experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA/Docs] Update known issues list

Browse Source 
PiperOrigin-RevId: 340502704
Change-Id: I0ddc1278852cc82622f4f4ff84faaf8613103d6a
This commit is contained in:
George Karpenkov 2020-11-03 12:37:37 -08:00 committed by TensorFlower Gardener
parent 6c315885df
commit 5feee72aff
1 changed files with 20 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
tensorflow/compiler/xla/g3doc/known_issues.md
View File

@ -5,7 +5,7 @@ the TensorFlow interop has a number of known sharp corners.
## TensorArray TF/XLA interconversion ## TensorArray TF/XLA interconversion
The problem manifests itself as an error message *Error message*:
`Support for TensorList crossing the XLA/TF boundary is not implemented`. `Support for TensorList crossing the XLA/TF boundary is not implemented`.
XLA supports `tf.TensorArray`. However, the _interconversion_ between TF and XLA supports `tf.TensorArray`. However, the _interconversion_ between TF and
@ -13,15 +13,31 @@ XLA representations is not implemented yet.
This error often arises when the `TensorArray` is used inside the compiled This error often arises when the `TensorArray` is used inside the compiled
block, but the derivative is taken outside. block, but the derivative is taken outside.
Workaround: compile the outermost scope which is taking the derivative. *Workaround*: compile the outermost scope which is taking the derivative.
## Dynamic `tf.TensorArray` is not supported ## TensorFlow while loops need to be bounded (or have backprop disabled)
*Error message*: `XLA compilation requires a fixed tensor list size. Set the max
number of elements. This could also happen if you're using a TensorArray in a
while loop that does not have its maximum_iteration set, you can fix this by
setting maximum_iteration to a suitable value`.
TF while [loops](https://www.tensorflow.org/api_docs/python/tf/while_loop)
created using `tf.while_loop` support backpropagation by accumulating all
intermediate results in a `TensorArray`, but XLA only supports bounded
`TensorArray`s.
*Workaround*: all compiled while loops need to either have `maximum_iterations`
parameter set to a constant value known at compile time, or backpropagation
disabled using `back_prop=False`.
## Dynamic `tf.TensorArray`
Writes into `tf.TensorArray(..., dynamic_size=True)` are not compilable with Writes into `tf.TensorArray(..., dynamic_size=True)` are not compilable with
XLA, as such writes require an unknown number of reallocations when the array XLA, as such writes require an unknown number of reallocations when the array
exceeds the original bound. exceeds the original bound.
Workaround: provide a statically known bound to your arrays. *Workaround*: provide a statically known bound to your arrays.
## Random number generation ## Random number generation
@ -29,13 +45,3 @@ XLA currently ignores TF seeds to random operations. This affects stateful TF
random operations, such as `tf.random.normal`, or `tf.nn.dropout`. XLA will random operations, such as `tf.random.normal`, or `tf.nn.dropout`. XLA will
behave as if the compilation was seeded with a new unique seed at each run. This behave as if the compilation was seeded with a new unique seed at each run. This
limitation does not apply to stateless random ops. limitation does not apply to stateless random ops.
## TensorFlow while loops need to be bounded (or have backprop disabled)
TF while [loops](https://www.tensorflow.org/api_docs/python/tf/while_loop)
created using `tf.while_loop` support backpropagation by accumulating all
intermediate results in a `TensorArray`.
Since XLA only supports bounded `TensorArray`s, all compiled while loops need to
either have `maximum_iterations` parameter set to a constant value known at
compile time, or backpropagation disabled using `back_prop=False`.