Skip to content

Fix handling of JAX tensor types without TVM FFI and leading dim deduction for all-1-size tensors #3583

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

Open
dextero wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix handling of JAX tensor types without TVM FFI and leading dim deduction for all-1-size tensors #3583

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a78599c
[CuTeDSL] Fix leading dimension deduction for tensors with dimensions…
Sep 4, 2026
3a59277
[CuTeDSL] Register lazy JIT argument adapter for JAX arrays
Sep 4, 2026
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
22 changes: 20 additions & 2 deletions python/CuTeDSL/cutlass/cute/runtime.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,6 +235,12 @@ def mark_layout_dynamic(self, leading_dim: Optional[int] = None) -> "_Tensor":
:rtype: _Tensor
"""
self.load_dltensor()
if leading_dim is None and self.shape and all(s <= 1 for s in self.shape):
leading_dim = len(self.shape) - 1
for dim in reversed(range(len(self.shape))):
if self.stride[dim] == 1:
leading_dim = dim
break
self._dltensor_wrapper.mark_layout_dynamic(leading_dim)
return self

Expand Down Expand Up @@ -931,11 +937,23 @@ def __get_mlir_types__(self) -> list[ir.Type]:


# -------------------------------------------------------------------------
# Register TensorAdapter for numpy/torch tensors lazily by type name, so
# importing this module never pays for importing numpy or torch itself.
# Register TensorAdapter for numpy/torch/jax tensors lazily by type name, so
# importing this module never pays for importing numpy, torch, or jax itself.
# -------------------------------------------------------------------------

JitArgAdapterRegistry.register_jit_arg_adapter("numpy.ndarray", lazy=True)(
TensorAdapter
)
JitArgAdapterRegistry.register_jit_arg_adapter("torch.Tensor", lazy=True)(TensorAdapter)

# Register JAX array implementation classes
for qualname in (
"jaxlib._jax.ArrayImpl",
"jax.jaxlib._jax.ArrayImpl",
"jaxlib.xla_extension.ArrayImpl",
"jax._src.array.ArrayImpl",
):
JitArgAdapterRegistry.register_jit_arg_adapter(qualname, lazy=True)(
TensorAdapter
)