Skip to content

Fix nvfp4 convert_and_update_tensor shape check #2670

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
skydoorkai wants to merge 2 commits into
base: main
Choose a base branch
from
+15 −1
Open

Fix nvfp4 convert_and_update_tensor shape check #2670

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a35de59
fix nvfp4 convert_and_update_tensor shape check
Feb 10, 2026
c7eaa05
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Feb 10, 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
16 changes: 15 additions & 1 deletion transformer_engine/pytorch/csrc/quantizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1260,6 +1260,18 @@ std::pair<TensorWrapper, py::object> NVFP4Quantizer::create_unquantized_tensor_w
return {std::move(out_cpp), std::move(out_py)};
}

std::vector<size_t> compressShapeTo2D(const std::vector<size_t>& data) {
// If 2 or fewer elements, return as-is
if (data.size() <= 2) {
return data;
}
// Multiply all elements except the last
size_t product = std::accumulate(data.begin(), data.end() - 1, static_cast<size_t>(1),
std::multiplies<size_t>());
// Return new vector of size 2: {product, last}
return std::vector<size_t>{product, data.back()};
}

std::pair<TensorWrapper, py::object> NVFP4Quantizer::convert_and_update_tensor(
py::object tensor) const {
NVTE_CHECK(detail::IsNVFP4Tensor(tensor.ptr()), "NVFP4Quantizer must output to IsNVFP4Tensor.");
Expand Down Expand Up @@ -1289,8 +1301,10 @@ std::pair<TensorWrapper, py::object> NVFP4Quantizer::convert_and_update_tensor(
shape = convert_shape_back_from_fp4(getTensorShape(*columnwise_data), true);
if (rowwise_data) {
auto expected_shape = convert_shape_back_from_fp4(getTensorShape(*rowwise_data), false);
NVTE_CHECK(shape == expected_shape, "NVFP4 row-wise data (shape=", expected_shape,
auto expected_shape_2d = compressShapeTo2D(expected_shape);
NVTE_CHECK(shape == expected_shape_2d, "NVFP4 row-wise data (2D shape=", expected_shape_2d,
") and column-wise data (shape=", shape, ") do not match");
shape = expected_shape;
}
} else { // Already checked columnwise_data_tensor == true
shape = convert_shape_back_from_fp4(getTensorShape(*rowwise_data), false);
Expand Down
Loading