cuda.FloatTensor'
+            grads.append(param.grad.detach())
+
+    # 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=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.
+            if grads_for_norm:
+                grad_norm, _ = multi_tensor_applier(
+                    amp_C.multi_tensor_l2norm,
+                    dummy_overflow_buf,
+                    [grads_for_norm],
+                    False # no per-parameter norm
+                )
+            else:
+                grad_norm = torch.cuda.FloatTensor([0])
+            # 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=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, model_parallel_group):
+
+    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 = torch.cuda.FloatTensor([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=model_parallel_group)
+
+    total_num_zeros = total_num_zeros.item()
+
+    return total_num_zeros

megatron/optimizer/distrib_optimizer.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Megatron distributed optimizer."""
+
+
+import math
+import torch
+
+from megatron import get_args
+from megatron import get_timers
+from megatron import print_rank_0
+from megatron.core import mpu, tensor_parallel
+from megatron.model.module import param_is_not_shared
+
+from .optimizer import MixedPrecisionOptimizer, _zero_grad_group_helper
+
+
+class Range:
+    """
+    A range represents a start and end points for indexing a shard
+    from a full tensor.
+    """
+    def __init__(self, start, end):
+        self.start = start
+        self.end = end
+        self.size = end - start
+    def normalize(self, start = 0):
+        return Range(start, start + self.size)
+    def __str__(self):
+        return "%d,%d [%d]" % (self.start, self.end, self.size)
+
+
+class DistributedOptimizer(MixedPrecisionOptimizer):
+    """Distributed optimizer, for all data types (fp16, bf16, and fp32).
+
+    Arguments:
+        optimizer: base optimizer such as Adam or SGD
+        clip_grad: clip gradeints with this global L2 norm. Note
+            that clipping is ignored if clip_grad == 0
+        log_num_zeros_in_grad: return number of zeros in the gradients.
+        params_have_main_grad: flag indicating if parameters have
+            a `main_grad` field. If this is set, we are assuming
+            that the model parameters are store in the `main_grad`
+            field instead of the typical `grad` field. This happens
+            for the DDP cases where there is a continuous buffer
+            holding the gradients. For example for bfloat16, we want
+            to do gradient accumulation and all-reduces in float32
+            and as a result we store those gradients in the main_grad.
+            Note that main grad is not necessarily in float32.
+        use_contiguous_buffers_in_local_ddp: if true, the local DDP model
+            is using a contiguous buffer to hold the model grads.
+        fp16: if true, the model is running in fp16.
+        bf16: if true, the model is running in bfloat16.
+        grad_scaler: used for scaling gradients. Note that this can be
+            None. This case happens when `bf16 = True` and we don't
+            use any loss scale. Note that for `bf16 = True`, we can have
+            a constnat gradient scaler. Also for `bf16 = False`, we
+            always require a grad scaler.
+        models: list of models (i.e., the virtual pipelining models). This
+            is used by the distributed optimizer for mapping parameters.
+    """
+
+    @classmethod
+    def build_model_gbuf_param_range_map(cls, model, dtype, gbuf_world_range):
+        """
+        Build mapping from param reference to grad buffer shard ranges.
+
+        This method builds a mapping from parameter references to grad
+        buffer shard ranges, specific to each data-parallel (DP) rank's
+        set of 'owned' parameters. Each grad buffer (padded to be an even
+        multiple of DP-world-size) is conceptually divided into DP-world-size
+        contiguous regions, where each DP rank 'owns' a contiguous regions.
+        Ownership in this sense means DP rank is responsible for reducing
+        the relevant subset of grads, and updating the relevant subset of
+        params.
+
+        This conceptual partitioning of the grad buffer does NOT respect
+        parameter boundaries, and as such it is assumed that each created
+        range references a shard (or subset) of the full parameter. It is
+        easiest to think of each DP rank as operating (i.e., reducing,
+        gathering) purely on views into the grad buffer, for all model-to-
+        main & main-to-model operations.
+
+        This method creates three ranges:
+        - The param's range within the entire grad buffer (i.e., world index).
+        - The param's range within the DP rank's local view of the grad buffer.
+        - The param's range within itself (i.e., its shard).
+        """
+
+        # Param range map.
+        param_world_index_map = model._grad_buffer_param_index_map[dtype]
+        param_range_map = {}
+        for param, param_world_indexes in param_world_index_map.items():
+
+            # Param range.
+            param_world_start, param_world_end = param_world_indexes
+            param_local_start = max(
+                0,
+                param_world_start - gbuf_world_range.start)
+            param_local_end = min(
+                gbuf_world_range.size,
+                param_world_end - gbuf_world_range.start)
+
+            # Add param, if within local gbuf range.
+            if param_local_end > param_local_start:
+                param_local_range = Range(param_local_start, param_local_end)
+                param_world_range = param_local_range.normalize(
+                    param_local_start + gbuf_world_range.start)
+                sub_param_start = max(0, gbuf_world_range.start-param_world_start)
+                sub_param_range = param_local_range.normalize(sub_param_start)
+                param_range_map[param] = {
+                    "gbuf_world" : param_world_range,
+                    "gbuf_local" : param_local_range,
+                    "param" : sub_param_range,
+                }
+
+        return param_range_map
+
+
+    @classmethod
+    def build_model_gbuf_range(cls, model, dtype):
+        """
+        Build mapping between params and their grad buffers.
+
+        This method does the initial setup for the method above. This setup
+        includes determining the shard ranges into the DDP's grad buffer for
+        each data-parallel (DP) rank. Each DP rank keeps range info for
+        all other DP ranks, for the purpose of creating args for
+        reduce-scatter and all-gather.
+        """
+
+        data_parallel_rank = mpu.get_data_parallel_rank()
+        data_parallel_world_size = mpu.get_data_parallel_world_size()
+
+        # Grad buffer range.
+        grad_buffer = model._grad_buffers[dtype]
+        gbuf_size = grad_buffer.numel
+        max_gbuf_range_size = int(math.ceil(gbuf_size / data_parallel_world_size))
+
+        # All world ranges. (i.e., across all data parallel ranks)
+        gbuf_world_all_ranges = []
+        for r in range(data_parallel_world_size):
+            gbuf_world_start = r * max_gbuf_range_size
+            gbuf_world_end = min(gbuf_size, gbuf_world_start+max_gbuf_range_size)
+            gbuf_world_range = Range(gbuf_world_start, gbuf_world_end)
+            gbuf_world_all_ranges.append(gbuf_world_range)
+
+        # Local DP's ranges.
+        gbuf_world_range = gbuf_world_all_ranges[data_parallel_rank]
+        gbuf_local_range = gbuf_world_range.normalize()
+
+        # Get each param's ranges.
+        param_range_map = cls.build_model_gbuf_param_range_map(model,
+                                                               dtype,
+                                                               gbuf_world_range)
+
+        # Group into dict.
+        data = {
+            "local" : gbuf_local_range,
+            "world" : gbuf_world_range,
+            "world_all" : gbuf_world_all_ranges,
+            "param_map" : param_range_map,
+            "max_range_size" : max_gbuf_range_size,
+        }
+
+        return data
+
+
+    @classmethod
+    def build_model_gbuf_range_map(cls, model):
+        """
+        Create param-to-grad-buffer mappings, for grad buffer data types
+        within a specific virtual model.
+        """
+        return {
+            dtype : cls.build_model_gbuf_range(model, dtype)
+            for dtype in model._grad_buffers
+        }
+
+
+    @classmethod
+    def build_model_param_gbuf_map(cls, model_gbuf_ranges):
+        """
+        Create a reverse of the model_gbuf_ranges, for referencing in
+        opposite direction.
+        """
+        param_gbuf_map = {}
+        for model_index, model_gbuf_range_map in enumerate(model_gbuf_ranges):
+            for dtype, gbuf_range_map in model_gbuf_range_map.items():
+                for param, param_range_map in gbuf_range_map["param_map"].items():
+                    param_gbuf_map[param] = (model_index, dtype)
+        return param_gbuf_map
+
+
+    @classmethod
+    def build_optimizer_group_ranges(cls, param_groups, model_gbuf_ranges):
+        """
+        Create optimizer groups.
+
+        Given the set of parameter shard ranges that are owned by the current
+        data-parallel (DP) rank, gather the set of parameters that will be
+        used (in the method below) to create the current DP's optimizer
+        groups.
+        """
+
+        num_groups = len(param_groups)
+
+        # Param group map.
+        param_group_map = {}
+        for group_index, group in enumerate(param_groups):
+            for param in group["params"]:
+                assert param.requires_grad
+                param_group_map[param] = group_index
+
+        # Optimizer group ranges.
+        group_ranges = [ {"params": []} for _ in param_groups ]
+        for model_gbuf_range_map in model_gbuf_ranges:
+            for dtype, gbuf_range_map in model_gbuf_range_map.items():
+                for param in gbuf_range_map["param_map"]:
+                    group_index = param_group_map[param]
+                    group_range = group_ranges[group_index]
+                    group_range["params"].append(param)
+
+        # Squeeze zero-size group ranges.
+        for group_index, group_range in enumerate(group_ranges):
+            group_range["orig_group"] = param_groups[group_index]
+        group_ranges = [ g for g in group_ranges if len(g["params"]) > 0 ]
+
+        return group_ranges
+
+
+    @classmethod
+    def build_model_and_main_param_groups(cls,
+                                          model_gbuf_ranges,
+                                          param_gbuf_map,
+                                          opt_group_ranges):
+        """
+        Create main parameter groups needed for the optimizer step.
+
+        These groups encompass both: 1) groups used by this class, for
+        reducing/gather, and 2) groups used by the inner optimizer for the
+        parameter update. Given that the conceptual grad buffer partitioning
+        (created in earlier method) doesn't respect parameter boundaries,
+        the optimizer operates on shards of the model parameters, rather than
+        the full parameters.
+        """
+
+        # Parameter groups:
+        #   model_float16_groups: original float16 parameters
+        #   model_fp32_groups: original fp32 parameters
+        #   shard_float16_groups: shards of original float16 parameters
+        #   shard_fp32_groups: shards of original fp32 parameters
+        #   shard_fp32_from_float16_groups: fp32 copy of float16 parameters
+        model_float16_groups = []
+        model_fp32_groups = []
+        shard_float16_groups = []
+        shard_fp32_groups = []
+        shard_fp32_from_float16_groups = []
+
+        # Allocate (or slice) each group's param shard.
+        for group_index, group_range in enumerate(opt_group_ranges):
+
+            # Params of this group.
+            model_float16_params_this_group = []
+            model_fp32_params_this_group = []
+            shard_float16_params_this_group = []
+            shard_fp32_params_this_group = []
+            shard_fp32_from_float16_params_this_group = []
+            model_float16_groups.append(model_float16_params_this_group)
+            model_fp32_groups.append(model_fp32_params_this_group)
+            shard_float16_groups.append(shard_float16_params_this_group)
+            shard_fp32_groups.append(shard_fp32_params_this_group)
+            shard_fp32_from_float16_groups.append(
+                shard_fp32_from_float16_params_this_group)
+
+            for model_param in group_range["params"]:
+
+                assert model_param.requires_grad
+
+                model_index, dtype = param_gbuf_map[model_param]
+                gbuf_range = model_gbuf_ranges[model_index][dtype]
+                param_range = gbuf_range["param_map"][model_param]["param"]
+
+                # fp16, bf16 params.
+                if model_param.type() in ['torch.cuda.HalfTensor',
+                                          'torch.cuda.BFloat16Tensor']:
+
+                    # Clone model -> main.
+                    shard_model_param = model_param.detach().view(-1) \
+                        [param_range.start:param_range.end]
+                    shard_main_param = shard_model_param.clone().float()
+                    tensor_parallel.copy_tensor_model_parallel_attributes(
+                        shard_model_param, model_param)
+                    tensor_parallel.copy_tensor_model_parallel_attributes(
+                        shard_main_param, model_param)
+                    if hasattr(model_param, 'shared'):
+                        shard_model_param.shared = model_param.shared
+                        shard_main_param.shared = model_param.shared
+
+                    # Add to group.
+                    model_float16_params_this_group.append(model_param)
+                    shard_float16_params_this_group.append(shard_model_param)
+                    shard_fp32_from_float16_params_this_group.append(shard_main_param)
+
+                # fp32 params.
+                elif model_param.type() == 'torch.cuda.FloatTensor':
+                    shard_model_param = model_param.view(-1) \
+                        [param_range.start:param_range.end]
+                    model_fp32_params_this_group.append(model_param)
+                    shard_fp32_params_this_group.append(shard_model_param)
+                    tensor_parallel.copy_tensor_model_parallel_attributes(
+                        shard_model_param, model_param)
+                    if hasattr(model_param, 'shared'):
+                        shard_model_param.shared = model_param.shared
+
+                else:
+                    raise TypeError('Wrapped parameters must be one of '
+                                    'torch.cuda.FloatTensor,  '
+                                    'torch.cuda.HalfTensor, or '
+                                    'torch.cuda.BFloat16Tensor. '
+                                    'Received {}'.format(param.type()))
+
+            # Update optimizer's params.
+            group_range["orig_group"]["params"] = [
+                *shard_fp32_params_this_group,
+                *shard_fp32_from_float16_params_this_group,
+            ]
+
+        return (
+            model_float16_groups,
+            model_fp32_groups,
+            shard_float16_groups,
+            shard_fp32_groups,
+            shard_fp32_from_float16_groups,
+        )
+
+
+    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
+                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+                 fp16, bf16, params_dtype, grad_scaler, models):
+        """
+        See top of class definition for argument descriptions.
+
+        The steps in this method create the core mapping between DDP grad
+        buffers, parameters, and parameter shard ranges, that is needed for
+        converting between model param indexes and main parameter shard
+        indexes. This method also updates the optimizer parameter groups
+        with the newly created shards.
+        """
+
+        super().__init__(
+            optimizer, clip_grad, log_num_zeros_in_grad,
+            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+            fp16, bf16, params_dtype, grad_scaler, models)
+
+        # Verify that contiguous buffers are being used.
+        # - Note: this should already be checked in arguments.py.
+        assert use_contiguous_buffers_in_local_ddp
+
+        # Model grad buffer ranges.
+        self.model_gbuf_ranges = []
+        for model_index, model in enumerate(self.models):
+            self.model_gbuf_ranges.append(self.build_model_gbuf_range_map(model))
+        self.model_param_gbuf_map = \
+            self.build_model_param_gbuf_map(self.model_gbuf_ranges)
+
+        # Optimizer ranges.
+        self.opt_group_ranges = self.build_optimizer_group_ranges(
+            self.optimizer.param_groups,
+            self.model_gbuf_ranges)
+
+        # Allocate main param shards.
+        (
+            self.model_float16_groups,
+            self.model_fp32_groups,
+            self.shard_float16_groups,
+            self.shard_fp32_groups,
+            self.shard_fp32_from_float16_groups,
+        ) = self.build_model_and_main_param_groups(self.model_gbuf_ranges,
+                                                   self.model_param_gbuf_map,
+                                                   self.opt_group_ranges)
+
+        # Initialize param buffers.
+        # - These are views on the DDP model's grad buffers, that share
+        #   storage & have their own dtype. This is safe because the param
+        #   dtype size is always <= grad dtype size.
+        self.param_buffers = []
+        for model_index, model in enumerate(self.models):
+            current_param_buffers = {}
+            for dtype, grad_buffer in model._grad_buffers.items():
+                param_buffer = torch.tensor(grad_buffer.data.storage()._untyped(),
+                                            dtype = params_dtype,
+                                            device = grad_buffer.data.device)
+                param_buffer = param_buffer[:grad_buffer.numel_padded]
+                current_param_buffers[dtype] = param_buffer
+            self.param_buffers.append(current_param_buffers)
+
+        # Update optimizer groups.
+        # - Also, leverage state_dict() and load_state_dict() to
+        #   recast preexisting per-param state tensors.
+        self.optimizer.param_groups = \
+            [ g["orig_group"] for g in self.opt_group_ranges ]
+        self.optimizer.load_state_dict(self.optimizer.state_dict())
+
+
+    def get_model_param_range_map(self, param):
+        """
+        Given a model param, get the index sub-range of the param that this
+        data-parallel rank owns.
+        """
+        model_index, dtype = self.model_param_gbuf_map[param]
+        gbuf_range_map = self.model_gbuf_ranges[model_index][dtype]
+        param_range_map = gbuf_range_map["param_map"][param]
+        return param_range_map
+
+
+    def get_model_parallel_group(self):
+        """
+        With the distributed optimizer, the model parallel group is the
+        entire world.
+        """
+        return None
+
+
+    def state_dict(self):
+        """
+        The state dict must contain the fp32-from-float16 shards.
+        """
+        state_dict = {}
+        state_dict['optimizer'] = self.optimizer.state_dict()
+        if self.grad_scaler:
+            state_dict['grad_scaler'] = self.grad_scaler.state_dict()
+        state_dict['shard_fp32_from_float16_groups'] = \
+            self.shard_fp32_from_float16_groups
+        return state_dict
+
+
+    def load_state_dict(self, state_dict):
+        """
+        Load the state dict.
+        """
+
+        # Optimizer.
+        optimizer_key = 'optimizer'
+        if optimizer_key not in state_dict:
+            optimizer_key = 'optimizer_state_dict'
+            print_rank_0('***WARNING*** loading optimizer from '
+                         'an old checkpoint ...')
+        self.optimizer.load_state_dict(state_dict[optimizer_key])
+
+        # Grad scaler.
+        if 'grad_scaler' not in state_dict:
+            if self.fp16:
+                print_rank_0('***WARNING*** found an old checkpoint, will not '
+                             'load grad scaler ...')
+        else:
+            if self.grad_scaler:
+                self.grad_scaler.load_state_dict(state_dict['grad_scaler'])
+            else:
+                print_rank_0('***WARNING*** fould the grad scaler in the '
+                             'checkpoint but it is None in the class. '
+                             'Skipping loading grad scaler ...')
+
+        # Copy data for the main params.
+        for current_group, saved_group in zip(
+                self.shard_fp32_from_float16_groups,
+                state_dict["shard_fp32_from_float16_groups"]):
+            for current_param, saved_param in zip(current_group, saved_group):
+                current_param.data.copy_(saved_param.data)
+
+
+    def zero_grad(self, set_to_none=True):
+        """
+        Zero grads.
+
+        We only need to zero the model related parameters, i.e.,
+        model_float16_groups & model_fp32_groups. We additionally zero
+        the remaining groups as a memory optimization to reduce
+        fragmentation; in the case of set_to_none==True, the space
+        used by this field can be safely deallocated at this point.
+        """
+        for groups in (
+                self.model_float16_groups,
+                self.model_fp32_groups,
+                self.shard_float16_groups, # grad empty/unused here?
+                self.shard_fp32_groups, # throws grad-access warning
+                self.shard_fp32_from_float16_groups):
+            for group in groups:
+                _zero_grad_group_helper(group, set_to_none)
+
+
+    @staticmethod
+    def get_model_buffer_dp_views(model_buffers):
+        """
+        Get shard views of each of the DDP's param/grad buffers.
+
+        In this nested list, the top level is grouped by the virtual model
+        index and the buffer's data type. The sub-level is a list of
+        shards of that buffer, where each shard in the list represents
+        a contiguous view of the buffer, that is owned by a data-parallel
+        rank. The shard boundary does not respect parameter boundaries, and
+        so the elements of some parameters are split across data parallel
+        ranks.
+
+        Additionally, return references to the entire buffers, for use
+        in _reduce_scatter_base and _all_gather_base.
+        """
+
+        data_parallel_world_size = mpu.get_data_parallel_world_size()
+
+        # Buffer views.
+        view_items = []
+        for model_index, buffers in enumerate(model_buffers):
+            for dtype, buf in buffers.items():
+
+                assert buf.numel() % data_parallel_world_size == 0
+                shard_size = int(buf.numel() / data_parallel_world_size)
+                buf_views = [buf[(r*shard_size):((r+1)*shard_size)]
+                             for r in range(data_parallel_world_size)]
+                view_items.append((model_index, dtype, buf, buf_views))
+
+        return view_items
+
+
+    def get_model_grad_buffer_dp_views(self):
+        return self.get_model_buffer_dp_views([
+            {dtype : mem_buffer.data}
+            for model in self.models
+            for dtype, mem_buffer in model._grad_buffers.items()])
+
+
+    def get_model_param_buffer_dp_views(self):
+        return self.get_model_buffer_dp_views(self.param_buffers)
+
+
+    def reduce_model_grads(self, args, timers):
+        """
+        Reduce-scatter model grads.
+
+        The DDP's grad buffer is used for the reduce-scatter, and thus no
+        tensors are dynamically allocated.
+
+        Note: this is a different order of reduction, versus the non-
+        distributed optimizer, which reduces: 1) layernorm grads, 2) all
+        grads, 3) embedding grads.
+        """
+
+        # All-reduce layer-norm grads (for sequence parallelism).
+        timers('layernorm-grads-all-reduce', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self.allreduce_layernorm_grads(args)
+        timers('layernorm-grads-all-reduce').stop()
+
+        # All-reduce embedding grads.
+        timers('embedding-grads-all-reduce', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self.allreduce_embedding_grads(args)
+        timers('embedding-grads-all-reduce').stop()
+
+        # Reduce-scatter setup.
+        timers('grads-reduce-scatter', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        data_parallel_rank = mpu.get_data_parallel_rank()
+        data_parallel_world_size = mpu.get_data_parallel_world_size()
+        data_parallel_group = mpu.get_data_parallel_group()
+
+        # Scale grad buffers by '1 / data_parallel_world_size'.
+        for model in self.models:
+            for dtype, gbuf in model._grad_buffers.items():
+                gbuf.data /= data_parallel_world_size
+
+        # Reduce-scatter all grads.
+        gbuf_view_items = self.get_model_grad_buffer_dp_views()
+        for index, (model_index, dtype, gbuf, gbuf_views) \
+            in enumerate(gbuf_view_items):
+
+            torch.distributed._reduce_scatter_base(
+                gbuf_views[data_parallel_rank],
+                gbuf,
+                group = data_parallel_group,
+            )
+
+        timers('grads-reduce-scatter').stop()
+
+
+    def gather_model_params(self, args, timers):
+        """
+        All-gather updated model params.
+
+        The DDP's param buffer is used for the all-gather, and thus no
+        tensors are dynamically allocated. After the all-gather, the params
+        can be copied from the param buffer to the param.
+        """
+
+        timers('params-all-gather', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+
+        data_parallel_rank = mpu.get_data_parallel_rank()
+        data_parallel_group = mpu.get_data_parallel_group()
+
+        # All-gather updated main params.
+        # - All param buffer views are guaranteed to have the same num elements
+        #   across all data parallel ranks, due to grad buffer padding that is
+        #   done in distributed.py, and extended to the param buffers. Thus,
+        #   all sub-views will have consistent start/end indexes across data
+        #   parallel ranks.
+        pbuf_view_items = self.get_model_param_buffer_dp_views()
+        for index, (model_index, dtype, pbuf, pbuf_views) \
+            in enumerate(pbuf_view_items):
+
+            torch.distributed._all_gather_base(
+                pbuf,
+                pbuf_views[data_parallel_rank],
+                group = data_parallel_group,
+            )
+
+        # Copy from param buffer to each param.
+        for model_id, model in enumerate(self.models):
+            for dtype, param_map in model._grad_buffer_param_index_map.items():
+                for param, buf_range in param_map.items():
+                    param_buf = self.param_buffers[model_id][dtype]
+                    param_buf_shard = param_buf[buf_range[0]:buf_range[1]]
+                    param.view(-1).detach().copy_(param_buf_shard)
+
+        timers('params-all-gather').stop()
+
+
+    def _collect_main_grad_data_for_unscaling(self):
+        """
+        Note: this should be equivalent to the float-16 optimizer's method,
+        but writtent differently, so the two should be combined.
+        """
+        return [
+            param.grad.data
+            for group in self.optimizer.param_groups
+            for param in group["params"]
+        ]
+
+
+    def _get_model_and_main_params_data_float16(self):
+        """
+        Get aligned list of model and main params.
+        """
+        model_data = []
+        main_data = []
+        for model_group, main_group in zip(self.shard_float16_groups,
+                                           self.shard_fp32_from_float16_groups):
+            for model_param, main_param in zip(model_group, main_group):
+                model_data.append(model_param.data)
+                main_data.append(main_param.data)
+        return model_data, main_data
+
+
+    def _copy_model_grads_to_main_grads(self):
+        """
+        Copy model grads to main grads.
+
+        Since this step follows a reduce-scatter through the DDP's grad
+        buffer, this method is responsible for copying the updated grads
+        from the grad buffer to the main shard's grad field.
+        """
+
+        # Utility method for copying group grads.
+        def copy_group_grads(model_groups, shard_main_groups):
+            for model_group, shard_main_group in zip(model_groups,
+                                                     shard_main_groups):
+                for model_param, shard_main_param in zip(model_group,
+                                                         shard_main_group):
+
+                    param_range_map = self.get_model_param_range_map(model_param)
+                    param_range = param_range_map["param"]
+                    assert param_range.size == shard_main_param.nelement()
+
+                    model_grad = model_param.main_grad
+                    shard_model_grad = model_grad.view(-1) \
+                        [param_range.start:param_range.end]
+                    shard_main_param.grad = shard_model_grad.float()
+
+        # Copy model groups to shard groups.
+        copy_group_grads(self.model_float16_groups,
+                         self.shard_fp32_from_float16_groups)
+        copy_group_grads(self.model_fp32_groups,
+                         self.shard_fp32_groups)
+
+
+    def _copy_main_params_to_model_params(self):
+        """
+        Copy main params to model params.
+
+        Since this step is followed by an all-gather through the DDP's grad
+        buffer, this method is responsible for copying the updated params
+        from the main shards into the correct position in the grad buffer.
+        """
+
+        # Utility method for copying group params.
+        def copy_group_params(shard_main_groups, model_groups):
+            for shard_main_group, model_group in zip(shard_main_groups,
+                                                     model_groups):
+                for shard_main_param, model_param in zip(shard_main_group,
+                                                         model_group):
+
+                    param_range_map = self.get_model_param_range_map(model_param)
+                    world_range = param_range_map["gbuf_world"]
+
+                    assert world_range.size == shard_main_param.nelement()
+
+                    model_id, dtype = self.model_param_gbuf_map[model_param]
+                    model_param_buffer = self.param_buffers[model_id][dtype]
+
+                    shard_model_param = model_param_buffer.view(-1) \
+                        [world_range.start:world_range.end]
+
+                    shard_model_param.data.copy_(shard_main_param)
+
+        # Copy shard groups to model groups.
+        copy_group_params(self.shard_fp32_from_float16_groups,
+                          self.model_float16_groups)
+        copy_group_params(self.shard_fp32_groups,
+                          self.model_fp32_groups)

megatron/optimizer/grad_scaler.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Megatron grad scaler."""
+
+from abc import ABC
+from abc import abstractmethod
+
+import torch
+
+
+class MegatronGradScaler(ABC):
+
+    def __init__(self, initial_scale):
+        """Initialize scale value with the input initial scale."""
+        assert initial_scale > 0.0
+        self._scale = torch.cuda.FloatTensor([initial_scale])
+
+    @property
+    def scale(self):
+        return self._scale
+
+    @property
+    def inv_scale(self):
+        return self._scale.double().reciprocal().float()
+
+    @abstractmethod
+    def update(self, found_inf):
+        pass
+
+    @abstractmethod
+    def state_dict(self):
+        pass
+
+    @abstractmethod
+    def load_state_dict(self, state_dict):
+        pass
+
+
+
+class ConstantGradScaler(MegatronGradScaler):
+
+    def update(self, found_inf):
+        pass
+
+    def state_dict(self):
+        return dict()
+
+    def load_state_dict(self, state_dict):
+        pass
+
+
+
+class DynamicGradScaler(MegatronGradScaler):
+
+    def __init__(self, initial_scale, min_scale,
+                 growth_factor, backoff_factor,
+                 growth_interval, hysteresis):
+        """"Grad scaler with dynamic scale that gets adjusted
+        during training."""
+        super(DynamicGradScaler, self).__init__(initial_scale)
+
+        # Lower bound on the scale.
+        assert min_scale > 0.0
+        assert min_scale <= initial_scale
+        self.min_scale = torch.cuda.FloatTensor([min_scale])
+        # Growth and backoff factors for the scale.
+        assert growth_factor > 1.0
+        self.growth_factor = torch.cuda.FloatTensor([growth_factor])
+        assert backoff_factor < 1.0
+        assert backoff_factor > 0.0
+        self.backoff_factor = torch.cuda.FloatTensor([backoff_factor])
+        # Interval over which if we don't see any inf/nan,
+        # we will scale the grad scale by the growth factor.
+        assert growth_interval > 0
+        self.growth_interval = growth_interval
+        # Number of inf/nans we should see before scaling down
+        # the grad scale by the backoff factor.
+        assert hysteresis > 0
+        self.hysteresis = hysteresis
+
+        # Trackers.
+        self._growth_tracker = 0
+        self._hysteresis_tracker = self.hysteresis
+
+
+    def update(self, found_inf):
+
+        # If we have an inf/nan, growth tracker is set to 0
+        # and hysterisis tracker is reduced by 1.
+        if found_inf:
+            self._growth_tracker = 0
+            self._hysteresis_tracker -= 1
+            # Now if we are out of hysteresis count, scale down the loss.
+            if self._hysteresis_tracker <= 0:
+                self._scale = torch.max(self._scale * self.backoff_factor,
+                                        self.min_scale)
+        else:
+            # If there is no nan/inf, increment the growth tracker.
+            self._growth_tracker += 1
+            # If we have had enough consequitive intervals with no nan/inf:
+            if self._growth_tracker == self.growth_interval:
+                # Reset the tracker and hysteresis trackers,
+                self._growth_tracker = 0
+                self._hysteresis_tracker = self.hysteresis
+                # and scale up the loss scale.
+                self._scale = self._scale * self.growth_factor
+
+
+    def state_dict(self):
+        state_dict = {}
+        state_dict['scale'] = self._scale
+        state_dict['growth_tracker'] = self._growth_tracker
+        state_dict['hysteresis_tracker'] = self._hysteresis_tracker
+        return state_dict
+
+
+    def load_state_dict(self, state_dict):
+        self._scale = state_dict['scale'].cuda(torch.cuda.current_device())
+        self._growth_tracker = state_dict['growth_tracker']
+        self._hysteresis_tracker = state_dict['hysteresis_tracker']

megatron/optimizer/optimizer.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Megatron optimizer."""
+
+from abc import ABC
+from abc import abstractmethod
+from apex.multi_tensor_apply import multi_tensor_applier
+import amp_C
+import torch
+from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
+from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
+
+from megatron import get_timers
+from megatron import print_rank_0
+from megatron.core import mpu, tensor_parallel
+from megatron.model import DistributedDataParallel as LocalDDP
+from megatron.model import Float16Module
+from megatron.model.module import param_is_not_shared
+from megatron.utils import unwrap_model
+
+from .clip_grads import clip_grad_norm_fp32, count_zeros_fp32
+
+
+def _zero_grad_group_helper(group, set_to_none):
+    """Zero out the gradient for a group of parameters.
+    Note: copied from torch.optim.optimizer."""
+    for param in group:
+        if param.grad is not None:
+            if set_to_none:
+                param.grad = None
+            else:
+                if param.grad.grad_fn is not None:
+                    param.grad.detach_()
+                else:
+                    param.grad.requires_grad_(False)
+                param.grad.zero_()
+
+
+def _multi_tensor_copy_this_to_that(this, that, overflow_buf=None):
+    """Use multi-tensor-applier to copy values from one list to another.
+    We don't have a blfoat16 implementation so for now if the overflow_buf
+    is not provided, we default back to simple loop copy to be compatible
+    with bfloat16."""
+    if overflow_buf:
+        overflow_buf.fill_(0)
+        # Scaling with factor `1.0` is equivalent to copy.
+        multi_tensor_applier(amp_C.multi_tensor_scale,
+                             overflow_buf,
+                             [this, that],
+                             1.0)
+    else:
+        for this_, that_ in zip(this, that):
+            that_.copy_(this_)
+
+
+
+class MegatronOptimizer(ABC):
+
+
+    def __init__(self, optimizer, clip_grad,
+                 log_num_zeros_in_grad,
+                 params_have_main_grad,
+                 use_contiguous_buffers_in_local_ddp,
+                 models):
+
+        """Input optimizer is the base optimizer for example Adam."""
+        self.optimizer = optimizer
+        assert self.optimizer, 'no optimizer is provided.'
+        # Set gradient clipping and logging params.
+        self.clip_grad = clip_grad
+        self.log_num_zeros_in_grad = log_num_zeros_in_grad
+        self.params_have_main_grad = params_have_main_grad
+        self.use_contiguous_buffers_in_local_ddp = use_contiguous_buffers_in_local_ddp
+
+        # 'models' are retained for access to the contiguous grad buffers.
+        # (see distributed optimizer)
+        self.models = models
+
+        if self.use_contiguous_buffers_in_local_ddp:
+            assert self.params_have_main_grad, \
+                "use of contiguous buffer requires that params have main grad"
+
+
+    def get_parameters(self):
+        params = []
+        for param_group in self.optimizer.param_groups:
+            for param in param_group['params']:
+                params.append(param)
+        return params
+
+
+    def get_main_grads_for_grad_norm(self):
+
+        # Filter parameters based on:
+        #   - grad should not be none
+        #   - parameter should not be shared
+        #   - should not be a replica due to tensor model parallelism
+        params = self.get_parameters()
+        grads_for_norm = []
+        for param in params:
+            grad = param.grad
+            grad_not_none = grad is not None
+            is_not_shared = param_is_not_shared(param)
+            is_not_tp_duplicate = tensor_parallel.param_is_not_tensor_parallel_duplicate(param)
+            if grad_not_none and is_not_shared and is_not_tp_duplicate:
+                grads_for_norm.append(grad)
+
+        return grads_for_norm
+
+
+    def get_model_parallel_group(self):
+        """Default returned here, but the distributed optimizer overrides this."""
+        return mpu.get_model_parallel_group()
+
+
+    def clip_grad_norm(self, clip_grad):
+        params = self.get_parameters()
+        grads_for_norm = self.get_main_grads_for_grad_norm()
+        return clip_grad_norm_fp32(
+            params, grads_for_norm, clip_grad,
+            model_parallel_group=self.get_model_parallel_group())
+
+
+    def count_zeros(self):
+        params = self.get_parameters()
+        return count_zeros_fp32(params,
+                                model_parallel_group=self.get_model_parallel_group())
+
+
+    @abstractmethod
+    def zero_grad(self, set_to_none=True):
+        pass
+
+
+    @abstractmethod
+    def get_loss_scale(self):
+        """The output should be a cuda tensor of size 1."""
+        pass
+
+
+    def scale_loss(self, loss):
+        """Simple scaling."""
+        return self.get_loss_scale() * loss
+
+
+    @abstractmethod
+    def reload_model_params(self):
+        """Refreshes any internal state from the current model parameters.
+        Call whenever the parameters are changed outside of the optimizer.
+        For example, when we load a model from a checkpoint  without loading
+        the optimizer, the model parameters are updated but for fp16 optimizer
+        with main parameters, the main parameters need to also be updated."""
+        pass
+
+
+    @abstractmethod
+    def state_dict(self):
+        pass
+
+
+    @abstractmethod
+    def load_state_dict(self, state_dict):
+        pass
+
+
+    # Promote state so it can be retrieved or set via
+    # "optimizer_instance.state"
+    def _get_state(self):
+        return self.optimizer.state
+
+    def _set_state(self, value):
+        self.optimizer.state = value
+
+    state = property(_get_state, _set_state)
+
+
+    # Promote param_groups so it can be retrieved or set via
+    # "optimizer_instance.param_groups"
+    # (for example, to adjust the learning rate)
+    def _get_param_groups(self):
+        return self.optimizer.param_groups
+
+    def _set_param_groups(self, value):
+        self.optimizer.param_groups = value
+
+    param_groups = property(_get_param_groups, _set_param_groups)
+
+
+    @abstractmethod
+    def step(self, args, timers):
+        pass
+
+
+    def gather_model_params(self, args, timers):
+        """
+        For the case of a non-distributed-optimizer, there is nothing to
+        do here.
+        """
+        pass
+
+
+    def allreduce_word_embedding_grads(self, args):
+        """
+        All-reduce word embedding grads.
+
+        Reduce grads across first and last stages to ensure that word_embeddings
+        parameters stay in sync. This should only run for models that support
+        pipelined model parallelism (BERT and GPT-2).
+        """
+
+        if mpu.is_rank_in_embedding_group(ignore_virtual=True) and \
+                mpu.get_pipeline_model_parallel_world_size() > 1:
+            if mpu.is_pipeline_first_stage(ignore_virtual=True):
+                unwrapped_model = self.models[0]
+            elif mpu.is_pipeline_last_stage(ignore_virtual=True):
+                unwrapped_model = self.models[-1]
+            else:  # We do not support the interleaved schedule for T5 yet.
+                unwrapped_model = self.models[0]
+            unwrapped_model = unwrap_model(
+                unwrapped_model, (torchDDP, LocalDDP, Float16Module))
+
+            if unwrapped_model.share_word_embeddings:
+                word_embeddings_weight = unwrapped_model.word_embeddings_weight()
+                if args.DDP_impl == 'local':
+                    grad = word_embeddings_weight.main_grad
+                else:
+                    grad = word_embeddings_weight.grad
+                torch.distributed.all_reduce(grad, group=mpu.get_embedding_group())
+
+
+    def allreduce_position_embedding_grads(self, args):
+        """
+        All-reduce position_embeddings grad across first (encoder) and
+        split (decoder) stages to ensure that position embeddings parameters
+        stay in sync. This should only run for T5 models with pipeline
+        parallelism.
+        """
+        if mpu.is_rank_in_position_embedding_group() and \
+                mpu.get_pipeline_model_parallel_world_size() > 1 and \
+                args.pipeline_model_parallel_split_rank is not None:
+            unwrapped_model = self.models[0]
+            unwrapped_model = unwrap_model(
+                unwrapped_model, (torchDDP, LocalDDP, Float16Module))
+            assert args.DDP_impl == 'local', \
+                'T5 model is only supported with local DDP mode'
+            grad = unwrapped_model.language_model.embedding.position_embeddings.weight.main_grad
+            torch.distributed.all_reduce(grad, group=mpu.get_position_embedding_group())
+
+
+    def allreduce_embedding_grads(self, args):
+        """All-reduce both word and position embeddings."""
+        self.allreduce_word_embedding_grads(args)
+        self.allreduce_position_embedding_grads(args)
+
+
+    def allreduce_layernorm_grads(self, args):
+        """All-reduce layernorm grads (for sequence parallelism)."""
+
+        # All-reduce layernorm parameters across model parallel nodes
+        # when sequence parallelism is used
+        if mpu.get_tensor_model_parallel_world_size() > 1 and \
+                args.sequence_parallel:
+            grads = []
+            for model_module in self.models:
+                unwrapped_model = unwrap_model( 
+                    model_module, (torchDDP, LocalDDP, Float16Module))
+                for param in unwrapped_model.parameters():
+                    if getattr(param, 'sequence_parallel', False):
+                        grad = param.main_grad if args.DDP_impl == 'local' else param.grad
+                        grads.append(grad.data)
+            coalesced = _flatten_dense_tensors(grads)
+            torch.distributed.all_reduce(
+                coalesced, group=mpu.get_tensor_model_parallel_group())
+            for buf, synced in zip(grads, _unflatten_dense_tensors(
+                    coalesced, grads)):
+                buf.copy_(synced)
+
+
+    def reduce_model_grads(self, args, timers):
+        """All-reduce all grads, and all-reduce embeddings."""
+
+        # All-reduce layer-norm grads (for sequence parallelism).
+        timers('layernorm-grads-all-reduce', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self.allreduce_layernorm_grads(args)
+        timers('layernorm-grads-all-reduce').stop()
+
+        # All-reduce if needed.
+        if args.DDP_impl == 'local':
+            timers('grads-all-reduce', log_level=1).start(
+                barrier=args.barrier_with_L1_time)
+            for model in self.models:
+                model.allreduce_gradients()
+            timers('grads-all-reduce').stop()
+
+        # All-reduce embedding grads.
+        timers('embedding-grads-all-reduce', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self.allreduce_embedding_grads(args)
+        timers('embedding-grads-all-reduce').stop()
+
+
+class MixedPrecisionOptimizer(MegatronOptimizer):
+    """Base class for both the float-16 and the distributed optimizer.
+
+    Arguments:
+        optimizer: base optimizer such as Adam or SGD
+        clip_grad: clip gradeints with this global L2 norm. Note
+            that clipping is ignored if clip_grad == 0
+        log_num_zeros_in_grad: return number of zeros in the gradients.
+        params_have_main_grad: flag indicating if parameters have
+            a `main_grad` field. If this is set, we are assuming
+            that the model parameters are store in the `main_grad`
+            field instead of the typical `grad` field. This happens
+            for the DDP cases where there is a continuous buffer
+            holding the gradients. For example for bfloat16, we want
+            to do gradient accumulation and all-reduces in float32
+            and as a result we store those gradients in the main_grad.
+            Note that main grad is not necessarily in float32.
+        use_contiguous_buffers_in_local_ddp: if true, the local DDP model
+            is using a contiguous buffer to hold the model grads.
+        fp16: if true, the model is running in fp16.
+        bf16: if true, the model is running in bfloat16.
+        params_dtype: used by distributed optimizer.
+        grad_scaler: used for scaling gradients. Note that this can be
+            None. This case happens when `bf16 = True` and we don't
+            use any loss scale. Note that for `bf16 = True`, we can have
+            a constnat gradient scaler. Also for `bf16 = False`, we
+            always require a grad scaler.
+        models: list of models (i.e., the virtual pipelining models). This
+            is used by the distributed optimizer for mapping parameters.
+    """
+
+    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
+                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+                 fp16, bf16, params_dtype, grad_scaler,
+                 models):
+
+        super().__init__(
+            optimizer, clip_grad, log_num_zeros_in_grad,
+            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+            models)
+
+        self.fp16 = fp16
+        self.bf16 = bf16
+        self.params_dtype = params_dtype
+        self.grad_scaler = grad_scaler
+
+        # None grad scaler is only supported for bf16.
+        if self.grad_scaler is None:
+            assert not self.fp16, 'fp16 expects a grad scaler.'
+
+        # Tensor used to determine if a nan/if has happend.
+        # Any non-zero value indicates inf/nan.
+        # Note that we keep this for the cases that grad scaler is none.
+        # We still record nan/inf if we have a bfloat16 with a grad scaler.
+        if self.grad_scaler:
+            self.found_inf = torch.cuda.FloatTensor([0.0])
+
+        # Dummy tensor needed for apex multi-apply tensor.
+        # For bfloat, we don't have multi-tensor apply and for now
+        # we set it to none so the multi-tensor apply gets ignored.
+        if bf16:
+            self._dummy_overflow_buf = None
+        else:
+            self._dummy_overflow_buf = torch.cuda.IntTensor([0])
+
+        # In case grad scaler is not passed, define the unity scale.
+        if self.grad_scaler is None:
+            self._scale_one = torch.cuda.FloatTensor([1.0])
+
+
+    def get_loss_scale(self):
+        if self.grad_scaler is None:
+            return self._scale_one
+        return self.grad_scaler.scale
+
+
+    def reload_model_params(self):
+        self._copy_model_params_to_main_params()
+
+
+    def _unscale_main_grads_and_check_for_nan(self):
+
+        # Collect main grads.
+        main_grads = self._collect_main_grad_data_for_unscaling()
+
+        # Reset found inf.
+        self.found_inf.fill_(0.0)
+
+        # Unscale and set found inf/nan
+        torch._amp_foreach_non_finite_check_and_unscale_(
+            main_grads, self.found_inf, self.grad_scaler.inv_scale)
+
+        # Update across all model parallel instances.
+        torch.distributed.all_reduce(self.found_inf,
+                                     op=torch.distributed.ReduceOp.MAX,
+                                     group=self.get_model_parallel_group())
+
+        # Check for nan.
+        found_inf_flag = (self.found_inf.item() > 0)
+
+        return found_inf_flag
+
+
+    @torch.no_grad()
+    def step(self, args, timers):
+
+        # Copy gradients from model params to main params.
+        timers('optimizer-copy-to-main-grad', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self._copy_model_grads_to_main_grads()
+        timers('optimizer-copy-to-main-grad').stop()
+
+        # Do unscale, check for inf, and update grad scaler only for
+        # the case that grad scaler is provided.
+        if self.grad_scaler:
+
+            # Unscale and check for inf/nan.
+            timers('optimizer-unscale-and-check-inf', log_level=1).start(
+                barrier=args.barrier_with_L1_time)
+            found_inf_flag = self._unscale_main_grads_and_check_for_nan()
+            timers('optimizer-unscale-and-check-inf').stop()
+
+            # We are done with scaling gradients
+            # so we can update the loss scale.
+            self.grad_scaler.update(found_inf_flag)
+
+            # If we found inf/nan, skip the update.
+            if found_inf_flag:
+                return False, None, None
+
+        # Clip the main gradients.
+        timers('optimizer-clip-main-grad', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        grad_norm = None
+        if self.clip_grad > 0.0:
+            grad_norm = self.clip_grad_norm(self.clip_grad)
+        timers('optimizer-clip-main-grad').stop()
+
+        # Count the zeros in the grads.
+        timers('optimizer-count-zeros', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        num_zeros_in_grad = self.count_zeros() if \
+                            self.log_num_zeros_in_grad else None
+        timers('optimizer-count-zeros').stop()
+
+        # Step the optimizer.
+        timers('optimizer-inner-step', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self.optimizer.step()
+        timers('optimizer-inner-step').stop()
+
+        # Update params from main params.
+        timers('optimizer-copy-main-to-model-params', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self._copy_main_params_to_model_params()
+        timers('optimizer-copy-main-to-model-params').stop()
+
+        # Successful update.
+        return True, grad_norm, num_zeros_in_grad
+
+
+class Float16OptimizerWithFloat16Params(MixedPrecisionOptimizer):
+    """Float16 optimizer for fp16 and bf16 data types.
+
+    Arguments:
+        optimizer: base optimizer such as Adam or SGD
+        clip_grad: clip gradeints with this global L2 norm. Note
+            that clipping is ignored if clip_grad == 0
+        log_num_zeros_in_grad: return number of zeros in the gradients.
+        params_have_main_grad: flag indicating if parameters have
+            a `main_grad` field. If this is set, we are assuming
+            that the model parameters are store in the `main_grad`
+            field instead of the typical `grad` field. This happens
+            for the DDP cases where there is a continuous buffer
+            holding the gradients. For example for bfloat16, we want
+            to do gradient accumulation and all-reduces in float32
+            and as a result we store those gradients in the main_grad.
+            Note that main grad is not necessarily in float32.
+        use_contiguous_buffers_in_local_ddp: if true, the local DDP model
+            is using a contiguous buffer to hold the model grads.
+        fp16: if true, the model is running in fp16.
+        bf16: if true, the model is running in bfloat16.
+        grad_scaler: used for scaling gradients. Note that this can be
+            None. This case happens when `bf16 = True` and we don't
+            use any loss scale. Note that for `bf16 = True`, we can have
+            a constnat gradient scaler. Also for `bf16 = False`, we
+            always require a grad scaler.
+        models: list of models (i.e., the virtual pipelining models). This
+            is used by the distributed optimizer for mapping parameters.
+    """
+
+    def __init__(self, optimizer, clip_grad, log_num_zeros_in_grad,
+                 params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+                 fp16, bf16, params_dtype, grad_scaler, models):
+
+        super().__init__(
+            optimizer, clip_grad, log_num_zeros_in_grad,
+            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+            fp16, bf16, params_dtype, grad_scaler, models)
+
+        # ======================
+        # main parameter stuff
+        # ======================
+
+        # Three groups of parameters:
+        #   float16_groups: original float16 parameters
+        #   fp32_from_float16_groups: fp32 copy of float16 parameters
+        #   fp32_from_fp32_groups: original fp32 parameters
+        self.float16_groups = []
+        self.fp32_from_float16_groups = []
+        self.fp32_from_fp32_groups = []
+
+        # For all the groups in the original optimizer:
+        for param_group in self.optimizer.param_groups:
+            float16_params_this_group = []
+            fp32_params_this_group = []
+            fp32_from_float16_params_this_group = []
+            # For all the parameters in this group:
+            for i, param in enumerate(param_group['params']):
+                if param.requires_grad:
+
+                    # float16 params:
+                    if param.type() in ['torch.cuda.HalfTensor',
+                                        'torch.cuda.BFloat16Tensor']:
+                        float16_params_this_group.append(param)
+                        # Create a copy
+                        main_param = param.detach().clone().float()
+                        # Copy tensor model parallel attributes.
+                        tensor_parallel.copy_tensor_model_parallel_attributes(main_param,
+                                                                              param)
+                        if hasattr(param, 'shared'):
+                            main_param.shared = param.shared
+                        # Replace the optimizer params with the new fp32 copy.
+                        param_group['params'][i] = main_param
+
+                        fp32_from_float16_params_this_group.append(main_param)
+                        # Reset existing state dict key to the new main param.
+                        if param in self.optimizer.state:
+                            self.optimizer.state[main_param] \
+                                = self.optimizer.state.pop(param)
+                    # fp32 params.
+                    elif param.type() == 'torch.cuda.FloatTensor':
+                        fp32_params_this_group.append(param)
+                        param_group['params'][i] = param
+
+                    else:
+                        raise TypeError('Wrapped parameters must be one of '
+                                        'torch.cuda.FloatTensor,  '
+                                        'torch.cuda.HalfTensor, or '
+                                        'torch.cuda.BFloat16Tensor. '
+                                        'Received {}'.format(param.type()))
+
+            self.float16_groups.append(float16_params_this_group)
+            self.fp32_from_float16_groups.append(
+                fp32_from_float16_params_this_group)
+            self.fp32_from_fp32_groups.append(fp32_params_this_group)
+
+
+    def zero_grad(self, set_to_none=True):
+        """We only need to zero the model related parameters, i.e.,
+        float16_groups & fp32_from_fp32_groups. We additionally zero
+        fp32_from_float16_groups as a memory optimization to reduce
+        fragmentation; in the case of set_to_none==True, the space
+        used by this field can be safely deallocated at this point."""
+        for group in self.float16_groups:
+            _zero_grad_group_helper(group, set_to_none)
+        for group in self.fp32_from_float16_groups:
+            _zero_grad_group_helper(group, set_to_none)
+        for group in self.fp32_from_fp32_groups:
+            _zero_grad_group_helper(group, set_to_none)
+
+
+    def _collect_main_grad_data_for_unscaling(self):
+
+        main_grads = []
+
+        # fp32 params from float16 ones.
+        for main_group in self.fp32_from_float16_groups:
+            for main_param in main_group:
+                if main_param.grad is not None:
+                    main_grads.append(main_param.grad.data)
+
+        # Append fp32 parameters.
+        for main_group in self.fp32_from_fp32_groups:
+            for main_param in main_group:
+                if main_param.grad is not None:
+                    main_grads.append(main_param.grad.data)
+        
+        return main_grads
+
+
+    def _get_model_and_main_params_data_float16(self):
+        model_data = []
+        main_data = []
+        for model_group, main_group in zip(self.float16_groups,
+                                           self.fp32_from_float16_groups):
+            for model_param, main_param in zip(model_group, main_group):
+                model_data.append(model_param.data)
+                main_data.append(main_param.data)
+        return model_data, main_data
+
+
+    def _copy_model_grads_to_main_grads(self):
+        # This only needs to be done for the float16 group.
+        for model_group, main_group in zip(self.float16_groups,
+                                           self.fp32_from_float16_groups):
+            for model_param, main_param in zip(model_group, main_group):
+                if self.params_have_main_grad and hasattr(model_param, 'main_grad'):
+                    main_param.grad = model_param.main_grad.float()
+                else:
+                    if model_param.grad is not None:
+                        main_param.grad = model_param.grad.float()
+
+                # Safe to deallocate model's grad/main_grad after copying.
+                # (If using contiguous buffers, main_grad's memory should
+                # persist and therefore should not be deallocated.)
+                model_param.grad = None
+                if self.params_have_main_grad and \
+                   not self.use_contiguous_buffers_in_local_ddp:
+                    model_param.main_grad = None
+
+        # For fp32 grads, we need to reset the grads to main grad.
+        if self.params_have_main_grad:
+            for model_group in self.fp32_from_fp32_groups:
+                for model_param in model_group:
+                    model_param.grad = model_param.main_grad
+
+                    # Safe to de-reference model's main_grad after copying.
+                    # (If using contiguous buffers, main_grad's memory should
+                    # persist and therefore should not be deallocated.)
+                    if not self.use_contiguous_buffers_in_local_ddp:
+                        model_param.main_grad = None
+
+
+    def _copy_main_params_to_model_params(self):
+        # Only needed for the float16 params.
+        model_data, main_data = self._get_model_and_main_params_data_float16()
+        _multi_tensor_copy_this_to_that(this=main_data, that=model_data,
+                                        overflow_buf=self._dummy_overflow_buf)
+
+
+    def _copy_model_params_to_main_params(self):
+        # Only needed for the float16 params.
+        model_data, main_data = self._get_model_and_main_params_data_float16()
+        _multi_tensor_copy_this_to_that(this=model_data, that=main_data,
+                                        overflow_buf=self._dummy_overflow_buf)
+
+
+    def state_dict(self):
+        state_dict = {}
+        state_dict['optimizer'] = self.optimizer.state_dict()
+        if self.grad_scaler:
+            state_dict['grad_scaler'] = self.grad_scaler.state_dict()
+        state_dict['fp32_from_fp16_params'] = self.fp32_from_float16_groups
+        return state_dict
+
+
+    def load_state_dict(self, state_dict):
+        # Optimizer.
+        optimizer_key = 'optimizer'
+        if optimizer_key not in state_dict:
+            optimizer_key = 'optimizer_state_dict'
+            print_rank_0('***WARNING*** loading optimizer from '
+                         'an old checkpoint ...')
+        self.optimizer.load_state_dict(state_dict[optimizer_key])
+
+        # Grad scaler.
+        if 'grad_scaler' not in state_dict:
+            if self.fp16:
+                print_rank_0('***WARNING*** found an old checkpoint, will not '
+                             'load grad scaler ...')
+        else:
+            if self.grad_scaler:
+                self.grad_scaler.load_state_dict(state_dict['grad_scaler'])
+            else:
+                print_rank_0('***WARNING*** fould the grad scaler in the '
+                             'checkpoint but it is None in the class. '
+                             'Skipping loading grad scaler ...')
+
+        # Copy data for the main params.
+        fp32_from_float16_params_key = 'fp32_from_fp16_params'
+        if fp32_from_float16_params_key not in state_dict:
+            fp32_from_float16_params_key = 'fp32_from_fp16'
+        for current_group, saved_group in zip(
+                self.fp32_from_float16_groups,
+                state_dict[fp32_from_float16_params_key]):
+            for current_param, saved_param in zip(current_group, saved_group):
+                current_param.data.copy_(saved_param.data)
+
+
+class FP32Optimizer(MegatronOptimizer):
+
+    def __init__(self, optimizer, clip_grad,
+                 log_num_zeros_in_grad,
+                 params_have_main_grad,
+                 use_contiguous_buffers_in_local_ddp,
+                 models):
+
+        super(FP32Optimizer, self).__init__(
+            optimizer, clip_grad, log_num_zeros_in_grad,
+            params_have_main_grad, use_contiguous_buffers_in_local_ddp,
+            models)
+
+        self._scale = torch.cuda.FloatTensor([1.0])
+
+
+    def zero_grad(self, set_to_none=True):
+        """Copied from torch.optim.optimizer"""
+        for group in self.optimizer.param_groups:
+            _zero_grad_group_helper(group['params'], set_to_none)
+
+
+    def get_loss_scale(self):
+        """FP32 optimizer does not do any scaling."""
+        return self._scale
+
+
+    @torch.no_grad()
+    def step(self, args, timers):
+        """Clip gradients (if needed) and step the base optimizer.
+        Always return successful since there is no overflow."""
+
+        # Copy main_grads to grads.
+        timers('optimizer-copy-to-main-grad', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        if self.params_have_main_grad:
+            for param_group in self.optimizer.param_groups:
+                for param in param_group['params']:
+                    param.grad = param.main_grad
+
+                    # Safe to de-reference model's main_grad after copying.
+                    # (If using contiguous buffers, main_grad's memory should
+                    # persist and therefore should not be deallocated.)
+                    if not self.use_contiguous_buffers_in_local_ddp:
+                        param.main_grad = None
+        timers('optimizer-copy-to-main-grad').stop()
+
+        # Clip gradients.
+        timers('optimizer-clip-main-grad', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        grad_norm = None
+        if self.clip_grad > 0.0:
+            grad_norm = self.clip_grad_norm(self.clip_grad)
+        timers('optimizer-clip-main-grad').stop()
+
+        # count the zeros in the grads
+        timers('optimizer-count-zeros', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        num_zeros_in_grad = self.count_zeros() if \
+                            self.log_num_zeros_in_grad else None
+        timers('optimizer-count-zeros').stop()
+
+        # Update parameters.
+        timers('optimizer-inner-step', log_level=1).start(
+            barrier=args.barrier_with_L1_time)
+        self.optimizer.step()
+        timers('optimizer-inner-step').stop()
+
+        # No overflow for FP32 optimizer.
+        return True, grad_norm, num_zeros_in_grad
+
+
+    def reload_model_params(self):
+        pass
+
+
+    def state_dict(self):
+        return self.optimizer.state_dict()
+
+
+    def load_state_dict(self, state_dict):
+        self.optimizer.load_state_dict(state_dict)

megatron/optimizer_param_scheduler.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Learning rate decay and weight decay incr functions."""
+
+import math
+
+from megatron import print_rank_0
+
+class OptimizerParamScheduler(object):
+    """Anneals learning rate and weight decay"""
+
+    def __init__(self, optimizer, max_lr, min_lr,
+                 lr_warmup_steps, lr_decay_steps, lr_decay_style,
+                 start_wd, end_wd, wd_incr_steps, wd_incr_style,
+                 use_checkpoint_opt_param_scheduler=True,
+                 override_opt_param_scheduler=False):
+
+        # Class values.
+        self.optimizer = optimizer
+
+        self.max_lr = float(max_lr)
+        self.min_lr = min_lr
+        assert self.min_lr >= 0.0
+        assert self.max_lr >= self.min_lr
+
+        self.lr_warmup_steps = lr_warmup_steps
+        self.num_steps = 0
+        self.lr_decay_steps = lr_decay_steps
+        assert self.lr_decay_steps > 0
+        assert self.lr_warmup_steps < self.lr_decay_steps
+
+        self.lr_decay_style = lr_decay_style
+
+        self.start_wd = start_wd
+        self.end_wd = end_wd
+        assert self.start_wd >= 0.0
+        assert self.end_wd >= self.start_wd
+        self.wd_incr_steps = wd_incr_steps
+        self.wd_incr_style = wd_incr_style
+
+        self.override_opt_param_scheduler = override_opt_param_scheduler
+        self.use_checkpoint_opt_param_scheduler = use_checkpoint_opt_param_scheduler
+        if self.override_opt_param_scheduler:
+            assert not self.use_checkpoint_opt_param_scheduler, 'both override and '\
+                'use-checkpoint are set.'
+
+        # Set the learning rate
+        self.step(0)
+        print_rank_0('> learning rate decay style: {}'.format(self.lr_decay_style))
+
+
+    def get_wd(self):
+        """ Weight decay incr functions"""
+        if self.num_steps > self.wd_incr_steps:
+            return self.end_wd
+
+        if self.wd_incr_style == 'constant':
+            assert self.start_wd == self.end_wd
+            return self.end_wd
+
+        incr_ratio = float(self.num_steps) / float(self.wd_incr_steps)
+        assert incr_ratio >= 0.0
+        assert incr_ratio <= 1.0
+        delta_wd = self.end_wd - self.start_wd
+
+        if self.wd_incr_style == 'linear':
+            coeff = incr_ratio
+        elif self.wd_incr_style == 'cosine':
+            coeff = 0.5 * (math.cos(math.pi * (1 - incr_ratio)) + 1.0)
+        else:
+            raise Exception('{} weight decay increment style is not supported.'.format(
+                self.wd_incr_style))
+
+        return self.start_wd + coeff * delta_wd
+
+
+    def get_lr(self):
+        """Learning rate decay functions from:
+              https://openreview.net/pdf?id=BJYwwY9ll pg. 4"""
+
+        # Use linear warmup for the initial part.
+        if self.lr_warmup_steps > 0 and self.num_steps <= self.lr_warmup_steps:
+            return self.max_lr * float(self.num_steps) / \
+                float(self.lr_warmup_steps)
+
+        # If the learning rate is constant, just return the initial value.
+        if self.lr_decay_style == 'constant':
+            return self.max_lr
+
+        # For any steps larger than `self.lr_decay_steps`, use `self.min_lr`.
+        if self.num_steps > self.lr_decay_steps:
+            return self.min_lr
+
+        # If we are done with the warmup period, use the decay style.
+        if self.lr_decay_style == 'inverse-square-root':
+            warmup_steps = max(self.lr_warmup_steps, 1)
+            num_steps = max(self.num_steps, 1)
+            lr = self.max_lr * warmup_steps ** 0.5 / (num_steps ** 0.5)
+            return max(self.min_lr, lr)
+
+        num_steps_ = self.num_steps - self.lr_warmup_steps
+        decay_steps_ = self.lr_decay_steps - self.lr_warmup_steps
+        decay_ratio = float(num_steps_) / float(decay_steps_)
+        assert decay_ratio >= 0.0
+        assert decay_ratio <= 1.0
+        delta_lr = self.max_lr - self.min_lr
+
+        if self.lr_decay_style == 'linear':
+            coeff = (1.0 - decay_ratio)
+        elif self.lr_decay_style == 'cosine':
+            coeff = 0.5 * (math.cos(math.pi * decay_ratio) + 1.0)
+        else:
+            raise Exception('{} decay style is not supported.'.format(
+                self.lr_decay_style))
+
+        return self.min_lr + coeff * delta_lr
+
+
+    def step(self, increment):
+        """Set lr for all parameters groups."""
+        self.num_steps += increment
+        new_lr = self.get_lr()
+        new_wd = self.get_wd()
+        for group in self.optimizer.param_groups:
+            group['lr'] = new_lr * group.get('lr_mult', 1.0)
+            group['weight_decay'] = new_wd * group.get('wd_mult', 1.0)
+
+
+    def state_dict(self):
+        state_dict = {
+            'max_lr': self.max_lr,
+            'lr_warmup_steps': self.lr_warmup_steps,
+            'num_steps': self.num_steps,
+            'lr_decay_style': self.lr_decay_style,
+            'lr_decay_steps': self.lr_decay_steps,
+            'min_lr': self.min_lr,
+            'start_wd': self.start_wd,
+            'end_wd': self.end_wd,
+            'wd_incr_style': self.wd_incr_style,
+            'wd_incr_steps': self.wd_incr_steps
+        }
+        return state_dict
+
+
+    def _check_and_set(self, cls_value, sd_value, name):
+        """Auxiliary function for checking the values in the checkpoint and
+        setting them."""
+        if self.override_opt_param_scheduler:
+            print_rank_0(' > overriding {} value to {}'.format(name, cls_value))
+            return cls_value
+
+        if not self.use_checkpoint_opt_param_scheduler:
+            assert cls_value == sd_value, \
+                f'OptimizerParamScheduler: class input value {cls_value} and checkpoint' \
+                f'value {sd_value} for {name} do not match'
+        print_rank_0(' > using checkpoint value {} for {}'.format(sd_value,
+                                                                  name))
+        return sd_value
+
+
+    def load_state_dict(self, sd):
+
+        if 'start_lr' in sd:
+            max_lr_ = sd['start_lr']
+        else:
+            max_lr_ = sd['max_lr']
+        self.max_lr = self._check_and_set(self.max_lr, max_lr_,
+                                          'learning rate')
+        
+        self.min_lr = self._check_and_set(self.min_lr, sd['min_lr'],
+                                          'minimum learning rate')
+
+        if 'warmup_iter' in sd:
+            lr_warmup_steps_ = sd['warmup_iter']
+        elif 'warmup_steps' in sd:
+            lr_warmup_steps_ = sd['warmup_steps']
+        else:
+            lr_warmup_steps_ = sd['lr_warmup_steps']
+        self.lr_warmup_steps = self._check_and_set(self.lr_warmup_steps,
+                                                lr_warmup_steps_,
+                                                'warmup iterations')
+
+        if 'end_iter' in sd:
+            lr_decay_steps_ = sd['end_iter']
+        elif 'decay_steps' in sd:
+            lr_decay_steps_  = sd['decay_steps']
+        else:
+            lr_decay_steps_ = sd['lr_decay_steps']
+        self.lr_decay_steps = self._check_and_set(self.lr_decay_steps, lr_decay_steps_,
+                                               'total number of iterations')
+
+        if 'decay_style' in sd:
+            lr_decay_style_ = sd['decay_style']
+        else:
+            lr_decay_style_ = sd['lr_decay_style']
+        self.lr_decay_style = self._check_and_set(self.lr_decay_style,
+                                               lr_decay_style_,
+                                               'learning rate decay style')
+
+        if 'num_iters' in sd:
+            num_steps = sd['num_iters']
+        else:
+            num_steps = sd['num_steps']
+        self.step(increment=num_steps)
+
+
+        if 'start_wd' in sd:
+            self.start_wd = self._check_and_set(self.start_wd,
+                                                sd['start_wd'],
+                                                "start weight decay")
+            self.end_wd = self._check_and_set(self.end_wd,
+                                                sd['end_wd'],
+                                                "end weight decay")
+            self.wd_incr_steps = self._check_and_set(self.wd_incr_steps,
+                                                sd['wd_incr_steps'],
+                                                "total number of weight decay iterations")
+            self.wd_incr_style = self._check_and_set(self.wd_incr_style,
+                                                sd['wd_incr_style'],
+                                                "weight decay incr style")
+            
+
+
+
+
+
+
+

megatron/static/index.html

+<!-- coding=utf-8-->
+<!-- Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.-->
+
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<title>Megatron</title>
+<style>
+
+.wrapper {
+max-width: 75%;
+margin: auto;
+}
+
+h1 { 
+margin: 3rem 0 1rem 0; 
+padding: 0;
+font-size: 1.5rem;
+}
+
+textarea {
+width: 100%;
+min-height: 300px;
+resize: none;
+border-radius: 8px;
+border: 1px solid #ddd;
+padding: 0.5rem;
+box-shadow: inset 0 0 0.25rem #ddd;
+&:focus {
+outline: none;
+border: 1px solid darken(#ddd, 5%);
+box-shadow: inset 0 0 0.5rem darken(#ddd, 5%);
+}
+}
+
+#the-count {
+float: right;
+padding: 0.1rem 0 0 0;
+font-size: 0.875rem;
+}
+/* Chat containers */
+.container {
+font-family: 'Arial', sans-serif;
+font-size: 16px;
+border: 2px solid #dedede;
+background-color: #f1f1f1;
+border-radius: 5px;
+padding: 15px;
+margin: 10px 0;
+}
+
+
+/* Clear floats */
+.container::after {
+content: "";
+clear: both;
+display: table;
+}
+
+/* Style images */
+.container img {
+float: left;
+max-width: 60px;
+width: 100%;
+margin-right: 20px;
+border-radius: 50%;
+}
+
+</style>
+</head>
+<body>
+<div class="wrapper">
+<h1>Prompt Megatron</h1>
+<textarea name="prompt" id="prompt" maxlength="1024" placeholder="Add prompt"autofocus></textarea>
+<label for="tokens_to_generate">Number tokens to generate (1-1024):</label>
+<input type="number" id="tokens_to_generate" name="tokens_to_generate" min="10" max="256", value=32>
+<button onclick="submit_query()">Submit</button>
+
+<div id="the-count">
+<span id="current">0</span>
+<span id="maximum">/ 1000</span>
+</div>
+<textarea name="response" id="response" maxlength="2048" placeholder="Megatron response..."></textarea>
+</div>
+<script type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
+<script type="text/javascript">
+
+	function submit_query() {
+		$("#response").val("Waiting for Megatron response...");
+		$.ajax({
+			url:"api",
+			type:"PUT",
+			data:JSON.stringify({prompts: [$("#prompt").val()], tokens_to_generate: parseInt($("#tokens_to_generate").val(),10)}),
+			contentType:"application/json; charset=utf-8",
+			dataType:"json",
+			success: function(data){
+				data.max_len=35;
+				$("#response").val(data.text);
+			}
+		});
+	}
+	
+$('textarea').keyup(function() {
+var characterCount = $(this).val().length,
+current = $('#current'),
+maximum = $('#maximum'),
+theCount = $('#the-count');
+
+current.text(characterCount);
+
+if (characterCount >= 800) {
+maximum.css('color', '#8f0001');
+current.css('color', '#8f0001');
+theCount.css('font-weight','bold');
+} else {
+maximum.css('color','#666');
+theCount.css('font-weight','normal');
+}
+});
+</script>
+</body>
+</html>
+

megatron/text_generation/__init__.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+
+from .api import (
+    generate,
+    generate_and_post_process,
+    beam_search_and_post_process)

megatron/text_generation/api.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Inference API."""
+
+
+import torch
+
+from megatron.core import mpu
+from .communication import broadcast_float_list
+from .generation import (
+        generate_tokens_probs_and_return_on_first_stage,
+        score_and_return_on_first_stage,
+        beam_search_and_return_on_first_stage)
+from .tokenization import (
+    tokenize_prompts,
+    detokenize_generations)
+
+def generate_and_post_process(model,
+                              prompts=None,
+                              tokens_to_generate=0,
+                              return_output_log_probs=False,
+                              top_k_sampling=0,
+                              top_p_sampling=0.0,
+                              top_p_decay=0.0,
+                              top_p_bound=0.0,
+                              temperature=1.0,
+                              add_BOS=False,
+                              use_eod_token_for_early_termination=True,
+                              stop_on_double_eol=False,
+                              stop_on_eol=False,
+                              prevent_newline_after_colon=False,
+                              random_seed=-1):
+    """Run inference and post-process outputs, i.e., detokenize,
+    move to cpu and convert to list."""
+
+    # Main inference.
+    tokens, lengths, output_log_probs = generate(
+        model,
+        prompts=prompts,
+        tokens_to_generate=tokens_to_generate,
+        return_output_log_probs=return_output_log_probs,
+        top_k_sampling=top_k_sampling,
+        top_p_sampling=top_p_sampling,
+        top_p_decay=top_p_decay,
+        top_p_bound=top_p_bound,
+        temperature=temperature,
+        add_BOS=add_BOS,
+        use_eod_token_for_early_termination=use_eod_token_for_early_termination,
+        stop_on_double_eol=stop_on_double_eol,
+        stop_on_eol=stop_on_eol,
+        prevent_newline_after_colon=prevent_newline_after_colon,
+        random_seed=random_seed)
+
+    # Only post-process on first stage.
+    if mpu.is_pipeline_first_stage():
+        tokens, prompts_plus_generations, prompts_plus_generations_segments = \
+            detokenize_generations(tokens, lengths, True)
+
+        if return_output_log_probs:
+            output_log_probs = output_log_probs.cpu().numpy().tolist()
+            for i, (prob, seg) in enumerate(zip(output_log_probs, prompts_plus_generations_segments)):
+                output_log_probs[i] = prob[:len(seg)-1]
+
+        return prompts_plus_generations, prompts_plus_generations_segments, \
+            output_log_probs, tokens
+
+    return None
+
+def generate(model,
+             prompts=None,
+             tokens_to_generate=0,
+             return_output_log_probs=False,
+             top_k_sampling=0,
+             top_p_sampling=0.0,
+             top_p_decay=0.0,
+             top_p_bound=0.0,
+             temperature=1.0,
+             add_BOS=False,
+             use_eod_token_for_early_termination=True,
+             stop_on_double_eol=False,
+             stop_on_eol=False,
+             prevent_newline_after_colon=False,
+             random_seed=-1):
+    """Given prompts and input parameters, run inference and return:
+       tokens: prompts plus the generated tokens.
+       lengths: length of the prompt + generations. Note that we can
+           discard tokens in the tokens tensor that are after the
+           corresponding length.
+       output_log_probs: log probs of the tokens.
+    """
+
+    # Make sure input params are avaialble to all ranks.
+    values = [tokens_to_generate,
+              return_output_log_probs,
+              top_k_sampling, top_p_sampling, top_p_decay, top_p_bound,
+              temperature, add_BOS, use_eod_token_for_early_termination,
+              stop_on_double_eol,
+              stop_on_eol,
+              prevent_newline_after_colon,
+              random_seed]
+    values_float_tensor = broadcast_float_list(len(values), float_list=values)
+    tokens_to_generate = int(values_float_tensor[0].item())
+    return_output_log_probs = bool(values_float_tensor[1].item())
+    top_k_sampling = int(values_float_tensor[2].item())
+    top_p_sampling = values_float_tensor[3].item()
+    top_p_decay = values_float_tensor[4].item()
+    top_p_bound = values_float_tensor[5].item()
+    temperature = values_float_tensor[6].item()
+    add_BOS = bool(values_float_tensor[7].item())
+    use_eod_token_for_early_termination = bool(values_float_tensor[8].item())
+    stop_on_double_eol = bool(values_float_tensor[9].item())
+    stop_on_eol = bool(values_float_tensor[10].item())
+    prevent_newline_after_colon = bool(values_float_tensor[11].item())
+    random_seed = int(values_float_tensor[12].item())
+
+    if random_seed != -1:
+        torch.random.manual_seed(random_seed)
+
+    # Tokenize prompts and get the batch.
+    # Note that these tensors are broadcaseted to all ranks.
+    if torch.distributed.get_rank() == 0:
+        assert prompts is not None
+    
+    context_tokens_tensor, context_length_tensor = tokenize_prompts(
+        prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)
+
+    if tokens_to_generate == 0:
+        return score_and_return_on_first_stage(
+            model, context_tokens_tensor, context_length_tensor)
+    
+    # Main inference function.
+    # Note that the outputs are available on the first stage.
+    return generate_tokens_probs_and_return_on_first_stage(
+        model, context_tokens_tensor, context_length_tensor,
+        return_output_log_probs=return_output_log_probs,
+        top_k=top_k_sampling,
+        top_p=top_p_sampling,
+        top_p_decay=top_p_decay,
+        top_p_bound=top_p_bound,
+        temperature=temperature,
+        use_eod_token_for_early_termination=use_eod_token_for_early_termination,
+        stop_on_double_eol=stop_on_double_eol,
+        stop_on_eol=stop_on_eol,
+        prevent_newline_after_colon=prevent_newline_after_colon)
+
+def beam_search_and_post_process(model,
+                                 prompts=None,
+                                 tokens_to_generate=0,
+                                 beam_size=0,
+                                 add_BOS=False,
+                                 stop_token=50256,
+                                 num_return_gen=1,
+                                 length_penalty=1,
+                                 prevent_newline_after_colon=False):
+    """Run beam search and post-process outputs, i.e., detokenize,
+    move to cpu and convert to list."""
+
+    # Main inference.
+    tokens, scores = beam_search(model,
+                                 prompts=prompts,
+                                 tokens_to_generate=tokens_to_generate,
+                                 beam_size=beam_size,
+                                 add_BOS=add_BOS,
+                                 stop_token=stop_token,
+                                 num_return_gen=num_return_gen,
+                                 length_penalty=length_penalty,
+                                 prevent_newline_after_colon=prevent_newline_after_colon)
+    # Only post-process on first stage.
+    if mpu.is_pipeline_first_stage():
+        lengths = tokens.size(1)*torch.ones(beam_size, dtype=torch.int64, device=torch.cuda.current_device()) 
+        tokens, prompts_plus_generations, prompts_plus_generations_segments = detokenize_generations(tokens, lengths, True)
+        scores = scores.cpu().numpy().tolist()
+        return prompts_plus_generations, prompts_plus_generations_segments, scores
+
+    return None
+
+def beam_search(model, prompts=None, tokens_to_generate=0, beam_size=0, add_BOS=False, stop_token=50256, num_return_gen=1, length_penalty=1, prevent_newline_after_colon=False):
+    # Make sure input params are avaialble to all ranks.
+    values = [tokens_to_generate,
+              beam_size,
+              add_BOS,
+              stop_token,
+              num_return_gen,
+              length_penalty,
+              prevent_newline_after_colon]
+    values_float_tensor = broadcast_float_list(len(values), float_list=values)
+    tokens_to_generate = int(values_float_tensor[0].item())
+    beam_size = int(values_float_tensor[1].item())
+    add_BOS = bool(values_float_tensor[2].item())
+    stop_token = int(values_float_tensor[3].item())
+    num_return_gen = int(values_float_tensor[4].item())
+    length_penalty = values_float_tensor[5].item()
+    prevent_newline_after_colon = values_float_tensor[6].item()
+
+    context_tokens_tensor, context_length_tensor = tokenize_prompts(
+        prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS)
+    
+    return beam_search_and_return_on_first_stage(model, context_tokens_tensor, context_length_tensor, 
+            beam_size, stop_token=stop_token, num_return_gen=num_return_gen, length_penalty=length_penalty,
+            prevent_newline_after_colon=prevent_newline_after_colon)