Merge triton-mlir branch - Complete rewrite of the backend from scratch (#1004)

This PR merges the `triton-mlir` branch, in which we have been quietly
rewriting the Triton backend from scratch to increase maintainability,
stability and ultimately performance. Changes to the runtime are
minimal, and this new version aims to remain backward-compatible with
the previous commit. The legacy backend is now officially deprecated,
but can still be accessed via the `legacy-backend` tag.

Co-authored-by: Keren Zhou <kerenzhou@openai.com>
Co-authored-by: Yan Chunwei <yanchunwei@outlook.com>
Co-authored-by: goostavz <109190422+goostavz@users.noreply.github.com>
Co-authored-by: Shintaro Iwasaki <siwasaki@fb.com>
Co-authored-by: Yan Da <dyanab@connect.ust.hk>
Co-authored-by: Jun Yang <yangjunpro@gmail.com>
Co-authored-by: Ian Bearman <ianb@microsoft.com>
Co-authored-by: Jason Ansel <jansel@jansel.net>
Co-authored-by: Qingyi Liu <qingyil@nvidia.com>
Co-authored-by: ben-zhang-609 <110140741+ben-zhang-609@users.noreply.github.com>
Co-authored-by: Chenggang Zhao <lyricz@yeah.net>
Co-authored-by: ben-zhang-609 <benzh609@gmail.com>
Co-authored-by: dongdongl <dongdongl@nvidia.com>
This commit is contained in:
Philippe Tillet
2022-12-21 01:30:50 -08:00
committed by GitHub
parent 8650b4d1cb
commit 20100a7254
285 changed files with 26312 additions and 50143 deletions

View File

@@ -80,7 +80,7 @@ def softmax_kernel(
row = tl.load(input_ptrs, mask=col_offsets < n_cols, other=-float('inf'))
# Subtract maximum for numerical stability
row_minus_max = row - tl.max(row, axis=0)
# Note that exponentials in Triton are fast but approximate (i.e., think __expf in CUDA)
# Note that exponentiation in Triton is fast but approximate (i.e., think __expf in CUDA)
numerator = tl.exp(row_minus_max)
denominator = tl.sum(numerator, axis=0)
softmax_output = numerator / denominator
@@ -188,4 +188,4 @@ benchmark.run(show_plots=True, print_data=True)
#
# - Triton is 4x faster than the Torch JIT. This confirms our suspicions that the Torch JIT does not do any fusion here.
# - Triton is noticeably faster than :code:`torch.softmax` -- in addition to being **easier to read, understand and maintain**.
# Note however that the PyTorch `softmax` operation is more general and will works on tensors of any shape.
# Note however that the PyTorch `softmax` operation is more general and will work on tensors of any shape.