Who contributed to PyTorch on July 20, 2026?

8 developers shipped this update, including Jean Schmidt, daisyden, cyy, RonitGandhi, orangeH25, calebmkim, jansel, and pytorchbot.

What were the notable PyTorch updates?

Fix cusparselt.cpp build when neither USE_CUSPARSELT nor USE_HIPSPARSELT is defined (#190496), [autotuner] promote pointwise seed to default on sm90/sm100, and [Test][Refactor] Decouple AotAutogradFallbackTests and HooksTests (#189373).

Jul 20, 2026

$ the-wire · showcase

PYTORCH SHIPS BUILD FIX WHILE HELION TUNES H100 KERNELS TO DEFAULT

By RepoJournal · Filed 06:58 UTC on July 20, 2026 · About PyTorch · Composed from the cited sources · methodology

8 people shipped this

jansel @jansel 3 cited

Jean Schmidt @jeanschmidt 1 cited

daisyden @daisyden 1 cited

cyy @cyyever 1 cited

RonitGandhi @RonitGandhi 1 cited

calebmkim @calebmkim 1 cited

orangeH25 @orangeH25

pytorchbot @pytorchbot

A critical build regression in cusparselt.cpp is now patched, while the kernel autotuner promotes its pointwise seed heuristic to production defaults on H100 and B200.

In brief

Helion promoted pointwise seed to default on sm90/sm100. H100 autotuner now ships tuned kernels out of the box instead of burning first-run cycles. that's the real win here.
CuTe flash attention got a full tuning pass on dense ops. Helion's basically hand-tuning the GPU math PyTorch users rely on. fewer surprises at scale.
PyTorch fixed fast-mode gradcheck crashing on zero-element inputs. edge case? sure. but ops that fail silently on empty tensors are worse than ops that crash loud.
PyTorch fixed cusparselt.cpp build when neither sparse library is set. build stability isn't flashy but it's the difference between a repo that works and one that doesn't.

PyTorch's core build pipeline had a lurking fault: #178737 removed the file-level guard around cusparselt.cpp but left the function registration unprotected, causing compilation to fail on systems without cuSPARSELt or HIP support [1]. That's fixed. Meanwhile, Helion's autotuner is maturing its pointwise seed heuristic from experimental to default behavior on sm90 and sm100 architectures [2], a validation that the bandwidth-tile constants actually work at scale on H100 and B200. The move stays conservative elsewhere, keeping block_size=32 as the safe baseline for older hardware. On the attention front, Helion shipped a stacked PR series hardening its CuTe flash attention benchmark tooling [5] and adding FA4 runtime helpers [4], setting the stage for dense attention tuning [6]. Back in core, test infrastructure continues cleaning house: the S1/S2 decoupling strategy landed for AotAutogradFallbackTests [3], separating accelerator-agnostic tests from device-specific ones. Docs are purging decade-old Caffe2 references that no longer exist [7], and fast-mode gradcheck now handles zero-element inputs without crashing on forward AD [8].

Action items

→ Rebuild PyTorch on systems without cuSPARSELt or HIP to verify the cusparselt.cpp fix compiles cleanly pytorch/pytorch [plan]
→ If targeting H100 or B200, verify pointwise kernel defaults are now active by checking autotune-off behavior pytorch/helion [monitor]
→ Run gradcheck tests with zero-element tensors to ensure forward AD no longer crashes pytorch/pytorch [plan]

References

[1] Fix cusparselt.cpp build when neither USE_CUSPARSELT nor USE_HIPSPARSELT is defined (#190496) pytorch/pytorch ↗
[2] [autotuner] promote pointwise seed to default on sm90/sm100 ↗ pytorch/helion
[3] [Test][Refactor] Decouple AotAutogradFallbackTests and HooksTests (#189373) pytorch/pytorch ↗
[4] [cutedsl] Add FA4 flash runtime helpers ↗ pytorch/helion
[5] [cutedsl] Harden attention benchmark tooling ↗ pytorch/helion
[6] [cutedsl] Tune dense CuTe flash attention ↗ pytorch/helion
[7] Remove stale Caffe2 references from docs (#190430) pytorch/pytorch ↗
[8] Fix fast-mode gradcheck forward AD crash for zero-element inputs (#190437) pytorch/pytorch ↗