huggingface/transformers

Pull Requests Commits

Merge branch 'main' into add-cellpose

ArthurZucker committed 3 hours ago

Verified 09f55c84

Cellpose: mark vendored post-processing with source and BSD-3 notice

ArthurZucker committed 3 hours ago

51ba14e4

Cellpose: unbloat modular, processor and loading tests

ArthurZucker committed 4 hours ago

e5816b12

Cellpose: fast post-processing on MPS (unique instead of contended bincount, integer center rounding)

ArthurZucker committed 4 hours ago

2c0a91a5

fix nested set_attn_implementation (#49138)

eustlb committed 5 hours ago

Verified 4e3fc5e8

feat: Add MTP support to Mellum2/2.1 (#49531)

shadeMe committed 5 hours ago

Verified 14300401

Cellpose: fill holes, vectorized flow-error QC, post-process on MPS (torch>=2.13)

ArthurZucker committed 5 hours ago

69e6da6d

Fall back to the AITER kernel for flash_attention_2 on ROCm (#49608)

Abdennacer-Badaoui committed 6 hours ago

Verified 2126f65b

Cellpose: fix device_map=auto (no-split layer, accelerate empty max) and set architectures in convert

ArthurZucker committed 6 hours ago

8c940a67

Cellpose: image-segmentation pipeline support and usage docs

ArthurZucker committed 6 hours ago

356fe35e

extend onnx exporters tests to non-cuda devices (#49627)

kaixuanliu committed 6 hours ago

Verified cd1d4736

Round DINOv3 rope periods to bf16 like the original checkpoints; simplify cellpose image processor

ArthurZucker committed 6 hours ago

b083a9c4

Document video processor backends (#49470)

DimensionSTP committed 8 hours ago

Verified b4c24288

Make sinusoidal position embeddings compileable (#49482)

jiqing-feng committed 8 hours ago

Verified 24a2afba

Update cellpose_dino integration values after rope fix

ArthurZucker committed 8 hours ago

6c0421f9

Load bf16 rope periods for cellpose_dino, repackage-only convert, point docs at checkpoints

ArthurZucker committed 11 hours ago

f2a52ad2

Add Cellpose (SAM + DINOv3 backbones)

ArthurZucker committed 12 hours ago

164e4f33

Fix a port allocation race in FSDP tests (#49455)

icavan committed 13 hours ago

Verified e24eb546

Save the PEFT adapter of a load-time sharded model from its own parameters (#49500)

qgallouedec committed 16 hours ago

Verified f0705f99

fix: Pass the MTP layer type to per-layer-type rotary embeddings (#49468)

shadeMe committed 21 hours ago

Verified 637b76e4

🚨 Remove deprecated cache utils methods (#49605)

guarin committed 22 hours ago

Verified 59a50426

Add LFM2 to the GGUF loading path (#49400)

francip committed 22 hours ago

Verified 680a17f4

Add `MtpModel` helpers for frameworks that own the drafting loop (#49544)

hmellor committed 22 hours ago

Verified 5201fa1b

Build docs with the Python version from .python-version (#49603)

tarekziade committed 22 hours ago

Verified 74cdc066

Expert-parallel dispatch hands the experts the received rows, not the all-to-all's pending result (#49421)

IlyasMoutawwakil committed 22 hours ago

Verified 0814cc79

Make the SDPA position_bias mask contiguous so T5-style models use fused kernels (#49517)

IvenHsu01 committed 23 hours ago

Verified 45729398

Use mlp_layer_types for AXK1 so the MoE layer layout honours moe_layer_freq (#49518)

hmellor committed 23 hours ago

Verified ca358790

Combine Nvidia+AMD run-slow CI comments into one (#49598)

ydshieh committed 23 hours ago

Verified e0db4024

Fix static-cache generate() with host-side inputs and under TP (#49317)

dacorvo committed 1 day ago

Verified 245167e0

Fix flaky streamer timeout tests (#49583)

tarekziade committed 1 day ago

Verified b611a013

Older