ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
6,601 commits 391 branches 58 tags 304 MiB
Commit graph

6,601 commits

This branch
This branch
All branches
Author SHA1 Message Date
Daniel Han
f39e997523 Merge branch 'diffusion-train-perf2' into diffusion-auto-policy 2026-07-05 04:56:22 +00:00
Daniel Han
d9d3ef462d Merge branch 'diffusion-krea2' into diffusion-train-perf2 2026-07-05 04:56:21 +00:00
Daniel Han
10d01cbc92 Merge branch 'diffusion-train-tab-2' into diffusion-krea2 
# Conflicts:
#	studio/backend/core/inference/diffusion.py
 2026-07-05 04:56:19 +00:00
Daniel Han
45f5ee7628 Merge branch 'diffusion-train-precision' into diffusion-train-tab-2 2026-07-05 04:55:07 +00:00
Daniel Han
83a0f15c07 Merge branch 'diffusion-train-perf' into diffusion-train-precision 2026-07-05 04:55:05 +00:00
Daniel Han
e0f7ac68a4 Merge branch 'image-generation' into diffusion-train-perf 2026-07-05 04:55:04 +00:00
Daniel Han
e3ad2263fb Merge branch 'image-generation' of https://github.com/unslothai/unsloth into image-generation 2026-07-05 04:53:43 +00:00
Daniel Han
c53a6cb65e Address review findings: dataset preflight, sd.cpp unload barrier, caption tombstone, ControlNet cache race 
- Run the trainer's caption discovery in the start route BEFORE freeing GPU
  residents, so a missing or uncaptionable dataset 400s without evicting the
  loaded chat/Images model.
- sd.cpp unload now waits out a cancelled one-shot generation on the generate
  lock before reporting the device free, matching the diffusers backend.
- Clearing a caption that came from metadata.jsonl writes an empty sidecar
  tombstone instead of unlinking (both readers treat an existing sidecar as
  authoritative), so the cleared label cannot resurface.
- The ControlNet wrapper pipe is only cached while its load is still current,
  closing the unload race the model cache already handled.
 2026-07-05 04:53:37 +00:00
pre-commit-ci[bot]
f6a7f81673 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 04:46:27 +00:00
pre-commit-ci[bot]
1a928afbec [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 04:45:54 +00:00
pre-commit-ci[bot]
f50e40c9d6 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 04:43:59 +00:00
Daniel Han
c9ce57f969 Merge branch 'diffusion-train-perf2' into diffusion-auto-policy 2026-07-05 04:43:43 +00:00
Daniel Han
d5f95f8e45 Merge branch 'diffusion-krea2' into diffusion-train-perf2 2026-07-05 04:43:42 +00:00
Daniel Han
176a544e1c Merge branch 'diffusion-train-tab-2' into diffusion-krea2 
# Conflicts:
#	studio/backend/core/inference/diffusion.py
 2026-07-05 04:43:41 +00:00
pre-commit-ci[bot]
546dc9dce1 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 04:43:25 +00:00
pre-commit-ci[bot]
694178a35e [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 04:42:51 +00:00
pre-commit-ci[bot]
919661ddf0 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 04:42:18 +00:00
Daniel Han
40723c3465 Merge branch 'diffusion-train-precision' into diffusion-train-tab-2 2026-07-05 04:41:58 +00:00
Daniel Han
7e3abefddd Merge branch 'diffusion-train-perf' into diffusion-train-precision 2026-07-05 04:41:57 +00:00
Daniel Han
39f9fa0fee Merge branch 'image-generation' into diffusion-train-perf 2026-07-05 04:41:56 +00:00
Daniel Han
62cae5fe71 Load image pipelines from the prefetched snapshot instead of re-sweeping the hub 
The prefetch already scopes the file list (no packaged root singles, no
dtype-variant twins, no ONNX/Flax exports), but from_pretrained was then
called with the hub id, and its own snapshot sweep re-downloaded the
skipped files anyway: 24 GB per FLUX.1 repo and 65 GB on FLUX.2-dev, as
found in the blob cache. Return the snapshot dir from the prefetch (keyed
on the pipeline manifest) and hand it to every pipeline-assembly
from_pretrained site; any prefetch failure keeps the hub id and the old
behavior.
 2026-07-05 04:41:43 +00:00
Daniel Han
26eeed371a Merge branch 'diffusion-train-perf2' into diffusion-auto-policy 2026-07-05 02:39:01 +00:00
Daniel Han
9416fbd73e Merge branch 'diffusion-krea2' into diffusion-train-perf2 2026-07-05 02:39:00 +00:00
Daniel Han
740ff8454e Merge branch 'diffusion-train-tab-2' into diffusion-krea2 2026-07-05 02:38:59 +00:00
Daniel Han
86b12b2213 Merge branch 'diffusion-train-precision' into diffusion-train-tab-2 2026-07-05 02:38:57 +00:00
Daniel Han
aaa9007c37 Merge branch 'diffusion-train-perf' into diffusion-train-precision 2026-07-05 02:38:56 +00:00
Daniel Han
ad8213ba70 Merge branch 'image-generation' into diffusion-train-perf 2026-07-05 02:38:55 +00:00
Daniel Han
53af22cf1e Merge branch 'diffusion-auto-policy' of https://github.com/unslothai/unsloth into diffusion-auto-policy 2026-07-05 02:14:36 +00:00
Daniel Han
68735819cd Keep transformer_quant tri-state through pre-eviction validation 2026-07-05 02:14:31 +00:00
pre-commit-ci[bot]
78abe26387 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 02:13:58 +00:00
pre-commit-ci[bot]
1bc77d0fd8 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 02:12:59 +00:00
Daniel Han
e049560a4a Merge branch 'diffusion-train-perf2' into diffusion-auto-policy 
# Conflicts:
#	studio/backend/core/inference/diffusion.py
 2026-07-05 02:12:30 +00:00
pre-commit-ci[bot]
20a3650108 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 02:12:27 +00:00
Daniel Han
551c38bd4a Merge branch 'diffusion-krea2' into diffusion-train-perf2 2026-07-05 02:11:26 +00:00
Daniel Han
dc36983562 Merge branch 'diffusion-train-tab-2' into diffusion-krea2 2026-07-05 02:11:25 +00:00
Daniel Han
a99b951c33 Merge branch 'diffusion-train-precision' into diffusion-train-tab-2 
# Conflicts:
#	studio/backend/tests/test_diffusion_training.py
 2026-07-05 02:11:15 +00:00
pre-commit-ci[bot]
91d7297d41 [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 02:10:58 +00:00
Daniel Han
32a77623ba Merge branch 'diffusion-train-perf' into diffusion-train-precision 
# Conflicts:
#	studio/backend/core/training/diffusion_train_common.py
 2026-07-05 02:10:23 +00:00
Daniel Han
c2ab1a0e61 Merge branch 'image-generation' into diffusion-train-perf 
# Conflicts:
#	studio/backend/core/training/diffusion_dit_trainer.py
#	studio/backend/core/training/diffusion_train_common.py
 2026-07-05 02:09:39 +00:00
pre-commit-ci[bot]
f6f198fd5f [pre-commit.ci] auto fixes from pre-commit.com hooks 
for more information, see https://pre-commit.ci
 2026-07-05 02:07:49 +00:00
Daniel Han
25d9cf9604 Stub diffusers.hooks too in the no-diffusers cache test 2026-07-05 02:04:16 +00:00
Daniel Han
e605075508 Fix diffusion training validation and honor lr_scheduler and batch size in the DiT trainer 2026-07-05 01:51:34 +00:00
Daniel Han
098809d2fe Reject sd-cli batch runs and clear stale output targets before a run 2026-07-05 01:49:19 +00:00
Daniel Han
76eee534ea Gate explicit attention kernels on NVIDIA CUDA and roll back partial FBCache hooks 2026-07-05 01:48:12 +00:00
Daniel Han
f1d9c88606 Validate load modes before eviction and wait out a cancelled denoise on unload 2026-07-05 01:47:02 +00:00
Daniel Han
6a8b0b47e7 Fix review findings on image generation: failed-load VRAM, API defaults, preflights 
- Free reserved VRAM in the diffusion load worker's failure path: a load-time OOM
  never commits _state and the next load's _unload_locked early-returns, so nothing
  else reclaimed the half-built pipeline's memory
- Use a monotonic clock for the denoise ETA rate
- Sync _GENERATION_DEFAULTS with the UI table: kontext, flux.2-dev, sdxl-turbo and
  SDXL base rows so /v1/images/generations stops falling back to 9 steps / CFG 0
- 400 (not sanitized 500) when /v1/images/generations hits an edit-only model
- Fail fast on pre-Ampere CUDA in the DiT trainer instead of dying in model load
- Run the trainer trust gate in the diffusion training route before freeing GPU
  residents so an untrusted base cannot tear down loaded chat/Images models
- Protect native sd.cpp companion VAE/text-encoder repos from cache deletion while
  a load is downloading them
- Exempt the task-scoped Images picker from the chat-only GGUF/MLX format gate so
  local diffusers pipelines stay selectable on no-GPU hosts
 2026-07-05 01:00:47 +00:00
Daniel Han
b4201e6390 Merge branch 'diffusion-train-perf2' into diffusion-auto-policy 2026-07-05 00:31:36 +00:00
Daniel Han
08674a297e Merge branch 'diffusion-krea2' into diffusion-train-perf2 2026-07-05 00:31:34 +00:00
Daniel Han
99eb248607 Merge branch 'diffusion-train-tab-2' into diffusion-krea2 2026-07-05 00:31:33 +00:00
Daniel Han
71c20ded19 Merge branch 'diffusion-train-precision' into diffusion-train-tab-2 2026-07-05 00:31:31 +00:00