Skip to content
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

NPUW: Fix SPATIAL execution mode #28862

Merged
merged 2 commits into from
Feb 7, 2025
+26 −4
Merged

NPUW: Fix SPATIAL execution mode #28862

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c00240a
NPUW: Make SPATIAL (dynamic prompt) work again
dmatveev Feb 7, 2025
3469240
NPUW Spatial: Eliminate copy in the strided view work-around
dmatveev Feb 7, 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
7 changes: 4 additions & 3 deletions src/plugins/intel_npu/src/plugin/npuw/just_sync_infer_request.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -733,9 +733,10 @@ void ov::npuw::JustInferRequest::unsafe_infer(std::size_t real_idx) {
// Collect spatial inputs for this offset
for (auto&& param : spatial.params) {
const auto& iport = comp_model_desc.compiled_model->inputs()[param.idx];
r->set_tensor(
iport,
ov::npuw::util::view(m_spatial_io[real_idx].inputs.at(param.idx), param.dim, offset, spatial.nway));
const auto& iview = ov::npuw::util::view(m_spatial_io[real_idx].inputs.at(param.idx),
param.dim, offset,
spatial.nway);
r->set_tensor(iport, iview);
} // for(params)

// Now set the spatial outputs
Expand Down
23 changes: 22 additions & 1 deletion src/plugins/intel_npu/src/plugin/npuw/util.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -360,11 +360,32 @@ ov::SoPtr<ov::ITensor> ov::npuw::util::view(const ov::SoPtr<ov::ITensor>& src,
std::size_t offset,
std::size_t len) {
const auto& shape = src->get_shape();
NPUW_ASSERT(dim < shape.size());
View view_start = View(shape.size(), 0u);
View view_end = shape;
view_start[dim] = offset;
view_end[dim] = offset + len;
return ov::npuw::util::view(src, view_start, view_end);

auto ret_view = ov::npuw::util::view(src, view_start, view_end);

// Check if a tensor view can be faked as "continuous"
// FIXME: This trick should be removed after strided tensor
// checks are relaxed
if (std::all_of(shape.begin(), shape.begin() + dim, [](std::size_t d) { return d == 1u; })) {
// If all dimensions up to the sub-ranged dimension are 1s,
// This tensor can be faked as continuous
const auto type = src->get_element_type();
const auto view_shape = ret_view->get_shape();
ov::Strides fake_strides(shape.size());
fake_strides.back() = type.size();
std::transform(view_shape.crbegin(),
view_shape.crend() - 1,
fake_strides.rbegin(),
fake_strides.rbegin() + 1,
std::multiplies<size_t>());
return ov::get_tensor_impl(ov::Tensor(type, view_shape, ret_view->data(), fake_strides));
}
return ret_view;
}

template <typename InT>
Expand Down
Loading