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

@@ -2,13 +2,13 @@ import pytest
import torch
import triton
import triton._C.libtriton.triton as _triton
@pytest.mark.parametrize("MODE", ["sdd", "dds", "dsd"])
@pytest.mark.parametrize("TRANS_A", [False, True])
@pytest.mark.parametrize("TRANS_B", [False, True])
@pytest.mark.parametrize("BLOCK", [16, 32, 64])
# TODO: float32 fails
@pytest.mark.parametrize("DTYPE", [torch.float16])
def test_matmul(MODE, TRANS_A, TRANS_B, BLOCK, DTYPE, Z=3, H=2, M=512, N=384, K=256):
seed = 0
@@ -32,9 +32,9 @@ def test_matmul(MODE, TRANS_A, TRANS_B, BLOCK, DTYPE, Z=3, H=2, M=512, N=384, K=
layout[1, 2, :] = 0
layout[1, :, 1] = 0
# create data
a_ref, a_tri = triton.testing.make_pair(a_shape, alpha=.1)
b_ref, b_tri = triton.testing.make_pair(b_shape, alpha=.1)
dc_ref, dc_tri = triton.testing.make_pair(c_shape)
a_ref, a_tri = triton.testing.make_pair(a_shape, alpha=.1, dtype=DTYPE)
b_ref, b_tri = triton.testing.make_pair(b_shape, alpha=.1, dtype=DTYPE)
dc_ref, dc_tri = triton.testing.make_pair(c_shape, dtype=DTYPE)
# compute [torch]
dc_ref = do_mask(dc_ref) if is_sdd else dc_ref
a_ref = do_mask(a_ref) if is_dsd else a_ref
@@ -126,8 +126,8 @@ def test_attention_fwd_bwd(
batch_size=2,
n_heads=2,
):
cc = _triton.runtime.cc(_triton.runtime.backend.CUDA, torch.cuda.current_device())
if cc < 70:
capability = torch.cuda.get_device_capability()
if capability[0] < 7:
pytest.skip("Only test tl.dot() on devices with sm >= 70")
# inputs

View File

@@ -2,20 +2,19 @@ import pytest
import torch
import triton
import triton._C.libtriton.triton as _triton
@pytest.mark.parametrize("M, N, dtype, mode",
[
(M, N, dtype, mode) for M in [1024, 821]
for N in [512, 857, 1871, 2089, 8573, 31000]
for dtype in ['bfloat16', 'float16', 'float32']
for dtype in ['float16', 'float32']
for mode in ['forward', 'backward']
]
)
def test_op(M, N, dtype, mode):
cc = _triton.runtime.cc(_triton.runtime.backend.CUDA, torch.cuda.current_device())
if cc < 80 and dtype == "bfloat16":
capability = torch.cuda.get_device_capability()
if capability[0] < 8 and dtype == "bfloat16":
pytest.skip("Only test bfloat16 on devices with sm >= 80")
dtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16, 'float32': torch.float32}[dtype]
# create inputs

View File

@@ -4,7 +4,6 @@ import pytest
import torch
import triton
import triton._C.libtriton.triton as _triton
@pytest.mark.parametrize(
@@ -67,10 +66,10 @@ import triton._C.libtriton.triton as _triton
),
)
def test_op(BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, NWARP, NSTAGE, M, N, K, AT, BT, DTYPE):
cc = _triton.runtime.cc(_triton.runtime.backend.CUDA, torch.cuda.current_device())
if cc < 70:
capability = torch.cuda.get_device_capability()
if capability[0] < 7:
pytest.skip("Only test tl.dot() on devices with sm >= 70")
if cc < 80 and DTYPE == "bfloat16":
if capability[0] < 8 and DTYPE == "bfloat16":
pytest.skip("Only test bfloat16 on devices with sm >= 80")
if DTYPE == "bfloat16" and SPLIT_K != 1:
pytest.skip("bfloat16 matmuls don't allow split_k for now")