RepoJournal
PyTorch

@pytorch

PyTorch and the broader machine-learning ecosystem

Pick a date

  1. Sat 16 may
  2. Sun 17 may
  3. Mon 18 may
  4. Tue 19 may
  5. Wed 20 may
  6. Thu 21 may
  7. Sat 23 may
  8. Sun 24 may
  9. Mon 25 may
  10. Tue 26 may
  11. Wed 27 may
  12. Thu 28 may
  13. Fri 29 may
  14. Sat 30 may
  15. Sun 31 may
  16. Mon 1 jun
  17. Mon 8 jun
  18. Tue 9 jun

The Wire · Showcase

DYNAMO GAINS ARITHMETIC OPERATORS AS FSDP2 CUTS BACKWARD COMPUTE STALLS

By RepoJournal · Filed · About PyTorch

PyTorch's JIT compiler now handles Python's core division and modulo operators natively, while FSDP2 adds buffering controls to stop reduce-scatter from blocking gradient computation.

Dynamo landed three operator implementations [1] [2] [3] that wire floor division (//), true division (/), and remainder (%) directly into torch.compile by adding CPython's number protocol slots to VariableTracker, letting compiled code use native Python arithmetic without fallback. On the distributed training front, FSDP2 shipped set_reduce_scatter_max_input_buffers [4] to decouple reduce-scatter blocking from backward compute, solving a critical latency bottleneck where compute streams stall waiting for buffer recycling every layer (measured at 37.6 ms per step per layer in production). ExecuTorch expanded MLX support with fused Q6_K quantized kernels [8] for Gemma 4 31B GGUF export, eliminating the slow dequant path and shrinking `.pte` size through kernel blob deduplication. Meanwhile, XNNPACK was removed from default builds [5] now that ExecuTorch is the recommended mobile inference path. TorchTitan consolidated FSDP configuration by deprecating the llama4 folder [6] [7] and unifying MoE and dense model setup into a single distributed/fsdp.py file. The AO team added int8/fp8 quantized QKV fusion for x86 [11], fusing three GEMMs and scaled dot product attention into a single kernel pair. Across ExecutorTorch, device tensor helpers got hardened with proper metadata preservation and error reporting [9], while test utilities moved to shared modules [10] to kill duplication. FBGEMM deployed deterministic seeding infrastructure [12] for legacy test reproducibility and optimized ROCm gradient accumulation with block-wise loop unrolling [13].

Action items

References

  1. [1] [Operators] Implement remainder operator in Dynamo (#185654) pytorch/pytorch
  2. [2] [Operators] Implement true division operator in Dynamo (#185653) pytorch/pytorch
  3. [3] [Operators] Implement floor division operator in Dynamo (#185652) pytorch/pytorch
  4. [4] [FSDP2] Add set_reduce_scatter_max_input_buffers to mitigate reduce-scatter blocking backward compute (#186000) pytorch/pytorch
  5. [5] Remove XNNPACK availability check from binary smoke test (#186662) pytorch/pytorch
  6. [6] [BE] deprecate llama4, move apply_fsdp to common file ↗ pytorch/torchtitan
  7. [7] [BE] deprecate llama4, move apply_fsdp to common file (#3573) pytorch/torchtitan
  8. [8] [MLX][Gemma4] Introduce Q6K kernels (#20004) pytorch/executorch
  9. [9] Address review feedback on device tensor helpers (#20078) (#20078) pytorch/executorch
  10. [10] Extract shared device test utilities to reduce redundancy (#20061) ↗ pytorch/executorch
  11. [11] add quantized qkv-fusion pass for x86 ↗ pytorch/ao
  12. [12] Add seed_all() deterministic-seed test helper (#5851) pytorch/FBGEMM
  13. [13] Add `PROCESS_BLOCK` macro for grad accumulation loop unrolling (#5835) pytorch/FBGEMM

FAQ

What changed in PyTorch on June 9, 2026?
PyTorch's JIT compiler now handles Python's core division and modulo operators natively, while FSDP2 adds buffering controls to stop reduce-scatter from blocking gradient computation.
What should PyTorch teams do about it?
Verify torch.compile handles your division/modulo ops without fallback in local tests • Pin FSDP2 code to new set_reduce_scatter_max_input_buffers API if running large-scale training with exposed reduce-scatter • Update torchtitan training scripts to use consolidated FSDP config from distributed/fsdp.py
Which PyTorch repositories shipped on June 9, 2026?
pytorch/pytorch, pytorch/torchtitan, pytorch/executorch, pytorch/ao, pytorch/FBGEMM

Related across the cluster

For your repos

The showcase is a teaser.
Your wire is the product.

Same engine. Different stack. Below: what changes when the wire is yours.

Showcase wire

  • •14 famous open source orgs
  • •One wire per day
  • •Public, generic
  • •Read on the web, when you remember

Your wire

  • →Up to 1,500 of your repos - orgs, deps, vendors
  • →Morning and evening briefs
  • →Action items routed to your team
  • →Slack delivery, email, breaking-news CVE alerts

Want a hands-on demo first? Ask a current user for an invite link.