RepoJournal
Hugging Face @huggingface
99 wires and counting

$ follow Hugging Face

Keep up with Hugging Face 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 2026-09-03
stories 78
GitHub RSS Archive Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

Jinja engine aligns with Jinja2; hub-kernel fallbacks now warn once

By RepoJournal · Filed · About Hugging Face

Quentin Gallouédec danieldk xenova mishig25 Albert Villanova del Moral github-actions[bot] sayakpaul
12 people shipped this
Quentin Gallouédec @qgallouedec 3 cited danieldk @danieldk 2 cited xenova @xenova 1 cited mishig25 @mishig25 1 cited Albert Villanova del Moral @albertvillanova 1 cited github-actions[bot] @github-actions[bot] sayakpaul @sayakpaul jiqing-feng @jiqing-feng sergereview[bot] @sergereview[bot] Harry Julian @harryjulian Cyril Vallez @Cyrilvallez kashif @kashif

The huggingface.js Jinja engine now supports Jinja2-style namespaces, `**` unpacking, and `**` exponentiation, while transformers warns once when a hub-kernel function silently falls back to its reference PyTorch path.

The huggingface.js Jinja engine now supports Jinja2-style namespaces, call/macro arguments, and `**` unpacking, along with `**` exponentiation [1]. It also tightens template formatting around operator precedence. This is a medium-risk change: existing templates that relied on old namespace or macro behavior may now evaluate differently, so reviewing templates that use calls or namespaces is advised. Additionally, a hub fix corrects parameter counts for packed MLX models, which previously under-reported logical parameters because packed weights are stored in U32 containers [2]. For example, `avlp12/Motif-3-Alis-MLX-8bit` reported 88.6B instead of the correct 314.8B, and the fix prevents valid `total_parameters` metadata from being rejected.

In transformers, hub-kernel functions now warn once when they fall back to their reference PyTorch path [3]. The warning names the pip distribution to install (for causal_conv1d too) and states how much slower the reference path is, replacing silent fallback with an explicit, single warning that stays out of traced graphs. Supporting this, TRL pins the invariant suite's FA2 hub kernel to the v2 branch after the nightly job went red on 2026-08-27 due to an upstream transformers 5.16.1 issue [4]. The stable-ABI builds of the kernel repo's v3 branch fail in the FA2 varlen backward whenever `num_heads != num_heads_kv`, which hits Qwen2.5-0.5B (GQA, 14 q / 2 kv heads) every run across torch 2.10 to 2.13.

Kernels now fail fast with a clear exception when a kernel version is incompatible, rather than warning and risking a hard-to-understand failure after a full download [5]. The nix-builder adds ROCm 7.14 support; upstream moved to a new build system and repository (TheRock), so the `rocm-packages` package set was forked into `rocm-packages` and `rocm-packages-old` to handle the non-1:1 repackaging [6]. TRL also removed the unused `FLASH_ATTENTION_VARIANTS` constant from the DPO trainer [7] and dropped vLLM 0.19.0 support [8].

Action items

References

  1. [1] [jinja] Fix namespaces and add support for ** syntax ↗ huggingface/huggingface.js
  2. [2] [hub] Fix parameter counts for packed MLX models ↗ huggingface/huggingface.js
  3. [3] Warn once when a hub-kernel function falls back to its reference PyTorch path (#48185) ↗ huggingface/transformers
  4. [4] Pin the invariant suite's FA2 hub kernel to the v2 branch ↗ huggingface/trl
  5. [5] kernels: make kernels version check a validator ↗ huggingface/kernels
  6. [6] nix-builder: add support for ROCm 7.14 ↗ huggingface/kernels
  7. [7] Remove the unused FLASH_ATTENTION_VARIANTS constant from the DPO trainer (#7013) ↗ huggingface/trl
  8. [8] Drop vLLM 0.19.0 support (#7002) ↗ huggingface/trl

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

the week in review Transformers deprecations cleaned, NeoMME and offload land the month in review Cache-billing leak fixed, and everything from robots to diffusers gets harder to break
Topics: Python AI / ML Full archive →

Quick answers

What shipped in Hugging Face on September 3, 2026?
The huggingface.js Jinja engine now supports Jinja2-style namespaces, `**` unpacking, and `**` exponentiation, while transformers warns once when a hub-kernel function silently falls back to its reference PyTorch path. In total, 39 commits, 36 pull requests, and 3 releases landed.
Who contributed to Hugging Face on September 3, 2026?
12 developers shipped this update, including github-actions[bot], xenova, mishig25, sayakpaul, danieldk, jiqing-feng, sergereview[bot], and Harry Julian, and 4 more.
What were the notable Hugging Face updates?
[jinja] Fix namespaces and add support for ** syntax, [hub] Fix parameter counts for packed MLX models, and Warn once when a hub-kernel function falls back to its reference PyTorch path (#48185).
♥

$ log @huggingface --recent

Transformers.js adds structured output and docs overhaul 29 shipped Transformers adds Kimi linear support 6 shipped TRL removes PPOTrainer, vendors Liger-Kernel losses 54 shipped TRL documents 1M-token training, diffusers fixes group offloading, kernels fix Intel heuristics 34 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.