Diffusers 0.42.0 adds FLUX.3 and requires Python 3.11

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

Transformers reshaped tensor parallelism for embeddings while TRL's trainers began warning about LoRA configs that leave MoE expert layers untrained.

210 commits 191 PRs merged 3 releases 7 briefings covered

all huggingface reviews →

← previous week huggingface wire →

Diffusers 0.42.0: FLUX.3, Fibo 2.0, Kandinsky 6, and more huggingface/diffusers

by sayakpaul

Alongside FLUX.3 Image Klein, Bria Fibo 2.0, Kandinsky 6, SANA-WM, and RFD4-Proteína, the release adds new adapter support and modular pipelines. The Python floor moves to 3.11, so pin the version before upgrading any pipeline that still runs on 3.10.

🚨 TP: give embeddings their own style, shard input embeddings by default, and shard widths that don't divide by the TP size (#49570) huggingface/transformers

by Ilyas Moutawwakil

`ColwiseParallel` and `RowwiseParallel` no longer special-case `nn.Embedding`, which get their own styles, and input embeddings now shard by default. Shard widths that don't divide evenly by the TP size are handled too, so expert and embedding plans written against the old behavior need another look.

Fix zero expert gradients in bfloat16 with expert parallelism and ep_size < fsdp_size huggingface/transformers

by qgallouedec

Expert parallelism with `ep_size` smaller than `fsdp_size` produced zero expert gradients in bfloat16. If you train MoE checkpoints in that configuration, this is the fix worth carrying.

Warn when LoRA skips the dense and shared expert layers of MoE models on transformers v5 with peft 0.19 to 0.21.2 huggingface/trl

by albertvillanova

With transformers v5 and peft 0.19 to 0.21.2, a LoRA config can silently skip the dense and shared expert layers of an MoE model. The trainers now warn instead of letting the run finish with those layers untrained; check your target modules against the warning.

Ernie4.5-VL-MoE: shard the text and vision experts under TP and EP huggingface/transformers

by IlyasMoutawwakil

Expert plans for Ernie4.5-VL-MoE now shard both the text and vision experts under TP and EP. The work is still in progress, but it targets the same parallel path the embedding changes above rework.

Add support for vLLM 0.32.0 huggingface/trl

by qgallouedec

TRL's vLLM 0.32.0 support absorbs one breaking change: `/inference/v1/generate` logprobs moved to integer token IDs, so `VLLMClient.generate` had to parse the new shape. Self-hosted vLLM deployments should match the version before the client change bites.

Security bumps (eMudhra compromise) (#918) huggingface/kernels

by Daniël de Kok

Security bumps land in kernels following an eMudhra compromise. Treat it as a bump, not a rewrite, and check whether the kernels you build pull the affected dependency.

Fix model imports when torchao is older than 0.15 (#15095) huggingface/diffusers

by Muhammad Usman Mateen

`torchao_quantizer.py` imported `FqnToConfig` unconditionally, but that symbol only exists from torchao 0.15, so imports failed on older installs. Alongside: XPU varlen LSE indexing was corrected for padded queries, the metal flash sdpa kernels gained `flash_attn_func`, and TRL documented training on AMD MI300X.

$ ls huggingface/week/ # the briefings behind this review

Mon Oct 19 Ernie4.5-VL-MoE experts shard under TP and EP Tue Oct 20 Tensor-parallel embeddings get their own styles, and LoRA silently skips MoE expert layers Wed Oct 21 TRL reads tokenizers for text-only models, chat-ui queues messages mid-run Thu Oct 22 Security bumps in kernels, a torchao import fix, and Grounding DINO heads that stay untied Fri Oct 23 Online DPO fixes Qwen-VL multimodal RoPE, and a CPU parity test for continuous batching Sat Oct 24 Diffusers 0.42.0 requires Python 3.11, lerobot cuts sidecar memory 91% Sun Oct 25 XPU flash-attention speed fix, FSDP2 dtype cast for async RL

all huggingface reviews →

← previous week huggingface wire →