Triton seems to support both 2D and 3D operands now, the latter case being a
batched matmul. We need more changes in the lowering to support 3D, so I will
leave it out of scope here.
Fixes#26013.
PiperOrigin-RevId: 733293299
When there are multiple dimensions, NumPy's semantics are as if the padding is applied to each dimension in order.
We lacked test coverage for this case because constant values ((0, 2),) and (0, 2) were handled by different code paths.
Fixes https://github.com/jax-ml/jax/issues/26888
Now the small tiling is always `(8, swizzle // bytewidth(dtype))`, no matter whether the input
is transposed or not. This should simply the follow-up refactoring of the code and make it easier
to enable small tiling for LHS too.
PiperOrigin-RevId: 732933005
This tests saving a module with one set of axis names, but loading it with another set of axis names.
This does also test the custom calls:
- `@Sharding`
- `@xla.sdy.GlobalToLocalShape`
- `@xla.sdy.LocalToGlobalShape`
But note that there are a bunch of other custom calls that will be tested in the Shardy and XLA codebases. The way the testing utils is tested here doesn't allow me to set `out_shardings` for example. So JAX can rely on the existence of those tests as stability guarantees just like for StableHLO.
PiperOrigin-RevId: 732893432
This CL only supports lowering a module with the exact same mesh, and loading it with either the exact same mesh or different meshes.
Note that we will be introducing some restrictions under Shardy for JAX export:
- You can only lower/save the module with meshes all of the same shape, but different axis names (this PR is right now only allowing the same axis names, but this will be relaxed in a follow-up)
- When loading the module, just like with GSPMD, you can use a different mesh with a different mesh shape and axis names. However, like with the restriction in the previous point, all shardings must use the same axis shapes, but can use different axis names (again this will be relaxed in a follow-up)
We may remove the restriction of having to use the exact same mesh shapes during export saving time and exact same mesh shaped during export loading time in the future. But for now we will keep this restriction while no one is using Shardy with JAX export.
PiperOrigin-RevId: 732878916
Add a mechanism for using the same Var names for Vars that
are aliased. In this PR, we use this for `pjit`, such that the
following `print(jax.make_jaxpr(lambda a: jax.jit(lambda a: a + 1)(a))(0.))`
prints:
```
{ lambda ; a:f32[]. let
b:f32[] = pjit[
name=<lambda>
jaxpr={ lambda ; a:f32[]. let b:f32[] = add a 1.0 in (b,) }
] a
in (b,) }
```
instead of the previous:
```
{ lambda ; a:f32[]. let
b:f32[] = pjit[
name=<lambda>
jaxpr={ lambda ; c:f32[]. let d:f32[] = add c 1.0 in (d,) }
] a
in (b,) }
```
The same mechanism could be used for other higher-order primitives,
e.g., cond, and others.
Also add some typing declarations and rename APIs to use "shared jaxpr"
in lieu of "top-level jaxpr" for those Jaxprs that are used multiple
times and are printed first. I presume that the term "top-level jaxpr"
was picked because these are printed first at top-level. But this is
confusing, because they are really subjaxprs. In fact, there was already
a function `core.pp_toplevel_jaxpr` for printing the top-level Jaxpr,
and there was also `core.pp_top_level_jaxpr` (which now is named
`core.pp_shared_jaxpr`.
that it more closely matches the CDF for low probably events (less than
2**-nmant).
Because -log(-log(x)) is more sensitive close to 1 than 0, we must use
-log(-logp1(-x)) instead to make better use of the extra range around 0.
PiperOrigin-RevId: 732757388
This change detects a situation where a gmem_memref is read via `async_load` and directly used in a wgmma. In such cases, we insert a cast before the load to add tile, transpose, and swizzle transformations.
PiperOrigin-RevId: 732618760
For example: Consider this einsum: `jnp.einsum('bthD, bthi, bthj->ijD', dy, i, j, out_sharding=P('data', None, None))`
This will decompose into 2 einsums where the intermediate einsum output will be of rank `5`:
* `'bthj,bthD->bthjD'`
* `'bthjD,bthi->ijD'`
The out_sharding specified (`P('data', None, None)`) is not compatible with the intermediate einsum: `'bthj,bthD->bthjD'` since the `length of spec (3) != out_aval.ndim (5)`.
This change makes it so that out_sharding is only applied to the contraction that leads to the final output. **If there are conflicts in intermediate einsums, then the user has to reshard the input or split into multiple einsums (and maybe provide out_sharding) so that conflicts don't exist.**
Note: We won't drop into auto mode for intermediate einsums. The user will have to split the einsum if any conflict is detected.
PiperOrigin-RevId: 732205849
Explicitly make the assumption that `runtime_smem` starts at `0` in the Pallas
module context---which should be enforced by Mosaic GPU.
This is in preparation of changes implementing transform inference.
PiperOrigin-RevId: 732091266
The CUDA 12.8 release significantly improved the MMA docs, letting us
improve upon the previously used "magic number" scheme. Sadly, the docs
are still incorrect, but at least I can begin to make some sense of those
parameters.
PiperOrigin-RevId: 732033585
The goal of this change is to avoid generating code to wrap negative indices back into range in cases where we know it doesn't matter. Change scan to pass allow_negative_indices=False to avoid emitting index wrapping code for each scan argument.
PiperOrigin-RevId: 731812827
* `_partitions` is now canonicalized and only contains `tuples`, `singular strings`, `None` or `UNCONSTRAINED`. No more empty tuples (`P((), 'x')`) and singleton tuples.
* Cache the creating of sharding on ShapedArray since it's expensive to do it a lot of times
* Change the `__hash__` and `__eq__` of `NamedSharding` to depend on `self.spec` instead of `self._parsed_pspec`.
PiperOrigin-RevId: 731745062
