Skip to content

Update dlpack to v1.2 #2193

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
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Update dlpack to v1.2 #2193

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
9ca567b
use vendored dlpack v1.2
ndgrigorian Nov 19, 2025
5df55b6
update Cython definitions of DLPack data type codes
ndgrigorian Nov 19, 2025
68c2167
Add AWS Trainium device
ndgrigorian Nov 19, 2025
043481b
change dlpack logic so shape and strides are null iff ndim == 0
ndgrigorian Nov 21, 2025
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
1 change: 1 addition & 0 deletions dpctl/tensor/_dlpack.pxd
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,7 @@ cdef extern from "dlpack/dlpack.h" nogil:
int device_WebGPU "kDLWebGPU"
int device_Hexagon "kDLHexagon"
int device_MAIA "kDLMAIA"
int device_Trn "kDLTrn"

cpdef object to_dlpack_capsule(usm_ndarray array) except +
cpdef object to_dlpack_versioned_capsule(
Expand Down
142 changes: 68 additions & 74 deletions dpctl/tensor/_dlpack.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@ from .._backend cimport (
DPCTLSyclDeviceRef,
DPCTLSyclUSMRef,
)
from ._usmarray cimport USM_ARRAY_C_CONTIGUOUS, USM_ARRAY_WRITABLE, usm_ndarray
from ._usmarray cimport USM_ARRAY_WRITABLE, usm_ndarray

import ctypes

Expand Down Expand Up @@ -76,6 +76,7 @@ cdef extern from "dlpack/dlpack.h" nogil:
kDLWebGPU
kDLHexagon
kDLMAIA
kDLTrn
Copy link
Collaborator

@antonwolfy antonwolfy Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It seems we need to update numpy_to_dlpack_versioned_capsule at line following dlpack changes added in v1.2
NumPy did that here.

Copy link
Collaborator Author

@ndgrigorian ndgrigorian Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good point, I didn't think we would need to make a change for that part, but I'll adjust it

Copy link
Collaborator Author

@ndgrigorian ndgrigorian Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

all logic should now be updated

shape and strides are both set to NULL for ndim == 0 only, and conversely, when converting from dlpack, shape and strides are simplified a bit if NULL is provided


ctypedef struct DLDevice:
DLDeviceType device_type
Expand All @@ -88,6 +89,17 @@ cdef extern from "dlpack/dlpack.h" nogil:
kDLBfloat
kDLComplex
kDLBool
kDLFloat8_e3m4
kDLFloat8_e4m3
kDLFloat8_e4m3b11fnuz
kDLFloat8_e4m3fn
kDLFloat8_e4m3fnuz
kDLFloat8_e5m2
kDLFloat8_e5m2fnuz
kDLFloat8_e8m0fnu
kDLFloat6_e2m3fn
kDLFloat6_e3m2fn
kDLFloat4_e2m1fn

ctypedef struct DLDataType:
uint8_t code
Expand Down Expand Up @@ -254,7 +266,6 @@ cpdef to_dlpack_capsule(usm_ndarray usm_ary):
cdef int64_t *shape_strides_ptr = NULL
cdef int i = 0
cdef int device_id = -1
cdef int flags = 0
cdef Py_ssize_t element_offset = 0
cdef Py_ssize_t byte_offset = 0
cdef Py_ssize_t si = 1
Expand All @@ -269,22 +280,21 @@ cpdef to_dlpack_capsule(usm_ndarray usm_ary):
raise MemoryError(
"to_dlpack_capsule: Could not allocate memory for DLManagedTensor"
)
shape_strides_ptr = <int64_t *>stdlib.malloc((sizeof(int64_t) * 2) * nd)
if shape_strides_ptr is NULL:
stdlib.free(dlm_tensor)
raise MemoryError(
"to_dlpack_capsule: Could not allocate memory for shape/strides"
)
shape_ptr = usm_ary.get_shape()
for i in range(nd):
shape_strides_ptr[i] = shape_ptr[i]
strides_ptr = usm_ary.get_strides()
flags = usm_ary.flags_
if strides_ptr:
if nd > 0:
shape_strides_ptr = <int64_t *>stdlib.malloc((sizeof(int64_t) * 2) * nd)
if shape_strides_ptr is NULL:
stdlib.free(dlm_tensor)
raise MemoryError(
"to_dlpack_capsule: Could not allocate memory for shape/strides"
)
shape_ptr = usm_ary.get_shape()
for i in range(nd):
shape_strides_ptr[nd + i] = strides_ptr[i]
else:
if not (flags & USM_ARRAY_C_CONTIGUOUS):
shape_strides_ptr[i] = shape_ptr[i]
strides_ptr = usm_ary.get_strides()
if strides_ptr:
for i in range(nd):
shape_strides_ptr[nd + i] = strides_ptr[i]
else:
si = 1
for i in range(0, nd):
shape_strides_ptr[nd + i] = si
Expand All @@ -300,11 +310,8 @@ cpdef to_dlpack_capsule(usm_ndarray usm_ary):
dl_tensor.data = <void*>(data_ptr - byte_offset)
dl_tensor.ndim = nd
dl_tensor.byte_offset = <uint64_t>byte_offset
dl_tensor.shape = &shape_strides_ptr[0]
if strides_ptr is NULL:
dl_tensor.strides = NULL
else:
dl_tensor.strides = &shape_strides_ptr[nd]
dl_tensor.shape = &shape_strides_ptr[0] if nd > 0 else NULL
dl_tensor.strides = &shape_strides_ptr[nd] if nd > 0 else NULL
dl_tensor.device.device_type = kDLOneAPI
dl_tensor.device.device_id = device_id
dl_tensor.dtype.lanes = <uint16_t>1
Expand Down Expand Up @@ -384,24 +391,24 @@ cpdef to_dlpack_versioned_capsule(usm_ndarray usm_ary, bint copied):
"to_dlpack_versioned_capsule: Could not allocate memory "
"for DLManagedTensorVersioned"
)
shape_strides_ptr = <int64_t *>stdlib.malloc((sizeof(int64_t) * 2) * nd)
if shape_strides_ptr is NULL:
stdlib.free(dlmv_tensor)
raise MemoryError(
"to_dlpack_versioned_capsule: Could not allocate memory "
"for shape/strides"
)
# this can be a separate function for handling shapes and strides
shape_ptr = usm_ary.get_shape()
for i in range(nd):
shape_strides_ptr[i] = shape_ptr[i]
strides_ptr = usm_ary.get_strides()
flags = usm_ary.flags_
if strides_ptr:
if nd > 0:
shape_strides_ptr = <int64_t *>stdlib.malloc((sizeof(int64_t) * 2) * nd)
if shape_strides_ptr is NULL:
stdlib.free(dlmv_tensor)
raise MemoryError(
"to_dlpack_versioned_capsule: Could not allocate memory "
"for shape/strides"
)
# this can be a separate function for handling shapes and strides
shape_ptr = usm_ary.get_shape()
for i in range(nd):
shape_strides_ptr[nd + i] = strides_ptr[i]
else:
if not (flags & USM_ARRAY_C_CONTIGUOUS):
shape_strides_ptr[i] = shape_ptr[i]
strides_ptr = usm_ary.get_strides()
flags = usm_ary.flags_
if strides_ptr:
for i in range(nd):
shape_strides_ptr[nd + i] = strides_ptr[i]
else:
si = 1
for i in range(0, nd):
shape_strides_ptr[nd + i] = si
Expand All @@ -419,11 +426,8 @@ cpdef to_dlpack_versioned_capsule(usm_ndarray usm_ary, bint copied):
dl_tensor.data = <void*>(data_ptr - byte_offset)
dl_tensor.ndim = nd
dl_tensor.byte_offset = <uint64_t>byte_offset
dl_tensor.shape = &shape_strides_ptr[0]
if strides_ptr is NULL:
dl_tensor.strides = NULL
else:
dl_tensor.strides = &shape_strides_ptr[nd]
dl_tensor.shape = &shape_strides_ptr[0] if nd > 0 else NULL
dl_tensor.strides = &shape_strides_ptr[nd] if nd > 0 else NULL
dl_tensor.device.device_type = kDLOneAPI
dl_tensor.device.device_id = device_id
dl_tensor.dtype.lanes = <uint16_t>1
Expand Down Expand Up @@ -503,10 +507,9 @@ cpdef numpy_to_dlpack_versioned_capsule(ndarray npy_ary, bint copied):
"for DLManagedTensorVersioned"
)

is_c_contiguous = npy_ary.flags["C"]
shape = npy_ary.ctypes.shape_as(ctypes.c_int64)
strides = npy_ary.ctypes.strides_as(ctypes.c_int64)
if not is_c_contiguous:
if nd > 0:
if npy_ary.size != 1:
for i in range(nd):
if shape[i] != 1 and strides[i] % itemsize != 0:
Expand All @@ -517,18 +520,14 @@ cpdef numpy_to_dlpack_versioned_capsule(ndarray npy_ary, bint copied):
"itemsize"
)
shape_strides_ptr = <int64_t *>stdlib.malloc((sizeof(int64_t) * 2) * nd)
else:
# no need to pass strides in this case
shape_strides_ptr = <int64_t *>stdlib.malloc(sizeof(int64_t) * nd)
if shape_strides_ptr is NULL:
stdlib.free(dlmv_tensor)
raise MemoryError(
"numpy_to_dlpack_versioned_capsule: Could not allocate memory "
"for shape/strides"
)
for i in range(nd):
shape_strides_ptr[i] = shape[i]
if not is_c_contiguous:
if shape_strides_ptr is NULL:
stdlib.free(dlmv_tensor)
raise MemoryError(
"numpy_to_dlpack_versioned_capsule: Could not allocate memory "
"for shape/strides"
)
for i in range(nd):
shape_strides_ptr[i] = shape[i]
shape_strides_ptr[nd + i] = strides[i] // itemsize

writable_flag = npy_ary.flags["W"]
Expand All @@ -540,11 +539,8 @@ cpdef numpy_to_dlpack_versioned_capsule(ndarray npy_ary, bint copied):
dl_tensor.data = <void *> npy_ary.data
dl_tensor.ndim = nd
dl_tensor.byte_offset = <uint64_t>byte_offset
dl_tensor.shape = &shape_strides_ptr[0]
if is_c_contiguous:
dl_tensor.strides = NULL
else:
dl_tensor.strides = &shape_strides_ptr[nd]
dl_tensor.shape = &shape_strides_ptr[0] if nd > 0 else NULL
dl_tensor.strides = &shape_strides_ptr[nd] if nd > 0 else NULL
dl_tensor.device.device_type = kDLCPU
dl_tensor.device.device_id = 0
dl_tensor.dtype.lanes = <uint16_t>1
Expand Down Expand Up @@ -816,12 +812,8 @@ cpdef object from_dlpack_capsule(object py_caps):
raise BufferError(
"Can not import DLPack tensor with lanes != 1"
)
offset_min = 0
if dl_tensor.strides is NULL:
for i in range(dl_tensor.ndim):
sz = sz * dl_tensor.shape[i]
offset_max = sz - 1
else:
if dl_tensor.ndim > 0:
offset_min = 0
offset_max = 0
for i in range(dl_tensor.ndim):
stride_i = dl_tensor.strides[i]
Expand Down Expand Up @@ -876,15 +868,17 @@ cpdef object from_dlpack_capsule(object py_caps):
(<c_dpctl.SyclQueue>q).get_queue_ref(),
memory_owner=tmp
)

py_shape = list()
for i in range(dl_tensor.ndim):
py_shape.append(dl_tensor.shape[i])
if (dl_tensor.strides is NULL):
py_strides = None
else:
if (dl_tensor.shape is not NULL):
for i in range(dl_tensor.ndim):
py_shape.append(dl_tensor.shape[i])
if (dl_tensor.strides is not NULL):
py_strides = list()
for i in range(dl_tensor.ndim):
py_strides.append(dl_tensor.strides[i])
else:
py_strides = None
if (dl_tensor.dtype.code == kDLUInt):
ary_dt = np.dtype("u" + str(element_bytesize))
elif (dl_tensor.dtype.code == kDLInt):
Expand Down
3 changes: 3 additions & 0 deletions dpctl/tensor/_usmarray.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,6 +86,8 @@ class DLDeviceType(IntEnum):
Qualcomm Hexagon DSP
``kDLMAIA``:
Microsoft MAIA device
``kDLTrn``:
AWS Trainium device
"""
kDLCPU = c_dlpack.device_CPU
kDLCUDA = c_dlpack.device_CUDA
Expand All @@ -101,6 +103,7 @@ class DLDeviceType(IntEnum):
kDLWebGPU = c_dlpack.device_WebGPU
kDLHexagon = c_dlpack.device_Hexagon
kDLMAIA = c_dlpack.device_MAIA
kDLTrn = c_dlpack.device_Trn


cdef class InternalUSMArrayError(Exception):
Expand Down
Loading
Loading