unslothai/unsloth v0.1.911-beta on GitHub

This release adds scheduled tasks, FLUX-3 edit support, perf improvements and more. Also new: MiniMax-M3.1-Flash, face touch-ups for images and faster performance + chats on Mac.

Highlights

Scheduled tasks: Run a prompt on a schedule with a local model or an API
MiniMax-M3.1-Flash: MiniMax's 428B MoE (23B active) Guide
FLUX 3 Editing: FLUX 3 now supports editing of images with prompts
- Perf improvements: Large Mac perf gains, less lag, reduced binary sizes
Saved presets: Save your Run settings as named presets for each model

Scheduled tasks

Open Scheduled from the sidebar's More menu to run a prompt once, daily, on weekdays, weekly or monthly; each answer is saved as a new chat.
Each task can run on the loaded model, a model on your device, or a model from a connected API provider.
When a task's model isn't loaded, Unsloth asks before loading it, unless you turn on Switch automatically when I'm away.
Each task can notify you after every run, only when it fails, or never.
Tasks run only while Unsloth is running, without tools or web search.

UI + UX

Save Run settings in the model picker as named presets for each model, then rename, delete or switch between them.
The preset you picked last is what that model loads with next time, including loads through the API.
Chat settings can remember your sampling settings for a model, or reset them to the model's recommended values.
GGUF chats now show progress and time left while processing a long prompt, before the reply starts.
The API page now shows each served model's context use, speed and request count.
Each kind of button and icon now looks the same everywhere in the app.
The Library and System pages in Settings show storage as a usage bar, split into models, caches and free space.

Mac

MLX chats now reply up to 16% faster with draft models.
MLX models that ship with a built-in draft model now use it automatically, with no separate download.
Image and video renders need far less memory at the end, so low-memory Macs rarely slow down from swapping.
On Apple Silicon, Z-Image GGUFs now make images from text with the MLX engine by default, 1.5-2x faster than before.

Images + video

Make videos from text or an image with NVIDIA's Cosmos3-Edge on the Video page.
On the Images page, turn on Detailer to redraw faces, hands or other things you name at higher resolution; it needs SAM 3 and a model that can inpaint.
On NVIDIA GPUs older than RTX 30-series, Wan2.1 1.3B text-to-video clips render about 5x faster.
On NVIDIA GPUs, many image and video models load faster after the first time, since Unsloth saves the version it builds on first load.
FLUX.3 Klein can now edit photos in Images > Edit, keeping the photo's dimensions.

Training

Split each layer of a model across GPUs for LoRA and QLoRA training, so each GPU holds less: set UNSLOTH_TP_SIZE to the number of GPUs and launch with torchrun.
Fine-tuning Unsloth's 4-bit gpt-oss checkpoints no longer crashes on the first training step.
GRPO that generates with vLLM runs up to 1.9x faster per step with 16-bit LoRA on Linux with NVIDIA GPUs.
GRPO that generates without vLLM runs up to 4.4x faster per step for Llama and Qwen3 on NVIDIA GPUs.

Coding agents + API

Apps using an Unsloth API key can no longer approve their own Python or terminal calls to run outside the sandbox.
Unsloth no longer trusts forwarded client IPs by default; behind your own reverse proxy, set UNSLOTH_STUDIO_TRUST_FORWARDED=1 or all users share one login rate limit.

Models + hardware

Run MiniMax-M3.1-Flash, MiniMax's MoE model for coding, tool use and agent workflows, with up to 1M tokens of context. GGUF
At 1-bit, MiniMax-M3.1-Flash needs at least 130 GB of RAM and VRAM combined.
On some Windows PCs, GGUF models load again instead of failing every time with "llama-server failed to start".

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

Platform	Link
Windows	Download
macOS	Download
Linux x64 / Ubuntu (deb)	Download
Linux ARM64 / Ubuntu 24.04+ (deb)	Download
Linux x64 (AppImage)	Download
Windows ARM64	Download

What's Changed

Bump install.sh / install.ps1 pins to unsloth>=2026.10.8, unsloth-zoo>=2026.10.8 by @danielhanchen in #14994
Studio: synchronise VAE decode and encode calls on Apple MPS by @Lyxot in #14940
Repair the studio-install checks red on main by @danielhanchen in #14995
MTP: keep the trained head when saving a named PEFT adapter by @danielhanchen in #14912
Fix test_every_resolver_probe_forwards_the_trust_decision on main by @danielhanchen in #14990
Format unsloth/models/speculative.py by @danielhanchen in #14997
Studio: use the MPS SDPA prefill kernel per call on torch 2.13+ by @Lyxot in #14941
Studio diffusion: measured activation rows for MOSS-Image and FLUX.3 Klein base; layer split prices Z-Image from its measured row by @danielhanchen in #14950
Repair four checks red on main by @danielhanchen in #14991
Repair two more backend checks red on main by @danielhanchen in #15003
Studio video: keep per-request residency fitting after the stream step graph is dropped (Wan2.2-TI2V-5B at 8 / 12 GB) by @danielhanchen in #14937
FP8 block dequant on ROCm: round fp32 to fp16 with integer ops by @danielhanchen in #14992
get_peft_model: pass a scalar layers_to_transform to PEFT as a list by @danielhanchen in #14910
Studio diffusion: fixed 2 GiB discrete-card reserve (ComfyUI-style fixed reserve plus activation estimate) by @danielhanchen in #14951
Studio diffusion: fp32 VAE upcasts keep the offloaded VAE's pinned host weights (SDXL-class force_upcast VAEs) by @danielhanchen in #14946
Studio video: MiniMax-H3 fused VAE decoder follows the mixed-precision load, and encode overlaps decode by @danielhanchen in #14986
Studio: truncate completion_mask / assistant_masks with pre-tokenized input_ids by @danielhanchen in #14884
Studio: import and export skills as a ZIP from the Skills dialog by @samihsq in #14181
Studio diffusion: cap the prefetcher's pinned staging so guided Z-Image-Turbo renders on a 31 GB host (Kaggle T4x2) by @danielhanchen in #14975
Studio diffusion: block-offload re-pins land at the addresses the step graph recorded by @danielhanchen in #14765
Studio diffusion: cut the int8 ConvRot overhead (one rotation per shared input, SwiGLU fused into the down quant, factored rotation kernel) by @danielhanchen in #14982
Pin CUDA availability in the step graph capture-stream release test by @danielhanchen in #15015
Agent guides CI: resolve the newest Node 24 for the openclaw legs by @danielhanchen in #15013
Studio diffusion: learned peaks per render phase, so later loads keep text encoders resident by @danielhanchen in #14979
Studio sd.cpp: route all-Q8_0 image GGUFs to BF16 cuBLAS, keep k-quants on MMQ by @danielhanchen in #14983
Studio: run the Qwen-Image-2.1 CFG branch on the replica with its own prefix KV cache by @danielhanchen in #14954
Studio MLX images: a LoRA load keeps sd.cpp or diffusers instead of routing into the MLX refusal by @danielhanchen in #14874
Check the GGUF load marker on load_model through the AST by @danielhanchen in #15017
Studio sd.cpp: pass --conv-direct off when Studio's policy keeps direct convolution off by @danielhanchen in #14859
Name the encoding on the trainer.py read in test_pretokenized_dataset by @danielhanchen in #15018
MLX backend tests: stub zoo's new fused_drafter_ops by @danielhanchen in #15020
Studio diffusion: route untuned archs (sm86 / 90 / 100 / 121) onto the factored ConvRot kernel behind a probe and latency check by @danielhanchen in #14988
Native FSDP2: faster native_clip_grad_norm_ that follows torch clip_grad_norm_ dtypes by @danielhanchen in #14973
Native FSDP2 float16: skip the step when every rank overflows under fused AdamW by @danielhanchen in #14996
Block compile: router_logits capture no longer recompiles every layer by @danielhanchen in #14998
Compiled blocks: keep the xformers stand-in SDPA off cuDNN attention by @danielhanchen in #15009
Studio Images: decide VAE tiling at the decode and stream resident groups there, so offloaded edits decode untiled by @danielhanchen in #15006
SFT: keep the router aux loss off for single-expert configs (Jamba) by @danielhanchen in #15010
Scan baseline: re-review the eight unsloth-zoo sites 2026.10.8 moved by @danielhanchen in #15031
Studio: keep the Z-Image CFG split on when group offload leaves every group resident by @danielhanchen in #14955
Studio Images: Detailer (face fix) pass with SAM 3 and inpaint by @danielhanchen in #14616
Studio diffusion: record CUDA graphs thread-locally and keep other threads' syncs out of a recording by @danielhanchen in #14964
Opt-in torchao float8 training for full finetuning by @danielhanchen in #15007
Keep frozen load_in_fp8 weights out of FSDP2 sharding and the DDP broadcast by @danielhanchen in #15001
Settings tabs smoke: hold the shortcut recorder box before pressing Backspace by @danielhanchen in #15032
Studio diffusion: correct the fused RMSNorm accuracy notes by @danielhanchen in #15035
FSDP2 QLoRA: reuse the planned NF4 dequant launch for all-gathered weights by @danielhanchen in #15011
Studio: route Apple Silicon image requests to the MLX engine by default, diffusers MPS for the rest by @danielhanchen in #14883
Name the encoding on two more diffusion.py reads in test_diffusion_zimage_cfg_split by @danielhanchen in #15040
FSDP2: keep bitsandbytes 8-bit AdamW state by stepping the local shards by @danielhanchen in #15008
Unsloth Studio / Desktop: image generation ignores the VRAM Budget setting when planning offload by @LeoBorcherding in #14816
Unsloth Studio / Desktop: remember the text encoder and VAE files per image model by @LeoBorcherding in #14845
Studio: return a video render's pinned host memory to the driver at render end by @danielhanchen in #15036
Block compile: keep a caller's flex BACKEND instead of adding FORCE_USE_FLEX_ATTENTION by @danielhanchen in #15044
FSDP2 optimizer_cls test: follow #15008, which keeps bitsandbytes Adam / AdamW on local shards by @danielhanchen in #15049
Decode LoRA cast cache: refresh after fused optimizer steps by @danielhanchen in #15052
FSDP2 native: fewer collectives under gradient accumulation and checkpoint recompute by @danielhanchen in #15023
Attention: keep training SDPA off cuDNN attention on SM100 by @danielhanchen in #15019
mxfp4_packed_matmul: exact gradients for grad-requiring inputs by @danielhanchen in #15060
Backend CI: split the 3.13 shards a-f / g-r instead of a-k / l-r by @danielhanchen in #15061
Native FSDP2 float16: start the loss scale where one GPU's scaled backward fits by @danielhanchen in #15033
Studio diffusion: drop stale pinned host layout entries after a dtype change in low_vram offload by @danielhanchen in #14857
Studio sd.cpp: use BF16 cuBLAS for Q8_0 only when every GGUF the process loads is Q8_0 by @danielhanchen in #15054
Studio diffusion: keep the VRAM Budget reserve free in the decode-time residency release by @danielhanchen in #15048
Train on CPU when no GPU is found, instead of refusing to import by @danielhanchen in #14717
Studio diffusion: key the CFG-batch verdict on the weight quant, not just the module layout by @danielhanchen in #15047
Studio diffusion: one bounded Hub probe per process when the Hub is unreachable by @danielhanchen in #15039
MOSS-Image: build the float64 timestep sinusoid on the CPU off CUDA by @danielhanchen in #14929
Studio: split guided Z-Image on the background-compile recording render by @danielhanchen in #15057
Studio: fp32 VAE upcast frees each 16-bit device tensor as it is cast by @danielhanchen in #15058
Studio: untiled VAE fit checks do not count CUDA graph pool cache as free by @danielhanchen in #15064
Studio MLX: retry a refused coexist diffusers fallback with MLX unloaded by @danielhanchen in #15056
MoE per-expert dW: count pipelined shared memory, fit tiles that overflow the card by @danielhanchen in #15070
FP8 dequant under compile: keep the torch path for bf16 on cards without native bf16 by @danielhanchen in #15071
Studio: leave rows with a null caption out of vision training by @vineethsaivs in #14726
Format the MoE per-expert dW shared memory files by @danielhanchen in #15074
Studio diffusion: do not demote a resident transformer to whole-module offload for the 2048 worst case by @danielhanchen in #15065
Native FSDP2: sharded resume reads for weights and an opt-in sharded optimizer checkpoint by @danielhanchen in #15042
FastDiffusionModel: record an explicit float32 request so T4 training stays float32 by @danielhanchen in #15062
Native FSDP2: sync initial trainable weights across ranks; shard FastSentenceTransformer and FastDiffusionModel by @danielhanchen in #15012
Forget the last NVML inventory between backend tests by @danielhanchen in #15072
Studio: CFG batch, H3 budget and stream residency fit checks do not count CUDA graph pool cache as free by @danielhanchen in #15067
Studio NVFP4: tag the compile cache in both switch states so a legacy cache never matches by @danielhanchen in #15055
Studio diffusion/video: share the Hub reachability verdict across cached-load calls by @danielhanchen in #15059
Studio Images: run Detailer inpaints eager and uncounted by @danielhanchen in #15063
Studio: run the image-load gates before the MLX router's diffusers fallback by @danielhanchen in #15066
Studio: name image and video loads so an unload ends only its own by @danielhanchen in #14827
Studio: multistage companion and CFG replica fit checks do not count CUDA graph pool cache as free by @danielhanchen in #15069
Studio: keep host memory unpinned on integrated GPUs by @danielhanchen in #15077
Read diffusion.py as UTF-8 in the NVFP4 act-quant test by @danielhanchen in #15086
Give the embedder unload race test room for a slow release by @danielhanchen in #15082
Context parallel (Ulysses) for DPO and GRPO training by @danielhanchen in #15041
Full finetune pack-quantized compressed-tensors checkpoints by @danielhanchen in #15075
Bring two diffusion route test fakes up to the load id routes by @danielhanchen in #15090
Do not let the gap worker lock test deadlock its own worker by @danielhanchen in #15091
Studio: stream an offloaded ControlNet through the group prefetcher by @danielhanchen in #15081
Studio diffusion: carry the next forward's first offload groups into the idle copy tail by @danielhanchen in #15088
FP8 tests: same automatic dequant route on sm89+, no Fp8Dequantize leak between files by @danielhanchen in #15078
Load Mamba / FalconMamba with eager instead of a refused sdpa by @danielhanchen in #15076
GRPO trainer: follow zoo's torch < 2.12 gate for persistent TMA matmul by @danielhanchen in #15100
Resume native FSDP2 optimizer checkpoints across torch AdamW and bnb 8-bit by @danielhanchen in #15092
Studio: free the Qwen-Image-2.1 prefix KV cache before VAE decode by @danielhanchen in #15094
Studio: contiguous k/v before SDPA on gfx1151 when they outgrow L2 by @danielhanchen in #15096
Studio diffusion: keep the denoiser resident ahead of the text encoders by @danielhanchen in #15099
FP8 to NF4 load tests: compare in the load dtype on GPUs without bf16 by @danielhanchen in #15103
Studio UI polish: projects, API page, audio inputs, settings layout by @shimmyshimmer in #15080
Block compile: mark the batch dim dynamic so an uneven last batch does not recompile every layer by @danielhanchen in #15106
Studio: one red shade for every red button in light and dark by @shimmyshimmer in #15085
Studio: language and platform tabs for the API page code examples by @shimmyshimmer in #15084
FP8 inductor fullgraph test: allow one ulp on GPUs where Inductor fuses the dequant by @danielhanchen in #15107
Find a LAN address's actions in its row, not the text's parent by @danielhanchen in #15109
Studio diffusion: pack up-front group-offload pins into slabs by @danielhanchen in #15105
Studio: cross-forward prefetch for video renders by @danielhanchen in #15108
Read the working folder picker labels from the component in project_folders by @danielhanchen in #15114
MXFP4 compressed-linear tests: fp16 where bf16 is unsupported by @danielhanchen in #15113
Fused LoRA: plain Linear projections after merge_and_unload / unload no longer crash by @danielhanchen in #15119
Wait for the sandbox pick to land instead of a fixed 5ms tick by @danielhanchen in #15120
Unwrap compressed-tensors offload caches after the lazy first-forward decompress by @danielhanchen in #15095
pre-commit: format with ruff-unsloth instead of three passes around stock ruff by @danielhanchen in #14993
Studio: shorter Agent permissions description, even card padding by @shimmyshimmer in #15124
Native FSDP2 optimizer.pt resume: parameters the saved run never stepped by @danielhanchen in #15050
Keep Mamba dt_proj out of 4-bit so QLoRA forward works by @danielhanchen in #15079
Preserve reasoning limits when unsloth start changes context by @NilayYadav in #15027
MXFP4 torch decode: fp16 in fp32 on torch < 2.12, where ldexp overflows by @danielhanchen in #15121
Studio video: pin MiniMax-H3's lazily streamed blocks in place so the stream prefetch engages by @danielhanchen in #15118
Restore the hub source and endpoint env between backend tests by @danielhanchen in #15131
Run Unsloth's patched forwards only on Unsloth-loaded models by @danielhanchen in #15101
Keep managed vLLM and SGLang environments usable across Docker updates by @NilayYadav in #15024
Fix Books & Documents recipe loading under CSP by @Imagineer99 in #15045
Studio: Fix model pickers opening an empty On Device tab by @NilayYadav in #14985
Sync before re-reading a buffer stored in an Inductor reduction loop (torch < 2.14) by @danielhanchen in #15000
Add the Unsloth provider to OpenClaw’s selected configuration by @NilayYadav in #15028
Let OpenClaw app setup read valid JSON5 by @NilayYadav in #15029
Explain disabled SentencePiece before loading MOSS-TTS-Nano by @Imagineer99 in #14693
Training SDPA: bucket sequence lengths where torch picks cuDNN, so padded batches stop rebuilding plans by @danielhanchen in #15134
Preserve padding masks in distillation backbone forwards by @taking-lying-flat in #15068
Studio: remember chat preferences per account across logins by @NilayYadav in #15025
Context parallel RL: single-tensor all-gather for hidden states and reduce-scatter in backward by @danielhanchen in #15132
Preserve runtime settings when unsloth start names the loaded model by @NilayYadav in #15026
Studio diffusion: keep whole-module offload for a bf16 denoiser past the streamed floor by @danielhanchen in #15129
FP8 block GEMM: TMA descriptor kernel with a 4-stage K loop on B200 by @danielhanchen in #15136
Remove hidden host-device syncs from per-step training paths by @danielhanchen in #15133
Cross entropy: 8192-wide chunks at 4 warps for the chunked forward on B200 by @danielhanchen in #15141
FSDP2: replicated grad sync without host syncs, once per gradient accumulation cycle by @danielhanchen in #15135
Studio: centralize GGUF identity rules and exclude incomplete GGUF splits by @oobabooga in #15073
Studio: show prompt progress, live generation details and where a GGUF loaded by @oobabooga in #14909
Studio diffusion: keep SDXL resident on 15-16 GB cards from its measured activations by @danielhanchen in #15144
Honour use_gradient_checkpointing=False over TRL's gradient_checkpointing=True default by @danielhanchen in #15145
Studio diffusion: read the local int8 / fp8 prequant cache back on pipeline loads by @danielhanchen in #15146
Studio video: decode the Wan VAE untiled or in the fewest tiles that fit, also when offloaded by @danielhanchen in #15148
FP8 block GEMM: find torchao 0.18 blockwise_fp8_gemm and route each call to the faster kernel by @danielhanchen in #15153
Studio: count the LTX-2.5 encoder worker context in every whole-module phase by @danielhanchen in #15151
Studio diffusion: move the deferred compile off the render for offloaded and Qwen-Image-2.1 denoisers by @danielhanchen in #15150
Tensor parallelism v1: apply_tensor_parallel for dense decoders (LoRA / QLoRA) by @danielhanchen in #15137
Studio video: keep Wan2.1-T2V-1.3B in fp16 on fp16-only GPUs by @danielhanchen in #15147
Studio diffusion: seed the dense bf16 placement from FLUX.1's measured peak by @danielhanchen in #15149
Padding-free: call flash varlen through a plain autograd.Function in eager by @danielhanchen in #15160
Opt-in CUDA graphs for compiled blocks under gradient checkpointing (UNSLOTH_BLOCK_CUDAGRAPHS=1) by @danielhanchen in #15154
Studio video: cache the offloaded native int8 DiT per expert (Wan2.2-A14B warm load 210 s to 21 s) by @danielhanchen in #15152
DDP: find unused params for a fully finetuned VLM on text-only batches; native FSDP2 names expert parallelism by @danielhanchen in #15156
FSDP2: free-memory fit check, MFU without recompute (HFU separately), nccl2 warning on small GPUs by @danielhanchen in #15158
Keep test_studio_root_resilience on the backend's utils and loggers by @danielhanchen in #15165
Clear the compile kill switch in the block CUDA graphs tests by @danielhanchen in #15167
Lift for_inference's tensor parallel helpers in the return-logits test by @danielhanchen in #15162
Studio diffusion: choose whole-module vs streamed offload per request from its step count by @danielhanchen in #15163
Block compile: per-region recompile limit with isolate_recompiles on torch 2.13+ by @danielhanchen in #15157
Decode: write K/V straight into the cache slot, and opt-in CUDA-graph replay of the decode step by @danielhanchen in #15171
Studio video: cache Wan text encodes per text so an unchanged negative prompt is encoded once by @danielhanchen in #15166
Studio: one audio icon everywhere, ai-speech for Text to Speech by @shimmyshimmer in #15112
Studio: pinned Run settings footer, model picker on the radius scale by @shimmyshimmer in #15104
Studio: restyle Library and System storage sections as a usage overview by @shimmyshimmer in #15083
Studio: API page fonts, Sampling footer in chat settings by @shimmyshimmer in #15111
Studio video: stream the Wan encode under model / sequential offload too by @danielhanchen in #15173
Bump setuptools from 84.0.0 to 84.1.0 in the python-security group across 1 directory by @dependabot[bot] in #15051
Bump github/codeql-action/init in / by @dependabot[bot] in #14675
Bump the cargo-tauri group across 1 directory with 2 updates by @dependabot[bot] in #14503
Bump vite, @modelcontextprotocol/sdk and dompurify in studio/frontend by @dependabot[bot] in #14442
Studio: Sampling footer waits for the saved-sampling read before Remember and Save by @shimmyshimmer in #15177
Find the Run settings footer by its rule, not by div count by @danielhanchen in #15181
Bump dompurify from 3.4.16 to 3.4.17 in /docker/jupyter/unsloth_labext in the npm_and_yarn group across 1 directory by @dependabot[bot] in #14932
Pipeline parallelism v1: apply_pipeline_parallel for dense decoders (GPipe, 1F1B) by @danielhanchen in #15175
Studio diffusion: skip the offload cost probe's pinned buffer on an integrated GPU by @danielhanchen in #15178
Studio: 20px for the xl corner radius by @shimmyshimmer in #15123
Give the single-row decode harness the real _decode_graphs helpers by @danielhanchen in #15182
Studio video: pin offloaded weights at their size, not a power of two by @danielhanchen in #15179
Bump the pip group across 1 directory with 2 updates by @dependabot[bot] in #15053
Studio: one padding for every main button by @shimmyshimmer in #15122
LayerNorm / RMSNorm: per-GPU num_warps tables by @danielhanchen in #15180
Studio diffusion: run an eager native int8 W8A8 linear as one compiled function (bit-identical, 1.33x streamed) by @danielhanchen in #15183
SFT: pick padded batches + block compile over padding-free when micro batches are host bound by @danielhanchen in #15185
Decode graphs on by default for RL trainer generation, released on return to training by @danielhanchen in #15188
Fused Qwen3 QK norm + RoPE: on by default on NVIDIA CUDA by @danielhanchen in #15189
Tensor parallelism v2: sharded loading, data parallel x tensor parallel, and the Trainer by @danielhanchen in #15174
Skip the norm warps CUDA case on a real CUDA probe by @danielhanchen in #15194
Gemma 3: keep flex attention only where its head_dim 256 template fits the card by @danielhanchen in #14670
Flex attention: tuned backward tiles on A100 for head_dim 256 (Gemma 3 LoRA 1.9-2.7x faster) by @danielhanchen in #15193
Clef: keep fp16 loss scaling on the float32 path (fixes Gemma 3 not training on T4) by @indrajeetapache in #14482
Studio: scheduled tasks that run a prompt on a schedule and save each answer as a new chat by @alkinun in #13871
FP8 block GEMM: route empty inputs to the Triton kernel by @danielhanchen in #15205
Studio diffusion: an int8 prequant read-back hits the compile bundle the runtime quantize saved by @danielhanchen in #15192
Name the encoding on the tensor parallel index reads; give the PPO unwrap test _decode_graphs by @danielhanchen in #15209
Packed metadata: keep the tensor path under torch.compile, fall back to pageable copies by @danielhanchen in #15199
Skip the torchao blockwise GEMM CUDA case on a real CUDA probe by @danielhanchen in #15211
Studio: an offline audio.cpp listing checks the account grant on the disk pool by @danielhanchen in #14829
Tests: skip the QAT MX and FP8 block GEMM GPU cases on the recorded real CUDA, not the spoofable probe by @danielhanchen in #14785
Studio: API keys cannot approve host-path tool calls out of the sandbox by @danielhanchen in #14757
adamw_flash: keep the unquantized bf16 Adam variance in fp32 so it can decay by @danielhanchen in #14749
llama.cpp installer: keep Qualcomm Adreno on the CPU prebuilt unless Vulkan is requested by @danielhanchen in #14709
Studio: keep enhancing prompts from an already downloaded FLUX.3 upsampler quant by @danielhanchen in #14836
Studio drafter: split TRL packed rows at seq_lengths, reattach a Hub full fine-tune's MTP head on export by @danielhanchen in #14911
Studio installer (AMD, Linux): CPU torch for a lone gfx90c instead of kernel-less ROCm wheels by @LeoBorcherding in #15002
Studio: refuse playlists at the entry of the remaining upload decoders by @danielhanchen in #14755
Studio: trust forwarded headers only behind an opted-in proxy or the managed tunnel by @danielhanchen in #14753
Tests: run the DDP no_sync accumulation steps' forward inside no_sync by @danielhanchen in #14711
DeepSeek-V4.1 context scratch: charge the dense KQ scores when flash attention is off by @danielhanchen in #14838
Studio: trim masked padding from pre-tokenized rows and range-check labels by @danielhanchen in #14818
Cross entropy: correct gradients under torch.compile for vocabs above 65536 by @danielhanchen in #14712
Studio GGUF placement: the dGPU wins Auto only without losing context; explicit picks keep host-only bytes off the dGPU by @danielhanchen in #14615
Padding choice: per compute capability thresholds, RTX PRO 6000 / RTX 5090 peaks by @danielhanchen in #15216
Installers: read lspci and the audio.cpp driver version once per run, one WMI query for both AMD scans by @danielhanchen in #14984
Cross entropy: an all -inf vocab chunk no longer turns the loss into NaN by @danielhanchen in #14713
Studio diffusion: price the compiled tier's requests from its own learned peak (offloaded bf16 keeps its residency after compile) by @danielhanchen in #15198
head_dim 512 attention: leave mask plus is_causal to SDPA, fall back on mask-scan OOM by @danielhanchen in #14691
Studio: keep the AMD gfx target and HIP id map until hardware re-detection by @danielhanchen in #14970
Studio: judge a weight name's .. where the filesystem sends it, not by its text by @danielhanchen in #14800
Linear input-cast hook: skip float8-weight Linears by @danielhanchen in #14690
gpt-oss: fix the unreadable lone .bin advice and pin the layout probe order to transformers' resolver by @danielhanchen in #14763
Prefetch: keep subfolder weights that a root safetensors index maps to by @danielhanchen in #15184
Studio sd.cpp: a lost Vulkan device is not a strike; Arc / ROCm keep Vulkan; a UHD iGPU beside a card no longer vetoes Vulkan by @danielhanchen in #14610
MoE grouped GEMM autotune: key, tune and fit tiles on the experts' own GPU by @danielhanchen in #14719
Studio: Mask from text runs SAM 3.1 in bf16 unless the card is short of memory by @danielhanchen in #14736
Studio diffusion: pin streamed image text encoders with the background pinner, and the prefix that fits by @danielhanchen in #15206
Generation keeps inference_mode on DeepSpeed ZeRO-1/2; only ZeRO-3 needs no_grad by @danielhanchen in #14673
RMSNorm / LayerNorm: do not hash a symbolic width in the launch settings cache by @danielhanchen in #14727
Studio: keep MiniMax-H3 PDD float64 math on the host on a device without float64 by @danielhanchen in #14848
Studio: keep LTX-2.5's bf16 text encoder on an unset request by @danielhanchen in #14737
Honour gradient_checkpointing assigned on the config after construction by @danielhanchen in #15200
Studio: keep one GGUF row per build in repos whose model name contains NVFP4 by @danielhanchen in #14733
Studio diffusion: cap the single-file host RAM refusal reserve at the fp8-storage reserve by @danielhanchen in #14758
RoPE: 64-bit cos / sin row offsets for tables past 2**31 elements by @danielhanchen in #14714
XPU: train an explicit bf16 load in bf16 under the fp16 idiom, and match vLLM's dtype when it refuses bf16 by @danielhanchen in #14606
NVFP4 QAT: accept fp16 autocast and non-contiguous activations by @danielhanchen in #14780
FSDP2 merged_4bit: keep the base bias on the GPU when lora_bias merges into it by @danielhanchen in #14801
Studio: read MiniMax-H3 card VRAM on ROCm without a torch context by @danielhanchen in #14732
Studio: keep a local folder's own text encoder when the text encoder setting is unset by @danielhanchen in #14738
Studio diffusion: write the runtime int8 / fp8 prequant cache after the load, off the first image by @danielhanchen in #15202
Stamp no bitsandbytes config on a dequantized fp8 load, and read a quantization_config object before stamping by @danielhanchen in #14803
FSDP2: restore the checkpoint's 4-bit QuantState on a full DCP load by @danielhanchen in #14747
Studio: load GGUF on CPU when the auto-selected Vulkan driver hangs or has no device by @danielhanchen in #14521
Installer: Vulkan llama.cpp for ROCm GPUs without a bundle or hipcc; keep Strix wheels below ROCm 7.13 by @danielhanchen in #14662
Fused LoRA: run PEFT for a partially merged block whose input needs no grad by @danielhanchen in #14748
RoPE cache: a cold build inside a compiled block traces again (no is_inference_mode_enabled under compile) by @danielhanchen in #14728
Context parallel gather: correct gradients under torch.compile, and a guard for in-place edits of returned tensors by @danielhanchen in #15231
Native FSDP2: restore the exact reshard-after-forward setting after generate by @danielhanchen in #14762
Native FSDP2: keep accelerate's FSDP2 for UNSLOTH_FSDP2=1 on XPU / NPU / other non-CUDA installs by @danielhanchen in #14760
Studio diffusion: streamed offload copies outside inference mode, so each compiled block traces once by @danielhanchen in #15223
Flex tuned backward: leave caller tiles and small sparse blocks to Inductor by @danielhanchen in #15228
Native FSDP2: write sub-mesh saves from the mesh's own first rank by @danielhanchen in #14761
Mixed NVIDIA hosts: bf16 and compile follow the current device; check where the weights landed after loading by @danielhanchen in #14622
Block CUDA graphs: copy parameter gradients out of the graph pool by @danielhanchen in #15207
Drop the duplicate has_real_cuda import in the torchao route test by @danielhanchen in #15239
Route ORPO, CPO, KTO, DPO and TRL 1.15's fused head to the zoo fused linear kernel by @danielhanchen in #14185
Studio: keep loading a downloaded retired FLUX.3 full copy from disk by @danielhanchen in #14835
Cross entropy: keep the loss and label gradient of confident tokens by @mlzoo in #15140
Inductor readback barrier: key the compile cache when a user pass is already set by @danielhanchen in #15208
Studio diffusion: read a pipeline's .fp16 weight files on fp16 loads (SDXL reads 6.5 GB, not 12.9 GB) by @danielhanchen in #15195
MoE dW: fall back to the legacy kernel when a per-expert tile overflows and no smaller tile is known by @danielhanchen in #15238
Studio: offer, size and pin every MIG GPU instance CUDA enumerates on R570+ drivers by @danielhanchen in #14981
Studio diffusion: price a released pipeline from its bf16 table while its cache is incomplete by @danielhanchen in #15236
Pipeline parallel x native FSDP2 data parallel x tensor parallel for dense decoders by @danielhanchen in #15204
Studio: read diffusion.py as utf-8 in the encoder pin structural test by @danielhanchen in #15242
Train the float8 FSDP2 reference in a fresh process like the sharded ranks by @danielhanchen in #15250
Compiled blocks: packed rows run varlen attention outside the graph by @danielhanchen in #14808
Decode graphs: grow the KV cache instead of reserving max_new_tokens, and capture on the model's GPU by @danielhanchen in #15237
Studio: make Keep multiple models loaded unload idle models before spilling a load to the CPU by @shimmyshimmer in #15246
Studio: show one composer button while a reply streams by @shimmyshimmer in #15248
RoPE: constexpr the strides and head counts fixed per model by @danielhanchen in #15243
Refuse accelerate config cp_size under Unsloth SFT, not only TrainingArguments by @danielhanchen in #15212
Refuse transformers / accelerate cp_size under Unsloth DPO and GRPO by @danielhanchen in #15229
SFT: compute loss_type="dft" on TRL's fused LM head instead of silently training NLL by @danielhanchen in #15213
Route PPO, RLOO, BCO, Online DPO, XPO, Nash-MD and GKD to the zoo fused linear kernels by @danielhanchen in #15244
MoE dW accumulate test: compare in-place against grad + dW, check the exact 1 on CUDA only by @danielhanchen in #15254
Studio: keep a Settings-picked embedder visible in the Hub by @danielhanchen in #15256
Define _recover_vulkan_stall below the CPU fallback it calls by @danielhanchen in #15261
TP x FSDP2 parity test: keep AdamW out of its eps cliff on the lm_head copy by @danielhanchen in #15264
GRPO truncated-support gate test: lift the block that binds _support_kwargs by @danielhanchen in #15267
Studio: stop refusing our own llama-server on Windows when psutil misreports its listener (#15022) by @debojyoti-dev in #15034
Studio: let --no-reasoning-preserve and a custom chat template turn off the Qwen3.8 preserve_thinking default by @debojyoti-dev in #15170
Studio: allow editing a refreshed reply when the tab Date disagrees by @Souravrajvi0 in #15087
Fused preference loss test: take TRL's own trainer as the reference by @danielhanchen in #15269
Studio: give web search engines a connect timeout so a blocked network fails fast (#15143) by @debojyoti-dev in #15161
Decode graph KV growth test: gate its CUDA skips on has_real_cuda() by @danielhanchen in #15270
SDPA on cuDNN: bucket masked and prefill calls, keep decode off cuDNN where torch would pick it by @danielhanchen in #15268
PP gloo parity test: give gemma3 its own per-element grad tolerance by @danielhanchen in #15273
Flex attention: fitting tiles for Gemma 3 on sm86 / sm89 (L4, A10, RTX 4090) instead of eager by @danielhanchen in #15271
DPO: UNSLOTH_FUSED_PREFERENCE_LOSS on TRL 1.x DPOTrainer._compute_loss by @danielhanchen in #15277
PP gloo parity test: size gemma3's grad bounds to its noise on AVX2 kernels by @danielhanchen in #15286
Studio: cache image native int8 builds in the local prequant cache by @danielhanchen in #15283
Studio backend tests: forget MCP discovery state between tests by @danielhanchen in #15287
Studio: build native int8 image transformers per tensor on a cache miss by @danielhanchen in #15289
Backend CI: run the multi-rank gloo suites in their own Repo tests shard by @danielhanchen in #15291
Tests: load the helpers ORPO rewriter constants are built with by @danielhanchen in #15296
Studio: let FLUX.3 Klein edit through the Edit workflow by @danielhanchen in #15263
Studio video: pin MiniMax-H3's lazily streamed blocks with the background pinner by @danielhanchen in #15297
Fused preference loss tests: build the CPU float64 model from transformers' own Llama by @danielhanchen in #15301
Studio GGUF: charge images in the fit and tool-result sizing counts (#14968) by @debojyoti-dev in #15014
Native FSDP2: keep gathered weights allocated so block CUDA graphs replay by @danielhanchen in #15215
Studio: control heights, composer fades, API page layout, settings polish by @shimmyshimmer in #15232
Studio: show newly downloaded quants of a loaded GGUF in the model picker by @shimmyshimmer in #15233
Studio: cache fp32-shard video components cast to float16 for later loads by @danielhanchen in #15303
Studio: one run settings dropdown per model, and the last setting used is what loads by @shimmyshimmer in #15247
Studio: show settings and the row menu on the loaded model in Recommended by @shimmyshimmer in #15249
Studio: one icon per job across the app by @shimmyshimmer in #15299
Studio: show the browser in Project chats by @Souravrajvi0 in #15089
Fix non-contiguous position indices in fused and unfused RoPE by @taking-lying-flat in #15235
Model picker contracts: read the multi-quant parent row's gutter through its aliases by @danielhanchen in #15319
Fix gradients through external inputs_embeds when applying attention masks by @taking-lying-flat in #15234
Read CSV text cells larger than 128 KB in RawTextDataLoader by @vineethsaivs in #14380
Keep Gemma embedding normalizer out of place for external inputs_embeds by @danielhanchen in #15308
Studio: answer 400 to out-of-range sampling on /v1/messages with a Connection model by @vineethsaivs in #15282
Studio: reject a NaN weight_decay in a training request by @vineethsaivs in #14721
GRPO kept-set refusals: stop saying UNSLOTH_GRPO_TRUNCATED_SUPPORT=1 was set by @danielhanchen in #15307
Studio: bring back the scheduled task model picker and model loading by @danielhanchen in #15255
Count prepared API prompts before GGUF admission by @Imagineer99 in #15294
Studio Images: one denoiser stream-out per prompt encode, and the Z-Image CFG negative encoded once by @danielhanchen in #15310
Studio diffusion: cast cache for fp32-shard components at bf16 too, and on the image path (Z-Image-Turbo reads 11.5 GiB, not 22.9) by @danielhanchen in #15321
TP / PP: every causal LM architecture trains like one process or is refused up front by @danielhanchen in #15230
Trim comments in unsloth, unsloth_cli and tests by @danielhanchen in #12968
Studio: reject a training random_seed outside the numpy range up front by @vineethsaivs in #15176
Studio: keep a system column when the advisor mapping carries metadata by @vineethsaivs in #14999
Studio diffusion: read hosted files we already ship (LTX-2 fp8 DiT, Qwen-Image-2.1 bf16 VAE), and name a missing int8 text encoder file by @danielhanchen in #15318
Honour YaRN passed as rope_parameters, and rope_scaling on FastModel by @danielhanchen in #15214
Fix MOSS-Image GGUF discovery in On Device by @Imagineer99 in #15168
Align System header controls and avoid the close button by @Imagineer99 in #15295
Fix GKD normalization after response generation by @taking-lying-flat in #15222
GKD: never pass a zero generated-token count by @danielhanchen in #15328
Studio: sample Vulkan drafts on the CPU when the vocab misaligns the logits rows by @danielhanchen in #15258
Studio video: write the offloaded native int8 build to the prequant cache after the load (Wan2.2-A14B first clip 97 s sooner) by @danielhanchen in #15331
Studio: Cosmos3-Nano and Cosmos3-Edge video families by @sbabamohamad in #14554
Studio diffusion: seed the small-host int8 placement from its measured peak (FLUX.2-klein on a 12.7 GB T4: 30.4 to 10.6 s per image) by @danielhanchen in #15330
Native FSDP2: fix generate after QAT convert (mixed torch.Tensor and DTensor) by @danielhanchen in #15311
Studio: open the tile's own menu on a right-click in the Images strip by @debojyoti-dev in #15276
Native FSDP2: memory-scalable 16-bit loading (each rank reads only its shard), and the fp16 start scale on sub-meshes by @danielhanchen in #15225
Studio: keep Images, Video and Hub search typing responsive by @danielhanchen in #15306
Fix TRL 0.27-0.28 GRPOTrainer import on transformers >= 5.3 (No module named 'weave') by @danielhanchen in #15343
Compiled blocks: packed rows run the flash varlen kernel inside the graph by @danielhanchen in #15325
Studio: warn when a custom Hub endpoint redirects downloads to huggingface.co by @danielhanchen in #15329
torchrun: plain Trainer DDP find_unused default, FSDP2-collective SentenceTransformer save by @danielhanchen in #15324
Intel XPU: fix 4bit dequantize and gemv on plain (not double-quantized) absmax by @danielhanchen in #15322
Trim comments in the Studio backend by @danielhanchen in #12967
Trim comments in the installers, scripts and CI config by @danielhanchen in #12969
Trim comments in the installer entry scripts by @danielhanchen in #15333
Studio diffusion: pin FLUX.2-klein's reduction-config filter on every arch (same seed, same image across processes) by @danielhanchen in #15340
Trim comments in the Studio frontend and Tauri shell by @danielhanchen in #12966
Studio: exclude NVIDIA NPU from GPU memory accounting by @danielhanchen in #15334
Import without Triton: define Unsloth's kernels against the zoo triton stub, visible to unsloth only by @danielhanchen in #15346
Tests: name the encoding in the gloo export guard, and treat a bare device index as CUDA in the YaRN load check by @danielhanchen in #15351
fix(studio): distinguish media RAM offloading from insufficient memory by @danielhanchen in #15341
Tensor parallel from UNSLOTH_TP_SIZE for unmodified training scripts, with Trainer checkpoints and 8-bit AdamW by @danielhanchen in #15326
Tests: follow TRL's renamed tiny Phi-3 and Qwen3.5 repos by @danielhanchen in #15355
Studio: identify models in download progress by @danielhanchen in #15335
Studio: show diffusion training preparation progress by @danielhanchen in #15337
Studio: exclude diffusion models from Chat's On Device list by @danielhanchen in #15336
Do not serve inference_mode packed metadata to a training forward by @danielhanchen in #15347
Studio: stop the macOS sandbox probe from leaking processes that spin forever by @shimmyshimmer in #15345
CI: rank apt mirrors by speed so a slow Azure mirror cannot time out apt steps by @danielhanchen in #15350
FP8 fp16 dequant: keep autograd for a trainable weight or scale by @danielhanchen in #15354
Studio: safer saved run settings, Always load, and API tab tidy-up by @shimmyshimmer in #15338
Bump trufflesecurity/trufflehog in / by @dependabot[bot] in #15366
Bump step-security/harden-runner in / by @dependabot[bot] in #15362
Studio diffusion: the prompt encode streams only the free-memory shortfall by @danielhanchen in #15348
Studio diffusion: cast-cache writes read block-streamed tensors from their host copy by @danielhanchen in #15357
Block compile: keep each compiled block whole under DDP (DeepSeek-OCR torchrun) by @danielhanchen in #15359
Studio: reserve the DSpark/DFlash sidecar in --fit-target and recover from a Metal out-of-memory by @danielhanchen in #14431
Studio: keep a row's control by its title when a description sits under it by @shimmyshimmer in #15349
Studio: one owner for every request to llama-server, so a crashed server recovers on every API route by @oobabooga in #15138
Studio video: keep the untiled decode gate visible under the stream encoder, skip the fp32 VAE rerun on non-finite latents, refuse NaN clips clearly by @danielhanchen in #15339
Studio: one request per chat model load, sent to both /validate and /load by @oobabooga in #15139
FSDP2: run GGUF, torchao, OpenVINO and GGML exports on rank 0 instead of refusing them by @danielhanchen in #15353
Studio diffusion: fuse the int8 SwiGLU MLP on block-streamed Qwen-Image-2.1 denoisers by @danielhanchen in #15370
Studio: prompt edit box, menu heading chips, hub Run, files panel, loaded row menu by @shimmyshimmer in #15375

New Contributors

@samihsq made their first contribution in #14181
@mlzoo made their first contribution in #15140
@sbabamohamad made their first contribution in #14554

Full Changelog: v0.1.910-beta...v0.1.911-beta

unslothai/unsloth v0.1.911-beta Scheduled + MiniMax-M3.1-Flash + Perf Gains on GitHub