[Security] Fix info disclosure via int32 truncation in GGUF dequantize kernels (#44971)

Signed-off-by: jperezde <jperezde@redhat.com>
This commit is contained in:
Juan Pérez de Algaba
2026-06-11 08:05:14 +00:00
committed by GitHub
parent 6e64c1bab1
commit f219788f91
3 changed files with 33 additions and 29 deletions
@@ -78,8 +78,8 @@ static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const in
}
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x);
static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
const int64_t i = 2*((int64_t)blockDim.x*blockIdx.x + threadIdx.x);
if (i >= k) {
return;
@@ -435,91 +435,91 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
}
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
const int64_t num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
template<typename dst_t>
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
}
@@ -1064,7 +1064,7 @@ typedef half dfloat; // dequantize float
typedef half2 dfloat2;
typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
template<typename dst_t>
using to_cuda_ggml_t = void (*)(const void * __restrict__ x, dst_t * __restrict__ y, int k, cudaStream_t stream);
using to_cuda_ggml_t = void (*)(const void * __restrict__ x, dst_t * __restrict__ y, int64_t k, cudaStream_t stream);
typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
typedef void (*load_tiles_cuda_t)(
@@ -79,6 +79,7 @@ torch::stable::Tensor ggml_dequantize(
W.get_device_index());
auto dtype_ = dtype.value_or(torch::headeronly::ScalarType::Half);
auto DW = torch::stable::empty({m, n}, dtype_, std::nullopt, W.device());
torch::stable::fill_(DW, 0.0);
cudaStream_t stream = get_current_cuda_stream();
VLLM_STABLE_DISPATCH_FLOATING_TYPES(DW.scalar_type(), "ggml_dequantize", [&] {
@@ -93,13 +94,14 @@ torch::stable::Tensor ggml_mul_mat_vec_a8(
torch::stable::Tensor W, // quant weight
torch::stable::Tensor X, // input
int64_t type, int64_t row) {
int col = X.sizes()[1];
int vecs = X.sizes()[0];
const int padded = (col + 512 - 1) / 512 * 512;
int64_t col = X.sizes()[1];
int64_t vecs = X.sizes()[0];
const int64_t padded = (col + 512 - 1) / 512 * 512;
const torch::stable::accelerator::DeviceGuard device_guard(
X.get_device_index());
auto Y = torch::stable::empty({vecs, row}, X.scalar_type(), std::nullopt,
W.device());
torch::stable::fill_(Y, 0.0);
cudaStream_t stream = get_current_cuda_stream();
auto quant_X = torch::stable::empty({vecs, padded / 32 * 9},
torch::headeronly::ScalarType::Int,
@@ -213,13 +215,14 @@ torch::stable::Tensor ggml_mul_mat_vec_a8(
torch::stable::Tensor ggml_mul_mat_a8(torch::stable::Tensor W, // quant weight
torch::stable::Tensor X, // input
int64_t type, int64_t row) {
int col = X.sizes()[1];
int padded = (col + 512 - 1) / 512 * 512;
int batch = X.sizes()[0];
int64_t col = X.sizes()[1];
int64_t padded = (col + 512 - 1) / 512 * 512;
int64_t batch = X.sizes()[0];
const torch::stable::accelerator::DeviceGuard device_guard(
X.get_device_index());
auto Y = torch::stable::empty({batch, row}, X.scalar_type(), std::nullopt,
W.device());
torch::stable::fill_(Y, 0.0);
cudaStream_t stream = get_current_cuda_stream();
auto quant_X = torch::stable::empty({batch, padded / 32 * 9},
torch::headeronly::ScalarType::Int,
@@ -291,12 +294,13 @@ torch::stable::Tensor ggml_moe_a8(torch::stable::Tensor X, // input
torch::stable::Tensor num_tokens_post_padded,
int64_t type, int64_t row, int64_t top_k,
int64_t tokens) {
int col = X.sizes()[1];
int padded = (col + 512 - 1) / 512 * 512;
int64_t col = X.sizes()[1];
int64_t padded = (col + 512 - 1) / 512 * 512;
const torch::stable::accelerator::DeviceGuard device_guard(
X.get_device_index());
auto Y = torch::stable::empty({tokens * top_k, row}, X.scalar_type(),
std::nullopt, W.device());
torch::stable::fill_(Y, 0.0);
cudaStream_t stream = get_current_cuda_stream();
auto quant_X = torch::stable::empty({tokens, padded / 32 * 9},
torch::headeronly::ScalarType::Int,
@@ -395,8 +399,8 @@ torch::stable::Tensor ggml_moe_a8_vec(
torch::stable::Tensor W, // expert weights
torch::stable::Tensor topk_ids, int64_t top_k, int64_t type, int64_t row,
int64_t tokens) {
int col = X.sizes()[1];
const int padded = (col + 512 - 1) / 512 * 512;
int64_t col = X.sizes()[1];
const int64_t padded = (col + 512 - 1) / 512 * 512;
const torch::stable::accelerator::DeviceGuard device_guard(
X.get_device_index());
auto Y = torch::stable::empty({tokens * top_k, row}, X.scalar_type(),