Merge pull request NVIDIA#17 from rohithkrn/fused_opt_bf16

Enable bfloat16 for optimizers

apex/optimizers/fused_adagrad.py

@@ -91,7 +91,7 @@ def step(self, closure=None):
                 if len(state) == 0:
                     # Exponential moving average of gradient values
                     state['sum'] = torch.zeros_like(p.data)
-                if p.dtype == torch.float16:
+                if p.dtype in {torch.float16, torch.bfloat16}:
                     g_16.append(p.grad.data)
                     p_16.append(p.data)
                     h_16.append(state['sum'])
@@ -100,7 +100,7 @@ def step(self, closure=None):
                     p_32.append(p.data)
                     h_32.append(state['sum'])
                 else:
-                    raise RuntimeError('FusedAdagrad only support fp16 and fp32.')
+                    raise RuntimeError('FusedAdagrad only support fp16, bfloat16 and fp32.')
 
             if(len(g_16) > 0):
                 multi_tensor_applier(self.multi_tensor_adagrad,

apex/optimizers/fused_adam.py

@@ -130,7 +130,7 @@ def step(self, closure=None, grads=None, output_params=None, scale=None, grad_no
                     # Exponential moving average of squared gradient values
                     state['exp_avg_sq'] = torch.zeros_like(p.data)
 
-                if p.dtype == torch.float16:
+                if p.dtype in {torch.float16, torch.bfloat16}:
                     g_16.append(p.grad.data)
                     p_16.append(p.data)
                     m_16.append(state['exp_avg'])
@@ -141,7 +141,7 @@ def step(self, closure=None, grads=None, output_params=None, scale=None, grad_no
                     m_32.append(state['exp_avg'])
                     v_32.append(state['exp_avg_sq'])
                 else:
-                    raise RuntimeError('FusedAdam only support fp16 and fp32.')
+                    raise RuntimeError('FusedAdam only support fp16, bfloat16 and fp32.')
 
             if(len(g_16) > 0):
                 multi_tensor_applier(self.multi_tensor_adam,

apex/optimizers/fused_lamb.py

@@ -130,7 +130,7 @@ def step(self, closure=None):
                     # Exponential moving average of gradient values
                     state['exp_avg_sq'] = torch.zeros_like(p.data)
 
-                if p.dtype == torch.float16:
+                if p.dtype in {torch.float16, torch.bfloat16}:
                     g_16.append(p.grad.data)
                     p_16.append(p.data)
                     m_16.append(state['exp_avg'])
@@ -141,7 +141,7 @@ def step(self, closure=None):
                     m_32.append(state['exp_avg'])
                     v_32.append(state['exp_avg_sq'])
                 else:
-                    raise RuntimeError('FusedLAMB only support fp16 and fp32.')
+                    raise RuntimeError('FusedLAMB only support fp16, bfloat16 and fp32.')
 
             if(len(g_16) > 0):
                 multi_tensor_applier(self.multi_tensor_lamb,

apex/optimizers/fused_novograd.py

@@ -142,7 +142,7 @@ def step(self, closure=None):
                     # Exponential moving average of gradient values
                     state['exp_avg'] = torch.zeros_like(p.data)
 
-                if p.dtype == torch.float16:
+                if p.dtype in {torch.float16, torch.bfloat16}:
                     g_16.append(p.grad.data)
                     p_16.append(p.data)
                     m_16.append(state['exp_avg'])
@@ -151,7 +151,7 @@ def step(self, closure=None):
                     p_32.append(p.data)
                     m_32.append(state['exp_avg'])
                 else:
-                    raise RuntimeError('FusedNovoGrad only support fp16 and fp32.')
+                    raise RuntimeError('FusedNovoGrad only support fp16, bfloat16 and fp32.')
 
             # we store per weight norm as one tensor for one group/precision combination
             # different from optim.Adam, we store norm here(not ^2) so we can unify calculation for norm types

csrc/layer_norm_cuda_kernel.cu

@@ -690,7 +690,7 @@ void cuda_layer_norm(
     double epsilon)
 {
     using namespace at;
-    DISPATCH_DOUBLE_FLOAT_AND_HALF(input->scalar_type(), 0, "layer_norm_cuda_kernel",
+    DISPATCH_DOUBLE_FLOAT_AND_HALF_AND_BFLOAT16(input->scalar_type(), 0, "layer_norm_cuda_kernel",
         using accscalar_t = at::acc_type<scalar_t_0, true>;
         HostApplyLayerNorm(
             output->DATA_PTR<scalar_t_0>(),
@@ -793,7 +793,7 @@ void cuda_layer_norm_gradient(
     at::Tensor* grad_beta)
 {
     using namespace at;
-    DISPATCH_FLOAT_AND_HALF(input->scalar_type(), 0, "cuComputeGradInput",
+    DISPATCH_FLOAT_AND_HALF_AND_BFLOAT16(input->scalar_type(), 0, "cuComputeGradInput",
         using accscalar_t = at::acc_type<scalar_t_0, true>;
         HostLayerNormGradient(
 	    dout->DATA_PTR<scalar_t_0>(),

csrc/multi_tensor_adagrad.cu

@@ -90,7 +90,7 @@ void multi_tensor_adagrad_cuda(
   using namespace at;
 
   // Assume single type across p,g,h now
-  DISPATCH_DOUBLE_FLOAT_AND_HALF(
+  DISPATCH_DOUBLE_FLOAT_AND_HALF_AND_BFLOAT16(
       tensor_lists[0][0].scalar_type(), 0, "adagrad",
       multi_tensor_apply<3>(BLOCK_SIZE, chunk_size, noop_flag, tensor_lists,
                             AdagradFunctor<scalar_t_0>(), epsilon, lr,

tests/L0/run_optimizers/test_adagrad.py

@@ -14,22 +14,28 @@ def setUp(self, max_abs_diff=1e-6, max_rel_diff=1, iters=7):
     def tearDown(self):
         pass
 
-    def gen_param_optim(self, tensors, adagrad_option):
+    def gen_param_optim(self, tensors, adagrad_option, apex_only=False):
         ref_param = []
         tst_param = []
         for tensor in tensors:
-            ref_param.append(torch.nn.Parameter(tensor.clone()))
+            if apex_only:
+                ref_param.append(torch.nn.Parameter(tensor.clone().float()))
+            else:
+                ref_param.append(torch.nn.Parameter(tensor.clone()))
             tst_param.append(torch.nn.Parameter(tensor.clone()))
 
-        ref_optim = torch.optim.Adagrad(ref_param, **adagrad_option)
+        if apex_only:
+            ref_optim = apex.optimizers.FusedAdagrad(ref_param, **adagrad_option)
+        else:
+            ref_optim = torch.optim.Adagrad(ref_param, **adagrad_option)
         tst_optim = apex.optimizers.FusedAdagrad(tst_param, **adagrad_option)
 
         return (ref_param, tst_param, ref_optim, tst_optim)
 
-    def gen_grad(self, ref_param, tst_param):
+    def gen_grad(self, ref_param, tst_param, apex_only=False):
         for p_ref, p_tst in zip(ref_param, tst_param):
-            p_ref.grad = torch.rand_like(p_ref)
-            p_tst.grad = p_ref.grad
+            p_tst.grad = torch.rand_like(p_tst)
+            p_ref.grad = p_tst.grad.detach().float() if apex_only else p_tst.grad
 
     def gen_mixed_grad(self, ref_param, tst_param, scale=1.0):
         half_grads = []
@@ -38,9 +44,11 @@ def gen_mixed_grad(self, ref_param, tst_param, scale=1.0):
             p_ref.grad = half_grads[-1].float() / scale
         return half_grads
 
-    def get_max_diff(self, ref_param, tst_param):
+    def get_max_diff(self, ref_param, tst_param, apex_only=False):
         max_abs_diff = max_rel_diff = 0
         for p_ref, p_tst in zip(ref_param, tst_param):
+            if apex_only:
+                p_tst = p_tst.float()
             max_abs_diff_p = (p_ref - p_tst).abs().max().item()
             max_rel_diff_p = ((p_ref - p_tst) / p_ref).abs().max().item()
 
@@ -51,23 +59,24 @@ def get_max_diff(self, ref_param, tst_param):
 
         return max_abs_diff, max_rel_diff
 
-    def gen_single_type_test(self, param_type=torch.float):
+    def gen_single_type_test(self, param_type=torch.float, apex_only=False):
         nelem = 278011
         adagrad_option = {"lr": 5e-4, "eps": 1e-08, "weight_decay": 1.0e-5}
 
         tensor = torch.rand(nelem, dtype=param_type, device="cuda")
         ref_param, tst_param, ref_optim, tst_optim = self.gen_param_optim(
-            [tensor], adagrad_option
+            [tensor], adagrad_option, apex_only=apex_only
         )
 
         for _ in range(self.iters):
-            self.gen_grad(ref_param, tst_param)
+            self.gen_grad(ref_param, tst_param, apex_only=apex_only)
             ref_optim.step()
             tst_optim.step()
-            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)
+            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param, apex_only=apex_only)
 
             self.assertLessEqual(max_abs_diff, self.max_abs_diff)
-            self.assertLessEqual(max_rel_diff, self.max_rel_diff)
+            if not apex_only:
+                self.assertLessEqual(max_rel_diff, self.max_rel_diff)
 
     def test_float(self):
         self.gen_single_type_test(param_type=torch.float)
@@ -76,6 +85,14 @@ def test_float(self):
     def test_half(self):
         self.gen_single_type_test(param_type=torch.float16)
 
+    # Compares bfloat16 computation against float32 as gold standard.
+    # Uses apex optimizers(controlled by apex_only flag) for both types.
+    # Doesn't use upstream optimizer like other tests as they seem to be
+    # numerically unstable for half types(see skip note for test above).
+    def test_bfloat16(self):
+        self.max_abs_diff = 1e-2
+        self.gen_single_type_test(param_type=torch.bfloat16, apex_only=True)
+
     def test_multi_params(self):
         sizes = [[4096, 1024], [4096], [4096, 2048], [32320, 1024], [1]]
         adagrad_option = {"lr": 5e-4, "eps": 1e-08, "weight_decay": 0}

tests/L0/run_optimizers/test_adam.py

@@ -15,22 +15,28 @@ def setUp(self, max_abs_diff=1e-3, max_rel_diff=1, iters=7):
     def tearDown(self):
         pass
 
-    def gen_param_optim(self, tensors, adam_option):
+    def gen_param_optim(self, tensors, adam_option, apex_only=False):
         ref_param = []
         tst_param = []
         for tensor in tensors:
-            ref_param.append(torch.nn.Parameter(tensor.clone()))
+            if apex_only:
+                ref_param.append(torch.nn.Parameter(tensor.clone().float()))
+            else:
+                ref_param.append(torch.nn.Parameter(tensor.clone()))
             tst_param.append(torch.nn.Parameter(tensor.clone()))
 
-        ref_optim = torch.optim.Adam(ref_param, **adam_option)
+        if apex_only:
+            ref_optim = apex.optimizers.FusedAdam(ref_param, **adam_option)
+        else:
+            ref_optim = torch.optim.Adam(ref_param, **adam_option)
         tst_optim = apex.optimizers.FusedAdam(tst_param, **adam_option)
 
         return (ref_param, tst_param, ref_optim, tst_optim)
 
-    def gen_grad(self, ref_param, tst_param):
+    def gen_grad(self, ref_param, tst_param, apex_only=False):
         for p_ref, p_tst in zip(ref_param, tst_param):
-            p_ref.grad = torch.rand_like(p_ref)
-            p_tst.grad = p_ref.grad
+            p_tst.grad = torch.rand_like(p_tst)
+            p_ref.grad = p_tst.grad.detach().float() if apex_only else p_tst.grad
 
     def gen_mixed_grad(self, ref_param, tst_param, scale=1.0):
         half_grads = []
@@ -39,9 +45,11 @@ def gen_mixed_grad(self, ref_param, tst_param, scale=1.0):
             p_ref.grad = half_grads[-1].float() / scale
         return half_grads
 
-    def get_max_diff(self, ref_param, tst_param):
+    def get_max_diff(self, ref_param, tst_param, apex_only=False):
         max_abs_diff = max_rel_diff = 0
         for p_ref, p_tst in zip(ref_param, tst_param):
+            if apex_only:
+                p_tst = p_tst.float()
             max_abs_diff_p = (p_ref - p_tst).abs().max().item()
             max_rel_diff_p = ((p_ref - p_tst) / p_ref).abs().max().item()
 
@@ -50,30 +58,39 @@ def get_max_diff(self, ref_param, tst_param):
 
         return max_abs_diff, max_rel_diff
 
-    def gen_single_type_test(self, param_type=torch.float):
+    def gen_single_type_test(self, param_type=torch.float, apex_only=False):
         nelem = 278011
         adam_option = {'lr':5e-4, 'betas':(0.9, 0.999), 'eps':1e-08,
             'weight_decay':0, 'amsgrad':False}
 
         tensor = torch.rand(nelem, dtype=param_type, device='cuda')
         ref_param, tst_param, ref_optim, tst_optim = \
-            self.gen_param_optim([tensor], adam_option)
+            self.gen_param_optim([tensor], adam_option, apex_only=apex_only)
 
         for i in range(self.iters):
-            self.gen_grad(ref_param, tst_param)
+            self.gen_grad(ref_param, tst_param, apex_only=apex_only)
             ref_optim.step()
             tst_optim.step()
-            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param)
+            max_abs_diff, max_rel_diff = self.get_max_diff(ref_param, tst_param, apex_only=apex_only)
 
             self.assertLessEqual(max_abs_diff, self.max_abs_diff)
-            self.assertLessEqual(max_rel_diff, self.max_rel_diff)
+            if not apex_only:
+                self.assertLessEqual(max_rel_diff, self.max_rel_diff)
 
     def test_float(self):
         self.gen_single_type_test(param_type=torch.float)
 
     def test_half(self):
         self.gen_single_type_test(param_type=torch.float16)
 
+    # Compares bfloat16 computation against float32 as gold standard.
+    # Uses apex optimizers(controlled by apex_only flag) for both types.
+    # Doesn't use upstream optimizer like other tests as they seem to be
+    # numerically unstable for half types
+    def test_bfloat16(self):
+        self.max_abs_diff = 1e-2
+        self.gen_single_type_test(param_type=torch.bfloat16, apex_only=True)
+
     @unittest.skip('Disable until 8/1/2019 adam/adamw upstream picked')
     def test_multi_params(self):
         sizes = [[4096, 1024], [4096], [4096, 2048], [32320, 1024], [1]]