Lixun Zhang 821e75a2b0 Improve FA fwd kernel with causal=True (#356) ...

* Attempt to absorb upstream's changes to improve causal=True

* Add autotuner

* Optimize for AMD MI250

- add pre_load_v as a tuning parameter
- do not define N_CTX as constexpr
- perform the second dot before sum
- remove qk_scale out of the inner loop
- add more configs in the autotuner

Note that bwd kernel is disabled for now. This is because we enabled
autotuning and grid becomes a function. So ctx.grid[0] no longer works.

* Enable bwd kernel

2023-10-12 12:34:27 -05:00

..

examples

[FORMAT] autopep8 now uses max-line-length=88 (#1410)

2023-03-25 15:46:50 -07:00

src

Add waves_per_eu as kernel parameter (#319)

2023-10-06 12:08:34 -05:00

test

ROCM IFU: Fix minimize_alloc

2023-10-03 05:34:44 +00:00

triton

enforce cc=None on PyTorch ROCm (#296)

2023-10-12 10:17:26 +01:00

tutorials

Improve FA fwd kernel with causal=True (#356)

2023-10-12 12:34:27 -05:00

MANIFEST.in

[TOOLS] improved ahead-of-time compiler (#1805)

2023-06-21 01:02:58 -07:00

pyproject.toml

[BUILD] use ninja (#2318)

2023-09-18 15:30:06 -05:00

setup.py

PyTorch triton branch synchronisation (#354)

2023-10-11 15:30:39 +01:00