Skip to content

Fix 32-bit index overflow and error-path leak in forced_align (CPU) #4209

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
happyarts wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix 32-bit index overflow and error-path leak in forced_align (CPU) #4209

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
df15555
Fix 32-bit loop index overflow in forced_align CPU init loops
happyarts Jul 22, 2026
362f1a8
Own the DP buffers, and size them only after validating the input
Aug 3, 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
22 changes: 11 additions & 11 deletions src/libtorchaudio/forced_align/cpu/compute.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,8 @@
#include <torch/headeronly/core/Dispatch_v2.h>
#include <torch/headeronly/core/ScalarType.h>

#include <vector>

namespace torchaudio {
namespace alignment {
namespace cpu {
Expand All @@ -28,15 +30,6 @@ void forced_align_impl(
const auto L = targets.size(1);
const auto S = 2 * L + 1;

auto alphas_a = new scalar_t[2 * S]; // scalar_t is just logProbs.dtype()
for (int i = 0; i < 2 * S; i++) {
alphas_a[i] = kNegInfinity;
}

auto backPtr_a = new int8_t[T * S];
for (int i = 0; i < T * S; i++) {
backPtr_a[i] = -1;
}
auto logProbs_a = torchaudio::accessor<scalar_t, 3>(logProbs);
auto targets_a = torchaudio::accessor<target_t, 2>(targets);
auto paths_a = torchaudio::accessor<target_t, 2>(paths);
Expand All @@ -54,6 +47,15 @@ void forced_align_impl(
L,
", and number of repeats: ",
R);

// scalar_t is just logProbs.dtype(). Both buffers are allocated after the
// length check: backPtr is T * S bytes, gigabytes for long audio, and the
// vectors own their memory so that a throw cannot leak them.
std::vector<scalar_t> alphas(2 * S, kNegInfinity);
auto alphas_a = alphas.data();

std::vector<int8_t> backPtr(T * S, -1);
auto backPtr_a = backPtr.data();
auto start = T - (L + R) > 0 ? 0 : 1;
auto end = (S == 1) ? 1 : 2;
for (auto i = start; i < end; i++) {
Expand Down Expand Up @@ -128,14 +130,12 @@ void forced_align_impl(
auto ltrIdx = alphas_a[S * idx1 + S - 1] > alphas_a[S * idx1 + S - 2]
? S - 1
: S - 2; // alphas_a[idx1][S - 1], alphas_a[idx1][S - 2]
delete[] alphas_a;
// path stores the token index for each time step after force alignment.
for (auto t = T - 1; t > -1; t--) {
auto lbl_idx = ltrIdx % 2 == 0 ? blank : targets_a[batchIndex][ltrIdx / 2];
paths_a[batchIndex][t] = lbl_idx;
ltrIdx -= backPtr_a[t * S + ltrIdx]; // backPtr_a[t][ltrIdx]
}
delete[] backPtr_a;
}

std::tuple<Tensor, Tensor> compute(
Expand Down
Loading