This is another attempt to land a rolled-back change https://github.com/google/jax/pull/14734 (cl/514070997).
See b/272154366 for more details.
The use case for call_tf with shape polymorphism is when we have a JAX program
that calls into TF function, and we want to serialize the JAX program with
some shapes unknown. Previously this use case did not work, except in the special
case when the output shape of the called TF function returns statically known
shapes.
The idea is that we allow the user of call_tf to specify the output shape.
This can be done even in presence of shape polymorphism, by writing the
output shape as an expression in terms of the input shapes. This is what
other JAX primitives do, e.g., concat, so we are simply enabling call_tf
to get the same behavior.
This change should be enough for old-style jax2tf, but will require more
work for native serialization.
We also removed some old code that was trying to workaround some limitations
in shape inference in TF. I think that those workarounds are ugly, and I am
prepared to give error messages rather than keep that code. So far no
tests fail.
PiperOrigin-RevId: 515137407
- jax.Array.broadcast: use lax.broadcast instead
- jax.Array.broadcast_in_dim: use lax.broadcast_in_dim instead
- jax.Array.split: use jnp.split instead
These are removed because they are not part of the np.ndarray API.
This changes the internals of JAX without affecting any public API.
Before, `jit` was a final style primitive. This means that the creation
of jaxpr was delayed as much as possible and transformations were stacked
on top of each other. With the `jit`-`pjit` implementation merge, `jit`
becomes an initial style primitive which means that we trace to jaxpr
as early as possible. For more information see [this section in autodidax](https://jax.readthedocs.io/en/latest/autodidax.html#on-the-fly-final-style-and-staged-initial-style-processing).
Moving to initial style should simplify JAX's internals and make
development of features like dynamic shapes, etc easier.
PiperOrigin-RevId: 508143501
Previously, division was only supported in certain situation, and this
led to errors, e.g., when using strides. Now we generalize the polynomials
to also include "floordiv(E, E)" and "mod(E, E)" as atoms, in addition
to dimension variables. A symbolic dimension is now a sum of products
of atoms. (We also changed the documentation to use symbolic dimension
instead of dimension polynomials).
Previously binary operations involving symbolic dimensions would
work only when the other operand is convertible to a symbolic dimension,
e.g., an integer. This resulted in errors when trying "x.shape[0] * 3.5"
and the recourse was to ask the user to add an explicit
"jnp.array(x.shape[0])".
Now we allow binary operations with any operand and the
"jnp.array" is added automatically if the other operand is not
an integer or a symbolic dimension. This means that instead
of an error they may be an error downstream if one tries to use
the result as a dimension. There is one known case where
JAX works with static shapes and with the previous behavior,
but will fail now. When you operate on `np.ndarray` and
symbolic dimension, previously this was kept as a `np.ndarray`
but not it is turned into a JAX array. The following
program will now fail if `x.shape[0]` is a symbolic dimension.:
`jnp.ones(np.arange(5) * x.shape[0])`
Instead you should write
`jnp.ones([i * x.shape[0] for i in range(5)])`
The CallTfEffect was added recently as an internal workaround for
DCE removing instances of call_tf. Here we add a parameter to
`call_tf` to be able to declare if the called computation is
effectful and should not be removed by DCE.
Currently when JAX config values are configured via ABSL, we use the ABSL flags as a source of truth: if we read or write the JAX config option, we read or write the corresponding ABSL flag. This works but has the unfortunate downside that ABSL flags are relatively slow to read, which slows down JAX every time we read a configuration option.
However, there's fundamentally no reason we are mirroring the JAX configuration options back to ABSL in the first place. We can use ABSL flag parsing as a way only to populate the JAX configuration values. The downside is that if someone changes the ABSL flag values after parsing, that change will not be reflected in JAX's config values. JAX config changes after ABSL flags have been parsed must be made via the `jax.config.update()` API.
This gives a decent improvement on the device_put benchmark:
```
name old cpu/op new cpu/op delta
device_put 79.5µs ± 6% 69.4µs ± 7% -12.73% (p=0.000 n=10+9)
name old time/op new time/op delta
device_put 79.5µs ± 6% 69.4µs ± 7% -12.73% (p=0.000 n=10+9)
```
PiperOrigin-RevId: 492519085
This requires some changes for abstract evaluation, when
JAX does not use a specific platform.
Also attempt to fix the case when the TF lowering fails because the TF computation
uses a tf.Variable on another device as that used for lowering.
PiperOrigin-RevId: 492112847
