-
Notifications
You must be signed in to change notification settings - Fork 579
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
RFC: Tensorflow Core as a product #331
base: master
Are you sure you want to change the base?
Conversation
rfcs/20201113-tf-core.md
Outdated
ExtensionType (as known as CompositeTensor, RFC): provides a pair of protocols for defining new types that can be directly supported by TensorFlow APIs. (This API is also known as the "CompositeTensor" API). Example ExtensionType in core includes IndexSlices. | ||
TypeSpec:(go/tf-type-spec) provides a standard way to define and reason about TensorFlow-specific types that are supported by TensorFlow APIs. | ||
|
||
Note: FullType (go/tf-fulltype) definitions to be added once the design is ready. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
need a public link for fulltype
|
||
|
||
## Objective | ||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This needs some more nuance.
When I first read "TF core" I thought it was about separating the C++ library from the python; but below it seems like you want to include some python API as part of core.
So for example when it comes to XLA/MLIR, either JITted or AOT (e.g., the functionality in saved_model_cli), do you plan to keep those in core, or do you expect that all XLA functionality be split off in separate symbolic libraries? If so, how do you expect to support jit_compile=True
extension of tf.function?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
+1 for separating C++ library from the Python one. The Python API might serve for prototyping purpose if we'd like to collect feedbacks from a broader audience though, e.g. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/distribute/v1/all_reduce.py
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'll try and clarify the proposal a bit here. There are couple of directions this RFC is pushing towards
-
Python API modularity - Currently TF core includes Keras, tf.data, distribution strategies etc. without any clear layering and dependency structure. We want to define a small core on which all these libraries can be layered on top of. We don't plan to remove any APIs from TF or add too many new APIs etc.
-
C++ first API - We want to think of the "Core API" as primarily a C++ one with a thin (probably 1-1 mapping) layer of python on top. Users can choose to interact with either API. Arguably, one can think of this as an implementation detail but its a real important implementation detail that we wanted to bring folks attention to.
Concretely, the first step we're taking now is defining the boundaries of what this small core API surface is going to look like and ensuring that the other TF APIs (tf.data, distribute, keras etc.) are dependent on it (and not the other way). This involves removing circular dependencies, BUILD file refactoring, moving some files around, creating right extension points / interfaces in TF core with no API surface changes expected whatsoever.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@rohan100jain +1
If there is design meeting kindly let us know. I would like to join the call.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, design review meeting is scheduled at 10AM PT Dec 15, 2020
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sorry, we need more time to figure out things better, the design review is cancelled.
|
||
### Keys to get there | ||
|
||
Understand users’ needs |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think this will require a SIG for the core.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Or "users" here refer to the high-level API owners, tf-keras/data/distribute etc. then one could argue internal communication might be more effective/efficient at the initial stage.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Please involve SIGs leads at least.
|
||
Performance | ||
|
||
TF Core will be extremely lightweight and performant. We’ll try to do as much computation as possible in C++ with a thin layer on top. For example, Python APIs will be one thin layer on top. This also involves building a clean C API layer where state of the art runtimes like TFRT and Graph Compilers like MLIR can plug in and improve performance. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
What Is the python layer here? A 1-to-1 wrapper? Or an higher level layer with a reduced surface?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We would like see layered Python APIs, with Keras, tf-data, distributed, etc built on core APIs and extension/composite mechanics. This core APIs are a reduced and complete surface. One high level API could be composed by a few core APIs.
|
||
* Is this API serving any use cases? | ||
|
||
When we add/remove APIs, the first thing we should ask is ‘what is the use case for this APIs’. This API should serve users’ purpose, not developers’ purpose. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think that we cannot do too much claims for users without a regular SIG.
Also what about custom c++ ops in SIGS?
Can we really analyze composabilty in python space before evaluate a core API change/extension?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
There are some essential and basic C++ Ops. Also a lot of custom c++ ops can be replaced with python code. currently one of main reason for these custom C++ ops is performance. We do have some plans to bring the python code performance closer to custom c++ op performance, so that custom c++ is not required, instead, same functions are implemented in the high level languages, this provides easier programability and flexibility to modify.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes this was my point and this seems to me to give a quite high responsibility to the core API set perimeter definition + compiler stack performance to not create a bottleneck.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
quick question, when I add a new hardware device(TensorFlow modular plugin), for ABI stable, I need register Ops/Kernels with C API, How could register them into TF-Core? also , I see (https://github.com/tensorflow/community/pull/77), if user need a new Python API(eg. for plugin device new Ops/Kernels), they need to interact TF-Core(C++ libraries) with C API, What do you think?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@sun51 In the meantime can you expose a little but more if and how this is going to interact with https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/mlir/tfr.
Cause I expect that reducing the core API surface is going to create more pressure over ecosystem repos that have the infrastructure for custom ops.
Are the error messages for the APIs easy to understand? Are the error messages clear enough to give the feedback on what is going wrong, and how to fix it? | ||
|
||
* Is this API performant? | ||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
On what reference runtime? TFRT?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
A general suggestion is to avoid over-emphasizing "performant" without going into detailed analysis to the trade-offs, e.g. why this API design enables high performance implementation and its alternatives do not.
|
||
* Is this API similar or redundant with existing ones? | ||
|
||
Corresponding to the principle ‘Primitive/Essential/Orthogonality’: If one API can be written by using existing a few APIs together with mechanics core provides, then the API is not essential. Instead of including that particular APIs, the primitive APIs that API is composed of should be included in the core. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think the "few" term Is different in the c++ domain instead of python domain
|
||
### Functional components of TF core | ||
* Control flow: tf.cond, tf.while | ||
* tf.function: Trace compile tensorflow programs to accelerate them |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can the quality of the compiler performance in a function have an impact on the evaluation about extending the core? Or we will be forwarded on compiler improvements?
Generally I think that this RFC will have some positive impacts if the SIGs leads will be involved in the process and not only in the RFC revision. What I don't understand is if the scope of this RFC is to have a TF-core only repository in the near future maintained by a specific TF-core team.
Are we going in a TF world where many SIGs will Need to maintain its own c++/c API to offer High level python API for all the cases not covered by python composabilty on TF-core or when the compiler stack (tf.function) is not efficient enought? Also, are we sure that we can still minimize duplications across the SIGs ecosystem if the "centralized" core is going to have a minimal surface? |
Edit: meeting has been postponed, see #331 (comment) |
Corresponding to the principle ‘Primitive/Essential/Orthogonality’: If one API can be written by using existing a few APIs together with mechanics core provides, then the API is not essential. Instead of including that particular APIs, the primitive APIs that API is composed of should be included in the core. | ||
|
||
* Is this API easy to use? | ||
Usability - representation |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Some formatting (especially line breaks) are broken in this doc when rendered. For example, you need a blank line here to break the line.
* Testing: APIs and functions for users to test Tensorflow programs. | ||
* Basic Utilities: Essential and primitive utilities functions, APIs. | ||
|
||
### Math/NN APIs |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
What about RNG ops (both stateless RNGs and tf.random.Generator)?
Sorry we need more time to figure out things better, the design review is cancelled for now, we will let you know if there is one later. |
Thanks everyone for the discussion and comments, we will need more time to figure out the design, so we will hold this design RFC and have cancelled the design review for now. Will update later. |
Do you mean #331 (comment) ? |
Would this TF-core have a C API as well as a C++ one, or is that planned to live elsewhere? |
Hope most things do in c/c++ |
@rohan100jain Are we still pursuing this RFC? |
It would be really useful if we could revamp and approve this RFC. |
Do we have an opportunity to re-evalute/refresh this and all the useful feedback we had collected? |
@joker-eph Do you think there is an opportunity to incorporate the collected feedback into the current refactoring projects? See @bhack's comment above. |
Tensorflow Core as a product RFC
Objective
This RFC proposes Tensorflow core as a standalone product and describes the high level components and the principles behind it.