RepoJournal
Hugging Face

@huggingface

Transformers, Datasets, and the open AI-model layer

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

One email a day. Unsubscribe in one click.

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

SERGE PHASE 3 GOES LIVE: POD-PER-TASK EXECUTION REPLACES IN-PROCESS MODEL

By RepoJournal · Filed · About Hugging Face

Serge's orchestration layer is shipping its biggest architectural shift: each task now runs in its own isolated Kubernetes pod, complete with the full LLM loop and normalization, while the main app becomes a thin HTTP-driven supervisor.

The pod-per-task execution model [1] is rolling to production after hitting a critical egress blocker that hung every checkout for 5 minutes [2]. The fix routes git operations through the allowlisting proxy, unblocking Phase 3 verification. To support this shift, CI now builds two new images: the task-runner pod (transformers-quality toolchain plus serge) and an egress-proxy sidecar [3], both pushed to ghcr.io on every commit. On the transformers side, FSDP tests wired into the dynamic PR CI caller [6] mirror the tensor-parallel test flow, while PEFT integration tests graduated to their own isolated job [5]. A new manual Serge exerciser workflow [4] tests the write-capable /tasks flow end-to-end against transformers-test-ci using in-VPC runners and OIDC minting. TRL is cleaning house: redundant `get_kbit_device_map()` calls removed from 29 files [9], KTO loss internals shed their `policy_` prefix [10], and PEFT models now work with Liger kernels in DPO [11]. Transformers exporters [7] landed partial support for unified PyTorch/ONNX/ExecuTorch export with rope-index fixes for get_rope_index models, while GLM-Moe-DSA now uses interleaved rope in its indexer [8].

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] One pod per task: run the whole LLM loop + normalize in an isolated per-task pod ↗ huggingface/serge
  2. [2] fix: route in-pod git checkout through the egress proxy (#39) huggingface/serge
  3. [3] CI: build the per-task runner + egress-proxy images for Phase 3 ↗ huggingface/serge
  4. [4] CI: add manual Serge /tasks exerciser ↗ huggingface/transformers-test-ci
  5. [5] Add PEFT integration tests (#74) huggingface/transformers-test-ci
  6. [6] Add tests_fsdp_ci job to PR CI workflow. (#80) huggingface/transformers-test-ci
  7. [7] [PoC] HF exporters (#41992) huggingface/transformers
  8. [8] [glm-mode-dsa] Indexer uses interleaved rope (#46842) huggingface/transformers
  9. [9] Remove redundant `get_kbit_device_map()` ↗ huggingface/trl
  10. [10] Drop `policy_` prefix from KTO loss-internal logps/logits (#6204) huggingface/trl
  11. [11] Support PEFT with Liger in DPO ↗ huggingface/trl

Quick answers

What shipped in Hugging Face on July 3, 2026?
Serge's orchestration layer is shipping its biggest architectural shift: each task now runs in its own isolated Kubernetes pod, complete with the full LLM loop and normalization, while the main app becomes a thin HTTP-driven supervisor. In total, 31 commits and 32 pull requests landed.
Who contributed to Hugging Face on July 3, 2026?
3 developers shipped this update, including tarekziade, qgallouedec, and albertvillanova.
What were the notable Hugging Face updates?
One pod per task: run the whole LLM loop + normalize in an isolated per-task pod, fix: route in-pod git checkout through the egress proxy (#39), and CI: build the per-task runner + egress-proxy images for Phase 3.

More from @huggingface

Daily updates, in your inbox

Follow Hugging Face

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

or

One email a day. Unsubscribe in one click. Read a past issue →

Elsewhere on the wire

Want every project, not just this one?

We use privacy-friendly analytics (Google Analytics, IP-anonymized) to see which pages help readers. No ads, and we never sell your data. See our Privacy Policy.