[CVS-182698] Fix tensor import failure caused by read-only blob tensor#971
[CVS-182698] Fix tensor import failure caused by read-only blob tensor#971n1harika merged 3 commits intoovep-developfrom
Conversation
There was a problem hiding this comment.
Pull request overview
Fixes OpenVINO EP compiled-model blob import failures by changing how native blob tensors are constructed so ov::Core::import_model() can access the underlying blob bytes when the blob comes from a file-mapped tensor.
Changes:
- Build native blob tensors from an explicit
u8pointer + shape instead of creating a sub-tensor view into the mapped file tensor. - Use
const_castto pass a mutable pointer to OpenVINO tensor construction for both mapped-file blobs and embedded-vector blobs.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| auto blob_offset = blob_container.serialized_info.file_offset; | ||
| auto blob_size = blob_container.serialized_info.size; | ||
| ov::Coordinate begin{blob_offset}; | ||
| ov::Coordinate end{blob_offset + blob_size}; | ||
| blob_container.tensor = ov::Tensor(mapped_bin_, begin, end); | ||
| const uint8_t* src = mapped_bin_.data<const uint8_t>() + blob_offset; | ||
|
|
||
| blob_container.tensor = ov::Tensor( | ||
| ov::element::u8, | ||
| ov::Shape{blob_size}, | ||
| const_cast<uint8_t*>(src)); |
There was a problem hiding this comment.
src is computed via pointer arithmetic without validating that blob_offset + blob_size stays within the mapped file tensor. If the BSON metadata is corrupt/mismatched (or a truncated external .bin is used), this can create an out-of-bounds pointer and later crash/UB when OpenVINO reads it. Add an ORT_ENFORCE range check against mapped_bin_.get_byte_size() (and also guard against integer overflow) before forming the pointer / tensor.
| const uint8_t* src = mapped_bin_.data<const uint8_t>() + blob_offset; | ||
|
|
||
| blob_container.tensor = ov::Tensor( | ||
| ov::element::u8, | ||
| ov::Shape{blob_size}, | ||
| const_cast<uint8_t*>(src)); | ||
| } else { | ||
| // Create a tensor from embedded data vector | ||
| blob_container.tensor = ov::Tensor( | ||
| ov::element::u8, | ||
| ov::Shape{blob_container.data.size()}, | ||
| blob_container.data.data()); | ||
| const_cast<uint8_t*>(blob_container.data.data())); |
There was a problem hiding this comment.
This introduces const_cast<uint8_t*>(...) to satisfy ov::Tensor's mutable-pointer constructor, but it drops const-safety for memory that may be genuinely read-only (mmap) and makes the intent unclear. Consider matching the existing pattern in ov_shared_context.cc by using a read-only ov::Tensor constructor when OPENVINO_VERSION_AT_LEAST(2026, 0) is available, and only using const_cast on older OpenVINO versions (with an explicit comment that the buffer must not be written).
| ov::element::u8, | ||
| ov::Shape{blob_container.data.size()}, | ||
| blob_container.data.data()); | ||
| const_cast<uint8_t*>(blob_container.data.data())); |
There was a problem hiding this comment.
blob_container.data.data() is already a mutable uint8_t* here, so the const_cast is unnecessary and makes it harder to reason about where const is being removed. Prefer passing blob_container.data.data() directly.
| const_cast<uint8_t*>(blob_container.data.data())); | |
| blob_container.data.data()); |
MayureshV1
left a comment
MayureshV1
left a comment
There was a problem hiding this comment.
@n1harika , Can you please look into CoPilot suggestions and if you prefer to make an implementation change,
n1harika
force-pushed
the
niharika/mutable_tensor
branch
from
b2052b7 to
94eca74
Compare
Fix tensor import failure in OVEP caused by blob tensors being created as read-only views.
Construct the tensor as a mutable pointer so OV core.import_model() can access the blob for file-mapped behavior.