mirror of
https://github.com/unslothai/unsloth.git
synced 2026-10-16 12:04:55 +00:00
Cover explicit RoPE position ids and compute the starting cache on CPU (#13976)
#18714 -Commit
815829f73b
pushed by
vrr
Studio: install causal-conv1d and mamba-ssm only when a model needs them (#13861)
#18635 -Commit
90ac3b7734
pushed by
vrr
Prebuilt CUDA wheels: compress the fatbins (flash-attn 1.49 GB -> ~285 MB) (#13854)
#18608 -Commit
7c235de6f9
pushed by
vrr
Add dtype option to save_pretrained_merged for float32 merges (#13823)
#18573 -Commit
f5f3c43606
pushed by
vrr
Studio passthrough: read the backend's identity through getattr in the cut scope (#13769)
#18461 -Commit
5e927b0fa3
pushed by
vrr
Studio: install the newest audio.cpp release and offer it in updates (#13589)
#18429 -Commit
7b6b2a8e48
pushed by
vrr
Studio: declare the Hub's /api/engines retry in the log budget (#13682)
#18300 -Commit
a413834cb1
pushed by
vrr
Studio: keep the Transcribe loaded state and Eject in step with the sidecar (#13514)
#18185 -Commit
6aaa1f36fa
pushed by
vrr
studio: restore file drops and composer drag feedback (#13387)
#18042 -Commit
5f38afbc2d
pushed by
vrr
Studio: remember Max Tokens on Max as Max, not as the old context length (#13223)
#17751 -Commit
2afec3cbf4
pushed by
vrr
Skip the fused LoRA kernels when a DoRA adapter is active (#13254)
#17663 -Commit
2308f6ef7c
pushed by
vrr
Studio: keep exponents when the model reads a web page (#13183)
#17541 -Commit
76dac379bc
pushed by
vrr
Fix the Backend CI test failures after the branch-restore and Omni routing merges (#13173)
#17338 -Commit
8002a80aaf
pushed by
vrr
Fix evaluate() under auto padding-free and keep compute_metrics rows per example (#13136)
#17242 -Commit
371da3d3bb
pushed by
vrr
Studio: head a decision score with its most likely level (#13111)
#17157 -Commit
8e3fd25af0
pushed by
vrr
Studio: reopen a chat on the branch it was left on (#13109)
#17079 -Commit
ee9c077088
pushed by
vrr
Studio: let Gemma 4 read web search and tool results on safetensors (#13096)
#17041 -Commit
20e12f5b58
pushed by
vrr
Decline the packed INT4 kernels when weight_scale has the wrong group count (#12957)
#16959 -Commit
0cb4066dae
pushed by
vrr