leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

update doc

Browse Source
This commit is contained in:
Fung Xie
2025-11-25 21:13:35 -08:00
parent 286781a1fb
commit 03aa211310
2 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
media/docs/pythonDSL/cute_dsl_general/compile_with_tvm_ffi.rst
View File

@@ -296,7 +296,9 @@ composed of the types that are supported by TVM FFI. The example below shows how
Limitations
-----------
The Fake Tensor flow is ONLY compatible with TVM FFI because TVM FFI support more flexible constraints on Tensor arguments.
The Fake Tensor flow supports more flexible constraints on Tensor arguments than the `from_dlpack` flow.
TVM FFI backend is recommended when fake tensor is used as TVM FFI support flexible constraints on Tensor arguments.
For instance, fake tensor can specify per-mode static shape or constraints on shape and strides which is not supported by
existing `from_dlpack` flow. It's expected that JIT function compiled with fake tensor may have different ABI with
tensor converted by `from_dlpack`.

3
python/CuTeDSL/cutlass/cute/runtime.py
View File

@@ -617,7 +617,8 @@ def make_fake_compact_tensor(
:param shape: Shape of the tensor.
:type shape: tuple[int, ...]
:param stride_order: Order in which strides (memory layout) are assigned to the tensor dimensions.
If None, the default layout is col-major. Otherwise, it should be a permutation of the dimension indices.
If None, the default layout is left-to-right order (known as column-major order for flatten layout).
Otherwise, it should be a permutation order of the dimension indices.
:type stride_order: tuple[int, ...], optional
:param memspace: Memory space where the fake tensor resides. Optional.
:type memspace: str, optional