Skip to content

Add tensor size validation for MatMulBnb4 to prevent OOB read via K/N attribute mismatch #27995

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

Merged
vraspar merged 4 commits into from
Apr 15, 2026
Merged

Add tensor size validation for MatMulBnb4 to prevent OOB read via K/N attribute mismatch #27995

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
3a8d28f
Add tensor size validation for MatMulBnb4 to prevent OOB read via K/N…
vraspar Apr 6, 2026
15e546f
Add same tensor size validation to CUDA MatMulBnb4 kernel and pin tes…
vraspar Apr 7, 2026
95e29fe
Add overflow-safe size checks and CUDA EP regression tests for MatMul…
vraspar Apr 8, 2026
dcd3d30
Implement overflow-safe ceiling division for expected sizes in MatMul…
vraspar Apr 13, 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
29 changes: 29 additions & 0 deletions onnxruntime/contrib_ops/cpu/quantization/matmul_bnb4.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,9 @@
ORT_ENFORCE(Status::OK() == info.GetAttr<int64_t>("N", &N_));
ORT_ENFORCE(Status::OK() == info.GetAttr<int64_t>("block_size", &block_size_));
ORT_ENFORCE(Status::OK() == info.GetAttr<int64_t>("quant_type", &quant_type_));
ORT_ENFORCE(K_ > 0, "K must be positive, got ", K_);
ORT_ENFORCE(N_ > 0, "N must be positive, got ", N_);
ORT_ENFORCE(block_size_ > 0, "block_size must be positive, got ", block_size_);
Comment thread
vraspar marked this conversation as resolved.
ORT_ENFORCE(
quant_type_ == FP4 || quant_type_ == NF4,
"Invalid quant_type, only 0 (FP4) and 1 (NF4) are supported.");
Expand Down Expand Up @@ -50,6 +53,32 @@
const uint8_t* b_quant_data = b_quant->Data<uint8_t>();
const float* absmax_data = absmax->Data<float>();

// Overflow-safe computation of expected tensor sizes.
// K_, N_, block_size_ are validated > 0 in the constructor.
if (K_ > std::numeric_limits<int64_t>::max() / N_) {

Check warning on line 58 in onnxruntime/contrib_ops/cpu/quantization/matmul_bnb4.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Add #include <limits> for numeric_limits<>  [build/include_what_you_use] [4]

Raw Output:
onnxruntime/contrib_ops/cpu/quantization/matmul_bnb4.cc:58:  Add #include <limits> for numeric_limits<>  [build/include_what_you_use] [4]
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"Overflow computing K * N for K=", K_, ", N=", N_, ".");
}
const int64_t numel = K_ * N_;
// Overflow-safe ceiling division: rewrite (a + b - 1) / b as ((a - 1) / b) + 1.
// Safe because numel > 0 (K_ > 0 and N_ > 0 validated in constructor).
const int64_t expected_b_quant_size = ((numel - 1) / 2) + 1;
const int64_t expected_absmax_size = ((numel - 1) / block_size_) + 1;

Comment thread
vraspar marked this conversation as resolved.
if (b_quant->Shape().Size() < expected_b_quant_size) {
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"b_quant tensor size (", b_quant->Shape().Size(),
") is too small for K=", K_, " and N=", N_,
". Expected at least ", expected_b_quant_size, " elements.");
}
if (absmax->Shape().Size() < expected_absmax_size) {
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"absmax tensor size (", absmax->Shape().Size(),
") is too small for K=", K_, ", N=", N_,
", block_size=", block_size_,
". Expected at least ", expected_absmax_size, " elements.");
}

AllocatorPtr allocator;
auto status = ctx->GetTempSpaceAllocator(&allocator);
ORT_RETURN_IF_ERROR(status);
Expand Down
29 changes: 29 additions & 0 deletions onnxruntime/contrib_ops/cuda/quantization/matmul_bnb4.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,9 @@
ORT_ENFORCE(Status::OK() == info.GetAttr<int64_t>("N", &N_));
ORT_ENFORCE(Status::OK() == info.GetAttr<int64_t>("block_size", &block_size_));
ORT_ENFORCE(Status::OK() == info.GetAttr<int64_t>("quant_type", &quant_type_));
ORT_ENFORCE(K_ > 0, "K must be positive, got ", K_);
ORT_ENFORCE(N_ > 0, "N must be positive, got ", N_);
ORT_ENFORCE(block_size_ > 0, "block_size must be positive, got ", block_size_);
ORT_ENFORCE(
quant_type_ == FP4 || quant_type_ == NF4,
"Invalid quant_type, only 0 (FP4) and 1 (NF4) are supported.");
Expand Down Expand Up @@ -51,6 +54,32 @@
const uint8_t* b_quant_data = b_quant->Data<uint8_t>();
const auto* absmax_data = absmax->Data<T>();

// Overflow-safe computation of expected tensor sizes.
// K_, N_, block_size_ are validated > 0 in the constructor.
if (K_ > std::numeric_limits<int64_t>::max() / N_) {

Check warning on line 59 in onnxruntime/contrib_ops/cuda/quantization/matmul_bnb4.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Add #include <limits> for numeric_limits<>  [build/include_what_you_use] [4]

Raw Output:
onnxruntime/contrib_ops/cuda/quantization/matmul_bnb4.cc:59:  Add #include <limits> for numeric_limits<>  [build/include_what_you_use] [4]
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"Overflow computing K * N for K=", K_, ", N=", N_, ".");
}
const int64_t numel = K_ * N_;
// Overflow-safe ceiling division: rewrite (a + b - 1) / b as ((a - 1) / b) + 1.
// Safe because numel > 0 (K_ > 0 and N_ > 0 validated in constructor).
const int64_t expected_b_quant_size = ((numel - 1) / 2) + 1;
const int64_t expected_absmax_size = ((numel - 1) / block_size_) + 1;

if (b_quant->Shape().Size() < expected_b_quant_size) {
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"b_quant tensor size (", b_quant->Shape().Size(),
") is too small for K=", K_, " and N=", N_,
". Expected at least ", expected_b_quant_size, " elements.");
}
if (absmax->Shape().Size() < expected_absmax_size) {
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"absmax tensor size (", absmax->Shape().Size(),
") is too small for K=", K_, ", N=", N_,
", block_size=", block_size_,
". Expected at least ", expected_absmax_size, " elements.");
}

typedef typename ToCudaType<T>::MappedType CudaT;

// TODO: find a better way to create the quant_map without using a buffer
Expand Down
82 changes: 82 additions & 0 deletions onnxruntime/test/contrib_ops/matmul_bnb4_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,88 @@ void RunTest(int64_t quant_type, int64_t M, int64_t N, int64_t K, int64_t block_
}
}

TEST(MatMulBnb4, RejectsUndersizedBQuantTensor) {
// K=32, N=2 → numel=64, expected b_quant size = (64+1)/2 = 32
// Provide only 4 bytes (valid for K=4, N=2) but claim K=32, N=2
OpTester test("MatMulBnb4", 1, kMSDomain);
test.AddAttribute<int64_t>("K", 32LL);
test.AddAttribute<int64_t>("N", 2LL);
test.AddAttribute<int64_t>("block_size", 32LL);
test.AddAttribute<int64_t>("quant_type", 1LL); // NF4

test.AddInput<float>("A", {1, 32}, std::vector<float>(32, 0.0f));
test.AddInput<uint8_t>("B", {4}, std::vector<uint8_t>(4, 0)); // too small
test.AddInput<float>("absmax", {2}, std::vector<float>(2, 1.0f));
test.AddOutput<float>("Y", {1, 2}, std::vector<float>(2, 0.0f));

std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.push_back(DefaultCpuExecutionProvider());
test.Run(OpTester::ExpectResult::kExpectFailure, "b_quant tensor size", {}, nullptr, &execution_providers);
}

TEST(MatMulBnb4, RejectsUndersizedAbsmaxTensor) {
// K=32, N=2, block_size=32 → numel=64, expected absmax size = (64+32-1)/32 = 2
// Provide only 1 absmax element
int64_t K = 32, N = 2, block_size = 32;
int64_t numel = K * N;
int64_t quantized_numel = (numel + 1) / 2;

OpTester test("MatMulBnb4", 1, kMSDomain);
test.AddAttribute<int64_t>("K", K);
test.AddAttribute<int64_t>("N", N);
test.AddAttribute<int64_t>("block_size", block_size);
test.AddAttribute<int64_t>("quant_type", 1LL); // NF4

test.AddInput<float>("A", {1, K}, std::vector<float>(K, 0.0f));
test.AddInput<uint8_t>("B", {quantized_numel}, std::vector<uint8_t>(quantized_numel, 0));
test.AddInput<float>("absmax", {1}, std::vector<float>(1, 1.0f)); // too small
test.AddOutput<float>("Y", {1, N}, std::vector<float>(N, 0.0f));

std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.push_back(DefaultCpuExecutionProvider());
test.Run(OpTester::ExpectResult::kExpectFailure, "absmax tensor size", {}, nullptr, &execution_providers);
}

Comment thread
vraspar marked this conversation as resolved.
#if defined(USE_CUDA)
TEST(MatMulBnb4, RejectsUndersizedBQuantTensorCuda) {
OpTester test("MatMulBnb4", 1, kMSDomain);
test.AddAttribute<int64_t>("K", 32LL);
test.AddAttribute<int64_t>("N", 2LL);
test.AddAttribute<int64_t>("block_size", 32LL);
test.AddAttribute<int64_t>("quant_type", 1LL); // NF4

test.AddInput<float>("A", {1, 32}, std::vector<float>(32, 0.0f));
test.AddInput<uint8_t>("B", {4}, std::vector<uint8_t>(4, 0)); // too small
test.AddInput<float>("absmax", {2}, std::vector<float>(2, 1.0f));
test.AddOutput<float>("Y", {1, 2}, std::vector<float>(2, 0.0f));

std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.push_back(DefaultCudaExecutionProvider());
test.Run(OpTester::ExpectResult::kExpectFailure, "b_quant tensor size", {}, nullptr, &execution_providers);
}

TEST(MatMulBnb4, RejectsUndersizedAbsmaxTensorCuda) {
int64_t K = 32, N = 2, block_size = 32;
int64_t numel = K * N;
int64_t quantized_numel = (numel + 1) / 2;

OpTester test("MatMulBnb4", 1, kMSDomain);
test.AddAttribute<int64_t>("K", K);
test.AddAttribute<int64_t>("N", N);
test.AddAttribute<int64_t>("block_size", block_size);
test.AddAttribute<int64_t>("quant_type", 1LL); // NF4

test.AddInput<float>("A", {1, K}, std::vector<float>(K, 0.0f));
test.AddInput<uint8_t>("B", {quantized_numel}, std::vector<uint8_t>(quantized_numel, 0));
test.AddInput<float>("absmax", {1}, std::vector<float>(1, 1.0f)); // too small
test.AddOutput<float>("Y", {1, N}, std::vector<float>(N, 0.0f));

std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.push_back(DefaultCudaExecutionProvider());
test.Run(OpTester::ExpectResult::kExpectFailure, "absmax tensor size", {}, nullptr, &execution_providers);
}
#endif

TEST(MatMulBnb4, DISABLED_Float32) {
for (auto qt : {0, 1}) {
for (auto M : {1, 2, 100}) {
Expand Down
Loading