Who contributed to PyTorch on July 18, 2026?

10 developers shipped this update, including calebmkim, norx1991, AmesingFlank, theap06, Tristan Rice, lxcxjxhx, Jeff Daily, and adityas-amd, and 2 more.

What were the notable PyTorch updates?

[autotuner][sm100] dedicated B200 reduction seed heuristics, [autotuner] reduction seed heuristic: one budget allocator over a Stage-1 reduction fact, and [c10d] Use PyTorch CUDA wrappers in NCCL2 (#190084).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

HELION SHIPS B200 AUTOTUNER, PYTORCH PLUGS CUDA DUPLICATION, TORCHRL ADDS PULL-BASED MONITORING

By RepoJournal · Filed 07:00 UTC on July 18, 2026 · About PyTorch

10 people shipped this

theap06 @theap06 3 cited

calebmkim @calebmkim 2 cited

AmesingFlank @AmesingFlank 2 cited

Benson Ma @q10 2 cited

Tristan Rice @d4l3k 1 cited

lxcxjxhx @lxcxjxhx 1 cited

norx1991 @norx1991

Jeff Daily @jeffdaily

adityas-amd @adityas-amd

mcw @Arkar-Hema

Helion's stacked reduction-seed redesign targets NVIDIA's B200 with dedicated SM100 heuristics, while PyTorch core eliminates duplicate CUDA wrappers in NCCL2 and TorchRL lands its first operational monitoring layer.

In brief

Helion ships B200 autotuner with dedicated reduction seed heuristics for SM100. No more guessing which warp layout cuts latency. That's the part that actually matters.
PyTorch pytorch/pytorch now uses its own CUDA wrappers in NCCL2 instead of duplicating them. Smaller surface, fewer sync bugs. Should've been this way.
TorchRL adds LoggerMonitor and pull-based operational monitoring. Your RL agents can now report stats without blocking the training loop. Real-time visibility that doesn't kill throughput.
ReplayBuffer.stats() lands in TorchRL. Lightweight snapshot monitoring for single, multiprocess, and Ray collectors. No instrumentation tax.
FBGEMM adds grid-stride loops and ROCm cap to jagged_bmm kernels. AMD hardware gets proper batched sparse matmul paths. Performance asymmetry closes a bit.

Helion landed the foundation for B200 optimization: a single budget allocator (ReductionKernelFact + ReductionDescriptor) that sizes every reduction axis, now retargeted at SM100 via two dedicated heuristic subclasses [1][2]. This redesign centralizes tuning logic and eliminates duplication across backends. Meanwhile, PyTorch's c10d layer plugged a systemic leak: NCCL2 was routing device, stream, event, and memory operations through its own backend-local CudaApi even when PyTorch's ownership-aware wrappers already existed [3]. That fix removes redundant guarding, stream pooling, and error handling. On the monitoring front, TorchRL shipped three coordinated pieces of the collector/replay-buffer RFC: LoggerMonitor + Every for pull-based observation without hot-path logging work [6], ReplayBuffer.stats() for lightweight snapshots [7], and collector stats() overrides for multiprocessing and Ray [8]. Helion also finished two backend cleanup refactors, moving Metal and CuTe codegen into their own compiler folders [4][5]. FBGEMM and inductor rounded out the day with ROCm grid-overflow fixes and code quality scrubbing [10][11][9].

Action items

→ Review Helion's stacked reduction redesign (PR #2996-#3036) for B200 tuning impact before next merge cycle pytorch/helion [plan]
→ Validate NCCL2 CUDA wrapper swap (c10d #190084) in cluster tests before rollout pytorch/pytorch [plan]
→ Monitor TorchRL monitoring RFC adoption - LoggerMonitor ready for experiments pytorch/rl [monitor]
→ Verify ROCm grid-stride fixes in FBGEMM jagged kernels before production use pytorch/FBGEMM [plan]

References

[1] [autotuner][sm100] dedicated B200 reduction seed heuristics ↗ pytorch/helion
[2] [autotuner] reduction seed heuristic: one budget allocator over a Stage-1 reduction fact ↗ pytorch/helion
[3] [c10d] Use PyTorch CUDA wrappers in NCCL2 (#190084) pytorch/pytorch
[4] refactor: move Metal-specific codegen into `helion/_compiler/metal/` ↗ pytorch/helion
[5] refactor: move CuTe-specific codegen into `helion/_compiler/cute/` ↗ pytorch/helion
[6] [Feature] Add LoggerMonitor and Every for pull-based operational monitoring ↗ pytorch/rl
[7] [Feature] Add ReplayBuffer.stats() for lightweight monitoring ↗ pytorch/rl
[8] [Feature] Add collector stats() snapshots (single, multiprocessing, Ray) ↗ pytorch/rl
[9] [inductor] Fix code quality issues in torch/_inductor (#190001) pytorch/pytorch
[10] Add grid-stride loops + ROCm cap to jagged_jagged_bmm_kernel (#6023) pytorch/FBGEMM
[11] Add grid-stride loops + ROCm cap to jagged_dense_bmm_kernel (#6019) pytorch/FBGEMM

Quick answers

What shipped in PyTorch on July 18, 2026?: Helion's stacked reduction-seed redesign targets NVIDIA's B200 with dedicated SM100 heuristics, while PyTorch core eliminates duplicate CUDA wrappers in NCCL2 and TorchRL lands its first operational monitoring layer. In total, 65 commits and 17 pull requests landed.
Who contributed to PyTorch on July 18, 2026?: 10 developers shipped this update, including calebmkim, norx1991, AmesingFlank, theap06, Tristan Rice, lxcxjxhx, Jeff Daily, and adityas-amd, and 2 more.
What were the notable PyTorch updates?: [autotuner][sm100] dedicated B200 reduction seed heuristics, [autotuner] reduction seed heuristic: one budget allocator over a Stage-1 reduction fact, and [c10d] Use PyTorch CUDA wrappers in NCCL2 (#190084).

EXECUTORCH SLASHES CONV LATENCY, PYTORCH CORE FIXES FLEXATTENTION SPECIALIZATION BUG

ExecutorTorch's ARM backend just folded scalar multiplication into convolution weights, eliminating post-conv ops entirely, while PyTorch core patches a FlexAttention bug that was over-specializing dynamic sequence lengths in block masks.

python 69 shipped 1-min read

@pytorch 4 days ago

INDUCTOR FIXES NESTED REDUCTION BLOCK POINTER BUG, TORCHRL UNIFIES CHECKPOINTING

Inductor's Triton codegen was silently corrupting fused broadcasts and reductions by losing track of block pointer advancement across outer loops, now patched [ref:3].

python 77 shipped 1-min read

@pytorch 5 days ago

PYTORCH FIXES CRITICAL MPS MATH BUG WHILE TAMING CI GHOSTS

PyTorch's Metal backend shipped an accurate erfc function that eliminates 100% relative error on upper tail computations, fixing a cascade of broken numerics in GELU and special functions.

python 14 shipped 1-min read

@pytorch 6 days ago

PYTORCH FIXES H100 FLASH ATTENTION CI BREAKAGE, HELION CONSOLIDATES PALLAS LAUNCHERS

PyTorch patched a critical H100 test failure caused by upstream flash-attention directory churn, while Helion merged a major Pallas launcher unification that simplifies the runtime API.

python 20 shipped 1-min read

— or follow the whole beat:

Elsewhere on the wire

@cachyos about 6 hours ago

PACKAGE UPDATES ROLL ACROSS AUR DERIVED AND PROTON CACHYOS

Claude Desktop jumped to 1.22209.0, Zen Browser patched to 1.21.8b, and documentation fixes landed across the stack overnight.

infra 43 shipped 1-min read

Elixir & Phoenix about 6 hours ago

ELIXIR CUTS EXPONENTIAL GROWTH IN BDD INTERSECTIONS, PHOENIX LIVE VIEW SKIPS CLONES ON STABLE LISTS

The compiler just got smarter about preventing exponential blowup in type intersection logic, while Phoenix Live View stops wastefully copying keyed comprehensions that aren't moving.

elixir 16 shipped 1-min read

@dotnet about 6 hours ago

ASPNETCORE TIGHTENS AUTHORIZATION METADATA EVERYWHERE, BLAZOR DEBUGGING EXITS

Authorization policies now work consistently across all routing paths, not just endpoint middleware, while Blazor templates drop their deprecated WebAssembly debugging hook.

dotnet 58 shipped 2-min read

@spring-projects about 6 hours ago

SPRING FRAMEWORK TIGHTENS RESOURCE SECURITY, NEO4J FIXES CONVERSION LOGIC

Spring Framework reinstated critical resource location validation checks that prevent unauthorized static resource resolution in sensitive directories.

java 4 shipped 1-min read