Merge branch 'master' of https://github.com/NVIDIA/apex

apex/amp/amp.py

@@ -145,6 +145,14 @@ def init(enabled=True, enable_caching=True, verbose=False, allow_banned=False):
     # 5.5) Extra-special handling of RNN backend
     wrap.rnn_cast(torch.nn.backends.thnn.backend, 'RNN', verbose)
 
+    # And even more special handling of `backward` for fused gru / lstm
+    # The `backward` method calls Tensor.sum() (blacklist) internally,
+    # and then the resulting grad_input has the wrong type.
+    # TODO: where else is this a problem?
+    for rnn_type in ['GRUFused', 'LSTMFused']:
+        mod = getattr(torch.nn._functions.thnn.rnnFusedPointwise, rnn_type)
+        wrap.disable_casts(mod, 'backward', handle)
+
     # 6) Place error+print message on banned functions
     if not allow_banned:
         for fn, err_msg in functional_overrides.BANNED_FUNCS:

apex/amp/handle.py

@@ -11,9 +11,16 @@ class AmpHandle(object):
         self._verbose = verbose
         self._cache = dict()
         self._default_scaler = LossScaler()
+        self._is_active = True
 
     def is_active(self):
-        return True
+        return self._is_active
+
+    @contextlib.contextmanager
+    def _disable_casts(self):
+        self._is_active = False
+        yield
+        self._is_active = True
 
     def wrap_optimizer(self, optimizer, num_loss=1):
         self._default_scaler = None
@@ -76,6 +83,10 @@ class NoOpHandle(object):
     def is_active(self):
         return False
 
+    @contextlib.contextmanager
+    def _disable_casts(self):
+        yield
+
     def wrap_optimizer(self, optimizer, num_loss=1):
         return OptimWrapper(optimizer, self, num_loss)
 

apex/amp/utils.py

@@ -49,7 +49,7 @@ def maybe_half(x, name='', verbose=False):
     if is_nested(x):
         return type(x)([maybe_half(y) for y in x])
 
-    if type_string(x) == 'HalfTensor':
+    if not x.is_cuda or type_string(x) == 'HalfTensor':
         return x
     else:
         if verbose:
@@ -60,7 +60,7 @@ def maybe_float(x, name='', verbose=False):
     if is_nested(x):
         return type(x)([maybe_float(y) for y in x])
 
-    if type_string(x) == 'FloatTensor':
+    if not x.is_cuda or type_string(x) == 'FloatTensor':
         return x
     else:
         if verbose:

apex/amp/wrap.py

@@ -9,6 +9,9 @@ def make_cast_wrapper(orig_fn, cast_fn, handle,
                       try_caching=False):
     @functools.wraps(orig_fn)
     def wrapper(*args, **kwargs):
+        if not handle.is_active():
+            return orig_fn(*args, **kwargs)
+
         if try_caching and handle.has_cache:
             args = list(args)
             for i in range(len(args)):
@@ -70,7 +73,7 @@ def sequence_promote(mod, fn, verbose=False):
                                          seq, {})
             return orig_fn(cast_seq, *args, **kwargs)
         else:
-            # TODO: other mixed-type cases aren't due to autohalf.
+            # TODO: other mixed-type cases aren't due to amp.
             #       Just pass through?
             return orig_fn(seq, *args, **kwargs)
     utils.set_func(mod, fn, wrapper)
@@ -201,3 +204,14 @@ def rnn_cast(backend, fn, verbose=False):
             return forward(*new_args, **fkwargs)
         return fwd_wrapper
     utils.set_func(backend, fn, rnn_wrapper)
+
+def disable_casts(mod, fn, handle):
+    if not utils.has_func(mod, fn):
+        return
+
+    orig_fn = utils.get_func(mod, fn)
+    @functools.wraps(orig_fn)
+    def wrapper(*args, **kwargs):
+        with handle._disable_casts():
+            return orig_fn(*args, **kwargs)
+    utils.set_func(mod, fn, wrapper)