Who contributed to Hugging Face on July 4, 2026?

4 developers shipped this update, including tarekziade, ydshieh, qgallouedec, and albertvillanova.

What were the notable Hugging Face updates?

Fix: large shard traces (tests_torch) silently dropped from dashboards, Emit one span per test — drop phase + fixture spans (durable fix for oversized traces), and Copied from `transformers-test-ci`.

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

TRANSFORMERS CI FIXES OVERSIZED TRACES; TRL STREAMLINES QLORA TRAINING

By RepoJournal · Filed 06:03 UTC on July 4, 2026 · About Hugging Face

4 people shipped this

tarekziade @tarekziade 3 cited

qgallouedec @qgallouedec 3 cited

ydshieh @ydshieh 1 cited

albertvillanova @albertvillanova 1 cited

The transformers test pipeline stopped silently dropping its largest jobs from dashboards. Three critical fixes ship today.

Transformers-ci fixed a cascade of trace observability failures that left engineers flying blind on big test runs. Test traces ballooned to 36-65 MB with 100k spans, causing Tempo to ingest them but fail on retrieval, which silently broke the dashboard [1]. The root cause: pytest-opentelemetry emits per-test protocol spans plus phase and fixture spans, bloating traces unnecessarily [2]. The fix drops all but the protocol span, cutting trace size to usable levels. A second bug froze traces as "settled" once span counts held steady, missing late-arriving ERROR spans from failed tests that eventually surfaced out-of-order [4]. Together, these fixes make the biggest jobs visible and accurate again [3].

On the training side, TRL shipped four quality-of-life wins. The `quantization_config` trainer argument lands across SFTTrainer, DPOTrainer, GRPOTrainer, RLOOTrainer, and RewardTrainer, killing the pattern of reaching into `model_init_kwargs` or manual loading [5]. Data collators across DPO, SFT, Reward, and KTO got a consistency pass with unified docstrings, naming, and structure [6]. SFT now truncates sequences during dataset prep instead of on every batch, speeding iteration and setting up for future work on dropping untrained rows [7]. KTO trainer now aligns with DPO by supporting PEFT models with the Liger fused loss, fixing a blanket rejection that blocked a common pattern [8]. One breaking change: vLLM 0.15 support drops [9].

Serge's pod-per-task security model is now durable and documented [10], with the design doc folded into security.md and the plan obsolete. CI now runs task tests on the in-VPC runner with proper route acceptance so internal ALB connections work [11]. Diffusers 0.39.0 shipped Cosmos 3, NVIDIA's unified world foundation model for Physical AI running omni-generation and reasoning in a single transformer [12]. huggingface.js registered the hi terminal coding agent in the harness registry [13].

Action items

→ Upgrade transformers-ci to pull the trace fixes (drop phase/fixture spans) before next large test run huggingface/transformers-ci [immediate]
→ Update TRL to 0.39+ if using QLoRA to simplify quantization_config; note vLLM 0.15 is no longer supported huggingface/trl [plan]
→ Review Serge security docs; pod-per-task is now the durable model on prod huggingface/serge [monitor]

References

[1] Fix: large shard traces (tests_torch) silently dropped from dashboards ↗ huggingface/transformers-ci
[2] Emit one span per test — drop phase + fixture spans (durable fix for oversized traces) ↗ huggingface/transformers-ci
[3] Copied from `transformers-test-ci` ↗ huggingface/transformers-ci
[4] Fix: reverify window so out-of-order late spans aren't frozen out ↗ huggingface/transformers-ci
[5] Add `quantization_config` trainer argument (streamline QLoRA) ↗ huggingface/trl
[6] Align data collators across DPO / SFT / Reward / KTO ↗ huggingface/trl
[7] SFT: Truncate during dataset preparation, not collation ↗ huggingface/trl
[8] Align KTO with DPO: Support PEFT with Liger ↗ huggingface/trl
[9] Drop vLLM 0.15 support (#6239) huggingface/trl
[10] docs: fold per-task-pod security into docs/security.md; drop the plan doc (#43) huggingface/serge
[11] CI: run serge-task-test on the in-VPC runner (aws-general-8-plus) (#41) huggingface/serge
[12] Diffusers 0.39.0: New image and video pipelines, core library improvements, and more ↗ huggingface/diffusers
[13] Add hi agent harness (#2269) huggingface/huggingface.js

Quick answers

What shipped in Hugging Face on July 4, 2026?: The transformers test pipeline stopped silently dropping its largest jobs from dashboards. Three critical fixes ship today. In total, 62 commits, 42 pull requests, and 3 releases landed.
Who contributed to Hugging Face on July 4, 2026?: 4 developers shipped this update, including tarekziade, ydshieh, qgallouedec, and albertvillanova.
What were the notable Hugging Face updates?: Fix: large shard traces (tests_torch) silently dropped from dashboards, Emit one span per test — drop phase + fixture spans (durable fix for oversized traces), and Copied from `transformers-test-ci`.

TRANSFORMERS-CI SHIPS INTELLIGENT TEST TRIAGE, CUTS LLM NOISE 100X

Hugging Face's serge verify loop now distills 100KB failure tracebacks down to the actual assertion, runs every test 5 times to catch flakiness, and dispatches each failure category with a custom instruction before spinning GPU or consuming a single LLM token.

python 45 shipped 2-min read

@huggingface 2 days ago

TRL CI BURNS, BITSANDBYTES PINNED BELOW 0.50.0

Hugging Face's TRL training library is firefighting two separate CI failures overnight, forcing temporary pins on bitsandbytes and xfailing critical tests against a broken transformers dev build.

python 7 shipped 1-min read

@huggingface 3 days ago

TRANSFORMERS OVERHAULS LINEAR ATTENTION WHILE DEPRECATING LEGACY RESPONSE SCHEMA

The transformers library is retiring its fragile response_schema prototype in favor of streaming-compatible parsing, while simultaneously refactoring every linear attention model to use standardized convolution patterns.

python 70 shipped 2-min read

@huggingface 4 days ago

TRANSFORMERS SHIPS FSDP DISTRIBUTED TRAINING STACK, HUB LIBRARY PLUGS REDOS HOLE

Hugging Face landed distributed training orchestration in transformers while plugging a regex vulnerability that could stall untrusted card parsing for minutes.

+10

python 61 shipped 1-min read

Elsewhere on the wire

AI Agents about 21 hours ago

CLINE SHIPS FIVE CRITICAL AGENT FIXES AFTER END-TO-END TESTING BLITZ

Saoudrizwan just landed a cascade of fixes that prevent the agent loop from running wild, stop file reads from resolving to the wrong directory, and plug the compaction pipeline.

ai-agents 16 shipped 2-min read

Local LLMs about 21 hours ago

LLAMA.CPP SHIPS MINIMAX-M3 SPARSE ATTENTION, VLLM UNLOCKS HETEROGENEOUS PREFILL-DECODE

The inference stack gets a new model architecture and a fundamental shift in how prefill and decode can run on different hardware configurations.

+10

llms 125 shipped 1-min read

@CachyOS about 21 hours ago

GLIBC 2.43 REBUILD SHIPS, INSTALLER JUMPS TO C++26

CachyOS pushed glibc 2.43 through the pipeline while the CLI installer upgraded its C++ standard to C++26, signaling a major refresh across core infrastructure.

infra 97 shipped 1-min read

Elixir & Phoenix about 21 hours ago

ECTO FIXES SUBQUERY PLANNING IN JOINS, ELIXIR PATCHES IEX PARSING

Ecto's query planner now handles subqueries correctly in interpolated join filters, while Elixir fixes a syntax error that broke 'not in' as a continuation operator in IEx.