$ git log --author="qgallouedec" --all --oneline

On the record

Quentin Gallouédec

@qgallouedec ↗ first byline June 2026

Every morning, RepoJournal's newsroom reads what shipped across the open-source projects developers depend on — and writes the wire. Your commits kept making the news. This is your clippings file: the stories where your work was the story.

78

stories filed

46

daily wires

$ grep -rn "author: qgallouedec" wires/ | sort -r

Hugging Face

wire 2026-11-04

Transformers renames gradient_checkpointing and drops rope deltas; diffusers adds two modular pipelines

Rename `gradient_checkpointing` to `activation_checkpointing` in the public API (#49248) huggingface/transformers

Hugging Face

wire 2026-10-30

Transformers 5.22.0 ships Mistral Large 4, TRL 1.17.0 fine-tunes it on one node

Shard the parameters kept in float32 in their own FSDP2 group (#49501) huggingface/transformers

Hugging Face

wire 2026-10-28

TRL fixes chunked loss precision, drops liger-kernel from dev deps

Drop liger-kernel from the dev dependencies, the Docker images and the tests huggingface/trl

Hugging Face

wire 2026-10-27

TRL fixes strided fused logprob backward, AsyncGRPO mitigates FP8 sampler collapse

Remove the experimental TPO trainer huggingface/trl

Hugging Face

wire 2026-10-21

TRL reads tokenizers for text-only models, chat-ui queues messages mid-run

Load the tokenizer for models that take no images huggingface/trl
Add support for vLLM 0.32.0 huggingface/trl
Document training on AMD GPUs huggingface/trl

Hugging Face

wire 2026-10-19

Ernie4.5-VL-MoE experts shard under TP and EP

Fix zero expert gradients in bfloat16 with expert parallelism and ep_size < fsdp_size huggingface/transformers
Declare the input modalities of the VLMs that report text only huggingface/transformers

Hugging Face

wire 2026-10-16

Transformers 5.20 ships Muse Spark Open, TRL adds AWS Trainium

v1.16.0 huggingface/trl
Add Muse Spark Open: tiny model, tests and 8-node SFT example huggingface/trl

Hugging Face

wire 2026-10-15

TRL defaults padding to "auto" on Neuron, drops vLLM 0.21.0

Default `pad_to_multiple_of` to `"auto"`: pad to a multiple of 128 on Neuron huggingface/trl
Save the PEFT adapter of a load-time sharded model from its own parameters huggingface/transformers
Run the fused LM head and distillation kernels on devices without Triton huggingface/trl

Hugging Face

wire 2026-10-14

AsyncGRPO packs rollouts as prefix trees, TRL types tool arguments by schema

Pass the tools to `parse_response` so tool-call arguments are typed by the tool's JSON schema huggingface/trl

Hugging Face

wire 2026-10-09

TRL 1.15 lands a fused LM head, transformers finishes its RoPE cleanup

v1.15.0 huggingface/trl

Hugging Face

wire 2026-10-08

TRL moves every trainer onto the fused LM head

Score SFT tokens with the fused LM head huggingface/trl
Score DPO and KTO tokens with the fused LM head huggingface/trl

Hugging Face

wire 2026-10-07

Multi-eos completions in TRL, Kandinsky 6 lands in diffusers

Refuse nn.DataParallel and drop the multi-GPU slow test job huggingface/trl

Hugging Face

wire 2026-10-06

Diffusers 0.41.0 ships Qwen-Image 2.1, TRL drops Python 3.10

Drop Python 3.10 support (#7497) huggingface/trl

Hugging Face

wire 2026-10-03

QLoRA keeps DoRA in float32, flash-attention kwargs stop leaking into vision encoders

Keep the DoRA magnitude vector in float32 under QLoRA huggingface/trl
Add Trainer.loss_is_scaled_for_ga to declare whether compute_loss already scales for gradient accumulation huggingface/transformers

Hugging Face

wire 2026-10-01

Transformers 5.18 ships Nemotron 3 diarization, TRL drops vLLM 0.20.0

Drop vLLM 0.20.0 support (#7435) huggingface/trl

Hugging Face

wire 2026-09-25

OpenEnv 0.6.0 drops implicit port 8000, TRL 1.14 removes trl.losses

Ship the logprob and entropy kernel in trl instead of the Hub huggingface/trl

Hugging Face

wire 2026-09-18

OpenEnv 0.5.0 ships new public surface, capture proxy gains an admin key

Remove the experimental BCO trainer (#7136) huggingface/trl

Hugging Face

wire 2026-09-13

FSDP2 plus expert parallelism, and a unified device mesh for PP + TP inference

Enable FSDP2 + expert parallelism via a 2-D (fsdp, tp) device mesh huggingface/transformers
Enable FSDP2 + expert parallelism via a 2-D (fsdp, tp) device mesh (#48516) huggingface/transformers

Hugging Face

wire 2026-09-12

HyperClovaX Vision lands in transformers, XPU test baselines reworked

fla: fix relative import depths in the vendored source and add smoke tests huggingface/kernels-community
fla: fix relative import depths in the vendored source and add smoke tests (#1129) huggingface/kernels-community
Add support for vLLM 0.29.0 (#7178) huggingface/trl

Hugging Face

wire 2026-09-10

Tau 0.4.2 trims session files, chat-ui gates deep links, TRL ships v1.13.0

v1.13.0 huggingface/trl
Add support for vLLM 0.28.0 huggingface/trl

Hugging Face

wire 2026-09-09

Fix expert-parallel NaN gradients and honor shift_labels

Fix expert-parallel training: NaN gradients and missing gradient contributions (#48205) huggingface/transformers
Honor `shift_labels` in decoder-only LLM/VLM losses huggingface/transformers
Require peft>=0.13.0 and drop the 0.12 version guards (#7102) huggingface/trl

Hugging Face

wire 2026-09-05

TRL removes PPOTrainer, vendors Liger-Kernel losses

Remove PPOTrainer (#7020) huggingface/trl
Vendor the fused linear losses from Liger-Kernel into `trl.losses` huggingface/trl

Hugging Face

wire 2026-09-04

TRL documents 1M-token training, diffusers fixes group offloading, kernels fix Intel heuristics

Document 1M-token training and add a context-parallelism example huggingface/trl
Rewrite the long context guide huggingface/trl

Hugging Face

wire 2026-09-03

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

Warn once when a hub-kernel function falls back to its reference PyTorch path (#48185) huggingface/transformers
Pin the invariant suite's FA2 hub kernel to the v2 branch huggingface/trl
Drop vLLM 0.19.0 support (#7002) huggingface/trl

Hugging Face

wire 2026-09-01

Transformers drops deprecated code, adds NeoMME and gradient checkpointing offload

Add offload to gradient checkpointing huggingface/transformers

Hugging Face

wire 2026-08-29

Diffusers pins revisions once, TRL fixes vLLM weight-sync hangs

Raise when a paged attention forward is called with no cache (#48297) huggingface/transformers

Hugging Face

wire 2026-08-28

SAGE ATTENTION GETS ITS SECOND WIND

Quiet continuous batching at default verbosity (#48314) huggingface/transformers

Hugging Face

wire 2026-08-25

Transformers MoE loss cut 99.7% as TRL gains tool calling

Compute MoE load-balancing loss per layer to avoid giant one-hot materialization −99.7% @ 128k huggingface/transformers
[DistillationTrainer refactor] Support tool calling huggingface/trl
Reorganize examples into self-contained per-example folders huggingface/trl
Emit position_ids from the SFT collator when sequence parallelism is enabled huggingface/trl

Hugging Face

wire 2026-08-20

TRL DROPS ASYNC DISTILLATION, DIFFUSERS GO TENSOR PARALLEL

Simplify AsyncGRPOTrainer: drop redundant metric guards, merge train-begin callbacks (#6378) huggingface/trl

Hugging Face

wire 2026-08-15

DIFFUSERS CLOSES LORA GAPS WHILE MOON-IDE PINS CACHE REPLICAS

Add support for vLLM 0.27.0 and 0.27.1 huggingface/trl

Hugging Face

wire 2026-08-01

TAU FIXES ANTHROPIC BILLING LEAK, UNLOCKS PROMPT CACHING

Reject unsupported `f_divergence_type` / loss type combinations in DPO huggingface/trl

Hugging Face

wire 2026-07-30

TRANSFORMERS DROPS TORCH 2.4 SUPPORT; DIFFUSERS FIXES LORA COMPILATION BUGS

[DistillationTrainer refactor] Switch generation to GRPO's stack; delete the buffer huggingface/trl
[DistillationTrainer refactor] Add `_get_last_hidden_state` (unwired) huggingface/trl

Hugging Face

wire 2026-07-25

TRANSFORMERS OVERHAULS LINEAR ATTENTION WHILE DEPRECATING LEGACY RESPONSE SCHEMA

Add Gemma4 response template huggingface/trl
[DistillationTrainer refactor] Add `_generate_single_turn` (unwired) (#6513) huggingface/trl
[DistillationTrainer refactor] Add `_generate` + `_generate_and_score_completions` (unwired) (#6521) huggingface/trl

Hugging Face

wire 2026-07-23

TRANSFORMERS HARDENS AGAINST PYTORCH FRAGMENTATION WHILE TRL SIMPLIFIES DISTILLATION

[DistillationTrainer refactor] Remove messages-format support and prompt-length config huggingface/trl
[DistillationTrainer refactor] Switch tests, docs, and example to prompt-only datasets huggingface/trl

Hugging Face

wire 2026-07-20

TAU FIXES SESSION CORRUPTION BUG, SERGE ADDS GPU VERIFICATION LOOP

Fix biased cross-rank aggregation of token-weighted metrics (GRPO, RLOO) huggingface/trl
Environment-owned dataset: make `train_dataset` optional huggingface/trl

Hugging Face

wire 2026-07-14

TRANSFORMERS ADDS MTP DECODING, TRL FIXES DEEPSEEP ZERO-3 CRASH

Fix DPO/KTO `use_liger_kernel` under DeepSpeed ZeRO-3 huggingface/trl
Drop fully-masked examples after truncation in SFT (#6320) huggingface/trl

Hugging Face

wire 2026-07-09

TRL STREAMLINES REWARD OWNERSHIP, TRANSFORMERS HARDENS SPECULATIVE DECODING

Environment-owned reward huggingface/trl
Remove post-training-toolkit integration huggingface/trl
Align KTO with DPO: `quantization_config` trainer argument huggingface/trl

Hugging Face

wire 2026-07-07

TRANSFORMERS CUTS DOC BUILD TIME, TRL DUMPS ZERO-USE PAPO TRAINER

Remove the PAPO trainer huggingface/trl

Hugging Face

wire 2026-07-04

TRANSFORMERS CI FIXES OVERSIZED TRACES; TRL STREAMLINES QLORA TRAINING

Add `quantization_config` trainer argument (streamline QLoRA) huggingface/trl
Align data collators across DPO / SFT / Reward / KTO huggingface/trl
SFT: Truncate during dataset preparation, not collation huggingface/trl

Hugging Face

wire 2026-07-03

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

Remove redundant `get_kbit_device_map()` huggingface/trl

Hugging Face

wire 2026-07-02

OPENENV ADDS HUGGING FACE SANDBOX PROVIDER, LEROBOT SHIPS FOXGLOVE VISUALIZATION

[async GRPO] Per-generation reset() observation huggingface/trl
Support multiple environments [1/2]: Pool and build environment tool dicts at batch time huggingface/trl

Hugging Face

wire 2026-07-01

TRANSFORMERS SHIPPING EXPORT STANDARDIZATION, GEMMA ATTENTION FIXES LAND

Integrate the new response parsing API huggingface/trl

Hugging Face

wire 2026-06-30

TRL FIXES SILENT TOKENIZATION CACHE DISABLING, ADDS SANDBOXES TO HUB

Fix dataset fingerprinting in DPO/SFT tokenization huggingface/trl

Hugging Face

wire 2026-06-17

TRAINING GOES PADDING-FREE WHILE DIFFUSERS EXPANDS TO SOUND AND PIXEL SPACE

Padding-free training in AsyncGRPO huggingface/trl

Hugging Face

wire 2026-06-15

TRL FIXES MEMORY LEAK IN REFERENCE MODEL HANDLING, DEFAULTS SFT LOSS TO FASTER CHUNKED ALGORITHM

Default SFT loss to chunked_nll huggingface/trl

Hugging Face

wire 2026-06-14

DISTILLATION TRAINER GRADIENT ACCUMULATION BUG FIXED, LEROBOT IMAGE STATS OVERFLOW PATCHED

[fix] GLM-4-MoE template: turn-terminating token to the turn itself huggingface/trl