I initially wanted to upgrade to 1.15, but it seems to have a bug in how
ternary expressions are type checked. For example,
def f(x: int) -> str: ...
def g(x: int) -> str: ...
callback = f if ... else g # has type object!
In https://github.com/jax-ml/jax/pull/23574, we added a new `algorithm` parameter to `lax.dot_general` with the goal of giving users explicit control over the specific algorithm used to control dot product accumulation. When using this feature in real use cases, we have found that the API is both too conservative (it required the user to pass the appropriate input types) and too restrictive for common use cases. In this change, I simplify the API to bring it more in line with user expectations, and generalize it to support a broader range of use cases.
The core change is to update the dot_general lowering rule to add explicit type casts to the inputs, making sure that they always have the appropriate storage types going into the `DotGeneral` StableHLO op. Before this change, some backends would implicitly cast for some algorithms (e.g. f32 -> bf16), but error for others. It seems more user friendly to include automatic casts in all cases where a specific algorithm is requested.
Another change in behavior is to (if needed) cast the result of the `DotGeneral` op (which is defined by the algorithm's `accumulation_type`) to match the input types. This means that, regardless of the algorithm choice, the output type will match the value that a user would expect from past use of `lax.dot_general`. The `preferred_element_type` parameter can now be used to control the output type, even when an algorithm is selected.
To summarize, the updated version of `dot_general` accepts _any_ input dtypes, and the output will always match the inputs (under the existing promotion rules if the LHS and RHS don't match) unless `preferred_element_type` is used to select a specific output type. The specified "algorithm" is now more of an implementation detail, rather than the defining feature of the API, and JAX will do whatever it can to satisfy the user's request. (If an algorithm is not supported on the current device, we will still get a compile time error.)
With the above changes in mind, it's no longer really necessary to have a `transpose_algorithm` parameter, because we can now use the same algorithm for the backwards pass. For users who need to customize the algorithm on the backwards pass, that is still possible using `custom_vjp`.
Given the above changes, @sbodenstein made the excellent point that we don't really need the `algorithm` parameter anymore: just accept `DotAlgorithm` inputs to `precision`. I think this is a really nice suggestion, so I have updated the interface to implement this.
One minor negative of this approach is that `preferred_element_type` isn't a great name for what that parameter does when it is used in conjunction with an algorithm. In the long run, I'd like to rename this parameter, but keeping it as is for now seems like the best short term approach.
PiperOrigin-RevId: 683302687
1. Add (limited) precision specifier handling to LSTM
Enables differentiating between TF32 and FP32 math. TF32 math had insufficient
precision to reliably pass LSTM correctness tests on A100 and H100.
2. Run the test using FP32
TF32 precision is not sufficient for the test to pass reliably on Ampere+ GPUs
such as A100 and H100.
