Skip to content

Make TMA tests compatible with older CUDA toolchains #5221

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
peterbell10 merged 2 commits into from
Nov 22, 2024
Merged

Make TMA tests compatible with older CUDA toolchains #5221

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
958d4f9
Gate TMA tests on cuda toolchain version
embg Nov 21, 2024
971eadb
nits
embg Nov 21, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 7 additions & 3 deletions python/test/unit/hopper/test_experimental_tma.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
import triton
import triton.language as tl
from triton.tools.experimental_descriptor import (create_1d_tma_descriptor, create_2d_tma_descriptor)
from triton._internal_testing import dtypes_with_bfloat16, numpy_random, to_triton, requires_tma
from triton._internal_testing import dtypes_with_bfloat16, numpy_random, to_triton, requires_tma, supports_tma, tma_skip_msg

from typing import Optional

Expand All @@ -29,9 +29,11 @@ def unwrap_tensor(t: torch.Tensor | triton.runtime.jit.TensorWrapper):
tma_dtypes = sorted(set(dtypes_with_bfloat16) - {"int64", "uint64", "float64"})


@requires_tma
@pytest.mark.parametrize("byval_tma", [True, False])
def test_experimetal_descriptor_load(byval_tma):
if not supports_tma(byval_tma):
pytest.skip(tma_skip_msg(byval_tma))
return
device = "cuda"
SIZE = 128

Expand Down Expand Up @@ -82,11 +84,13 @@ def matmul_kernel_tma(a_desc_ptr, b_desc_ptr, c_desc_ptr, #
tl._experimental_descriptor_store(c_desc_ptr, accumulator, [offs_am, offs_bn])


@requires_tma
@pytest.mark.parametrize("num_stages", [1, 4])
@pytest.mark.parametrize("BLOCK_M, BLOCK_N, BLOCK_K", [(32, 32, 32), (128, 64, 64), (128, 128, 64), (128, 256, 64)])
@pytest.mark.parametrize("byval_tma", [True, False])
def test_experimental_tma_matmul(num_stages, BLOCK_M, BLOCK_N, BLOCK_K, byval_tma):
if not supports_tma(byval_tma):
pytest.skip(tma_skip_msg(byval_tma))
return
device = "cuda"
M, N, K = 8192, 8192, 1024
torch.manual_seed(42)
Expand Down
16 changes: 13 additions & 3 deletions python/triton/_internal_testing.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
import torch
import triton
import triton.language as tl
from triton.backends.nvidia.compiler import _path_to_binary
import pytest

from numpy.random import RandomState
Expand Down Expand Up @@ -140,8 +141,17 @@ def to_numpy(x):
raise ValueError(f"Not a triton-compatible tensor: {x}")


def supports_tma():
return is_cuda() and torch.cuda.get_device_capability()[0] >= 9
def supports_tma(byval_only=False):
_, cuda_version = _path_to_binary("ptxas")
min_cuda_version = 12.0 if byval_only else 12.3
return is_cuda() and torch.cuda.get_device_capability()[0] >= 9 and float(cuda_version) >= min_cuda_version


def tma_skip_msg(byval_only=False):
if byval_only:
return "Requires __grid_constant__ TMA support (NVIDIA Hopper or higher, CUDA 12.0 or higher)"
else:
return "Requires advanced TMA support (NVIDIA Hopper or higher, CUDA 12.3 or higher)"


requires_tma = pytest.mark.skipif(not supports_tma(), reason="Requires TMA support (NVIDIA Hopper or higher)")
requires_tma = pytest.mark.skipif(not supports_tma(), reason=tma_skip_msg())
Loading