RepoJournal
PyTorch @pytorch
77 wires and counting

$ follow PyTorch

Keep up with PyTorch in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-09-05
stories 160
GitHub RSS Archive Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

executorch fixes dropped edge-manager passes; TorchTitan overhauls MxFP8

By RepoJournal · Filed · About PyTorch

narendasan vmoens anijain2305 RJ Ascani theap06 oscarandersson8218 emmakujala
15 people shipped this
narendasan @narendasan 5 cited vmoens @vmoens 2 cited anijain2305 @anijain2305 2 cited RJ Ascani @rascani 1 cited theap06 @theap06 1 cited oscarandersson8218 @oscarandersson8218 emmakujala @emmakujala jirioc @jirioc Jonah Saltzman @jonah-saltzman drisspg @drisspg Isalia20 @Isalia20 Arsh Zahed @azahed98 Jason Ansel @jansel Yeonwoo Sung @YeonwooSung IvanKobzarev @IvanKobzarev

Executorch now runs a recipe's edge-manager passes after partitioning, TorchTitan replaced its MxFP8 training stack, and TensorRT's Dynamo integration ships a batch of correctness fixes.

Executorch's `LoweringRecipe.edge_manager_transform_passes` was accepted but never scheduled in a default pipeline, so passes declared by a recipe were silently dropped. The stage that applies them is now scheduled after partitioning, so recipes declaring such passes execute them. This is a breaking change for pipelines that relied on the previous behavior, which only allowed these passes to follow `TO_EDGE`, never the partitioning stage. [1]

TorchTitan overhauled its MxFP8 linear training to own autograd, FSDP cache management, and model policy, relying on TorchAO only for quantization and layout kernels. FSDP now owns the quantized weights and scales, so their lifetime follows the normal unshard/reshard cycle. A related fix handles unevenly sharded expert weights in FSDP compute weights: the mxfp8 linear's pre/post all-gather hooks now pad during pre-gather and unpad (narrow) during post-gather, which matters for `efsdp` where expert counts may not divide the FSDP degree. [2][3]

The pytorch/TensorRT Dynamo integration landed five fixes this cycle. Scalar partition inputs are materialized as zero-dimensional CUDA tensors before the Tensor[] execute_engine schema, and scalar outputs are restored with `aten.item`. `truncate_double` metadata now comes from partition FX metadata instead of executing subgraphs, matching engine precision. Explicit device placement is preserved across partitions, scatter and index updates with data-dependent extents are skipped from decomposition, and engine output symbols are remapped into the active ShapeEnv during fake execution so symbols aren't reused across traces. [4][5][6][7][8]

TorchRL added `MicroDuckEnv`, a commanded longitudinal-velocity locomotion task for Pollen Robotics' MicroDuck biped, plus MuJoCo contact and site accessors so tasks can compute contact-based metrics without reaching into backend internals. The new walker reproduction script for DreamerV3 reaches final evaluation returns of 978.7 / 971.8 / 973.7 at 1.1M steps on a single MI300X, with a `--smoke` mode that validates the pipeline in minutes. A related change seeds batched native MuJoCo workers distinctly so each worker no longer replays the same reset noise. [9][10][11]

Action items

References

  1. [1] Run a recipe's edge-manager passes after partitioning (#22474) ↗ pytorch/executorch
  2. [2] Overhaul MxFP8 support ↗ pytorch/torchtitan
  3. [3] [mxfp8] Handle unevenly sharded expert weights in FSDP compute weights ↗ pytorch/torchtitan
  4. [4] fix(dynamo): normalize scalar values at engine boundaries ↗ pytorch/TensorRT
  5. [5] fix(dynamo): make truncate_double metadata match engine precision ↗ pytorch/TensorRT
  6. [6] fix(dynamo): preserve explicit device placement across partitions ↗ pytorch/TensorRT
  7. [7] fix(dynamo): handle dynamic scatter and index updates ↗ pytorch/TensorRT
  8. [8] fix(dynamo): remap engine output symbols during fake execution ↗ pytorch/TensorRT
  9. [9] [Feature] Add MicroDuckEnv and MuJoCo contact and site accessors ↗ pytorch/rl
  10. [10] [Algorithm] Add DMC Walker reproduction script for DreamerV3 ↗ pytorch/rl
  11. [11] [BugFix] Seed batched native MuJoCo workers distinctly ↗ pytorch/rl

$ ls pytorch/reviews/ # the step-back reads

the week in review PyTorch 2.14.0 ships, security docs clarified the month in review ExecuTorch MoE quantization, Helion runtime splits, and a broader cleanup wave
Topics: Python AI / ML Full archive →

Quick answers

What shipped in PyTorch on September 5, 2026?
Executorch now runs a recipe's edge-manager passes after partitioning, TorchTitan replaced its MxFP8 training stack, and TensorRT's Dynamo integration ships a batch of correctness fixes. In total, 105 commits and 55 pull requests landed.
Who contributed to PyTorch on September 5, 2026?
15 developers shipped this update, including RJ Ascani, oscarandersson8218, emmakujala, jirioc, vmoens, theap06, Jonah Saltzman, and drisspg, and 7 more.
What were the notable PyTorch updates?
Run a recipe's edge-manager passes after partitioning (#22474), Overhaul MxFP8 support, and [mxfp8] Handle unevenly sharded expert weights in FSDP compute weights.
♥

$ log @pytorch --recent

MPS complex solves land, Arm round() matches torch, dynamo fixes repro fidelity 24 shipped Torchtitan trains MoE routers in fp32 with bf16x9 on Blackwell 73 shipped Inductor fixes stale captured-tensor mutations in while_loop 117 shipped PyTorch 2.14.0 ships, import reload guarded, AI reviewer unblocks merges 125 shipped

$ log --all # elsewhere on the wire

Meta React Native adds Stable API guards to renderer attributed string FastAPI & Pydantic Pydantic fixes Mypy plugin crash, tightens datetime constraints

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.