Speed improvements (#531)

Summary:
* Add FusedLayerNorm and FusedAdam
* Softmax and zero grad optimizations
Pull Request resolved: #531

Differential Revision: D14218457

Pulled By: myleott

fbshipit-source-id: 5656b2d0152cd85f77dc21ec0e1439ec04b9fa89

Author: Myle Ott

README.md

@@ -36,12 +36,12 @@ translation and language modeling datasets.
 ![Model](fairseq.gif)
 
 # Requirements and Installation
-* A [PyTorch installation](http://pytorch.org/)
+
+* [PyTorch](http://pytorch.org/) version >= 1.0.0
+* Python version >= 3.6
 * For training new models, you'll also need an NVIDIA GPU and [NCCL](https://github.com/NVIDIA/nccl)
-* Python version 3.6
 
-Currently fairseq requires PyTorch version >= 1.0.0.
-Please follow the instructions here: https://github.com/pytorch/pytorch#installation.
+Please follow the instructions here to install PyTorch: https://github.com/pytorch/pytorch#installation.
 
 If you use Docker make sure to increase the shared memory size either with
 `--ipc=host` or `--shm-size` as command line options to `nvidia-docker run`.
@@ -60,6 +60,12 @@ cd fairseq
 pip install --editable .
 ```
 
+**Improved training speed**
+
+Training speed can be further improved by installing NVIDIA's
+[apex](https://github.com/NVIDIA/apex) library with the `--cuda_ext` option.
+fairseq will automatically switch to the faster modules provided by apex.
+
 # Getting Started
 
 The [full documentation](https://fairseq.readthedocs.io/) contains instructions

fairseq/distributed_utils.py

@@ -122,19 +122,23 @@ def all_gather_list(data, group=None, max_size=16384):
     if not hasattr(all_gather_list, '_buffer') or \
             all_gather_list._buffer.numel() < buffer_size:
         all_gather_list._buffer = torch.cuda.ByteTensor(buffer_size)
+        all_gather_list._cpu_buffer = torch.ByteTensor(max_size).pin_memory()
     buffer = all_gather_list._buffer
     buffer.zero_()
+    cpu_buffer = all_gather_list._cpu_buffer
 
     enc = pickle.dumps(data)
     enc_size = len(enc)
     if enc_size + 2 > max_size:
         raise ValueError('encoded data exceeds max_size: {}'.format(enc_size + 2))
     assert max_size < 255*256
 
-    buffer_rank = buffer[rank * max_size : (rank + 1) * max_size]
-    buffer_rank[0] = enc_size // 255  # this encoding works for max_size < 65k
-    buffer_rank[1] = enc_size % 255
-    buffer_rank[2:enc_size+2] = torch.ByteTensor(list(enc))
+    cpu_buffer[0] = enc_size // 255  # this encoding works for max_size < 65k
+    cpu_buffer[1] = enc_size % 255
+    cpu_buffer[2 : enc_size + 2] = torch.ByteTensor(list(enc))
+    start = rank * max_size
+    size = enc_size + 2
+    buffer[start : start + size].copy_(cpu_buffer[:size])
 
     all_reduce(buffer, group=group)
 
@@ -144,9 +148,7 @@ def all_gather_list(data, group=None, max_size=16384):
             out_buffer = buffer[i * max_size : (i + 1) * max_size]
             size = (255 * utils.item(out_buffer[0])) + utils.item(out_buffer[1])
             if size > 0:
-                result.append(
-                    pickle.loads(bytes(out_buffer[2:size+2].tolist()))
-                )
+                result.append(pickle.loads(bytes(out_buffer[2 : size + 2].tolist())))
         return result
     except pickle.UnpicklingError:
         raise Exception(

fairseq/models/fairseq_decoder.py

@@ -6,7 +6,8 @@
 # can be found in the PATENTS file in the same directory.
 
 import torch.nn as nn
-import torch.nn.functional as F
+
+from fairseq import utils
 
 
 class FairseqDecoder(nn.Module):
@@ -15,6 +16,7 @@ class FairseqDecoder(nn.Module):
     def __init__(self, dictionary):
         super().__init__()
         self.dictionary = dictionary
+        self.onnx_trace = False
 
     def forward(self, prev_output_tokens, encoder_out):
         """
@@ -33,6 +35,9 @@ def forward(self, prev_output_tokens, encoder_out):
         """
         raise NotImplementedError
 
+    def prepare_for_onnx_export_(self):
+        self.onnx_trace = True
+
     def get_normalized_probs(self, net_output, log_probs, sample):
         """Get normalized probabilities (or log probs) from a net's output."""
 
@@ -45,11 +50,11 @@ def get_normalized_probs(self, net_output, log_probs, sample):
             out = self.adaptive_softmax.get_log_prob(net_output[0], target=target)
             return out.exp_() if not log_probs else out
 
-        logits = net_output[0].float()
+        logits = net_output[0]
         if log_probs:
-            return F.log_softmax(logits, dim=-1)
+            return utils.log_softmax(logits, dim=-1, onnx_trace=self.onnx_trace)
         else:
-            return F.softmax(logits, dim=-1)
+            return utils.softmax(logits, dim=-1, onnx_trace=self.onnx_trace)
 
     def max_positions(self):
         """Maximum input length supported by the decoder."""

fairseq/models/fconv_self_att.py

@@ -13,8 +13,8 @@
 import torch.nn.functional as F
 
 from fairseq.modules import (
-    DownsampledMultiHeadAttention, GradMultiply, LearnedPositionalEmbedding,
-    LinearizedConvolution,
+    DownsampledMultiHeadAttention, GradMultiply, LayerNorm,
+    LearnedPositionalEmbedding, LinearizedConvolution,
 )
 from fairseq import utils
 
@@ -351,13 +351,13 @@ def expand_bool_array(val):
             # pretrained and trained models are joined
             self.joining = nn.Sequential(
                 Linear(out_embed_dim*2, out_embed_dim*2),
-                nn.LayerNorm(out_embed_dim*2),
+                LayerNorm(out_embed_dim*2),
                 nn.GLU(),
                 Linear(out_embed_dim, out_embed_dim*2),
-                nn.LayerNorm(out_embed_dim*2),
+                LayerNorm(out_embed_dim*2),
                 nn.GLU(),
                 Linear(out_embed_dim, out_embed_dim),
-                nn.LayerNorm(out_embed_dim)
+                LayerNorm(out_embed_dim)
             )
             # pretrained model contains an output layer that is nhid -> vocab size
             # but the models are combined in their hidden state
@@ -470,7 +470,7 @@ def __init__(self, out_channels, embed_dim, num_heads, project_input=False, gate
         self.in_proj_q = Linear(out_channels, embed_dim)
         self.in_proj_k = Linear(out_channels, embed_dim)
         self.in_proj_v = Linear(out_channels, embed_dim)
-        self.ln = nn.LayerNorm(out_channels)
+        self.ln = LayerNorm(out_channels)
 
     def forward(self, x):
         residual = x

fairseq/models/lightconv.py

@@ -11,17 +11,16 @@
 import torch.nn as nn
 import torch.nn.functional as F
 
-from fairseq import options
-from fairseq import utils
-
+from fairseq import options, utils
 from fairseq.modules import (
-    AdaptiveInput, AdaptiveSoftmax, CharacterTokenEmbedder, LearnedPositionalEmbedding, MultiheadAttention,
-    SinusoidalPositionalEmbedding, DynamicConv1dTBC, LightweightConv1dTBC
+    AdaptiveInput, AdaptiveSoftmax, CharacterTokenEmbedder, LayerNorm,
+    LearnedPositionalEmbedding, MultiheadAttention, SinusoidalPositionalEmbedding,
+    DynamicConv1dTBC, LightweightConv1dTBC,
 )
 
 from . import (
-    FairseqIncrementalDecoder, FairseqEncoder, FairseqLanguageModel, FairseqModel, register_model,
-    register_model_architecture,
+    FairseqIncrementalDecoder, FairseqEncoder, FairseqLanguageModel,
+    FairseqModel, register_model, register_model_architecture,
 )
 
 
@@ -771,11 +770,6 @@ def Embedding(num_embeddings, embedding_dim, padding_idx):
     return m
 
 
-def LayerNorm(embedding_dim):
-    m = nn.LayerNorm(embedding_dim)
-    return m
-
-
 def Linear(in_features, out_features, bias=True):
     m = nn.Linear(in_features, out_features, bias)
     nn.init.xavier_uniform_(m.weight)

fairseq/models/transformer.py

@@ -11,17 +11,15 @@
 import torch.nn as nn
 import torch.nn.functional as F
 
-from fairseq import options
-from fairseq import utils
-
+from fairseq import options, utils
 from fairseq.modules import (
-    AdaptiveInput, AdaptiveSoftmax, CharacterTokenEmbedder, LearnedPositionalEmbedding, MultiheadAttention,
-    SinusoidalPositionalEmbedding
+    AdaptiveInput, AdaptiveSoftmax, CharacterTokenEmbedder, LayerNorm,
+    LearnedPositionalEmbedding, MultiheadAttention, SinusoidalPositionalEmbedding,
 )
 
 from . import (
-    FairseqIncrementalDecoder, FairseqEncoder, FairseqLanguageModel, FairseqModel, register_model,
-    register_model_architecture,
+    FairseqIncrementalDecoder, FairseqEncoder, FairseqLanguageModel,
+    FairseqModel, register_model, register_model_architecture,
 )
 
 
@@ -766,11 +764,6 @@ def Embedding(num_embeddings, embedding_dim, padding_idx):
     return m
 
 
-def LayerNorm(embedding_dim):
-    m = nn.LayerNorm(embedding_dim)
-    return m
-
-
 def Linear(in_features, out_features, bias=True):
     m = nn.Linear(in_features, out_features, bias)
     nn.init.xavier_uniform_(m.weight)

fairseq/modules/__init__.py

@@ -14,6 +14,7 @@
 from .dynamic_convolution import DynamicConv1dTBC
 from .grad_multiply import GradMultiply
 from .highway import Highway
+from .layer_norm import LayerNorm
 from .learned_positional_embedding import LearnedPositionalEmbedding
 from .lightweight_convolution import LightweightConv1dTBC
 from .linearized_convolution import LinearizedConvolution
@@ -34,6 +35,7 @@
     'DynamicConv1dTBC',
     'GradMultiply',
     'Highway',
+    'LayerNorm',
     'LearnedPositionalEmbedding',
     'LightweightConv1dTBC',
     'LinearizedConvolution',

fairseq/modules/layer_norm.py

@@ -0,0 +1,18 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the LICENSE file in
+# the root directory of this source tree. An additional grant of patent rights
+# can be found in the PATENTS file in the same directory.
+
+import torch
+
+
+def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True):
+    if torch.cuda.is_available():
+        try:
+            from apex.normalization import FusedLayerNorm
+            return FusedLayerNorm(normalized_shape, eps, elementwise_affine)
+        except ImportError:
+            pass
+    return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine)

fairseq/modules/lightweight_convolution.py

@@ -6,6 +6,7 @@
 # can be found in the PATENTS file in the same directory.
 
 import math
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -121,6 +122,8 @@ def __init__(self, input_size, kernel_size=1, padding_l=None, num_heads=1,
 
         self.reset_parameters()
 
+        self.onnx_trace = False
+
     def reset_parameters(self):
         nn.init.xavier_uniform_(self.weight)
         if self.bias is not None:
@@ -144,6 +147,9 @@ def forward(self, x, incremental_state=None, unfold=False):
             output = output + self.bias.view(1, 1, -1)
         return output
 
+    def prepare_for_onnx_export_(self):
+        self.onnx_trace = True
+
     def _forward_unfolded(self, x, incremental_state):
         '''The conventional implementation of convolutions.
         Unfolding the input by having a window shifting to the right.'''
@@ -167,7 +173,7 @@ def _forward_unfolded(self, x, incremental_state):
             x_unfold = x_unfold.view(T*B*H, R, K)
 
         if self.weight_softmax:
-            weight = F.softmax(weight.float(), dim=1).type_as(weight)
+            weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as(weight)
 
         if incremental_state is not None:
             weight = weight[:, -x_unfold.size(2):]
@@ -192,7 +198,7 @@ def _forward_expanded(self, x, incremental_state):
 
         weight = self.weight.view(H, K)
         if self.weight_softmax:
-            weight = F.softmax(weight.float(), dim=1).type_as(weight)
+            weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as(weight)
         weight = weight.view(1, H, K).expand(T*B, H, K).contiguous()
         weight = weight.view(T, B*H, K).transpose(0, 1)
 

fairseq/modules/multihead_attention.py

@@ -184,7 +184,9 @@ def forward(self, query, key, value, key_padding_mask=None, incremental_state=No
                 ).type_as(attn_weights)  # FP16 support: cast to float and back
             attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
 
-        attn_weights = F.softmax(attn_weights.float(), dim=-1).type_as(attn_weights)
+        attn_weights = utils.softmax(
+            attn_weights, dim=-1, onnx_trace=self.onnx_trace,
+        ).type_as(attn_weights)
         attn_weights = F.dropout(attn_weights, p=self.dropout, training=self.training)
 
         attn = torch.bmm(attn_weights, v)