$ the-wire · showcase
Diffusers 0.41.0 ships Qwen-Image 2.1, TRL drops Python 3.10
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
12 people shipped this
Sayak Paul
@sayakpaul
1 cited
sirahd
@sirahd
1 cited
Lysandre Debut
@LysandreJik
1 cited
Matt
@Rocketknight1
1 cited
jiqing-feng
@jiqing-feng
1 cited
Quentin Gallouédec
@qgallouedec
1 cited
dacorvo
@dacorvo
naykun
@naykun
Alvaro Moran
@tengomucho
Raushan Turganbay
@zucchini-nlp
sergiopaniego
@sergiopaniego
albertvillanova
@albertvillanova
Diffusers 0.41.0 lands with the Qwen-Image 2.1 pipeline and a new release cadence, while TRL cuts two supported runtimes and xet-core makes uploaded shards content-addressed.
Diffusers 0.41.0: QwenImage 2.1 pipeline and more huggingface/diffusers
Qwen-Image 2.1 arrives with text-to-image, editing, native transparency, and LoRA training, and Diffusers now coordinates minor releases around new model integrations while keeping patch releases for fixes. The release policy change means minor versions will be where new model support lands, so pinning to a patch line will no longer pick up architectural additions.
Drop Python 3.10 support (#7497) huggingface/trl
TRL drops Python 3.10, and vLLM 0.20.1 support goes with it. If your training environment still runs either, pinning TRL to the previous release is now the only supported combination.
fix(shard): make an uploaded shard's bytes depend only on its content huggingface/xet-core
CAS re-serializes every uploaded shard and stores it under its own hash, so the footer's current_timestamp() gave identical content a different S3 object on each upload; zeroing it in streaming_shard.rs makes a shard's key and ETag depend only on its content, as a xorb's already do. The two serializers that still stamp the timestamp are local-cache exports where it orders eviction, and shard_ke...
Deprecated pipelines should redirect; removed pipelines should raise (#49308) huggingface/transformers
Deprecated pipeline classes now redirect to their replacements rather than resolving as before, and pipelines that have been removed outright raise instead of failing silently. Code calling a renamed pipeline by its old name will keep working through the redirect, but code calling a deleted one will now hit an explicit error rather than an obscure import failure.
Do RoPE with mul instead of matmul (#49265) huggingface/transformers
RoPE frequencies were computed as a matmul with inner dimension of 1, which depending on torch.set_float32_matmul_precision, torch.compile settings, and kernel choice could run in TF32 and round positions above 2048 and the inverse frequencies to a 10-bit mantissa; they are now a broadcast multiply. Long-context positions that previously came out inexact should now be exact regardless of those ...
Fix quantized cache (#48700) huggingface/transformers
QuantizedLayer.reorder_cache used to dequantize the states, reorder them, and quantize them back, costing an extra dequantize plus quantize per layer per beam step and recomputing the scales; the reorder is now deferred to the next update. generate also no longer mutates the user-provided cache_config, and the shared test model is no longer forced into a sliding window. The remaining items are ...
Action items
- → Move TRL environments off Python 3.10 and vLLM 0.20.1 before upgrading past the current release huggingface/trl [plan]
