Revert "supporting different intermediate sizes other than 4 * hidden_dim (#389)" (#404)

This reverts commit e549be60.

csrc/transformer/ds_transformer_cuda.cpp

@@ -20,14 +20,13 @@ template <typename T>
 size_t get_workspace_size(int maxBatchSize,
                           int seq_len,
                           int hidden_size,
-                          int intermediate_size,
                           int heads,
                           bool training,
                           bool gelu_checkpoint)
 {
     size_t workSpacesize = 4 * (size_t(maxBatchSize) * seq_len * hidden_size);
     if (training) {
-        workSpacesize += (std::max((size_t(maxBatchSize) * seq_len * intermediate_size),
+        workSpacesize += (std::max((4 * size_t(maxBatchSize) * seq_len * hidden_size),
                                    2 * (size_t(maxBatchSize) * heads * seq_len * seq_len)));
         if (gelu_checkpoint) workSpacesize += 2 * (size_t(maxBatchSize) * seq_len * hidden_size);
     }
@@ -93,15 +92,15 @@ BertTransformerLayer<T>::BertTransformerLayer(int layer_id,
                                                        false,
                                                        !normalize_invertible)),
       _ff1(typename FeedForward<T>::Config(batch_size * seq_length,
-                                           intermediate_size,
+                                           4 * hidden_size,
                                            hidden_size,
                                            gemm_algos[1])),
       _ff2(typename FeedForward<T>::Config(batch_size * seq_length,
                                            hidden_size,
-                                           intermediate_size,
+                                           4 * hidden_size,
                                            gemm_algos[2])),
       _softmax(typename Softmax<T>::Config(batch_size, num_heads, seq_length)),
-      _gelu(typename Gelu<T>::Config(_batch_size, _seq_length, intermediate_size)),
+      _gelu(typename Gelu<T>::Config(_batch_size, _seq_length, _intermediate_size)),
       _attn_prob_dropout(typename Dropout<T>::Config(attn_prob_dropout_ratio,
                                                      _batch_size * _heads * _seq_length,
                                                      _seq_length)),
@@ -144,13 +143,8 @@ BertTransformerLayer<T>::~BertTransformerLayer()
 template <typename T>
 void BertTransformerLayer<T>::Initialize()
 {
-    Context::Instance().GenWorkSpace(get_workspace_size<T>(_batch_size,
-                                                           _seq_length,
-                                                           _hidden_size,
-                                                           _intermediate_size,
-                                                           _heads,
-                                                           _training,
-                                                           _gelu_checkpoint));
+    Context::Instance().GenWorkSpace(get_workspace_size<T>(
+        _batch_size, _seq_length, _hidden_size, _heads, _training, _gelu_checkpoint));
 
     if (std::is_same<T, __half>::value) cublasSetMathMode(_cublasHandle, CUBLAS_TENSOR_OP_MATH);
 }
@@ -349,7 +343,7 @@ void BertTransformerLayer<T>::Backward(int bsz,
     T* buf_2 = buf_1 + small_buf_size;
     T* buf_3 = buf_2 + small_buf_size;
 
-    T* ff2_buf = (_gelu_checkpoint ? (buf_2 + _intermediate_size) : (buf_3 + small_buf_size));
+    T* ff2_buf = buf_3 + (_gelu_checkpoint ? 3 : 1) * small_buf_size;
     T* ctx_bufB_ptr_recomp = ff2_buf + (_seq_length * _seq_length * bsz * _heads);
 
     cudaStream_t streams[2] = {_stream, _stream};

deepspeed/ops/transformer/transformer.py

@@ -18,7 +18,6 @@ class TransformerConfig():
                  batch_size,
                  max_seq_length,
                  hidden_size,
-                 intermediate_size,
                  heads,
                  attn_dropout_ratio,
                  hidden_dropout_ratio,
@@ -27,7 +26,6 @@ class TransformerConfig():
         self.layer_id = -1
         self.batch_size = batch_size
         self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size if intermediate_size > 0 else 4 * hidden_size
         self.max_seq_length = max_seq_length
         self.heads = heads
         self.attn_dropout_ratio = attn_dropout_ratio
@@ -90,7 +88,6 @@ class DeepSpeedTransformerConfig(TransformerConfig):
                  batch_size=-1,
                  max_seq_length=-1,
                  hidden_size=-1,
-                 intermediate_size=-1,
                  heads=-1,
                  attn_dropout_ratio=-1,
                  hidden_dropout_ratio=-1,
@@ -109,7 +106,6 @@ class DeepSpeedTransformerConfig(TransformerConfig):
               self).__init__(batch_size,
                              max_seq_length,
                              hidden_size,
-                             intermediate_size,
                              heads,
                              attn_dropout_ratio,
                              hidden_dropout_ratio,
@@ -436,12 +432,12 @@ class DeepSpeedTransformerLayer(nn.Module):
             self.attn_nw = nn.Parameter(torch.Tensor(self.config.hidden_size))
             self.attn_nb = nn.Parameter(torch.Tensor(self.config.hidden_size))
             self.inter_w = nn.Parameter(
-                torch.Tensor(self.config.intermediate_size,
+                torch.Tensor(4 * self.config.hidden_size,
                              self.config.hidden_size))
-            self.inter_b = nn.Parameter(torch.Tensor(self.config.intermediate_size))
+            self.inter_b = nn.Parameter(torch.Tensor(4 * self.config.hidden_size))
             self.output_w = nn.Parameter(
                 torch.Tensor(self.config.hidden_size,
-                             self.config.intermediate_size))
+                             4 * self.config.hidden_size))
             self.output_b = nn.Parameter(torch.Tensor(self.config.hidden_size))
             self.norm_w = nn.Parameter(torch.Tensor(self.config.hidden_size))
             self.norm_b = nn.Parameter(torch.Tensor(self.config.hidden_size))
@@ -489,7 +485,7 @@ class DeepSpeedTransformerLayer(nn.Module):
                           self.config.batch_size,
                           self.config.hidden_size,
                           self.config.heads,
-                          self.config.intermediate_size,
+                          4 * self.config.hidden_size,
                           self.config.max_seq_length,
                           self.config.attn_dropout_ratio,
                           self.config.hidden_dropout_ratio,

tests/unit/test_cuda_backward.py

@@ -142,7 +142,7 @@ def create_models(ds_config):
                              hidden_size=ds_config.hidden_size,
                              num_hidden_layers=ds_config.num_hidden_layers,
                              num_attention_heads=ds_config.heads,
-                             intermediate_size=ds_config.intermediate_size,
+                             intermediate_size=4 * ds_config.hidden_size,
                              hidden_act="gelu",
                              hidden_dropout_prob=ds_config.hidden_dropout_ratio,
                              attention_probs_dropout_prob=ds_config.attn_dropout_ratio,
@@ -162,12 +162,12 @@ def create_models(ds_config):
     weights.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
     weights[4].data.fill_(1.0)
     weights.append(
-        nn.Parameter(torch.Tensor(ds_config.intermediate_size,
+        nn.Parameter(torch.Tensor(4 * ds_config.hidden_size,
                                   ds_config.hidden_size)))
     weights[5].data.normal_(mean=0.0, std=ds_config.initializer_range)
     weights.append(
         nn.Parameter(torch.Tensor(ds_config.hidden_size,
-                                  ds_config.intermediate_size)))
+                                  4 * ds_config.hidden_size)))
     weights[6].data.normal_(mean=0.0, std=ds_config.initializer_range)
     weights.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
     weights[7].data.fill_(1.0)
@@ -177,7 +177,7 @@ def create_models(ds_config):
     for i in range(4):
         biases.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
         biases[i + 1].data.zero_()
-    biases.append(nn.Parameter(torch.Tensor(ds_config.intermediate_size)))
+    biases.append(nn.Parameter(torch.Tensor(4 * ds_config.hidden_size)))
     biases[5].data.zero_()
     biases.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
     biases[6].data.zero_()
@@ -274,7 +274,6 @@ def test_backward(batch_size,
     ds_config.layer_id = None
     ds_config.batch_size = batch_size
     ds_config.hidden_size = hidden_size
-    ds_config.intermediate_size = hidden_size
     ds_config.max_seq_length = seq_len
     ds_config.heads = heads
     ds_config.attn_dropout_ratio = 0.0

tests/unit/test_cuda_forward.py

@@ -109,7 +109,7 @@ def create_models(ds_config):
                              num_hidden_layers=ds_config.num_hidden_layers,
                              num_attention_heads=ds_config.heads,
                              batch_size=ds_config.batch_size,
-                             intermediate_size=ds_config.intermediate_size,
+                             intermediate_size=4 * ds_config.hidden_size,
                              hidden_act="gelu",
                              hidden_dropout_prob=ds_config.hidden_dropout_ratio,
                              attention_probs_dropout_prob=ds_config.attn_dropout_ratio,
@@ -130,12 +130,12 @@ def create_models(ds_config):
     weights.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
     weights[4].data.fill_(1.0)
     weights.append(
-        nn.Parameter(torch.Tensor(ds_config.intermediate_size,
+        nn.Parameter(torch.Tensor(4 * ds_config.hidden_size,
                                   ds_config.hidden_size)))
     weights[5].data.normal_(mean=0.0, std=ds_config.initializer_range)
     weights.append(
         nn.Parameter(torch.Tensor(ds_config.hidden_size,
-                                  ds_config.intermediate_size)))
+                                  4 * ds_config.hidden_size)))
     weights[6].data.normal_(mean=0.0, std=ds_config.initializer_range)
     weights.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
     weights[7].data.fill_(1.0)
@@ -145,7 +145,7 @@ def create_models(ds_config):
     for i in range(4):
         biases.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
         biases[i + 1].data.zero_()
-    biases.append(nn.Parameter(torch.Tensor(ds_config.intermediate_size)))
+    biases.append(nn.Parameter(torch.Tensor(4 * ds_config.hidden_size)))
     biases[5].data.zero_()
     biases.append(nn.Parameter(torch.Tensor(ds_config.hidden_size)))
     biases[6].data.zero_()
@@ -242,7 +242,6 @@ def test_forward(batch_size,
     ds_config.layer_id = None
     ds_config.batch_size = batch_size
     ds_config.hidden_size = hidden_size
-    ds_config.intermediate_size = 4 * hidden_size
     ds_config.max_seq_length = seq_len
     ds_config.heads = heads
     ds_config.attn_dropout_ratio = 0.0
@@ -279,7 +278,6 @@ def test_forward_with_small_bsz(batch_size,
     ds_config.layer_id = None
     ds_config.batch_size = batch_size
     ds_config.hidden_size = hidden_size
-    ds_config.intermediate_size = 4 * hidden_size
     ds_config.max_seq_length = seq_len
     ds_config.heads = heads
     ds_config.attn_dropout_ratio = 0.0
@@ -314,7 +312,6 @@ def test_forward_stochastic(batch_size,
     ds_config.layer_id = None
     ds_config.batch_size = batch_size
     ds_config.hidden_size = hidden_size
-    ds_config.intermediate_size = hidden_size
     ds_config.max_seq_length = seq_len
     ds_config.heads = heads
     ds_config.attn_dropout_ratio = 0.0