feat(magi2): add fused BF16 routed MoE path (#7206)
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
This commit is contained in:
1 parent
79a46a4f54
commit
29b818a785
8 files changed
+1818
-299
No files matched your search
@@ -0,0 +1,43 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
||||
|
||||
"""No vLLM/model/MATE dependency: python -m pytest -q this_file.py."""
|
||||
|
||||
import importlib
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
if hasattr(torch.version, "musa") and torch.version.musa is not None:
|
||||
importlib.import_module("torchada")
|
||||
|
||||
from vllm.triton_utils import tl, triton
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.musa]
|
||||
|
||||
|
||||
@triton.jit
|
||||
def _atomic_sum(values, output, count: tl.constexpr, block: tl.constexpr):
|
||||
offsets = tl.program_id(0) * block + tl.arange(0, block)
|
||||
increments = tl.load(values + offsets, mask=offsets < count, other=0)
|
||||
tl.atomic_add(output + offsets % 32, increments, mask=offsets < count)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("dtype", [torch.float32, torch.bfloat16])
|
||||
def test_atomic_add(dtype):
|
||||
cuda = getattr(torch, "cuda")
|
||||
available = cuda.is_available() or (hasattr(torch, "musa") and torch.musa.is_available())
|
||||
if not available:
|
||||
pytest.skip("No CUDA/MUSA device")
|
||||
if dtype == torch.bfloat16 and getattr(torch.version, "musa", None) is not None:
|
||||
if tuple(int(part) for part in triton.__version__.split(".")[:2]) < (3, 6):
|
||||
pytest.skip("Triton-MUSA < 3.6 does not support BF16 atomic_add")
|
||||
# Each destination receives eight exactly representable increments from
|
||||
# two independent blocks. Reset between runs to exercise real atomics.
|
||||
values = torch.ones(256, device="cuda", dtype=dtype)
|
||||
output = torch.zeros(32, device="cuda", dtype=dtype)
|
||||
for _ in range(5):
|
||||
output.zero_()
|
||||
_atomic_sum[(2,)](values, output, 256, 128)
|
||||
cuda.synchronize()
|
||||
torch.testing.assert_close(output, torch.full_like(output, 8), rtol=0, atol=0)
|
||||
Reference in new issue
Block a user