DeepSpeed seems to be incompatible with the current version of Triton. (https://github.com/openai/triton)
In sparse_attention/matmul.py
triton = importlib.import_module('triton')
...
_sparse_matmul.sdd_cache[key] = triton.kernel(
But there is no such thing.
>>> import triton
>>> triton.kernel
AttributeError: module 'triton' has no attribute 'kernel'
To reproduce, fix the test:
diff --git a/tests/unit/test_sparse_attention.py b/tests/unit/test_sparse_attention.py
index 128e93b..84c554f 100755
--- a/tests/unit/test_sparse_attention.py
+++ b/tests/unit/test_sparse_attention.py
@@ -240,7 +240,7 @@ def init_softmax_inputs(Z, H, M, N, scale, rho, block, dtype, dense_x=True, layo
def _skip_on_cuda_compatability():
#pytest.skip("Skip these tests for now until we get our docker image fixed.")
- if torch.cuda.get_device_capability()[0] >= 7:
+ if torch.cuda.get_device_capability()[0] <= 7:
pytest.skip("needs higher compute capability than 7")
cuda_major = int(torch.version.cuda.split('.')[0]) * 10
cuda_minor = int(torch.version.cuda.split('.')[1])
and run
pytest ./tests/unit/test_sparse_attention.py -v
The result:
FAILED tests/unit/test_sparse_attention.py::test_softmax[dtype1-576-16] - AttributeError: module 'triton' has no attribute 'kernel'
FAILED tests/unit/test_sparse_attention.py::test_softmax[dtype1-576-32] - AttributeError: module 'triton' has no attribute 'kernel'
FAILED tests/unit/test_sparse_attention.py::test_matmul[16-dtype0-sdd-False-False] - AttributeError: module 'triton' has no attribute 'kernel'
FAILED tests/unit/test_sparse_attention.py::test_matmul[16-dtype1-sdd-False-True] - AttributeError: module 'triton' has no attribute 'kernel'
FAILED tests/unit/test_sparse_attention.py::test_matmul[16-dtype2-sdd-True-False] - AttributeError: module 'triton' has no attribute 'kernel'
DeepSpeed seems to be incompatible with the current version of Triton. (https://github.com/openai/triton)
In
sparse_attention/matmul.pyBut there is no such thing.
To reproduce, fix the test:
and run
The result: